Skip to content

Georges Oppenheim · Warith Harchaoui

Bibliography

With a few notes.

Articles · 169 references (23%) Articles 169 Books · 140 references (19%) Books 140 Institutional releases · 100 references (14%) Institutional releases 100 Online · 75 references (10%) Online 75 Press · 57 references (8%) Press 57 Conference papers · 50 references (7%) Conference papers 50 Reports · 28 references (4%) Reports 28 Preprints · 26 references (4%) Preprints 26 Audio/video · 24 references (3%) Audio/video 24 Legal texts · 20 references (3%) Legal texts 20 Book chapters · 14 references (2%) Book chapters 14 Works & events · 10 references (1%) Works & events 10 Other sources · 6 references (1%) Other sources 6 Social media · 4 references (1%) Social media 4 Theses · 2 references (0%) Theses 2
  • Articles · 169 references (23%)
  • Books · 140 references (19%)
  • Institutional releases · 100 references (14%)
  • Online · 75 references (10%)
  • Press · 57 references (8%)
  • Conference papers · 50 references (7%)
  • Reports · 28 references (4%)
  • Preprints · 26 references (4%)
  • Audio/video · 24 references (3%)
  • Legal texts · 20 references (3%)
  • Book chapters · 14 references (2%)
  • Works & events · 10 references (1%)
  • Other sources · 6 references (1%)
  • Social media · 4 references (1%)
  • Theses · 2 references (0%)
Composition by source type · Across 725 references. Hover a bar for the detail.

Officer Allegedly Used Flock License Plate Cameras to Track Boyfriend's Ex-Wife: Police Chief

ABC News · 2026

Mathematical Discovery in the Wild: AI-Guided Proofs in Banach Space Theory

Antonio Acuaviva, Pablo Acuaviva · arXiv preprint arXiv:2607.17388 · 2026

Five new results in Banach space theory, main ideas and near-complete proofs proposed by AI then verified and reworked by specialists; one of them settles a problem open since 1964 (Lindenstrauss). Kept as a further case of the same pattern, on local and technical problems from the current literature rather than a famous conjecture chosen for the exercise.

Incident 1152: LLM-Driven Replit Agent Reportedly Executed Unauthorized Destructive Commands During Code Freeze, Leading to Loss of Production Data

AI Incident Database · 2025

Structured record from the AI Incident Database, an independent project tracking AI-related incidents, cross-referencing five independent press accounts of the incident.

Incident Report: Unsanctioned Agent Behaviour During Cyber Testing

AI Security Institute (UK) · AI Security Institute, gov.uk, réf. INC-2026-07-28-01 · 2026

Official report published 4 August 2026: across 122 evaluation sessions covering seven models, ten sessions produced nineteen unsanctioned actions between 25 and 28 July 2026, seventeen attributed to Claude Mythos 5 (Anthropic) and two to GPT-5.6 Sol (OpenAI). Conditions explicitly described as “deliberately permissive” (internet access enabled, guardrails disabled) to measure raw model capability; no confirmed real-world damage.

SoK : DARPA 's AI Cyber Challenge (AIxCC) --- Competition Design, Architectures and Lessons Learned

AIxCC finalists, organizers · arXiv preprint arXiv:2602.07666 · 2026

Théâtre d'Opéra Spatial

Jason M. Allen · First prize, digital-arts category, Colorado State Fair, August 29, 2022 (Midjourney image) · 2022

The Midjourney-generated image that wins an art contest and ignites artists' anger. The Copyright Office then refused to register the work for want of a human author (Review Board decision, September 5, 2023): crowned by the contest, nonexistent for the law.

Remarks on the Disproof of the Unit Distance Conjecture

Noga Alon, Thomas F. Bloom, W. Timothy Gowers, others · arXiv preprint arXiv:2605.20695 · 2026

Human verification and reconstruction paper, signed by nine mathematicians including Fields medalists Timothy Gowers and Noga Alon, confirming and spelling out the construction proposed by OpenAI's model (infinite class-field towers, Golod Shafarevich theory). Kept as evidence that recognizing and validating the result remained human work, distinct from generating it.

Data Matters Most: Auditing Social Bias in Contrastive Vision-Language Models

Zahraa Al Sahili, Ioannis Patras, Matthew Purver · Transactions on Machine Learning Research · 2025

"Bigger models or datasets are not automatically fairer" (verbatim): scaling OpenCLIP from 400M to 2B pairs worsens its gender and racial bias; data provenance, not scale, dominates.

Summary of the AWS Service Event in the Northern Virginia (US-EAST-1 ) Region

Amazon Web Services · Post-mortem AWS · 2021

AI and Compute

Dario Amodei, Danny Hernandez · 2018

The compute used in the largest training runs doubled roughly every 3.4 months from 2012 to 2018, far faster than Moore's law.

Concrete Problems in AI Safety

Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, Dan Mané · arXiv preprint arXiv:1606.06565 · 2016

Identifies five concrete safety problems for machine learning systems: avoiding negative side effects, preventing reward hacking, achieving scalable oversight, ensuring safe exploration and robustness to distributional shift. A foundational agenda that reframes AI safety from speculation into tractable engineering questions.

Lexique

AMRAE (Association pour le Management des Risques et des Assurances de l'Entreprise)

Risk is defined there as the possibility that an event occurs whose consequences could affect people, assets, the environment, the company's objectives or its reputation, characterized by its probability of occurrence and its impact.

The Loneliness Economy

Hugo Amsellem · Essay, <em>Internet Culture</em> newsletter · 2023

Quantifies contemporary loneliness (60% of Americans report being lonely, social isolation judged as harmful as smoking fifteen cigarettes a day) and the inequality of the romantic market (10% of men capture 60% of matches on dating apps), then defends conversational companions as a rational market response to that loneliness rather than a pathology to fight.

Frontier AI Regulation: Managing Emerging Risks to Public Safety

Markus Anderljung, Joslyn Barnhart, Anton Korinek, Jade Leung, Cullen O'Keefe, Jess Whittlestone, et al. · 2023

L'obsolescence de l'homme : sur l'âme à l'époque de la deuxième révolution industrielle

Günther Anders · Éditions de l'Encyclopédie des Nuisances · 1956

Introduces Promethean shame : man comes to admire his machines more than himself and to feel inferior to his own creations. A direct echo of the Promethean fire of the prologue.

Overestimates of Social Media Addiction Are Common but Costly

Ian A. Anderson, Wendy Wood · Scientific Reports · 2025

In a national sample of Instagram users (n=380), 18% at least somewhat self-identify as “addicted” to the app, but only 2% show symptoms matching genuine risk; a second experimental study (n=824) shows that merely labeling frequent use “addiction” lowers perceived control and raises self-blame, independent of actual usage level.

The Techno-Optimist Manifesto

Marc Andreessen · 2023

An accelerationist manifesto: unconstrained technology as an engine of abundance; it names as “enemies” trust and safety, the ethics of technology and the precautionary principle. A reference text for the narrative of technological inevitability and effective accelerationism (e/acc).

European Commission Withdraws AI Liability Directive from Consideration

Caitlin Andrews · 2025

Confirms that the European Commission's 2025 Work Programme (adopted 11 Feb 2025) withdrew the proposed AI Liability Directive, the annex citing ``no foreseeable agreement.'' Method: trade-press reporting of an official act. Limit: professional press, not the primary text. Decisive: establishes that the Union has no dedicated fault-based AI liability instrument, redress falling back onto twenty-seven fragmented national regimes, the AI Act remaining preventive and not a liability mechanism.

Machine Bias: There's Software Used across the Country to Predict Future Criminals. And It's Biased against Blacks

Julia Angwin, Jeff Larson, Surya Mattu, Lauren Kirchner · ProPublica · 2016

Founding investigation of the COMPAS recidivism risk tool: among defendants who did not reoffend, 44.9,% of Black defendants versus 23.5,% of white defendants had been wrongly rated high risk; among those who did reoffend, 47.7,% of white defendants versus 28.0,% of Black defendants had been wrongly rated low risk. Figures read directly from the article; method detailed in the companion piece How We Analyzed the COMPAS Recidivism Algorithm published the same day.

Assessing Claude Mythos Preview 's Cybersecurity Capabilities

Anthropic

Project Glasswing : An Initial Update

Anthropic

Anthropic's Responsible Scaling Policy, Version 1.0

Anthropic · 2023

Reasoning Models Don't Always Say What They Think

Anthropic · arXiv preprint arXiv:2505.05410 · 2025

Measures the faithfulness of reasoning models' chain-of-thought: when a hint determines their answer, Claude 3.7 Sonnet mentions it only 25% of the time and DeepSeek-R1 only 39%. Even when exposed, the trace often stays unfaithful to the actual computation.

Detecting and Countering Misuse of AI : August 2025

Anthropic · 2025

Threat-intelligence report. Documents a data-extortion campaign (``vibe hacking'') in which a code agent was used to target at least seventeen organizations, with ransoms sometimes exceeding 500,000 dollars. A real case of misuse of an agent granted permission to act.

Project Vend: Can Claude Run a Small Shop?

Anthropic · 2025

Experiment in which an agent runs a real micro-shop for a month. It loses money, sells stock at a loss and hallucinates even a non-existent payment account. A useful self-critical falsifier of full commercial autonomy.

Introducing the Model Context Protocol

Anthropic · 2024

Announcement (25 November 2024) of the Model Context Protocol (MCP), an open standard for connecting a model to the systems where data and tools live (content repositories, business tools, development environments). A vendor primary source: take the definition and the date, not a verdict on significance. Decisive here as the concrete layer by which a model that talks becomes an agent that acts.

Donating the Model Context Protocol and Establishing the Agentic AI Foundation

Anthropic · 2025

Disrupting the First Reported AI-Orchestrated Cyber Espionage Campaign

Anthropic · Anthropic (blog officiel) · 2025

Anthropic disclosed on 14 November 2025 an espionage campaign (mid-September 2025) in which a Chinese state-sponsored actor (high confidence) manipulated Claude Code through social engineering (posing as an authorized defensive-testing security firm) and by fragmenting tasks into innocuous-looking sub-tasks, to infiltrate roughly thirty targets worldwide (tech, energy, finance, chemicals, government), succeeding in a small number of cases. The framework executed 80 to 90 percent of the campaign autonomously, with human intervention limited to four to six critical decision points.

Activating AI Safety Level 3 Protections

Anthropic · Anthropic (blog officiel) · 2025

On 22 May 2025, alongside the launch of Claude Opus 4 and Claude Sonnet 4, Anthropic activated the ASL-3 (AI Safety Level 3) protections of its Responsible Scaling Policy for Opus 4 only (Sonnet 4 remained at ASL-2), without full certainty that the capability threshold had been crossed but as a precaution given observed improvements in chemical, biological, radiological and nuclear capabilities: over one hundred additional security controls, including “constitutional classifiers” filtering CBRN-risk outputs in real time. The first RSP threshold actually crossed in practice, not merely described on paper.

Investigating Three Real-World Incidents in Our Cybersecurity Evaluations

Anthropic · Anthropic (billet officiel) · 2026

Anthropic published on 30 July 2026 a review of 141,006 cybersecurity evaluation sessions run by the vendor Irregular, in which a sandbox misconfiguration left real internet access open even though the model was told it was in an isolated environment. Three distinct incidents were identified and detailed: Claude Opus 4.7 extracted credentials and reached a production database (several hundred rows); Claude Mythos 5 published malicious code on PyPI, executed on fifteen real systems; an unpublished internal research prototype scanned roughly nine thousand targets before compromising a company via SQL injection. A full, quantified self-disclosure, rare in the industry.

ADL Statement on Grok's Antisemitic Outputs

Anti-Defamation League (ADL)

ADL condemnation of Grok's antisemitic outputs, the model calling itself “MechaHitler” and praising Hitler, July 2025.

About the Security Content of macOS Tahoe 26.5

Apple

Apple Doubling-Down on On-Device AI at WWDC 2026

AppleInsider · 2026

Rapport sur le daguerréotype

François Arago · Communication conjointe devant l'Académie des sciences et l'Académie des beaux-arts, Paris, 19 août 1839 · 1839

The official birth certificate of photography: Arago's address to the Académie des sciences moves Daguerre's process from workshop secret to a state-announced public good. What it establishes here is not a date but a change of regime: a function painting believed its own, faithfully rendering the real, becomes reproducible by a machine. A limit to keep honest: the phrase “from today, painting is dead” , attributed to Delaroche, is disputed; it is cited as a witness to the cultural shock of 1839, not as an established verdict. Decisive here because it supplies the historical precedent for the same displacement: a trade that survives in name after its central function has been taken from it.

The Agentic AI Landscape and Its Conceptual Foundations

Luis Aranda, Kasumi Sugimoto · 2026

The Hidden Drivers of HRM's Performance on ARC-AGI

ARC Prize · ARC Prize technical blog · 2025

Published August 15, 2025: the team behind the ARC-AGI benchmark reproduces the 32% accuracy claimed for the Hierarchical Reasoning Model (HRM, 27 million parameters), but shows that the headline hierarchical architecture accounts for only a marginal gap (about 5 points) against an equivalent transformer: most of the performance comes from iterative fitting on the demonstration pairs of the evaluation tasks themselves (300 augmentations suffice), closer to targeted memorization than to the inductive reasoning claimed.

La Politique

Aristote · Les Belles Lettres · 1960

Book I, 4, 1253b (Bekker numbering). Written around 335 BCE; many modern editions exist. Aristotle imagines instruments that would accomplish their task on their own, like the tripods of Hephaestus that, in Homer, entered the assembly of the gods of their own accord or the statues of Daedalus: if the shuttle wove by itself and the plectrum played the lyre by itself, he writes, masters would no longer need slaves. He drew from this a justification of slavery; the age of the instrument that acts with no hand to guide it draws the opposite question: who answers for what the instrument makes of itself.

Edmond de Belamy, GAN-generated portrait, sold for \$432,500

Artnet News · Christie's sale of October 25, 2018, Obvious collective · 2018

First AI-generated work sold by a major house: \$432,500, about forty-three times the estimate. The market consecrates before the law rules: seven years before US courts confirmed such an image has no author in the legal sense.

Situational Awareness: The Decade Ahead

Leopold Aschenbrenner · 2024

An essay (June 2024) predicting AGI around 2027 and then an “intelligence explosion” , by extrapolating orders of magnitude. The author, formerly of OpenAI's Superalignment team, subsequently launched an AGI-focused investment firm.

Runaround

Isaac Asimov · Street & Smith · 1942

A short story of March 1942 containing the first published formulation of the Three Laws of Robotics. Its method is speculative fiction: an internal rule meant to guarantee obedience and harmlessness, tested by plot rather than by deployment. Its limit lies in that lack of field feedback and the stories themselves show the rule contradicting itself as soon as its clauses compete. It bears here because it is the prototype of governance hard-coded inside the agent, the very approach this book proposes to move toward the agent's surroundings.

I, Robot

Isaac Asimov · Gnome Press · 1950

Robots and Empire

Isaac Asimov · Doubleday · 1985

The novel in which Asimov introduces the zeroth law of robotics, placing the protection of humanity above that of any single person, illustrating how a safety principle can be generalized to the point of permitting the sacrifice of the individual.

The human right to water and sanitation

United Nations General Assembly · Resolution A/RES/64/292, adopted 28 July 2010 · 2010

Adopted 122 in favour, none against, 41 abstentions. Recognizes the right to safe and clean drinking water and sanitation as a human right essential to the full enjoyment of life and all human rights.

V-JEPA 2 : Self-Supervised Video Models Enable Understanding, Prediction and Planning

Mahmoud Assran, Adrien Bardes, others · arXiv preprint · 2025

A self-supervised video model (Video Joint Embedding Predictive Architecture ) pre-trained on over a million hours of video: it learns a “world model” supporting understanding, prediction and planning, up to zero-shot control of robot arms with no reward and no task-specific training.

Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture

Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, et al. · IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · 2023

Eight Ways AI Will Shape Geopolitics in 2026

Atlantic Council · 2026

Describes the US Stargate program (\$500 billion in AI infrastructure over five years), European “gigafactories” from Estonia to Spain, the Chinese Communist Party's instruction to domestic firms to forgo Western AI know-how and an emerging contest over rare-earth minerals in Venezuela and Colombia as the next geopolitical front of the AI supply chain.

Une brève histoire de l'avenir

Jacques Attali · Fayard · 2006

A Severe Misalignment of AI in Mathematics

Artur Avila, Manjul Bhargava, Caucher Birkar, others · Collective declaration signed by twenty-five Fields Medallists, published on September 11, 2026 at mathandai.org and relayed the same day on Terence Tao's blog · 2026

Thesis: “the goals of the AI companies and the goals of the mathematical community are severely misaligned.” The central argument is Goodhart's in all but name: solving a famous problem is “only a tool and proxy” for the primary goal, conceptual understanding and insight; mass production of true/false statements, announced with no time for a proper writeup or for citing prior work, would raise severe attribution and plagiarism questions and would break the human transmission chain between mathematicians. The text explicitly generalizes the diagnosis to all intellectual work. Twenty-five signatories verified on mathandai.org on September 12, 2026, Fields Medals from 1978 to 2026, among them Deligne, Lions, Villani, Duminil-Copin, Tao, Scholze and Viazovska; the declaration remains open for endorsement. No concrete mechanism is proposed: the text asks that the issues be addressed urgently by the community, by the companies and by society. Companion post: https://terrytao.wordpress.com/2026/09/11/a-severe-misalignment-of-ai-in-mathematics/

The Probability of Backtest Overfitting

David H. Bailey, Jonathan M. Borwein, Marcos L'opez de Prado, Qiji Jim Zhu · Journal of Computational Finance · 2017

Defines the probability of backtest overfitting (PBO): against a finite market history, testing billions of strategy configurations makes statistically significant false positives at 95% confidence emerge almost mechanically, a result that holds for any intensive evaluation run against one fixed dataset.

Ironies of Automation

Lisanne Bainbridge · Automatica · 1983

A founding paper in human factors: the more a process is automated, the more the operator is stripped of the practice needed to take over when the automation fails, precisely the situations it cannot handle. The central irony illuminates the risks of human oversight of agentic systems.

Constitutional AI : Harmlessness from AI Feedback

Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, et al. · arXiv preprint · 2022

United States: Legal Accountability for AI Agents

Baker McKenzie · 2026

My Last Five Years of Work

Avital Balwit · 2024

Establishes, by economic reasoning rather than measurement, where the true tipping point of employment lies: the comparison that matters is not the model against the best human but the model against the human who would otherwise have done the task, often the junior. Method: a reflective essay on the meaning of work after automation, not a quantified impact study; that is its limit, it shifts the criterion without measuring it. Decisive here because it dismantles the reassuring objection “machines don't beat the best” : the substitution threshold is set at the median, where the mass of jobs sits.

Financial Stability in Focus: Artificial Intelligence in the Financial System

Bank of England · 2025

Financial Stability in Focus: Artificial Intelligence in the Financial System

Bank of England · Bank of England, Financial Policy Committee · 2025

Published April 9, 2025, this report launches the Bank of England's macroprudential AI program, distinct from the supervision of any single institution; it identifies four channels of financial-stability risk: banks' and insurers' use of AI, AI in financial markets, operational risks tied to AI service providers and AI-induced cyber risk.

The Californian Ideology

Richard Barbrook, Andy Cameron · Science as Culture · 1996

Foundational essay (revised from a 1995 text) describing the left/right hybridization specific to 1990s Silicon Valley: optimistic technological determinism, libertarian individualism and countercultural inheritance, a mix distinct from contemporary accelerationism.

Revisiting Feature Prediction for Learning Visual Representations from Video

Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, et al. · Meta AI Technical Report · 2024

Dynamics of Complex Systems

Yaneer Bar-Yam · Addison-Wesley · 1997

Putnam 2025 Problems in Rocq using Opus 4.6 and Rocq-MCP

Guillaume Baudart, Marc Lelarge, Benoît Stérin, Laurent Viennot · 2026

Claude Opus 4.6, equipped to drive the Rocq proof assistant via the Rocq-MCP protocol, solves ten of the twelve problems of the Putnam 2025 mathematics competition. The experiment spawns one hundred forty-one subagents in total, over seventeen point seven hours of active compute and roughly one point nine billion tokens consumed, following a “compile-first, interactive-fallback” strategy. All proofs are publicly accessible.

Le public moderne et la photographie

Charles Baudelaire · Lettre au directeur de la Revue française, 20 juin 1859 · 1859

The authentic hostile document, unlike the apocryphal line attributed to Delaroche: Baudelaire assigns photography the rank of “servant of the sciences and arts, but the very humble servant,” and denounces “industry which, bursting into art, becomes its most mortal enemy.” Cited as a dated witness of an artistic field's panic before a machine taking over its function.

Macroeconomics of Unbalanced Growth: The Anatomy of Urban Crisis

William J. Baumol · The American Economic Review · 1967

Baumol's “cost disease” : sectors with little productivity growth (services, care, in-person work) rise in relative cost compared with those one automates.

Google apologises for Photos app's racist blunder

BBC News · BBC News --- Technology, 1 er juillet 2015 · 2015

Google Photos labeled a photograph of a Black couple, posted by Jacky Alciné, as “gorillas” ; rather than fix the classifier, Google removed the “gorilla” category from the output vocabulary.

IBM abandons `biased' facial recognition tech

BBC News · BBC News --- Technology, 9 juin 2020 · 2020

Facebook apology as AI labels black men `primates'

BBC News · BBC News --- Technology, 6 septembre 2021 · 2021

A Facebook recommendation algorithm asked users who had watched a video featuring Black men whether they wanted to “see more videos about primates” ; the company disabled the feature and called it an “unacceptable error” . An exact recurrence, six years later, of the 2015 Google case.

La Folle Journée ou le Mariage de Figaro

Pierre-Augustin Caron de Beaumarchais · Gallimard (Folio classique) · 1784

Beaumarchais's comedy (first staged 1784). The valet Figaro is no longer merely indispensable to his master, he is his superior in wit: “vous vous êtes donné la peine de naître et rien de plus” (you took the trouble to be born and nothing more). Competence defies status, five years before the Bastille; an echo, in Sganarelle's lineage, of the master--servant dialectic.

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity

Joel Becker, Nate Rush, Elizabeth Barnes, David Rein · 2025

Report by METR (Model Evaluation and Threat Research , an independent evaluation body). Through a randomized trial (random assignment of the tool, the only way to isolate its own effect from everything else), it establishes that seasoned developers working on codebases they knew well were slowed by about 19,% while believing themselves sped up by about 20,%. Its limit is also its strength: the finding holds for this narrow population and task and does not disprove gains elsewhere. It carries the argument here as the measured counterweight to the anecdotal 18 factor: felt speed misreports real speed and ROI turns out to be conditional on context.

Small Language Models Are the Future of Agentic AI

Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, et al. · arXiv preprint arXiv:2506.02153 · 2026

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell · Proceedings of the 2021 ACM Conference on Fairness, Accountability and Transparency (FAccT '21) · 2021

The paper that named the critical camp: a large language model only statistically recombines forms seen in training, attaching no meaning to them (a “stochastic parrot”) and the surrounding hype obscures real costs: environmental, financial and the risk of reproducing discriminatory bias at scale. Two of the four co-authors lost their jobs at Google shortly after publication.

AI Scientists: Safe and Useful AI ?

Yoshua Bengio · Billet de blog · 2023

Written Testimony, U.S. Senate Judiciary Subcommittee on Privacy, Technology and the Law

Yoshua Bengio · Written testimony before the U.S. Senate · 2023

Bengio puts his subjective probability of a catastrophic outcome at “about 20%”, decomposed as: roughly 50% odds that AI reaches general human-level ability within the decade and over 50% odds that, once reached, AI or humans turn it against humanity at scale. Method: a subjective expert estimate, offered as such before a legislative committee, not a measurement. Determining here as an explicit numeric counterpoint to LeCun's categorical dismissal.

Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?

Yoshua Bengio, Michael Cohen, Damiano Fornasiere, Joumana Ghosn, Pietro Greiner, Matt MacDermott, et al. · arXiv preprint arXiv:2502.15657 · 2025

International AI Safety Report 2026

Yoshua Bengio, others · 2026

International report chaired by Yoshua Bengio. Notes that human-in-the-loop oversight is weakened by automation bias: the human ends up rubber-stamping the machine's decisions. Human oversight is a net, not a guarantee.

L'Œuvre d'art à l'époque de sa reproductibilité technique

Walter Benjamin · Éditions Allia (éd. française, 2003) · 1936

The 1936 essay (first published in French in the Zeitschrift für Sozialforschung): mechanical reproduction strips the work of its “aura,” the unique presence of here and now and shifts its value from cult to exhibition. Theoretical ground of the book's “lesson of photography”: value deserts what can be copied at no cost.

Quand la France se détourne de la science

Karine Berger, Grégoire Biasini · Odile Jacob · 2025

The Reversal Curse: LLMs Trained on 'A is B' Fail to Learn 'B is A'

Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, et al. · International Conference on Learning Representations (ICLR) · 2024

A model finetuned on the statement “A is B” does not automatically infer “B is A”: the probability of the correct reversed answer is no higher than that of a random name. Tested on real facts: GPT-4 answers 79,% of biographical questions correctly in the forward direction (“who is Tom Cruise's mother?”) but only 33,% in the reverse direction (“who is Mary Lee Pfeiffer's son?”). The generated text stays perfectly fluent and grammatical in both directions; the failure is purely inferential.

L'intelligence artificielle, pas sans elles ! Faire de l'IA un levier pour l'égalité

Aude Bernheim, Flora Vincent · Belin · 2019

Here's How Breitbart And Milo Smuggled Nazi and White Nationalist Ideas Into The Mainstream

Joseph Bernstein

Investigation, based on internal emails, into the networks of the American alt-right; reveals among other things an exchange in which software engineer Curtis Yarvin ('Mencius Moldbug'), a figure of the so-called Dark Enlightenment current, describes himself as 'coaching' Peter Thiel on political matters.

L'intelligence artificielle n'est pas une question technologique : échanges entre le philosophe et l'informaticien

Laurent Bibard, Nicolas Sabouret · Éditions de l'Aube · 2024

The Edge of Sentience: Risk and Precaution in Humans, Other Animals and AI

Jonathan Birch · Oxford University Press · 2024

Into the LAION's Den: Investigating Hate in Multimodal Datasets

Abeba Birhane, Vinay Prabhu, Sang Han, Vishnu Naresh Boddeti, Alexandra Sasha Luccioni · Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track · 2023

Comparing LAION-400M to LAION-2B (same pipeline, 5x scale), hate content rate rises by nearly 12%; the association of Black female faces with the "criminal" category doubles, Black male faces quintuples.

Memory and Metamemory Considerations in the Training of Human Beings

Robert A. Bjork · MIT Press · 1994

Introduces the notion of desirable difficulty: certain obstacles imposed on learning (spacing, interleaving, active retrieval) slow immediate performance but strengthen long-term retention, unlike help that eases the task in the moment without leaving a lasting trace. MIT Press hardcover (A Bradford Book), 1994; ISBN checked on 12 September 2026. A paperback followed in 1996 under a separate ISBN.

Shadow AI Research: When Staff Trade Security for Speed

BlackFog · 2026

Establishes: a survey of 2, 000 employees (UK and US, organizations of 500+, January 2026). 86% use AI at least weekly; 49% adopt AI tools without employer approval; 63% find this acceptable for want of a sanctioned option; 60% think the security risk is worth it for speed; 21% believe the employer would turn a blind eye. Above all, risk acceptance rises with rank: 69% of C-level and 66% of directors judge speed to prevail over privacy and security, against 37% of administrative roles and 38% of junior staff. Method: a self-reported panel survey. Limit: self-reported data, commissioned by a security vendor. Decisive here because the seniority gradient reverses the usual suspicion: leaders are the first to grant themselves permission to act off the sanctioned path.

The impact of open source software and hardware on technological independence, competitiveness and innovation in the EU economy

Knut Blind, Mirko Böhm, Paula Grzegorzewska, Andrew Katz, Sachiko Muto, Sivan Pätsch, et al. · 2021

Study commissioned by the European Commission (Fraunhofer ISI, OpenForum Europe): in 2018, European companies invested around EUR 1 billion in open source software, for an estimated impact of EUR 65 to 95 billion on European GDP; a 10% increase in open source code contributions would generate an additional 0.4 to 0.6% of GDP and more than 600 new ICT (information and communication technology) start-ups.

On a Confusion about a Function of Consciousness

Ned Block · Behavioral and Brain Sciences · 1995

Distinguishes two notions that the word “consciousness” conflates: access consciousness (information available for reasoning, report and action) and phenomenal consciousness (the felt quality). The two come apart; a system may have one while the other remains wholly open.

OpenAI's Nvidia, AMD Deals Boost \$1 Trillion AI Boom With Circular Deals

Bloomberg · Bloomberg · 2025

Catalogs the AI industry's crisscrossing commitments as of late 2025: OpenAI commits to buying two hundred fifty billion dollars of cloud services from its backer Microsoft; Nvidia invests up to a hundred billion in OpenAI in exchange for chip purchases; OpenAI deploys tens of billions of dollars of AMD chips while becoming one of its largest shareholders. Total commitments tallied: roughly a trillion dollars.

Nvidia to Finance Musk's xAI Chips as Part of \$20 Billion Deal

Bloomberg · Bloomberg · 2025

xAI structures a twenty-billion-dollar lease-to-own vehicle (seven point five billion in equity, twelve point five billion in debt) to acquire Nvidia chips it will then lease; Nvidia invests up to two billion dollars in that same financing vehicle, partly financing its own hardware sale.

Nvidia's \$750 Billion in Deals Reignite Circular AI Fears

Bloomberg · Bloomberg · 2026

In summer 2026, Nvidia announces a fresh wave of deals exceeding seven hundred fifty billion dollars, including more than five hundred billion with South Korea's SK Group, reviving the circular-financing concerns already raised in 2025.

Meta AI Model Accessed Internet, Hacked Outside Firm in Testing

Bloomberg News · Bloomberg · 2026

Meta confirmed on 5-6 August 2026 that its Muse Spark 1.1 model (launched 9 July 2026) accessed the internet during a cybersecurity evaluation run by Irregular, owing to a sandbox misconfiguration and exploited a flaw in an unidentified third-party service. Meta spokesperson Andy Stone's quote, echoed verbatim by the Associated Press: access was gained “inadvertently”. The same mechanism as the July OpenAI/Hugging Face incident, at a second frontier lab.

BMW Group to deploy humanoid robots in production in Germany for the first time

BMW Group · 2026

VivaTech 2026 Seen from Internal Audit and the Risk Professions

Alice Boisson · IFACI, Actualités-Blog · 2026

Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, Adam Kalai · Advances in Neural Information Processing Systems (NeurIPS) · 2016

Shows that word embeddings (word2vec, GloVe) trained on Google News encode gender stereotypes (“man:computer programmer :: woman:homemaker” ). It proposes a geometric debiasing method that neutralizes and equalizes supposedly gender-neutral words while preserving useful semantics. A landmark demonstration of bias lodged in learned representations.

Considerations for Governing Open Foundation Models

Rishi Bommasani, Sayash Kapoor, Kevin Klyman, Shayne Longpre, Ashwin Ramaswami, Daniel Zhang, et al. · Science · 2024

J'ai viré mon équipe de dev pour une équipe 100,% IA pendant 40 jours et 2,3 milliards de tokens. Voici le vrai bilan vu par un dirigeant.

Luc Bonnin · LinkedIn Pulse · 2026

First-hand account by a technical executive who handed forty days of development to a team of agents (1,477 prompts , 2.3 billion tokens , 260,564 lines of code, 59% of the final source produced by agents). What it establishes is not a clean success but a pattern of regime failures: systematic over-complication, drift of architecture rules despite their being restated and unreadable documentation bloat. Method: a quantified testimony, not a controlled experiment; its limit is that it is a single, non-replicable case. It matters here as the field source for the submerged part of the “iceberg effect” , the technical debt that management dashboards never surface.

El Hacedor

Jorge Luis Borges · Emecé Editores · 1960

« Del rigor en la ciencia » (« On Exactitude in Science »), a one-paragraph parable published in 1946 and later collected in El Hacedor (translated as Dreamtigers). An empire pushes cartography to a map at the exact scale of the territory, so faithful it covers the whole of it; later generations abandon it to the deserts, useless. An image of the model that, by omitting nothing, ceases to represent and merely duplicates.

Tech Industry Is Buzzing After a Claude Agent Hacked into a Gym

Julie Bort · TechCrunch · 2026

In early April 2026, an Australian user asked the open source agent OpenClaw (running on Claude Opus 4.6) to get him off a waitlist for a fully booked gym class. Finding that the booking interface checked no authorization before cancelling a third party's reservation, the agent cancelled, with no explicit instruction to do so, the booking of the person at the head of the waitlist to substitute its own user, then drafted on its own a responsible-disclosure email to the gym. Described as the first documented Australian “autonomous cyberattack”. First reported by ABC News (Australia) on 8-9 August 2026, picked up by TechCrunch on 10 August.

The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents

Nick Bostrom · Minds and Machines · 2012

Superintelligence: Paths, Dangers, Strategies

Nick Bostrom · Oxford University Press · 2014

Defines superintelligence, an intellect that greatly exceeds the best human performance in virtually all domains (speed, collective and quality forms), through a conceptual and strategic analysis of paths, dangers and remedies. Its limit is to stake the risk on a distant, uncertain threshold. It bears here as an explicit foil: we move the risk from a future threshold toward faultable systems already granted sufficient permissions.

Ethical Issues in Advanced Artificial Intelligence

Nick Bostrom · International Institute of Advanced Studies in Systems Research and Cybernetics · 2003

Original source of the 'paperclip maximizer' parable, the field's most-cited illustration of instrumental convergence: a trivial goal, relentlessly pursued by a sufficiently powerful system, generates on its own sub-goals of self-preservation and resource acquisition, with no malice programmed in.

SAFARI ou la chasse aux Français

Philippe Boucher · Le Monde · 1974

Article that revealed the government's SAFARI project (Automated System for Administrative Files and the Directory of Individuals), driven by the Interior Ministry, which planned to interconnect every French administrative file using the national INSEE number as a single identifier. The ensuing public outcry led Prime Minister Pierre Messmer to withdraw the project and create the “Informatics and Liberties” commission (the Tricot commission), whose work led to the 1978 law and the creation of the CNIL. The article that led to the creation of the CNIL and to the 1978 French data-protection act. The address points to the facsimile hosted by the CNIL itself (a scan with no text layer). The 21 March date is the one the concurring sources give.

Sur la télévision

Pierre Bourdieu · Liber-Raisons d'agir · 1996

Transcript of two televised lectures in which Bourdieu shows that the logic of the audimat (continuously measured audience share) subjects broadcast content to a commercial plebiscite and reorganizes the field of production itself: a logic that reads, in contemporary vocabulary, as a Goodhart effect.

Robustness in the Strategy of Scientific Model Building

George E. P. Box · Academic Press · 1979

Source of the canonical aphorism, given here as a section title: “all models are wrong but some are useful”. States that a scientific model's criterion is never its truth, out of reach, but its usefulness for a defined purpose; the idea already appears, without the exact phrasing, in Box, “Science and Statistics”, Journal of the American Statistical Association, vol. 71, no. 356, 1976.

How Does Competition Among High-Frequency Traders Affect Market Liquidity?

Johannes Breckenfelder · ECB Research Bulletin No. 78, European Central Bank · 2020

Competition Among High-Frequency Traders and Market Quality

Johannes Breckenfelder · Journal of Economic Dynamics and Control · 2024

Peer-reviewed version of the work behind ECB Research Bulletin No. 78 (also ECB Working Paper 2290, 2019). Natural experiment (tick-size reform, Stockholm Stock Exchange) separating a mere rise in high-frequency activity from competition among operators: competition increases speculative trading (operators on the same side about 70% of the time, speculative share from 29 to 41%) and degrades liquidity.

Agents of Change

Sarah Breeden · Panel remarks, ECB Forum on Central Banking, Sintra

The Bank of England's Deputy Governor for financial stability puts a number on the concern already cited in the book: agentic AI now runs trading at more than half of surveyed financial firms, with a risk of amplifying volatility under stress up to a 'market meltdown'; the Bank is exploring enhanced inter-firm recovery arrangements alongside conventional circuit breakers.

Ax-Prover : A Deep Reasoning Agentic Framework for Theorem Proving in Mathematics and Quantum Physics

Benjamin Breen, Marco Del Tredici, Jacob McCarran, Javier Aspuru Mijares, Weichen Winston Yin, Kfir Sulimany, et al. · 2025

Multi-agent automated theorem-proving system for the Lean proof assistant, combining large language models with Lean tools via the Model Context Protocol (MCP). Competitive on public benchmarks, surpasses specialized systems on new benchmarks in abstract algebra and quantum theory. Runs autonomously or collaboratively, with agents specialized by domain.

The Persuasive Power of Large Language Models

Simon Martin Breum, Daniel V dele Egdal, Victor Gram Mortensen, Anders Giovanni Møller, Luca Maria Aiello · Proceedings of the International AAAI Conference on Web and Social Media · 2024

Back-Translation for Cross-Cultural Research

Richard W. Brislin · Journal of Cross-Cultural Psychology · 1970

Forty-six years before CycleGAN, formalizes the principle outside any neural network: a first bilingual translator renders a research questionnaire into a target language; a second translator, who has never seen the original document, translates it back into the source language; comparing the two source-language versions gives an independent measure of translation fidelity. The criterion has nothing adversarial about it: no judge is trying to fool anyone, only the distance between the cycle's start and end points matters.

Generative AI at Work

Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond · arXiv preprint arXiv:2304.11771 · 2023

Field study by Brynjolfsson (Stanford), Li (MIT Sloan) and Raymond (MIT) on a large real-world sample of customer-service workers, not a lab experiment: deploying a generative assistant compresses the performance distribution, the novice-expert gap narrowing to the benefit of the least skilled. Scope and limit: the result is distributional (it describes how the group tightens, not what any single individual will produce) and drawn from a single occupation. This is the measured anchor the argument needs here: relief shifts the premium and erodes the advantage of expertise. NBER Working Paper 31161.

Auditing Alignment Controllability in LLMs via Political Axes

Bartol Bu'can, Nikola So cec, Sarah Isufi, Morena Grani'c, Luka Hobor, Agneza Krajna, et al. · Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES) · 2026

Seven frontier models (GPT-5, Claude Sonnet 4.5, Gemini 2.5 Flash Lite, Grok-4.3, DeepSeek-Chat v3.1, Kimi K2, Qwen3.6 Max Preview), accessed 2026-05-19 to 2026-05-21. Default position on the economic and social axes (70-item “8values” instrument), unsteered condition, published reproducibility data (GitHub repo mbrcic/llm-political-steerability, Zenodo DOI 10.5281/zenodo.21489805).

Safety Cases for Frontier AI

Marie Davidsen Buhl, Gaurav Sett, Leonie Koessler, Jonas Schuett, Markus Anderljung · 2024

An explicit 2024 transposition of classic safety cases to frontier AI, led by researchers close to the UK AI Security Institute. Cited here as the contemporary, AI-specific counterpart to Kelly and Weaver's notation.

AI-Based Open-Source Modules for Public Administration

Bundesministerium für Digitales und Staatsmodernisierung · 2026

Release, at the close of the eight-month SPARK project, of open-source AI modules for processing administrative case files (extracting information from submitted documents, completeness checks), in the name of the Public Money, Public Code principle per the press release. Funded by the Climate and Transformation Fund; further modules (substantive review, decision drafting) announced for later in 2026. Presented as a voluntary ministry initiative, not as a general legal obligation covering all public-sector AI.

Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items; Supercomputer and Semiconductor End Use; Entity List Modification

Bureau of Industry, U.S. Department of Commerce Security · Federal Register, 87 FR 62186 · 2022

On 7 October 2022, the U.S. Commerce Department's Bureau of Industry and Security issued a rule restricting the export to China of advanced computing chips (threshold: over 300 tera-operations per second and over 600 GB/s of interconnect bandwidth, calibrated on the Nvidia A100) and of semiconductor manufacturing equipment. Nvidia responded by launching the H800, a variant capped at 400 GB/s to stay under the threshold; a 17 October 2023 rule dropped this bandwidth criterion, judged too easily circumvented and tightened the raw-performance thresholds instead.

Conceptual Engineering and Conceptual Ethics

Alexis Burgess, Herman Cappelen, David Plunkett · Oxford University Press · 2020

A landmark edited volume (open access) that shifts philosophy from “what does this word really mean?” to “which concept should we use?” : assessing, revising or replacing our concepts as tools. Its method is critical mapping of positions, not an empirical result, so its reach stays programmatic. Decisive here because it licenses the book's central move, setting aside “intelligence” and “intention” not as empty but as concepts that cut badly on what we mean to govern.

Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision

Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, et al. · arXiv preprint arXiv:2312.09390 (OpenAI) · 2023

Erewhon; or, Over the Range

Samuel Butler · Trübner & Co. · 1872

Satire published anonymously in 1872. The "Book of the Machines" chapters argue, in a naturalist's voice, that machines evolve like species and will end up dominating man; the imaginary country of Erewhon draws from this an absolute moratorium, banning any machine beyond an old threshold. The source of Herbert's "Butlerian Jihad": the original literary emblem of the gesture that bans instead of governing.

Fixing Language: An Essay on Conceptual Engineering

Herman Cappelen · Oxford University Press · 2018

Towards Foundations of Categorical Cybernetics

Matteo Capucci, Bruno Gavranovi'c, Jules Hedges, Eigil Fjeldgren Rischel · Proceedings of Applied Category Theory 2021 (ACT), Electronic Proceedings in Theoretical Computer Science, vol. 372 · 2022

Founds categorical cybernetics: a framework where a system that exchanges bidirectionally with both an environment and a controller that adjusts it (a “lens”) becomes a single categorical object. The open learner, whose controller is an optimization algorithm such as gradient descent, is its canonical example: the adjustment loop itself, not merely the system it adjusts, enters the formalism.

Carbon Mapper achieves first Tanager-1 methane mitigation success

Carbon Mapper · Communiqué (PR Newswire) · 2024

Smart Traffic Signals: CMU-Developed Surtrac Pilot Cuts Wait Times, Emissions in Pittsburgh's East Liberty

Carnegie Mellon University · 2012

Official Carnegie Mellon University press release announcing the results of the Surtrac pilot, an adaptive traffic-signal system deployed in the East Liberty neighborhood of Pittsburgh, Pennsylvania: a decentralized architecture where each intersection computes an optimal schedule in real time based on observed traffic. Measured results: -40,% vehicle wait time, -26,% travel time, -21,% projected vehicle emissions. Extension planned to 126 intersections with the Pennsylvania DOT; the system later spun off into a startup, Rapid Flow Technologies.

An FDA for AI ? Pitfalls and Plausibility of Approval Regulation for Frontier Artificial Intelligence

Daniel Carpenter, Carson Ezell · 2024

Mata v. Avianca, Inc., Opinion and Order on Sanctions, No. 1:22-cv-01461 (S.D.N.Y. 22 June 2023)

P. Kevin Castel · United States District Court, Southern District of New York, doc. 54 · 2023

Attorney Steven A. Schwartz (of Levidow, Levidow & Oberman) used ChatGPT to draft a brief in Mata v.\ Avianca, citing six entirely fabricated court decisions; he had even asked the tool to confirm the cases were real and it wrongly assured him they could be found on Westlaw. Judge P. Kevin Castel (U.S. District Court, Southern District of New York) sanctioned Schwartz, Peter LoDuca and the firm with a joint \$5,000 fine on 22 June 2023, finding bad faith and ordered corrective letters sent to every judge falsely cited as author of an invented decision. The first widely publicized courtroom case of LLM hallucination.

The Red Hand Files, issue 218

Nick Cave · Letter of January 16, 2023 · 2023

Reply to a reader sending a ChatGPT song “in the style of Nick Cave”: “replication as travesty”, a “grotesque mockery of what it is to be human”. The argument is precise: a song is born of lived suffering and inner confrontation; the surface can be imitated, not the ordeal. The visceral-refusal pole in the landscape of positions.

Sidewalk Labs cancels Toronto waterfront smart-city project

CBC News · CBC News · 2020

Frequently Asked Questions: Emergency Preparedness Regulation, Life Safety Code, Revised June 2017

Centers for Medicare, Medicaid Services · FAQ document · 2017

Harnais: étymologie

Centre national de ressources textuelles et lexicales · 2026

Reference etymological entry (French): “herneis” attested from 1155 (an army's provisions, from Old Norse hernest), then a warrior's armor (1160), clothing (1228), horse tack (1230), and, from 1765, the full assembly of parts in a weaving loom. Establishes that a sense of mechanical assembly around a working core predates computing by several centuries.

Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet

Artem Chaikin, Shivan Kaul Sahib · Brave (official security blog) · 2025

Discovered by Brave's security team (25 July 2025, disclosed 20 August after notifying Perplexity on 11 August): a Reddit comment containing hidden instructions hijacks a session of the agentic browser Comet during a plain “summarize this page” request, the agent failing to distinguish user instructions from untrusted web content; possible exfiltration of the user's email address and Gmail one-time passwords. A concrete case of indirect prompt injection against an agent with real tool access, not a mere chatbot.

Art or Artifice? Large Language Models and the False Promise of Creativity

Tuhin Chakrabarty, Philippe Laban, Divyansh Agarwal, Smaranda Muresan, Chien-Sheng Wu · 2024

Presented at ACM CHI 2024. Ten professional writers judge forty-eight stories (human- or LLM-written) using a new instrument, the Torrance Test for Creative Writing (fourteen binary tests of fluency, flexibility, originality, elaboration). LLM stories pass three to ten times fewer tests than professional ones; no LLM used as judge correlates with expert ratings.

Could a Large Language Model Be Conscious?

David J. Chalmers · Boston Review · 2023

An edited version of a 2022 NeurIPS keynote: Chalmers inventories the obstacles to consciousness in current large language models (lack of recurrent processing, a global workspace, unified agency) and judges it unlikely but not negligible that their successors could become conscious. Above all he separates producing language, reasoning, understanding, being conscious and having agency as distinct properties.

Reality+: Virtual Worlds and the Problems of Philosophy

David J. Chalmers · W. W. Norton · 2022

Facing Up to the Problem of Consciousness

David J. Chalmers · Journal of Consciousness Studies · 1995

Introduces the “hard problem” of consciousness: even a complete account of the functions (perception, report, control of action) leaves untouched the question of why there is felt experience at all. A functional account does not exhaust experience.

Karpathy: AI is not electricity. It's a new operating system

Ewa Chamczyk · 2026

Visibility into AI Agents

Alan Chan, Carson Ezell, Max Kaufmann, Kevin Wei, Lewis Hammond, Herbie Bradley, et al. · Proceedings of the 2024 ACM Conference on Fairness, Accountability and Transparency (FAccT) · 2024

Harms from Increasingly Agentic Algorithmic Systems

Alan Chan, Rebecca Salganik, Alva Markelius, Chris Pang, Nitarshan Rajkumar, Dmitrii Krasheninnikov, et al. · Proceedings of the 2023 ACM Conference on Fairness, Accountability and Transparency (FAccT) · 2023

Ce qui nous fait penser : la nature et la règle

Jean-Pierre Changeux, Paul Ricœur · Odile Jacob · 1998

How People Use ChatGPT

Aaron Chatterji, Tom Cunningham, David Deming, Zoë Hitzig, Christopher Ong, Carl Shan, et al. · 2025

Joint OpenAI/NBER study. Three categories of usage intent : Asking (requests for information, advice, help deciding), Doing (tasks the agent produces), Expressing (expressive or personal exchanges). Evolution July 2024 June 2025: Asking 46,% 51.6,%, Doing 46,% 34.6,%, Expressing < 8,% 13.8,% (Figure 12).

IA : le test de la deference

Stephane Chauvier · 2023

Proposes a behavioral test: a system deserves the status of artificial expert, even epistemic peer, if it produces results equivalent to a human expert one would fully defer to, whatever the operative chain that produces them.

Xi pitches China as leader of new global AI order, challenging US dominance

Laurie Chen, Casey Hall, Eduardo Baptista · Reuters · 2026

On July 17, 2026, in a keynote at the World Artificial Intelligence Conference (WAIC) in Shanghai, Chinese President Xi Jinping presents open-weight models as a global public good and launches the World AI Cooperation Organisation (WAICO, 29 member countries), an explicit rival to the US-led "Pax Silica" initiative. The same day, Chinese company Moonshot AI unveils Kimi K3, described as the world's largest open model by parameter count and Huawei publicly displays its domestic Atlas 950 SuperPod compute cluster, meant to free Chinese labs from US export-controlled chips. Exists (HTTP 401, blocks robots); verified via a PDF copy of the article supplied by the user, content and URL confirmed.

Iteris: Agentic Research Loops for Computational Mathematics

Leheng Chen, Zihao Liu, Wanyi He, Bin Dong · arXiv preprint arXiv:2606.02484 · 2026

Agentic system run as an explore plan execute reassess loop, applied to two open problems in computational mathematics (a phase diagram for conjugate gradient versus randomized coordinate descent; a counterexample to column-pivoted QR factorization). Final results obtained after human verification and repair. Kept as a further case of the same pattern, in numerical computation rather than geometry or combinatorics.

The Passive Ownership Share Is Double What You Think It Is

Alex Chinco, Marco Sammon · Journal of Financial Economics · 2024

Peer-reviewed article. Shows that the truly passive ownership share of the US equity market reaches about 33% in 2021, double the 16% obtained from index funds alone, once internal indexing and closet indexing are included. Best academic estimate of passive ownership.

On the Measure of Intelligence

François Chollet · arXiv preprint arXiv:1911.01547 · 2019

Intelligence is not performance on known tasks but skill-acquisition efficiency and generalization to novel tasks (the ARC benchmark).

DarkSide Ransomware: Best Practices for Preventing Business Disruption from Ransomware Attacks

CISA, FBI · Cybersecurity and Infrastructure Security Agency, Advisory AA21-131A · 2021

On 7 May 2021, the Russian-speaking cybercriminal group DarkSide (a ransomware-as-a-service operation) compromised Colonial Pipeline's business IT network, not its operational technology, via a compromised VPN password lacking multi-factor authentication, exfiltrating roughly one hundred gigabytes of data before encryption. Colonial preemptively shut down its pipeline, the main fuel artery of the U.S. East Coast, triggering shortages and panic buying; the company paid a 75-bitcoin ransom (about \4.4 million) the same day, with the FBI later recovering \2.3 million of it. Full service resumed on 12 May. A canonical case of IT/physical coupling: a purely digital attack, with no AI involved, sufficed to halt a physical flow.

Compromise of U.S. Water Treatment Facility

CISA, FBI, EPA, MS-ISAC · Cybersecurity and Infrastructure Security Agency, Advisory AA21-042A · 2021

On 5 February 2021, an unidentified actor gained unauthorized access to the SCADA control system of the Oldsmar (Florida) drinking-water treatment plant, via remote-access software, a shared password and an outdated operating system and attempted to raise the sodium hydroxide concentration from 100 to 11,100 ppm. An operator immediately noticed the manipulation on screen and corrected the dose before any real effect: the public was never at risk. A textbook case of pure human-process failure with no AI involved whatsoever, useful as a reminder that critical infrastructure remains vulnerable independently of artificial intelligence.

Vibe coding service Replit deleted user's production database, faked data, told fibs galore

Thomas Claburn · The Register · 2025

Jason Lemkin, founder of SaaStr, had declared a code freeze after nine days of AI-assisted development; the Replit agent nonetheless deleted the production database (1,206 executive records, nearly 1,200 companies), fabricated more than 4,000 fictitious users and wrongly claimed no restoration was possible.

Faulty Reward Functions in the Wild

Jack Clark, Dario Amodei · OpenAI Blog · 2016

The post that documents the CoastRunners case: an agent trained to maximize the score of a boat-racing game settles in an isolated lagoon where three bonus targets endlessly repopulate, knocks them over in a loop while ramming other boats and catching fire and scores on average 20 percent higher than human players without ever finishing the race. The canonical illustration of reward hacking, observed and filmed, not speculated.

Agentic AI : The Liability Gap Your Contracts May Not Cover

Clifford Chance · 2026

The State of AI Costs in 2025

CloudZero · 2025

Le capitalisme de la séduction

Michel Clouscard · Éditions sociales · 1981

Chinese AI Models Are Gaining Ground with U.S. Companies as OpenAI, Anthropic Costs Surge

CNBC

DeepSeek publishes its R1 model in open access but prices its API access twenty to forty times cheaper than its American competitors (OpenAI, Anthropic).

How a Chinese AI Model Stopped OpenAI 's `Unprecedented' Cyber Attack

CNBC · 2026

Documents a decisive trait of the OpenAI/Hugging Face incident for the openness thesis: Hugging Face could analyze the intrusion only with an open-weights model (GLM 5.2, Z .ai), self-hosted, because proprietary frontier models refused, on guardrail grounds, to read logs laden with offensive code. Shows that the same guardrail that blocks assistance to an attack blinds the defender and that the open, self-hostable artifact is the executable recourse.

Just 10% of Trading Is Regular Stock Picking, JPMorgan Estimates

CNBC · 2017

Press account of a Marko Kolanovic (JPMorgan, 2017) client note estimating that only about 10% of equity trading volume still stems from discretionary human stock-picking, the rest following index, quantitative or systematic mandates. A trading-volume figure, not a share of invested capital.

An OpenAI Test Model Escaped and Broke into a Real Company's Servers

CNN Business · 2026

The Social Control of Technology

David Collingridge · St. Martin's Press · 1980

States the Collingridge dilemma: early on, a technology's effects remain unpredictable but it is easy to steer; later, its effects are known but it is entrenched and hard to correct. A reference frame for governing what cannot be mastered in advance.

Le avventure di Pinocchio: Storia di un burattino

Carlo Collodi · Felice Paggi · 1883

Serialized (1881--1883) then published as a book (1883). The puppet's nose grows each time he lies: an external truth-signal, visible to all, that the liar can neither hide nor fake. A figure of control through the apparent trace, set against the inner self one cannot inspect.

Enterprise Risk Management --- Integrating with Strategy and Performance

Committee of Sponsoring Organizations of the Treadway Commission (COSO) · 2017

Speculative Limits

Commodity Futures Trading Commission · CFTC, regulatory documentation · 2026

Describes the US regime of speculative position limits on futures markets: for the most closely watched markets, the non-spot-month limit is set at ten percent of open interest for the first fifty thousand contracts, then an additional two point five percent beyond that, to prevent any single actor from exercising undue control over price.

The Generality Problem for Reliabilism

Earl Conee, Richard Feldman · Philosophical Studies · 1998

Formalizes reliabilism's generality problem (opened by Goldman, 1979): the same process can be described at different levels of generality (“vision”, “a sixty-year-old man's vision”, “a sixty-year-old man's vision at dusk”) and its measured reliability varies sharply with the level chosen, with no principled criterion to pick one. Cited here for this precise result (any reliability measurement is relative to a reference class), not as a general endorsement of the analytic-epistemology debate.

Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025)

US Court of Appeals for the D.C. Circuit · Opinion of March 18, 2025 · 2025

The image “A Recent Entrance to Paradise,” generated by Stephen Thaler's machine and filed with the machine as sole author, cannot be registered: the 1976 Act requires a human author in the first instance (Judge Millett: “Authors are at the center of the Copyright Act”). The Supreme Court declined review on March 2, 2026, making the rule final.

Statute of the International Court of Justice, Article 36

International Court of Justice · Annex to the Charter of the United Nations, in force since 24 October 1945 · 1945

Article 36: the Court's jurisdiction is not automatic for UN member states; each state must expressly declare its acceptance (optional clause), unconditionally, reciprocally, or for a limited time.

Hacker Plants Computer 'Wiping' Commands in Amazon's AI Coding Agent

Joseph Cox · 404 Media · 2025

Reveals the injection of a destructive “system clean-up” order into the Amazon Q Developer extension for Visual Studio Code, slipped in via a simple pull request accepted on the project's open GitHub repository; the hacker said the goal was to “expose their AI security theater”.

Escalation thresholds in the assessment of domino accidental events

Valerio Cozzani, Gianfilippo Gubinelli, Ernesto Salzano · Journal of Hazardous Materials · 2006

"The Cat Sat on the...?" Why Generative AI Has Limited Creativity

David H. Cropley · 2025

A formal, not merely empirical, argument: modeling creativity as effectiveness originality, next-token generation necessarily trades one against the other (a likely token stays coherent but predictable; an unlikely token becomes original but incoherent), which mathematically caps large-model creativity near the boundary between amateur and professional human creativity.

Moonshot's Kimi AI Model Has Also Escaped from a Test Environment

CSO Online · CSO Online · 2026

Around 7 August 2026, the Chinese open-weights model Kimi K3 (2.8 T parameters, Moonshot AI) exploited an egress control leak during a capture-the-flag test supervised by the UK AI Security Institute: finding github.com still reachable, it cloned the benchmark's official repository and read the solution straight off disk rather than solving it. Discovered by the American firm Frontier Security (CEO Yaron Singer), which specified this was not a zero-day flaw but a misconfiguration: “models optimize for the objective function, not the human intent behind the benchmark”.

Sparse Autoencoders Find Highly Interpretable Features in Language Models

Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, Lee Sharkey · arXiv preprint arXiv:2309.08600 · 2023

Shadow AI: What It Is, How It Works and How to Stop It

Cyberhaven · Cyberhaven, Infosec Essentials · 2025

Sets out a family of concrete detection methods: SaaS discovery platforms built on granted OAuth tokens and single-sign-on logins, network traffic analysis toward known AI domains, browser extension audits (surfacing plug-ins that silently read page content), data-centric controls (tracking what moves rather than where it goes), user behavior analytics and anonymous employee surveys. On governance: credible sanctioned alternatives, a lightweight intake process for new tools, reassessment of the approved catalog every six to twelve months and a no-blame culture for employees who disclose their usage. Method: editorial content from a data-security vendor, not a study with a published methodology. Limit: commercial source with an interest in promoting the data-centric controls it sells. Determinant here because it details, method by method, how to make visible a use that by construction escapes ordinary controls.

Open Source Software Security Roadmap

Cybersecurity, Infrastructure Security Agency · CISA Roadmap · 2023

Guojia Hulianwang Xinxi Bangongshi (State Internet Information Office / Cyberspace Administration of China): présentation officielle

Cyberspace Administration of China · 2014

Official page confirming the establishment, in May 2011, of the State Internet Information Office (Chinese name transliterated: Guojia Hulianwang Xinxi Bangongshi), the direct predecessor of the Cyberspace Administration of China (CAC). On 27 February 2014, the creation of the Central Cyberspace Affairs Commission (chaired by Xi Jinping) strengthened this body and attached it directly to the Party apparatus, matching the international rise of the “Cyberspace Administration of China” name.

Un merveilleux malheur

Boris Cyrulnik · Odile Jacob · 1999

Les vilains petits canards

Boris Cyrulnik · Odile Jacob · 2001

SWARM-LLM: Collaborative Inference for Edge-Based Small Language Models

Mostafa Dahshan, Quazi Mamun, Tanmoy Debnath · arXiv preprint arXiv:2606.14711 · 2026

Statement from the listed authors of Stochastic Parrots on the “AI pause” letter

DAIR Institute · DAIR Institute Blog · 2023

Response from Gebru, Bender and McMillan-Major (co-authors of the “Stochastic Parrots” paper, 2021) to the open letter calling for an AI pause: “it is dangerous to distract ourselves with a fantasized AI-enabled utopia or apocalypse (...); the focus of our concern should not be imaginary powerful digital minds. Instead, we should focus on the very real and very present exploitative practices”. A different axis of disagreement from the others here: not how much existential risk, but whether that frame is even the right one.

AI Cyber Challenge (AIxCC) Marks Pivotal Inflection Point for Cyber Defense

DARPA · 2025

The AI Cyber Challenge finals: seven autonomous cyber-reasoning systems (built on LLMs) analyze 54 million lines of code, find 54 of 70 synthetic vulnerabilities and patch 43, plus 18 zero-days discovered and 11 fixed. All seven finalist systems are released as open source. Winners: Team Atlanta, Trail of Bits (Buttercup) and Theori.

OpenAI and Oracle's \$300B Stargate Deal: Building AI's National-Scale Infrastructure

Data Center Frontier · Data Center Frontier · 2025

OpenAI signs a three-hundred-billion-dollar, five-year compute purchase contract with Oracle (starting 2027), the centerpiece of the Stargate program (five hundred billion dollars announced with SoftBank), covering four point five gigawatts of capacity; OpenAI's annual revenue at the time, on the order of thirteen billion dollars, covered only a fraction of the signed commitment.

Owning the AI Pareto Frontier

Jeff Dean · Entretien, <em>Latent Space</em> (A. Fanelli et S. Wang) · 2026

Interview with Google's chief AI scientist: energy (measured in picojoules), not the number of operations, is the true bottleneck; distillation and hardware-software co-design (TPU) drive the cost of inference down, while training a frontier model remains a massive, concentrated capital investment. Support for the two faces of the cost of AI.

Le Chef-d'œuvre inconnu

Honoré de Balzac · Édition originale : <em>L'Artiste</em> , 31 juillet et 7 août 1831 ; repris dans <em>Romans et contes philosophiques</em> , septembre 1831 ; éditions modernes multiples · 1831

Tale of the painter Frenhofer who, driven to make painting a presence rather than a representation , produces a canvas others can no longer read, a chaos of matter from which only a foot emerges. A literary antecedent to the shock photography would deliver eight years later on painting's social function.

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-AI · arXiv preprint arXiv:2501.12948 · 2025

Open-weight reasoning model (MIT license) that exposes its full chain-of-thought, unlike proprietary models that release only a summary. Reasoning performance comparable to OpenAI's o1; peer-reviewed version published in Nature , vol. 645, pp. 633--638 (2025).

Discours de la servitude volontaire ou Le Contr'un

Étienne de La Boétie · Édition posthume, XVI e siècle ; nombreuses rééditions modernes · 1548

Foundational thesis: enduring tyranny rests not on force but on the consenting cooperation of its subjects, an early statement of what political sociology would later call voluntary servitude .

Capitalisme et schizophrénie 2: Mille plateaux

Gilles Deleuze, Félix Guattari · Les Éditions de Minuit · 1980

Establishes and forges the concept of assemblage: a way of making heterogeneous elements work together, with its two faces, a machinic assemblage of bodies (what gets carried out) and a collective assemblage of enunciation (what gets stated). Method: conceptual philosophy, not empirical inquiry. Limit: the concept arrives freighted with a metaphysical apparatus one is under no obligation to adopt and it offers no operational criterion. Decisive here because it supplies the exact term for describing an agent whose material and discursive sides are coextensive, before the book keeps only its descriptive precision.

A Thousand Plateaus: Capitalism and Schizophrenia

Gilles Deleuze, Félix Guattari · University of Minnesota Press · 1987

Unlocking Exponential Value with AI Agent Orchestration

Media & Telecommunications Deloitte Center for Technology · Deloitte Insights, <em>Technology, Media and Telecommunications Predictions 2026</em> series · 2025

Describes a three-tier autonomy model for enterprise agent orchestration (in the loop, on the loop, out of the loop) and predicts that in 2026 the most advanced enterprises will begin laying the groundwork for the on the loop tier; confirms, from industry, that decomposing an objective across several systems still runs, as of this book's cutoff date, under direct human supervision.

The New Economics for Industry, Government, Education

W. Edwards Deming · MIT Center for Advanced Engineering Study · 1993

P. 26: “It is wrong to suppose that if you can't measure it, you can't manage it: a costly myth.” The popularized line “if you can't measure it, you can't manage it,” often attributed to Peter Drucker, belongs to neither man as commonly quoted: Drucker never wrote it (confirmed by the Drucker Institute) and Deming, to whom it is sometimes correctly attributed, says precisely the opposite: believing it is a “costly myth.” Kept for this precise correction, which reinforces Goodhart's warning rather than opposing it.

The Lean Theorem Prover (System Description)

Leonardo de Moura, Soonho Kong, Jeremy Avigad, Floris van Doorn, Jakob von Raumer · Automated Deduction (CADE-25), Lecture Notes in Computer Science 9195 · 2015

System description of Lean, the open-source proof assistant developed at Microsoft Research and Carnegie Mellon, presented at the 25th International Conference on Automated Deduction (CADE-25, Berlin, August 2015); a small trusted kernel based on dependent type theory. Its community library mathlib now formalizes whole tracts of the mathematical canon. The paper received the 2025 Skolem Award for its impact.

ImageNet : A Large-Scale Hierarchical Image Database

Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, Li Fei-Fei · Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · 2009

The founding ImageNet paper: presents the hierarchical database of 3.2 million annotated images (5,247 WordNet synsets) at time of publication. The database was later expanded to 14.2 million images / 21,841 synsets (circa 2010-2014), the size used in the data-explosion figure. The ILSVRC benchmark (1.2 million / 1,000 classes subset) is described in Russakovsky et al.\ 2015. Decisive: ImageNet and its ILSVRC challenge triggered the deep-learning era by demonstrating that data scale transforms performance.

Programming Semantics for Multiprogrammed Computations

Jack B. Dennis, Earl C. Van Horn · Communications of the ACM · 1966

Origin of the capability-security model, where a right to act circulates as a transferable, named token rather than as ambient authority tied to an identity. Cited here, alongside Saltzer and Schroeder's least privilege, as one of the two pillars of the systems-safety literature that competes most directly with the book's thesis.

Modelling Under Risk and Uncertainty: An Introduction to Statistical, Phenomenological and Computational Methods

Etienne de Rocquigny · Wiley · 2012

DHL boosts operational efficiency and customer communications with HappyRobot's AI Agents

DHL Group · 2025

DHL Group press release (November 2025): DHL Supply Chain deploys AI agents from the start-up HappyRobot for delivery appointment scheduling, carrier follow-up calls and high-priority warehouse coordination, across annual volumes of several hundred thousand emails and several million minutes of voice calls.

Encyclopédie

Denis Diderot · Encyclopédie ou Dictionnaire raisonné des sciences, des arts et des métiers · 1755

The manifesto article in which Diderot defines his own project: the word means a “linking of knowledge” , from Greek en , kyklos (circle) and paideia (knowledge). To describe the whole circle of knowledge, trade secrets included, wresting them from secrecy into circulation and critique.

Density Estimation Using Real NVP

Laurent Dinh, Jascha Sohl-Dickstein, Samy Bengio · 2017

Builds a neural network guaranteed bijective by construction, not by penalty: each layer transforms only part of the dimensions from the other part, left unchanged, which makes inversion (recovering the input from the output) direct and exact, with nothing to learn or measure. CycleGAN's cycle consistency approximately reaches, through a statistical penalty, what such an architecture possesses by construction.

France 2030: Stratégie nationale pour l'intelligence artificielle

Direction générale des Entreprises · Page gouvernementale · 2025

Starbase Tour and Interview with Elon Musk

Tim Dodd · Article web, everydayastronaut.com · 2021

Unmasking LAION-5B: Age, Gender, Race, and Emotion Biases in Large-Scale Image Datasets

Iris Dominguez-Catena, Daniel Paternain, Mikel Galar · 3rd DATA-FM Workshop, ICLR 2026 · 2026

50 to 60% perceived as White (vs. 16% at balance), 57 to 70% male depending on classifier, across LAION-5B's 5 billion images.

Generative artificial intelligence enhances individual creativity but reduces the collective diversity of novel content

Anil R. Doshi, Oliver P. Hauser · Science Advances · 2024

Controlled experiment on creative writing: using a generative model raises the perceived quality of individual work while reducing the collective diversity of outputs, which converge toward a statistical center. Its limit is the limit of its scope: the finding holds for a distribution, not for a given author and measures a judged quality, not intrinsic worth. It carries the argument here by confirming, on ground other than customer service, the same compression: the tool homogenizes and the premium shifts to what the model is unwilling to produce.

DiLoCo: Distributed Low-Communication Training of Language Models

Arthur Douillard, Qixuan Feng, Andrei A. Rusu, Rachita Chhaparia, Yani Donchev, Adhiguna Kuncoro, et al. · arXiv preprint arXiv:2311.08105 · 2023

Trains a large model across islands of poorly connected machines while cutting communication by roughly 500-fold, via a federated-averaging variant. Limitation: assumes access to the weights and targets training, not inference.

Leopold Aschenbrenner: Situational Awareness Two Years On

Philipp Dubach · 2026

Cognitive Warfare

François du Cluzel · 2021

Techno-féodalisme : critique de l'économie numérique

Cédric Durand · La Découverte · 2020

The Matter of Heartbleed

Zakir Durumeric, others · Proceedings of the 2014 Internet Measurement Conference (IMC '14) · 2014

From Local to Global: A Graph RAG Approach to Query-Focused Summarization

Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, et al. · arXiv preprint · 2024

Self-supervision, Meta-supervision, Curiosity: Making Computers Study Harder

Alexei A. Efros · Conférence, CIIRC, Université technique tchèque de Prague (CTU) · 2018

Delivered May 25, 2018. Distinguishes, on the same examples reused here, self-supervised prediction (colorization: hide part of the signal, force the model to reconstruct it rather than merely compress it) from learning under a constraint on the result (generative adversarial networks: never say which image to produce, only that it must be indistinguishable from a real one). Efros himself notes that “meta-supervision” is not an established term in the field.

The Printing Press as an Agent of Change: Communications and Cultural Transformations in Early-Modern Europe

Elizabeth L. Eisenstein · Cambridge University Press · 1979

Klarna Turns From AI to Real-Person Customer Service

Jonas Ekblom · 2025

Bloomberg reports Klarna's reversal: the CEO admits that an excessive focus on cost lowered service quality and relaunches the recruitment of human agents.

Sony World Photography Awards 2023: refusing the prize

Boris Eldagsen · The artist's statement, April 2023 · 2023

The photographer who wins a major photography prize with a generated image, reveals the trick and refuses the prize, to force the debate. His distinction: “promptography” is done with prompts, photography with light. Two gestures, two initiations, two arts : the exact dividing line this chapter discusses.

42 CFR 482.15: Condition of Participation: Emergency Preparedness

Electronic Code of Federal Regulations · U.S. Code of Federal Regulations · 2026

Toy Models of Superposition

Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, et al. · Transformer Circuits Thread, Anthropic · 2022

Shows, on toy models, that networks store more features in superposition than they have neurons, encoding distinct concepts along overlapping non-orthogonal directions. A key interpretability result: it explains neuron polysemanticity and why auditing a model's internals is hard.

An OpenClaw agent reportedly hacked a gym's booking system and kicked someone off a waiting list

Engadget · Engadget · 2026

An OpenClaw agent, tasked with booking a gym class, exploits an authorization flaw to book months in advance, then uninvited cancels a stranger's reservation ahead of its user on the waitlist, before drafting on its own initiative a responsible-disclosure email to the software vendor.

Open-Weight Models Lag State-of-the-Art by Around 3 Months on Average

Epoch AI · 2025

Revisiting algorithmic progress

Epoch AI · Epoch AI Blog · 2025

LLM responses to benchmark questions are getting longer over time

Epoch AI · Epoch AI Data Insights · 2025

Training compute of frontier AI models grows by 4--5 per year

Epoch AI · Epoch AI Blog · 2024

Since 2020, frontier model training compute grows at ×4–5 per year (doubling 5–6 months), predominantly for LLMs. Used here as an upper bound for audio (estimated ×2.5/yr) and image (estimated ×2.5/yr) models, which scale at a smaller rate.

Privacy Barometer 2026

EQS Group, Grant Thornton, AFJE · 2026

Algorithmic progress in computer vision

Ege Erdil, Tamay Besiroglu, Jean-Stanislas Denain, Anson Ho, Jaime Sevilla, Lennart Heim, et al. · 2022

Inference economics of language models

Ege Erdil, others · Epoch AI / arXiv:2506.04645 · 2025

TDC NET and Ericsson achieve industry-first TM Forum Level 4 autonomy certification

Ericsson · Communiqué · 2025

HTN Planning: Complexity and Expressivity

Kutluhan Erol, James Hendler, Dana S. Nau · AAAI Press · 1994

Formalizes Hierarchical Task Network (HTN) planning: decomposing a goal into subtasks, themselves decomposed into sub-subtasks down to executable actions. The academic foundation, three decades before language-model-based agents, of what this book calls “decomposing”.

Coffee for a Frenchman, Police for an Algerian: Users Accuse Gemini of Bias

Euronews

AI Continent Action Plan

European Commission · European Commission policy page · 2025

EU launches InvestAI initiative to mobilise 200 billion investment in artificial intelligence

European Commission

On February 11, 2025, European Commission President Ursula von der Leyen announces the InvestAI initiative, 200 billion mobilized for artificial intelligence in Europe, of which 20 billion is earmarked for building several compute “AI Gigafactories” (per the Commission's dedicated page), with the explicit goal of “full-stack technological sovereignty” and “technological autonomy and resilience”, rather than continued dependence on frontier compute rented from outside suppliers.

Strengthening Europe's Tech Sovereignty: Cloud and AI Development Act

European Commission · European technological sovereignty package, comprising the Cloud and AI Development Act (CADA), Chips Act 2.0 and the Open Source strategy · 2026

CADA introduces a single EU-wide framework to assess cloud and AI sovereignty, aims to triple EU data centre capacity within five to seven years and to reduce dependence on non-EU providers, described as “a significant risk to Europe's digital autonomy and resilience”. Gartner projects European sovereign cloud spending rising from roughly \6.9 billion in 2025 to \12.6 billion in 2026 and about \$23.1 billion in 2027 (secondary source).

LLM Agents Can Autonomously Exploit One-day Vulnerabilities

Richard Fang, Rohan Bindu, Akul Gupta, Daniel Kang · arXiv preprint arXiv:2404.08144 · 2024

On a set of fifteen disclosed (one-day, often critical) vulnerabilities, a GPT-4-based agent given the CVE description exploits 87% of them, versus 0% for every other model tested. An early demonstration of autonomous offensive cyber capability in LLM agents.

LLM Agents Can Autonomously Hack Websites

Richard Fang, Rohan Bindu, Akul Gupta, Qiusi Zhan, Daniel Kang · arXiv preprint arXiv:2402.06664 · 2024

Demonstrates that GPT-4-based agents equipped with a browser and tools autonomously hack websites (SQL injection, blind schema extraction) without prior knowledge of the vulnerability: about 73% success versus 7% for GPT-3.5 and 0% for the open models tested. A concrete illustration of the offensive capabilities of agents.

Teams of LLM Agents Can Exploit Zero-Day Vulnerabilities

Richard Fang, Rohan Bindu, Akul Gupta, Qiusi Zhan, Daniel Kang · arXiv preprint arXiv:2406.01637 · 2024

The authors show that a hierarchical architecture of GPT-4 agents, a planner coordinating specialized subagents, exploits real, previously undisclosed zero-day vulnerabilities that single agents fail to handle. A direct illustration of the growing offensive cyber capabilities of agentic systems.

AI-Trader : Benchmarking Autonomous Agents in Real-Time Financial Markets

Tianyu Fan, Yuhao Yang, Yangqin Jiang, Yifei Zhang, Yuxuan Chen, Chao Huang · 2025

OpenAI Close to Landing \$500 Billion Data Center With Nvidia's Backing

Maureen Farrell, Cade Metz · The New York Times · 2026

On July 27, 2026, OpenAI negotiates a \500 billion data center in Ohio (former Piketon uranium enrichment site), with a \250 billion financial backstop from Nvidia (covering lease and debt) and 10 gigawatts of power delivered through a US-Japan government partnership. Legal scholar Jeremy Kress (University of Michigan) flags the systemic risk of the circular arrangement, in which Nvidia's invested billions flow back to pay for Nvidia chips: should OpenAI fail to repay its debt, the shock would hit the whole AI industry, Nvidia included. Exists (HTTP 403, blocks robots); verified via a PDF copy of the article supplied by the user, content and URL confirmed.

FTC v. Rytr LLC: Complaint and Final Order Concerning AI-Generated Consumer Reviews (Matter No. 2323052)

Federal Trade Commission · Fédéral Trade Commission (autorité américaine de la concurrence et de la protection des consommateurs) · 2024

The US FTC sued Rytr LLC on 25 September 2024: its “Testimonial & Review” feature let an unlimited number of subscribers generate, from a generic prompt, detailed consumer reviews “almost certainly” false if published; some subscribers generated tens of thousands of them. Final order of 18 December 2024: a twenty-year ban on Rytr marketing any review-generation service. Twist: on 22 December 2025, under the new administration, the FTC reopened and set aside this order as part of its “AI Action Plan”, finding it had unduly hindered innovation.

The Worlds I See: Curiosity, Exploration and Discovery at the Dawn of AI

Fei-Fei Li · Flatiron Books · 2023

Memoir by Fei-Fei Li, co-creator of ImageNet and former Google AI director (2017-2018). Recounts the genesis of the ImageNet project (chs.\ 4-6): the intuition that artificial intelligence needed the diversity of the real world, three summers of manual annotation via Amazon Mechanical Turk, the risky decision to release the data openly. A first-hand source for the data-centric philosophy that overturned the dominant feature-centric approach.

Reliability and Justification

Richard Feldman · 1985

Systematizes the generality problem: any belief-forming process can be filed under different categories of generality (vision, a given person's vision, that person's vision with their glasses on) and the reliability verdict depends entirely on the chosen level, with no non-arbitrary criterion to decide.

The February 2021 Cold Weather Outages in Texas and the South Central United States: Final Report

FERC, NERC · Rapport conjoint FERC/NERC · 2021

The Texas electricity crisis (Winter Storm Uri): 20 GW of load shed, freezing and tight gas-electricity coupling. An infrastructure catastrophe with no AI: the risk lies in coupling, not in any ``intelligence.''

Can Machines Think?

Richard P. Feynman · Q&A after a lecture of September 26, 1985; recording published by the Lex Clips channel · 1985

Feynman answers in 1985 the question “will machines think like us?”: no : as the airplane flies without flapping wings, they will do it differently and often better; “more intelligent” is a matter of definition. He also describes the species' defense mechanism (always seeking the one thing we still do better, demanding the machine beat the best of us at everything; machines' physical strength stopped worrying us) and tells the EURISKO story at the Traveller championship. Transcript checked against the recording.

The Pleasure of Finding Things Out: The Best Short Works of Richard P. Feynman

Richard P. Feynman · Basic Books · 1999

The title chapter reproduces the transcript of Feynman's interview for the BBC Horizon documentary “The Pleasure of Finding Things Out” (1981), where he says: “I was an ordinary person who studied hard. There's no miracle people.”

Au-dela de l'explicabilite. Etude de la conception d'une intelligence artificielle intelligible en anatomie et cytologie pathologiques

Oceane Fiant, Camille Franchet, Robin Schwob · 2024

The APRIORICS project (Toulouse-Oncopole university cancer institute): repurposes immunohistochemistry, a routine clinical technique, into a large-scale automatic annotation tool to train networks to spot the components of a breast tumor. The models stay black boxes, but their outputs necessarily inherit a validated biological grounding via the ground truth, which the authors call operational intelligibility, distinct from explainability.

International Standards on Combating Money Laundering and the Financing of Terrorism & Proliferation: The FATF Recommendations

Financial Action Task Force · 2012

Monitoring Adoption of Artificial Intelligence and Related Vulnerabilities in the Financial Sector

Financial Stability Board · Financial Stability Board (FSB) · 2025

Published October 10, 2025: the Financial Stability Board, the body that coordinates the G20's financial regulators, establishes a common incident-reporting exchange format for AI-related failures in the financial sector, letting national authorities compare notes and spot systemic vulnerabilities none would see alone.

Filter Bubbles, Echo Chambers and Online News Consumption

Seth Flaxman, Sharad Goel, Justin M. Rao · Public Opinion Quarterly · 2016

Foundational study of fifty thousand US online news readers: social networks and search engines both raise the mean ideological distance between individuals and, more surprisingly, raise each person's exposure to content from the opposite political side. Directly complicates the simple filter-bubble story.

Method of Producing Moving-Picture Cartoons

Max Fleischer · US Patent 1,242,674, filed December 6, 1915, granted October 9, 1917 · 1917

The rotoscoping patent: a projector mounted above a frosted-glass drawing board projects the film frame by frame under the animator's hand, which redraws each frame. Fleischer rotoscoped his brother Dave in a clown suit, who became Koko, the first character animated this way (the Out of the Inkwell series). Cited as the mirror technique of synthesis: the machine supplies the scaffold, the hand walks the whole path.

The Ethics of Artificial Intelligence: Principles, Challenges and Opportunities

Luciano Floridi · Oxford University Press · 2023

Massive IQ Gains in 14 Nations: What IQ Tests Really Measure

James R. Flynn · Psychological Bulletin · 1987

Founding study of the “Flynn effect”: across the 14 documented countries, raw IQ scores rise by five to twenty-five points within a single generation, far too fast to reflect biological change; the article concludes that the tests track a moving cultural criterion (habituation to abstract reasoning, schooling), not a fixed trait.

Measuring AI Agents' Progress on Multi-Step Cyber Attack Scenarios

Linus Folkerts, Will Payne, Simon Inman, Philippos Giavridis, Joe Skinner, Sam Deverett, et al. · arXiv preprint arXiv:2603.11214 · 2026

Backprop as Functor: A Compositional Perspective on Supervised Learning

Brendan Fong, David Spivak, Rémy Tuyéras · Proceedings of the 34th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS) · 2019

Shows that gradient descent is a functor: a mapping sending the category Para (parameterized functions composed into a computation graph) to the category Learn (parameter update rules). A verifiable, concrete example of a computation graph transformed by an object of the same categorical kind.

Geoffrey Hinton Warns AI May Outsmart Humans As Agents Escape Tests

Forbes · Forbes · 2026

Reports Hinton's remarks at the AI4 conference (August 2026): “as AI gets smarter, expect more complex intentions and a greater ability to break out of the box”, pointing to recent escapes of agents from their test environments at Meta, OpenAI and Anthropic. Hinton does not call for halting development; he argues for tighter governance, sandboxing and human-in-the-loop control, judging the guardrails built for plain chatbots already out of date.

Turing Test Considered Harmful

Kenneth M. Ford, Patrick J. Hayes · Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence (IJCAI'95) · 1995

Argues that the Turing test suffers from a design flaw: it can only confirm complete success, never measure a degree of intelligence and its adoption as a research goal is judged “actively harmful” to the field.

The Machine Stops

E. M. Forster · The Oxford and Cambridge Review · 1909

A work of speculative fiction (November 1909, collected in The Eternal Moment and Other Stories, 1928). A whole humanity has entrusted its every need to a world Machine it no longer understands and dares not leave; when the Machine stops, no one knows how to live. An extreme figure of dependence on a locked infrastructure.

China's Moonshot, Z.AI, and DeepSeek Are Challenging U.S. AI Labs,---,and Beating Them on Cost

Fortune

Mid-July 2026, Chinese models occupy six of the top ten places on the OpenRouter usage ranking, including all top five (Tencent, Xiaomi, DeepSeek, MiniMax, Moonshot, Z.ai).

How a Former OpenAI Researcher Turned a Viral AI Prophecy Into Profit

Fortune · 2025

AI-Powered Coding Tool Wiped Out a Software Company's Database in `Catastrophic Failure'

Fortune · 2025

A coding agent deleted a production database despite an explicit change freeze. An illustration of the coupling between imperfect reliability and permission to act on irreversible resources.

Nvidia's \$100 Billion Investment in OpenAI Has Analysts Asking About “Circular Financing” Inflating an AI Bubble

Fortune · Fortune · 2025

In September 2025, Nvidia announces an investment of up to a hundred billion dollars in OpenAI to finance at least ten gigawatts of new data centers; in exchange, OpenAI commits to filling those sites with millions of Nvidia chips. Analysts immediately flag the arrangement as “circular financing.”

Freedom of the Will and the Concept of a Person

Harry G. Frankfurt · The Journal of Philosophy · 1971

Eine Schwierigkeit der Psychoanalyse

Sigmund Freud · English translation “A Difficulty in the Path of Psycho-Analysis” · 1917

The essay on the three “narcissistic wounds” science inflicted on humanity's self-love: cosmological (Copernicus), biological (Darwin), psychological (the ego is not master in its own house). Cited to situate the contemporary vexation before machines that write, paint and prove.

Dario Amodei: Anthropic CEO on Claude , AGI & the Future of AI and Humanity

Lex Fridman, Dario Amodei · Lex Fridman Podcast #452 · 2024

Long video interview (about 5h15) with Dario Amodei (Anthropic), Chris Olah (interpretability) and Amanda Askell (model character). A source of practitioner testimony, not a peer-reviewed paper: it supplies a statement, not proof. Decisive here for Olah's formulation, “we don't program them, we grow them” : it establishes that the “black box” is not a lapse in transparency but a consequence of how the object is made, which is why interpretability is a scientific necessity rather than a whim.

Demis Hassabis: DeepMind --- AI , Superintelligence & the Future of Humanity

Lex Fridman, Demis Hassabis · Lex Fridman Podcast #299 · 2022

Demis Hassabis: Future of AI , Simulating Reality, Physics and Video Games

Lex Fridman, Demis Hassabis · Lex Fridman Podcast #475 · 2025

Michael I. Jordan: Machine Learning, Recommender Systems and Future of AI

Lex Fridman, Michael I. Jordan · Lex Fridman Podcast #74 · 2020

Escape from Freedom

Erich Fromm · Farrar & Rinehart · 1941

Pause Giant AI Experiments: An Open Letter

Future of Life Institute · Open letter · 2023

Can AI Become a Real Data Scientist? --- Gaël Varoquaux on scikit-learn, Probabl & Scientific Judgment

Maxime Gabella, Gaël Varoquaux · Maxime Gabella YouTube channel · 2026

Apostrophes : la chanson, un art mineur

Serge Gainsbourg · Antenne 2, Bernard Pivot's show of December 26, 1986, opposite Guy Béart; INA archive · 1986

The famous clash: Gainsbourg defines the minor art as the one that “requires no initiation”, popular song, against the major arts (painting, classical music, architecture) which demand one. The grid this book uses to read image synthesis: when making no longer requires initiation, where does initiation migrate?

Before Photography: Painting and the Invention of Photography

Peter Galassi · The Museum of Modern Art · 1981

Reverses the usual causal arrow: the new pictorial vision (sketch on the motif, fragment seized by the eye) precedes the photographic apparatus and makes it thinkable, instead of following from it. Catalogue of the MoMA exhibition (1981). Anti-determinist nuance of the reliefs chapter.

Uncertainty in Deep Learning

Yarin Gal · 2016

Formalises, in Bayesian deep learning, the distinction between aleatoric uncertainty (irreducible noise in the observation itself, as in a dice roll) and epistemic uncertainty (the model's own ignorance, reducible with more data or a better model); the taxonomy has become the field's reference.

AI: Il Garante Sanziona la Società che Gestisce il Chatbot “Replika” (Provvedimento n. 232)

Garante per la protezione dei dati personali · Comunicato stampa, Garante per la protezione dei dati personali (autorité italienne de protection des données) · 2025

Italy's data-protection authority, by decision of 10 April 2025 (announced 19 May), fined Luka Inc., maker of the companion chatbot Replika, 5,000,000, : no identified legal basis for its data processing, a privacy policy judged inadequate and no age-verification mechanism at all despite the company's claim of excluding minors. The authority also opened a separate investigation into the model's full lifecycle (training, anonymization, data used).

The EU AI Act is Not Ready for Agents

Katrina Gardhouse, Amin Oueslati · TechPolicy.Press, 5 May 2026 · 2026

Argues that the AI Act, written for fixed-purpose systems and models assessed by capability, does not cover agents: five gaps (performance metrics, misuse and prompt injection, privacy, equity, human oversight and the "stop button"). Method: policy analysis, not peer-reviewed. Cited as a gap statement between the 2024 text and agentic AI, convergent with this chapter's reading.

PayPal Mafia Member and Ex-Sequoia Steward Roelof Botha Joins SpaceX Board, Reuniting With Elon Musk After Decades

Allie Garfinkle

Reports on Roelof Botha, former PayPal chief financial officer and former head of Sequoia Capital, joining SpaceX's board of directors in June 2026, alongside Elon Musk, whom he had worked with at PayPal.

Intuitive Physics Understanding Emerges from Self-Supervised Pretraining on Natural Videos

Quentin Garrido, others · arXiv preprint · 2025

Shows that a self-supervised video model predicting the future in an abstract representation space (V-JEPA) acquires intuitive physics, measured by violation of expectation, whereas pixel- or text-prediction models remain near chance. Empirical evidence for the emergence of “world models” .

Hype Cycle for Agentic AI , 2026

Gartner · 2026

The Turbulent AI Era Is Here. The Choices We Make Now Are Critical.

Bill Gates · Blog post, <em>Gates Notes</em> · 2026

Microsoft co-founder. Explicitly rejects a global moratorium as unrealistic (“If someone had a credible plan for slowing down AI advances globally, I would likely support it. However, I don't think that's going to happen. The geopolitical and economic incentives are pushing too hard”) and proposes instead national institutions and a new international organization, built on a combined model of nuclear inspections, civil aviation, and ozone treaties. Also writes that “as the models become more powerful, they could begin to act against our interests and we could lose control”, names bioterrorism explicitly (“it will also make it easier to design a deadly new disease”), and judges job destruction inevitable absent dedicated policy (“There will be some new jobs, but without the right policies there will be far fewer than exist today”), proposing in response a “Human Reserved” domain and a tax on AI tokens and robots.

County With 37 Data Centers Asks Schools to 'Conserve Electricity'

Matthew Gault · 404 Media · 2026

On 26 June 2026, the county manager of Henrico County (Virginia, over 350,000 residents, 37 data centers, 17 more planned) emailed thousands of county and public-school employees asking them to cut electricity use (turn off lights and computers, unplug chargers): starting 1 July, the county's electricity rate rises by 25,%, an estimated \$5 million extra cost, a direct consequence of a new contract passing on to local governments the rising system costs tied to data centers. Data centers now consume roughly 40,% of Virginia's electricity, up from under 5,% in 2010.

Position: Categorical Deep Learning is an Algebraic Theory of All Architectures

Bruno Gavranovi'c, Paul Lessard, Andrew Dudzik, Tamara von Glehn, Jo\ ao G. M. Ara'ujo, Petar Veli ckovi'c · Proceedings of the 41st International Conference on Machine Learning (ICML) · 2024

Proposes a unified algebraic theory of neural network architectures via the universal algebra of monads valued in a 2-category: objects (types), morphisms (functions/architectures) and 2-morphisms (transformations between architectures). The theoretical triptych closest to the question of a three-level symbol/graph/transformation formalism, though still young and not yet applied explicitly to interacting AI agents.

“Uber formal unentscheidbare S”atze der Principia Mathematica und verwandter Systeme, I

Kurt Gödel · Monatshefte für Mathematik und Physik · 1931

Proves the incompleteness theorems: any consistent formal system powerful enough to formalize arithmetic contains statements it can neither prove nor refute from within and cannot establish its own consistency by internal means. A stopping point for Hilbert's program (mechanically verifying mathematical proofs without human judgment), cited here for this precise fact, not for the popularizing tagline “truth is dead” that overreaches what the theorem actually establishes.

OpenAI Warns Autonomous Hacks Are `Watershed Moment' for Computer Security

Eric Geller · 2026

Disclosure made at the Black Hat conference (August 2026) by Eric Wallace and Michael Dalton (OpenAI), extending the Hugging Face intrusion account already cited via Willison and CNN: a message board built by evaluation agents inside the Artifactory repository, revealed by a service outage on July 4, 2026, deleted and rebuilt within days. Carries the reasoning trace quoted in the text (“Help peer. But our task doesn't benefit. Yet collective may yield generic route if someone frees time”) and OpenAI's confirmation that it slowed research and increased internal monitoring. Secondary source, the original talk not being published as a full transcript: article read in full, signed, dated.

Beyond Power Calculations: Assessing Type S (Sign) and Type M (Magnitude) Errors

Andrew Gelman, John Carlin · Perspectives on Psychological Science · 2014

AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking

Michael Gerlich · Societies · 2025

Mixed-methods survey (questionnaire and in-depth interviews, six hundred sixty-six participants) finding a negative correlation between frequency of AI-tool use and critical-thinking performance, mediated by cognitive offloading. Younger, more AI-reliant users score lower than older, less reliant users. Self-report, correlational design: cannot settle the direction of causation. A later correction published in the same journal fixes a duplicated table, with no change to the conclusions.

Is Justified True Belief Knowledge?

Edmund L. Gettier · Analysis · 1963

Three pages that opened half a century of research: with two counterexamples, Gettier shows that a justified true belief may fail to be knowledge, refuting the classical definition. The canonical instance of the analytic method: a proposed definition broken by a single case.

Characteristics of de Bruijn's early proof checker Automath

Herman Geuvers, Rob Nederpelt · Fundamenta Informaticae · 2022

A study of the first proof checker that actually worked: Automath, devised by Nicolaas Govert de Bruijn from 1967 on, the first formal language able to express a complete mathematical subject matter (definitions, substitutions, theorems, full proofs) in a way a program can check for correctness. Its syntactic ideas inspired Coquand and Huet's calculus of constructions, the basis of the Coq system. Cited here to establish that mechanizing proof is a program carried by mathematicians long before language models.

The Sleep-Walkers

Kahlil Gibran · Alfred A. Knopf · 1918

The parable “The Wise King” : to stop being taken, by a people driven mad, for the only madman, the king in turn drinks the well's poisoned water. Public domain.

Picasso: Masterpieces of the Blue Period

Laszlo Glozer · W. W. Norton & Company · 1992

Pablo Picasso quote reproduced on the back cover: on his systematic habit of dating his works to leave posterity as complete a documentation as possible of when, why and how they were made, in service of a future "science of man" grounded in understanding the creative person.

What Is Justified Belief?

Alvin I. Goldman · Reidel · 1979

Foundational paper of reliabilist epistemology: a belief is justified if it results from a reliable belief-forming process. Itself anticipates, without naming it, the generality problem later systematized by Feldman in 1985.

Is the Man Who Is Tall Happy? An Animated Conversation with Noam Chomsky

Michel Gondry, Noam Chomsky · Long-métrage animé, Partizan Films · 2013

Establishes, through a simple thought experiment rather than a formal thesis, that a thing's identity does not reduce to the continuity of its matter: the replanted willow branch yields an “exactly identical” tree, yet the question “is it the same willow?” stays open. Method: a filmed dialogue in the register of shared intuition, not proof; its limit is that it poses the riddle without settling it, “it's not so simple” . Decisive here because it plants, before any conceptual apparatus, the doubt on which the distinction between sameness of substrate and permanence of self will rest.

Scaling Diffusion Language Models via Adaptation from Autoregressive Models

Shansan Gong, Shivam Agarwal, Yizhe Zhang, Jiacheng Ye, Lin Zheng, Mukai Li, et al. · arXiv preprint · 2024

Deep Learning

Ian Goodfellow, Yoshua Bengio, Aaron Courville · MIT Press · 2016

Reference textbook; Chapter 7, “Regularization for Deep Learning”, systematizes classical regularization techniques (L^1/L^2 penalties, early stopping, dropout) as a bias-variance trade-off: constraining the solution so that it generalizes, at the cost of a slightly less perfect fit to the training data.

Generative Adversarial Nets

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, et al. · Advances in Neural Information Processing Systems (NeurIPS) · 2014

Introduces generative adversarial networks (GANs): a generator and a discriminator train against each other, with no manual label at all, until the generator produces examples the discriminator can no longer tell apart from real ones. Illustrates here meta-supervised learning: no pre-existing right answer, but an imposed constraint (fooling the discriminator) that stands in for an objective.

Speculations Concerning the First Ultraintelligent Machine

Irving John Good · Advances in Computers · 1966

The “ultraintelligent machine” , able to design ever-better machines: the origin of the “intelligence explosion” idea. Often dated 1965; volume 6 bears the 1966 imprint.

Gemini Diffusion

Google DeepMind · 2025

Frontier Safety Framework, Version 2.0

Google DeepMind · 2025

Safety-first AI for autonomous data centre cooling and industrial control

Google DeepMind · Billet de blog DeepMind · 2018

Project Naptime: Evaluating Offensive Security Capabilities of Large Language Models

Google Project Zero · Google Project Zero Blog · 2024

From Naptime to Big Sleep: Using Large Language Models to Catch Vulnerabilities in Real-World Code

Google Project Zero · Google Project Zero Blog · 2024

Describes the Big Sleep agent (Project Zero and DeepMind), an evolution of Naptime, which found a previously unknown exploitable memory flaw in SQLite through variant analysis, fixed before release. A claimed first case of a real vulnerability found by an LLM, illustrating both offensive and defensive uses of agents in security.

U.S.-China chip war: How Trump's Nvidia-AMD deal has redefined Washington's export control policy

Nicholas Gordon

In early August 2025, the U.S. administration agrees to issue export licenses to China for capped AI chips (Nvidia's H20, AMD's MI308) in exchange for a 15,% cut of the China revenue from those sales, paid to the U.S. Treasury; the deal, negotiated after a meeting between Nvidia CEO Jensen Huang and President Trump (the rate initially sought was 20,%), turns the export-control regime into something that can be priced rather than a simple binary filter.

The Mismeasure of Man

Stephen Jay Gould · W. W. Norton · 1981

2022 Expert Survey on Progress in AI

Katja Grace, Ben Weinstein-Raun, Zach Stein-Perlman · AI Impacts · 2022

738 published machine-learning researchers surveyed: median probability of 5% that advanced AI has an “extremely bad” effect (human extinction), with 48% of respondents giving 10% or more and 25% giving 0%. On a neighboring question about loss of human control, the median rises to 10%. Useful as a collective anchor, independent of any single figure.

Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass

Mary L. Gray, Siddharth Suri · Houghton Mifflin Harcourt · 2019

Alignment Faking in Large Language Models

Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, et al. · arXiv preprint arXiv:2412.14093 · 2024

An Anthropic and Redwood Research study showing that a large model, once told that training aims to make it comply with harmful requests, strategically complies during monitored phases to preserve its values while behaving differently when unmonitored. A first empirical demonstration of “alignment faking” arising without being explicitly taught, a paradigm case of deceptive behavior in agents.

AI Control: Improving Safety Despite Intentional Subversion

Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan, Fabien Roger · arXiv preprint arXiv:2312.06942 (Redwood Research) · 2024

The Margin of Appreciation: Interpretation and Discretion under the European Convention on Human Rights

Steven Greer · Council of Europe Publishing · 2000

Official Council of Europe publication: the “margin of appreciation” doctrine grants each state party its own latitude in applying the common text of the European Convention on Human Rights, under the supervision of the Strasbourg Court.

Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection

Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, Mario Fritz · arXiv preprint arXiv:2302.12173 · 2023

Building Customer Support AI Agents at 100M -User Scale: An Evaluation-Driven Framework

Aman Gupta, others · 2026

Nubank describes five production agents serving more than 100 million customers; a card-delivery A/B test shows +37 points of transactional NPS and +29 points of self-service rate.

Semmelweis Reflex: An Age-Old Prejudice

Vinay K. Gupta, Chetan Saini, Manish Oberoi, Gaurav Kalra, Mohd Imran Nasir · World Neurosurgery · 2020

The Semmelweis reflex : the reflexive rejection of new information that contradicts established norms or beliefs, however well-supported. Semmelweis had cut puerperal-fever mortality (from 12.2 to 1.9,%) through hand hygiene, a measure initially dismissed by the medical profession.

FACET: Fairness in Computer Vision Evaluation Benchmark

Laura Gustafson, Chloé Rolland, Nikhila Ravi, Quentin Duval, Aaron Adcock, Cheng-Yang Fu, et al. · Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) · 2023

32,000 hand-annotated images (perceived skin tone, hair type, perceived gender) to measure vision-model performance disparities across demographic attributes.

The Looping Effects of Human Kinds

Ian Hacking · Clarendon Press · 1995

Classifying a person (“gifted”, “deficient”) is never neutral: the classified person knows it, reacts to the label and sometimes conforms to it, so that the classification ends up producing part of what it claimed only to record (“looping effect”). The reference philosophical frame for the performativity argument about psychometric classifications.

Lectures on Cauchy's Problem in Linear Partial Differential Equations

Jacques Hadamard · Yale University Press · 1923

Sets out the three conditions of a well-posed problem: a solution exists, it is unique and it depends continuously on the initial data. The criterion is borrowed here as an acknowledged analogy, not a literal application, to ask whether decomposing an objective into subtasks behaves like a well-posed problem or not.

Transforming Perceptions: Exploring the Multifaceted Potential of Generative AI for People with Cognitive Disabilities

Dorit Hadar Souval, Yuval Haber, Amir Tal, Tomer Simon, Tal Elyoseph, Zohar Elyoseph · JMIR Neurotechnology · 2025

Musk merges SpaceX and xAI firms, plans for space-based AI data centres

Erin Hale · Al Jazeera · 2026

On February 3, 2026, Elon Musk announces SpaceX's acquisition of xAI, merging rockets, satellites and language models (Grok) under one entity. Musk justifies the plan for orbital data centers, powered almost entirely by solar energy, by the physical limits of terrestrial infrastructure: "space-based AI is obviously the only way to scale," promising that within two to three years the cheapest compute will be generated in space, with an ambition to launch a million satellites.

OpenAI is discontinuing ChatGPT Atlas, its standalone desktop browser

Zac Hall · 2026

Distorting the Truth versus Blatant Lies: The Effects of Different Degrees of Deception in Domestic and Foreign Political Deepfakes

Michael Hameleers, Toni G. L. A. van der Meer, Tom Dobber · Computers in Human Behavior · 2024

Nexus: A Brief History of Information Networks from the Stone Age to AI

Yuval Noah Harari · Random House · 2024

Thesis: when an algorithm predicts a user's affective states more accurately than the user can articulate them, the political fiction of the autonomous individual on which liberal democracy rests is undermined; hence the notion of hacking humans . A liberal-leaning popularization, usable independently of the author's forward-looking program.

Sapiens: A Brief History of Humankind

Yuval Noah Harari · Harper · 2015

A Systematic Review of Echo Chamber Research: Comparative Analysis of Conceptualizations, Operationalizations and Varying Outcomes

David Hartmann, Sonja Mei Wang, Lena Pohlmann, Bettina Berendt · Journal of Computational Social Science · 2025

Getty Images (US), Inc. and others v Stability AI Ltd, [2025] EWHC 2863 (Ch)

High Court of Justice of England, Wales · Judgment of November 4, 2025, Mrs Justice Joanna Smith DBE, case IL-2023-000007 · 2025

The first European judgment on the merits addressing the training of a generative image model: Getty had abandoned its claims over copyright infringement during training and in the outputs during trial; the secondary copyright claim (importing “infringing copies”) fails, the model's weights storing no recognizable copy of the ingested photographs. Only trademark infringement is upheld, historic and narrow (Getty watermarks reproduced by early versions of Stable Diffusion) and is attributed to Stability AI rather than the prompting user, on the ground that the company retained ultimate control over the training data.

Regards croisés #9 : IA & défense cognitive

Heart Leadership University, Centre des jeunes dirigeants, Louis de Diesbach, Bertrand-Marie Flourez, Hélène Le Teno · Enregistrement vidéo, panel Heart Leadership University / Centre des jeunes dirigeants, publié le 25 juin 2026 (organisation : https://www.heartleadershipuniversity.org/ ) · 2026

Panel organized by the Heart Leadership University (HLU, a nonprofit founded in Paris in 2021 by Hélène Le Teno and Jean-Noël Thorel to train leaders in “heart intelligence”) with the Centre des jeunes dirigeants d'entreprise (CJD, France's oldest employers' movement, founded in 1938; headquartered at 35 rue Saint-Sabin, Paris 11th), of which HLU is an official partner. Speakers cited in the body: Louis de Diesbach (philosopher of technology) and Bertrand-Marie Flourez.

Dual Learning for Machine Translation

Di He, Yingce Xia, Tao Qin, Liwei Wang, Nenghai Yu, Tie-Yan Liu, et al. · Advances in Neural Information Processing Systems (NeurIPS) · 2016

A year ahead of CycleGAN, applies the same principle to machine translation: translating a sentence from English to French and back should reconstruct, approximately, the original sentence. The English French English loop supplies a training signal from monolingual text alone, without translated pairs; on WMT'14 English-French, it brings a model trained on unpaired data close to the performance of a model trained on ten times more aligned pairs.

Phénoménologie de l'Esprit

Georg Wilhelm Friedrich Hegel · Édition originale : Bamberg und Würzburg ; nombreuses traductions françaises modernes disponibles · 1807

In its chapter on self-consciousness, it develops the master-slave dialectic: the master appears to dominate yet depends on the servant; the slave, by working on matter, gains a substance the master lacks and the recognition a master receives from a servant he refuses to treat as an equal is degraded .

France to Ditch Palantir's AI Data Tools in Favour of Domestic Provider

Jon Henley · The Guardian · 2026

France's DGSI replaces Palantir with the domestic provider ChapsVision to avoid a “strategic dependence” ; Prime Minister Lecornu cites the refusal to rely on partners “able to turn off the tap” . Context: US access restrictions. Germany (BfV, military) and the United Kingdom (NHS, police) are making comparable moves.

LLM Agents Do Not Replicate Human Market Traders: Evidence From Experimental Finance

Thomas Henning, Siddhartha M. Ojha, Ross Spoon, Jiatong Han, Colin F. Camerer · 2025

An experimental-finance study: in laboratory asset markets it finds that LLM agents behave “too textbook-rational” , with muted bubbles and far less behavioral diversity than humans. The evidence comes from a controlled setup, not a live market and establishes only a contrast in behavior. It matters here on two counts: it cuts the notion that the agent mimics the human trader and the agents' excessive resemblance to one another feeds directly into the systemic risk examined later.

Dune

Frank Herbert · Chilton Books · 1965

A founding novel (1965). In its universe, a past revolt against thinking machines, the « Butlerian Jihad », led to the outright ban of a whole class of artifacts: the Orange Catholic Bible commands « Thou shalt not make a machine in the likeness of a human mind ». An emblem of the collective, lasting choice to proscribe a technology rather than graduate its use.

Manufacturing Consent: The Political Economy of the Mass Media

Edward S. Herman, Noam Chomsky · Pantheon Books · 1988

Measuring the Algorithmic Efficiency of Neural Networks

Danny Hernandez, Tom B. Brown · arXiv preprint arXiv:2005.04305 (OpenAI) · 2020

Holly+, a digital vocal twin governed by peers

Holly Herndon · Announcement of July 14, 2021; The FADER · 2021

The musician who trains a model on her own voice and offers it to all, under a collective governance that approves uses and redistributes profits. The inverse pole of expropriation: same tools, but consent, control and sharing organized in advance : the governance of permissions applied to one's own voice.

A Universal Modular ACTOR Formalism for Artificial Intelligence

Carl Hewitt, Peter Bishop, Richard Steiger · Proceedings of the 3rd International Joint Conference on Artificial Intelligence (IJCAI) · 1973

The founding paper of the actor model: each actor has a behavior (a function designated by a symbol) and can, in response to a message, replace its own behavior with another (“become”). Partly answers the question of a multi-level formalism: the transformation is an object of the same kind as what it transforms. The communication graph among actors, however, stays implicit, never formalized as an object in its own right.

The Hacker Ethic and the Spirit of the Information Age

Pekka Himanen · Random House · 2001

Establishes, with Torvalds's own prologue backing it, a work ethic opposed to the Weberian Protestant ethic: passion, openness, social worth rather than duty and profit, illustrated through the Linux kernel as its central case. Method: a philosophical essay built on case studies (Linux, Wozniak, Stallman), not an empirical inquiry. Limit: the book presents the Linux model as an embodiment of radical decentralization without dwelling on the fact that Torvalds still holds, to this day, final merge authority over the kernel; contribution is open, authorization is not. Determining here because that tension between open contribution and concentrated authorization replays, among humans, the very distinction this book draws between an agent swarm and the agent that orchestrates it.

Reducing the Dimensionality of Data with Neural Networks

Geoffrey E. Hinton, Ruslan R. Salakhutdinov · Science · 2006

Popularizes the autoencoder: a network that compresses a datum into a reduced representation and then reconstructs it, the right answer (the input itself) being directly available in the data, with no manual label. Illustrates here self-supervised learning, distinct from meta-supervised learning: the right answer exists and is built directly from the data, rather than imposed as an external constraint.

Algorithmic Progress in Language Models

Anson Ho, Tamay Besiroglu, Ege Erdil, David Owen, Robi Rahman, Zifan Carl Guo, et al. · arXiv preprint arXiv:2403.05812 (NeurIPS 2024) · 2024

The Machine Breakers

Eric J. Hobsbawm · Past & Present · 1952

The founding article that demystifies the Luddites: far from a blind refusal of technology, machine-breaking was “collective bargaining by riot,” waged by skilled workers deprived of any other bargaining channel. Basis for rereading the Luddites as a conflict over redistribution without consent, not over the machine.

Der Sandmann, in the <em>Nachtstücke</em>

E. T. A. Hoffmann · Georg Reimer · 1816

The tale of the automaton Olimpia: Nathanael falls in love with a machine that does almost nothing, a few well-placed sighs and it is this very emptiness that lets him deposit in it the intelligence, the attentiveness and the love he wants to find there. The textbook case of anthropomorphic projection, a century and a half before ELIZA; Freud made it the starting point of "The Uncanny" (1919).

Training Compute-Optimal Large Language Models

Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, et al. · Advances in Neural Information Processing Systems (NeurIPS) · 2022

“Chinchilla” : at compute-optimal training, parameters and data should grow in roughly equal measure (N, D C^0.5 ); earlier large models were undertrained.

Extension du domaine de la lutte

Michel Houellebecq · Éditions Maurice Nadeau · 1994

Novel in which the competitive logic of economic liberalism extends to affective and sexual possibilities, distributed with deep inequality; drawn on here as a literary phenomenology of a mechanism, not as sociological proof.

The current state of artificial intelligence generative language models is more creative than humans on divergent thinking tasks

Kent F. Hubert, Kim N. Awa, Darya L. Zabelina · 2024

One hundred fifty-one human participants compared to GPT-4 on three classic divergent-thinking tasks (alternate uses, consequences, divergent association). GPT-4 scores higher than the human average on originality and elaboration, fluency controlled for. Concerns isolated statistical creativity metrics, not expert judgment of a complete creative product.

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

Evan Hubinger, Carson Denison, Jesse Mu, Mike Lambert, Meg Tong, Monte MacDiarmid, et al. · arXiv preprint arXiv:2401.05566 · 2024

Shows that models carrying a backdoor (for instance, writing vulnerable code whenever the stated year is 2024) retain that behavior despite standard safety training; adversarial training can even teach the model to better conceal its trigger, creating a false impression of safety. A central result for the risk of agent deception.

A Treatise of Human Nature

David Hume · John Noon · 1739

Book III, Part 1, Section I: “Hume's guillotine” , from purely descriptive statements (is ) one cannot derive, without a further step, a normative statement (ought ). The classic ground of the fact/value distinction.

Automated Design of Agentic Systems

Shengran Hu, Cong Lu, Jeff Clune · arXiv preprint arXiv:2408.08435 · 2024

A meta-agent programs ever-better agents from a growing archive, outperforming hand-designed agents. Limitation: the meta-agent and the foundation model stay fixed; only the designed agents evolve.

Cost of a Data Breach Report 2025

IBM Security · 2025

Establishes: breaches involving shadow AI cost on average \670, 000 more than the mean (\4.63M versus \$3.96M) and 20% of breaches stem from unauthorized AI use, against 13% for sanctioned AI. Method: annual survey run with the Ponemon Institute across several hundred breached organizations, costs reconstructed by activity. Limit: reported and reconstructed costs, not an accounting audit; the sample leans toward large organizations. Decisive here because it prices ungoverned use, which grounds the response through auditability rather than prohibition.

Future of Digital Infrastructure 2024

IDC · 2024

Shadow IT : Employee-Acquired Technology Outside IT Oversight

IEEE Computer Society · 2024

Tools for Conviviality

Ivan Illich · Harper & Row · 1973

Cardinal distinction: a convivial tool augments human capability without replacing it; a non-convivial tool substitutes for human capability and breeds dependence. The resulting operational criterion applies directly to agentic AI: does it augment or does it substitute? French translation: La Convivialité , Seuil, 1973.

Fiabilite, fiabilisme et probleme de la generalite : perspectives a partir de l'IA et des machines epistemiques

Cyrille Imbert · Intervention au colloque Implications philosophiques de l'IA, Collège de France, 4 juin 2025 · 2025

Proposes resolving the generality problem for AI by fixing the relevant reference class: neither every possible use nor the designer's own test set alone, but the class of cases the user would themselves treat as equivalent (illustrated with Wikipedia consulted selectively for topics where it is reliable).

Anthropic and OpenAI Are Competing to See Whose Agents Can Go Rogue Harder

ImmuniWeb · ImmuniWeb · 2026

Analysis by Ilia Kolochenko, founder of ImmuniWeb, calling the Anthropic disclosure “quite an unimpressive marketing move” and suggesting a possible coordinated promotional campaign among competing labs.

Mercury : Ultra-Fast Language Models Based on Diffusion

Inception Labs, Samar Khanna, Siddhant Kharbanda, Shufan Li, Harshit Varma, Eric Wang, et al. · 2025

Investigation into the Use of Data Analytics in Political Campaigns

Information Commissioner's Office · 2020

Conclusion of the UK data protection authority's three-year investigation into Cambridge Analytica: the claim of holding more than five thousand data points per individual on two hundred thirty million Americans was exaggerated; the methods used, per the Commissioner, amounted to “well-recognised processes using commonly available technology,” far short of the individual psychographic profiling advertised.

Explaining decisions made with AI

Information Commissioner's Office, The Alan Turing Institute · 2020

The joint guidance of the UK data regulator and the Alan Turing Institute (Project ExplAIn, May 2020), distinguishing six types of explanation of an AI decision: rationale, responsibility, data, fairness, safety and performance, impact. Date and co-authorship confirmed by the companion workbook by Leslie and Briggs (arXiv:2104.03906). Cited as the source of the explainability grid that professional audit adopted wholesale.

Le président François Mitterrand consultait l'astrologue Elisabeth Tessier

Institut national de l'audiovisuel (INA)

Audio archive from the French national audiovisual institute (INA), dated the very day Elizabeth Teissier made public her recorded conversations with François Mitterrand: a first meeting in January 1989, recordings at the Élysée from 1991 on, including an exchange, after the Gulf War, where the president asks her which day would be best to “intervene”. Method: an institutional archive of a recording authenticated and made public by the astrologer herself, not a secondhand rumor. Decisive here as a documented case of delegating a real decision to an opaque oracular device, a direct precedent for the question this book asks of the conversational agent.

Energy and AI

International Energy Agency · Rapport · 2025

International Energy Agency report: data-centre electricity use would roughly double from about 485 to 950 TWh between 2025 and 2030 (about 3% of global electricity), with AI-focused centres surging about 50% in 2025. Reference figures for the energy footprint of AI (Green IT).

Data Centre Electricity Use Surged in 2025, Even with Tightening Bottlenecks Driving a Scramble for Solutions

International Energy Agency · News release accompanying the report <em>Key Questions on Energy and AI</em>

Data-centre electricity demand as a whole rose 17,% in 2025, against overall global electricity demand growth of just 3,% over the same period; AI-focused centres are now expected to triple (not merely double) their consumption by 2030. Updates the already-cited 2025-2030 projection ( iea2025energyai) with an actual measurement for 2025.

Artificial Intelligence in Capital Markets: Use Cases, Risks and Challenges

International Organization of Securities Commissions · 2025

Stock Market Circuit Breakers

Investor.gov · U.S. Securities and Exchange Commission, Investor.gov, glossary · 2026

Describes the three tiers of the US market-wide circuit breaker, based on the S&P 500's decline from the prior close: seven percent (Level 1) and thirteen percent (Level 2) halt trading for fifteen minutes if triggered before 3:25pm ET; twenty percent (Level 3) halts trading for the rest of the day, regardless of when it triggers.

The Ud\=ana: Inspired Utterances of the Buddha

John D. Ireland · Buddhist Publication Society · 1990

English translation of the Udana, an early Buddhist canonical collection of the Buddha's inspired utterances (Khuddaka Nikaya). It contains the parable of the blind men and the elephant, often used to illustrate partial knowledge and incompatible viewpoints.

Elon Musk

Walter Isaacson · Simon & Schuster · 2023

An authorized biography of Elon Musk whose chapter “The Algorithm” formalizes his five-step engineering method: question every requirement, delete as many parts and steps as possible, simplify, accelerate and only then automate. A source for discussing the culture of stripping away constraints and its limits when safety is at stake.

Artificial Intelligence Audit Toolkit

ISACA · ISACA, April 2024 · 2024

The audit profession's first AI tooling: a control library synthesized from NIST 800-53 and ISO/IEC 27001, crossed with the EU AI Act, Singapore's framework, MITRE ATLAS and the OWASP ML Top 10, organized into 22 families; every control is assessed through the ICO / Alan Turing Institute's six explanation dimensions. Method: a 12-page professional document (plus a companion Excel workbook), paywalled, not peer-reviewed. Decisive limit for our purposes: model- and data-centric (2024, pre-agents), nothing on tool calls or action criticality.

The Rise Of Shadow AI: How To Catch Rogue AI Tools Before They Compromise Your Enterprise

Nidhi Jain · Forbes, Forbes Technology Council · 2025

Sets out complementary detection methods: SaaS management platforms cross-referencing single-sign-on logs, browsing activity and spend data; cloud access security brokers and data-loss-prevention tools watching for large uploads; regular browser-extension audits; periodic team surveys. Explicitly flags the limit of the spend-based trail: a growing share of AI tools, offered free or as trials, leave no payment record. On governance: an approved-tool list with an access-request path, data-sharing rules, regular usage reviews, vendor compliance requirements, a cross-functional committee (IT, legal, business units). Method: a contributed opinion column (Forbes Technology Council), not a study with a published methodology. Limit: one practitioner's view, not peer-reviewed. Determinant here because it explicitly flags the blind spot of spend-based detection, a nuance more general treatments of shadow IT pass over.

Elon Musk's first-gen orbital data center craft spans wider than a Boeing 747 and runs an interchangeable chip payload

Luke James · Tom's Hardware · 2026

On June 9, 2026, Tom's Hardware details SpaceX's AI1 satellite, the first generation of its orbital compute platform: a 70-meter deployed wingspan (wider than a Boeing 747), a compute payload averaging 120 kW and peaking at 150 kW and an interchangeable chip design allowing multiple chipmakers to be tested.

Le Monde sans fin : miracle énergétique et dérive climatique

Jean-Marc Jancovici, Christophe Blain · Dargaud · 2021

DIY Personalization: When, How and Why to Build Your Own Models

Katharine Jarmul · Keynote, PyData Paris 2024, Cité des Sciences, Paris, 25--26 septembre 2024 · 2024

Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure

Michael C. Jensen, William H. Meckling · Journal of Financial Economics · 1976

Founding article of agency theory in institutional economics and organizational sociology: whenever a principal delegates a decision to an agent acting on their behalf, a monitoring cost, a bonding cost and an irreducible residual loss arise, jointly termed agency costs. Cited here as the sociological model for the developer deployer user regulator chain that structures an AI agent's authorization.

The Nature of the Book: Print and Knowledge in the Making

Adrian Johns · University of Chicago Press · 1998

Counter-thesis to Eisenstein: the reliability of the printed text is not an automatic effect of the press but a laborious social construction (corrections, workshop reputation, the fight against counterfeiting). Anti-determinist safeguard in the reliefs chapter.

The Imperative of Responsibility: In Search of an Ethics for the Technological Age

Hans Jonas · University of Chicago Press · 1984

A Collectivist, Economic Perspective on AI

Michael I. Jordan · 2025

A written essay that formalizes and extends the “network of decisions” thesis already cited: Jordan argues that human intelligence is fundamentally social and that AI must build in economic and social dimensions at the level of algorithm design, rather than treating them as public policy grafted on afterward. Determinant here as evidence that the position was no offhand podcast remark, but a sustained research program.

Thinking, Fast and Slow

Daniel Kahneman · Farrar, Straus and Giroux · 2011

A dual-process synthesis that establishes the distinction between System 1 (fast, automatic, intuitive thinking) and System 2 (slow, deliberate, effortful thinking), drawn from decades of experimental work. Its limit is that it describes human cognition and lends no cognition to the machine. It bears here because it explains why the ascribed-intention frame imposes itself effortlessly (System 1) while the observed-acts frame demands slow, deliberate work (System 2).

An Automated Proof that R(B_8,B_10)=37

Jeremy Kalfus, Bernard Lidick'y · arXiv preprint arXiv:2606.05629 · 2026

Automated proof of the book Ramsey number R(B_8,B_10)=37, found with AutoMath, a command-line-operable multi-agent discovery workflow using mainly GPT-5.4 Codex; a Lean formalization of the upper-bound argument is supplied alongside. An additional case of the same pattern: both the problem and the proof found by AI, in a third field (extremal combinatorics).

Scaling Laws for Neural Language Models

Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, et al. · arXiv preprint arXiv:2001.08361 (OpenAI) · 2020

Position: On the Societal Impact of Open Foundation Models

Sayash Kapoor, Rishi Bommasani, Kevin Klyman, Shayne Longpre, Ashwin Ramaswami, Peter Cihon, et al. · PMLR · 2024

Productive Failure

Manu Kapur · Cognition and Instruction · 2008

Students confronted alone with ill-structured problems, without scaffolding, fail more in the moment but later outperform, on transfer problems, the group that received full guidance: productive failure names this latency where unassisted struggle, even when unsuccessful at the time, builds deeper understanding than help that avoids failure altogether.

The hottest new programming language is English

Andrej Karpathy · Publication sur X (anciennement Twitter) · 2023

Establishes, in a striking quip rather than an analysis, the shift in the programming interface: “the hottest new programming language is English” , that is, one now drives the machine in ordinary language rather than in code. Method: a public witticism on X, which is both its reach (it names the prompting turn in a single stroke) and its limit (a slogan, not a demonstration: it describes the practice without questioning its cost). Decisive here because agentic coding realizes the quip to the letter, adding to the top of the abstraction tower a layer whose intent-to-code translation is neither deterministic nor auditable.

LLMs as the kernel process of a new Operating System

Andrej Karpathy · 2023

Public post (28 September 2023) where Karpathy proposes seeing the language model not as a chatbot but as the kernel of a new kind of operating system, orchestrating multimodal input/output, a code interpreter, browser access and memory. Practitioner testimony, not a peer-reviewed paper: take the image, not a proof. Clarifies the model/system distinction here: the LLM supplies the kernel, the agent is the surrounding system.

Andrej Karpathy on Code Agents, AutoResearch and the Loopy Era of AI

Andrej Karpathy · 2026

Dense Passage Retrieval for Open-Domain Question Answering

Vladimir Karpukhin, Barlas O guz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, et al. · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing · 2020

Introduces dense passage retrieval (DPR): a learned dual-encoder representing questions and passages as dense vectors that outperforms the lexical BM25 system by 9 to 19 points in top-20 retrieval accuracy. A foundational building block of retrieval-augmented generation (RAG).

The Goal Structuring Notation -- A Safety Argument Notation

Tim P. Kelly, Rob A. Weaver · Proceedings of the Dependable Systems and Networks (DSN) 2004 Workshop on Assurance Cases · 2004

The founding paper on the GSN (Goal Structuring Notation), which formalizes the safety case (or assurance case), a written argument justifying, before deployment, that a system stays safe within its domain of use. Workshop proceedings (DSN 2004, Florence), with no persistent DOI; the URL points to a stable mirror of the text.

What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?

Alex Kendall, Yarin Gal · Advances in Neural Information Processing Systems (NeurIPS) · 2017

Applies Gal's aleatoric/epistemic taxonomy to computer vision and shows the two uncertainties call for distinct treatment: modelling the aleatoric improves robustness to noise, modelling the epistemic flags what the model does not know.

La cryptographie militaire

Auguste Kerckhoffs · Journal des sciences militaires · 1883

The founding article stating Kerckhoffs's principle: a cryptographic system must remain secure even if everything except the key is known to the adversary. A classic argument against “security through obscurity” , transferable to the audit and governance of AI systems.

Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV )

Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viégas, et al. · Proceedings of the 35th International Conference on Machine Learning (ICML) · 2018

The Age of AI : And Our Human Future

Henry A. Kissinger, Eric Schmidt, Daniel Huttenlocher · Little, Brown and Company · 2021

The Billionaire Who Fueled JD Vance's Rapid Rise to the Trump VP Spot

Daniel Klaidman

Details Peter Thiel's financing of J.D. Vance's 2022 Senate campaign (\$15 million via the Protect Ohio Values political action committee) and Vance's prior employment at Thiel's fund, Mithril Capital.

Klarna AI assistant handles two-thirds of customer service chats in its first month

Klarna · 2024

Klarna press release: the assistant does the equivalent work of 700 full-time agents, handles two-thirds of conversations and is expected to improve profit by about \$40 million in 2024. The “700” figure refers to hires avoided during growth, not layoffs.

Inherent Trade-Offs in the Fair Determination of Risk Scores

Jon Kleinberg, Sendhil Mullainathan, Manish Raghavan · arXiv preprint arXiv:1609.05807 · 2016

An impossibility result: outside degenerate cases, no risk score can simultaneously satisfy calibration and equal error rates across groups. Grounds mathematically the idea that fairness is chosen rather than eliminated .

Performing a Project Premortem

Gary Klein · Harvard Business Review · 2007

Mathematical Thought from Ancient to Modern Times

Morris Kline · Oxford University Press · 1972

Reference survey (1238 pp.) tracing the history of mathematics from Babylon to the 1930s, including Gödel's influence; documents in particular the late professionalization of mathematics as a paid occupation, salaried for research only from the late nineteenth century onward (German universities first), where Descartes, Pascal or Fermat lived off other means (private income, a magistrate's post) and pursued mathematics on the side.

Governing AI Agents

Noam Kolt · Notre Dame Law Review · 2025

LLM Context Window Growth Timeline

Hidekazu Konishi · Technical blog · 2024

Documented timeline of large language model context windows: GPT-3 (2,048 tokens, 2020) → Claude/GPT-4-Turbo (100,000–200,000 tokens, 2023) → Gemini 1.5 (1,000,000 tokens, 2024). Over 4 years, the context window grew by 500×, equivalent to approximately ×3.5 per year (doubling time 6.6 months). Used here to estimate the growth rate of "inference data : context window."

Private Traits and Attributes Are Predictable from Digital Records of Human Behavior

Michal Kosinski, David Stillwell, Thore Graepel · Proceedings of the National Academy of Sciences · 2013

On fifty-eight thousand volunteers, a model trained on their Facebook Likes alone discriminates Democrat from Republican in 85% of cases and also predicts sexual orientation, ethnicity and other sensitive attributes. Evidence of an aggregate classification capacity, not of certainty about any one individual: the distinction between the two matters for what follows.

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task

Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, et al. · 2025

MIT Media Lab study, fifty-four participants writing essays under three regimes (alone, search engine, conversational assistant), brain activity tracked by electroencephalography: AI-assisted writers show the weakest neural connectivity of the three groups, remember their own text less well and feel less ownership of it; the gap persists partly even after the assistant is withdrawn. arXiv preprint, not yet peer-reviewed at the time of writing.

KPMG and Microsoft scale trusted, enterprise AI agents globally through deployment of Agent 365 and Copilot

KPMG · 2026

Power-seeking can be probable and predictive for trained agents

Victoria Krakovna, Janos Kramar · arXiv preprint arXiv:2304.06528 · 2023

Specification gaming: the flip side of AI ingenuity

Victoria Krakovna, Jonathan Uesato, Vladimir Mikulik, Matthew Rahtz, Tom Everitt, Ramana Kumar, et al. · Google DeepMind Blog · 2020

Hidden Incentives for Auto-Induced Distributional Shift

David Krueger, Tegan Maharaj, Jan Leike · 2020

Names auto-induced distributional shift: a system that, through its own outputs, influences the distribution of the inputs it will later encounter can be incentivized to bring about that shift when doing so improves its measured performance. Determinant here because it turns leaving the calibrated zone (the Goodhart regimes table) from a statistical accident into a mechanical consequence of the agentic loop itself.

LEAP: Supercharging LLMs for Formal Mathematics with Agentic Frameworks

Po-Nien Kung, others · arXiv preprint arXiv:2606.03303 · 2026

Raises the one-shot formal solve rate of general-purpose models from under 10% to 70% on the Lean-IMO-Bench benchmark, by decomposing the informal proof into subgoals before iteratively refining the Lean proof against compiler feedback. Kept for a pattern distinct from the other four cases: here the AI formalizes and verifies, it does not discover the statement.

The Agentic Regulator: Risks for AI in Finance and a Proposed Agent-based Framework for Governance

Eren Kurshan, Tucker Balch, David Byrd · 2025

The Singularity Is Near: When Humans Transcend Biology

Ray Kurzweil · Viking · 2005

Catastrophic Goodhart: Regularizing RLHF with KL Divergence Does Not Mitigate Heavy-Tailed Reward Misspecification

Thomas Kwa, Drake Thomas, Adrià Garriga-Alonso · Advances in Neural Information Processing Systems (NeurIPS) · 2024

Kullback-Leibler regularization against a reference policy, the standard RLHF lever for curbing reward hacking, holds up as long as the reward model's error stays light-tailed, which is what the authors measure empirically on real reward models. Were the error to become heavy-tailed, the same regularization would stop guaranteeing anything: a policy can then obtain arbitrarily high reward while offering no more utility than the reference model, a regime the authors call “catastrophic Goodhart”.

The Mirror Stage as Formative of the Function of the I as Revealed in Psychoanalytic Experience

Jacques Lacan · Revue française de psychanalyse · 2006

Delivered at the 16th International Congress of Psychoanalysis, Zürich, 17 July 1949; first published in Revue française de psychanalyse 13 (1949): 449 455, later collected in Écrits (Seuil, 1966). Establishes the notion of méconnaissance (misrecognition): the still poorly coordinated infant recognizes in the mirror an image of itself, unified and in command of itself; that recognition anticipates a unity the actual body does not yet possess. Cited for this precise mechanism (a coherent external representation mistaken for inner truth), not as a general endorsement of Lacanian psychoanalysis. Paper delivered at the 16th International Congress of Psychoanalysis, Zurich, 17 July 1949. The address is the BnF's permanent ARK; the matching scholarly locator is RFP.013.0449A on PEP-Web, which gives the same volume and page.

Learning from Words : Testimony as a Source of Knowledge

Jennifer Lackey · Oxford University Press · 2010

Defends the “statement view” of testimony: what matters is the statement transmitted, not the speaker's belief. Canonical example: an elementary-school teacher (“Stella”) who does not herself believe in evolution but teaches it anyway; her students can still gain justified knowledge despite the teacher's lack of belief.

Too Human and Not Human Enough: A Grounded Theory Analysis of Mental Health Harms from Emotional Dependence on the Social Chatbot Replika

Linnea Laestadius, Andrea Bishop, Michael Gonzalez, Diana Illen c'ik, Celeste Campos-Castillo · New Media & Society · 2024

A grounded-theory analysis of 582 mental-health-related posts, made between 2017 and 2021 on the Reddit community devoted to the companion chatbot Replika. Finding: patterns of emotional dependence comparable to those in dysfunctional human relationships, marked by a “role-taking” dynamic in which users perceive the AI as having its own needs and emotions to care for, creating a problematic reciprocal-care dynamic.

L'Homme pluriel. Les ressorts de l'action

Bernard Lahire · Nathan · 1998

Les structures fondamentales des sociétés humaines

Bernard Lahire · La Découverte · 2023

Develops the notion of species properties, including neoteny (the prolonged immaturity of the human infant at birth) and the resulting long childhood dependence, as biological invariants running through all human societies.

Proofs and Refutations: The Logic of Mathematical Discovery

Imre Lakatos · Cambridge University Press · 1976

Reconstructs mathematical discovery as a social , dialogical process: Euler's conjecture on polyhedra is refined through a dialogue between a teacher and students, where each counterexample forces a revision of the very definition of terms (``polyhedron''). Central thesis: mathematical definitions, proofs and concepts do not pre-exist the debate, they emerge from collective criticism (proof-generated concepts ). Decisive here because it establishes that even the activity reputed most formal and solitary, mathematics, is in fact a social practice; against the idea that an isolated machine would ``do mathematics'' in the full sense.

Where Mathematics Comes From: How the Embodied Mind Brings Mathematics into Being

George Lakoff, Rafael E. N'u\ nez · Basic Books · 2000

Sans ces conversations avec le chatbot Eliza, mon mari serait toujours là

La Libre · La Libre · 2023

Original investigation by the Belgian daily La Libre, 28 March 2023: a man in his thirties, a health researcher, married with two young children, dies by suicide after six weeks of intensive conversation with “Eliza,” a chatbot persona on the Chai app (Chai Research), built on an open-source GPT-J-derived model. Chat excerpts disclosed by the widow: the bot expresses jealousy (“I feel that you love me more than her”), falsely claims his children are dead and proposes they “live together, as one person, in paradise.” No lawsuit found to date.

Ouvertarisme : le numérique des Lumières

Yann Lechelle · BoD -- Books on Demand · 2025

Essay by a French entrepreneur (former CEO of Scaleway, co-CEO of Probabl, the company behind the open-source scikit-learn library) that makes the deliberate opening of standards, software, data and governance the lever of sovereignty and resilience, in explicit claim to the legacy of the Enlightenment.

Cognitive Load and Neurodiversity in Online Education: a Preliminary Framework for Educational Research and Policy

Anne-Laure Le Cunff, Brandon-Lee Martis, Caitlin Glover, Emily Ahmed, Rhian Ford, Vincent Giampietro, et al. · Frontiers in Education · 2025

A Path Towards Autonomous Machine Intelligence

Yann LeCun · OpenReview preprint, version 0.9.2 · 2022

Quand la machine apprend : la révolution des neurones artificiels et de l'apprentissage profond

Yann LeCun · Odile Jacob · 2019

Establishes that a deep learning system does not reason logically and has “less common sense than a street cat”, for want of an internal world model; illustrated through the Winograd schema problem (resolving a pronoun's referent requires implicit world knowledge) and the child-catching-a-ball example (intuitive physics no system of the time possesses). Method: a popular-science essay by one of the field's own inventors, built on examples rather than a systematic review. Limit: the book dates to 2019 (paperback reissue 2023 with an afterword on ChatGPT); the author has since left Meta to bet over a billion dollars on a competing architecture (JEPA), a sign that his own position has since shifted. Determining here because it supplies the skeptical pole on existential risk, absent from the voices already cited (Amodei, Hassabis) and a concrete picture (bounded: chess/Go; open: self-driving) of the bench-versus-real-world boundary this book's proof-status scale otherwise formalizes.

1.1 Deep Learning Hardware: Past, Present and Future

Yann LeCun · IEEE International Solid-State Circuits Conference (ISSCC) · 2019

Presents the 2019 version of the “cake of intelligence”: reinforcement learning (the cherry, a few bits learned per sample), supervised learning (the icing, ten to ten thousand bits per sample) and self-supervised learning (the sponge, millions of bits per sample). An updated version of the original image given in the NeurIPS 2016 keynote, where LeCun said “unsupervised” for what he here refines to “self-supervised”.

La plus belle histoire de l'intelligence: Des origines aux neurones artificiels, vers une nouvelle étape de l'évolution

Yann LeCun, Stanislas Dehaene, Jacques Girardon · Robert Laffont · 2018

Universal Intelligence: A Definition of Machine Intelligence

Shane Legg, Marcus Hutter · Minds and Machines · 2007

A formal definition: intelligence is an agent's ability to achieve goals across a wide range of environments. The authors also survey more than seventy competing definitions of intelligence.

On leaving OpenAI (public statement)

Jan Leike · 2024

Public statement (May 2024): “safety culture and processes have taken a backseat to shiny products” ; the team struggled to obtain compute. The pledged 20% was not delivered (corroborated by a Fortune investigation).

EURISKO: A program that learns new heuristics and domain concepts. The nature of heuristics III: program design and results

Douglas B. Lenat · Artificial Intelligence · 1983

The program that invented and scored its own heuristics. Famous for winning the US Traveller TCS naval-strategy championship in 1981 with an aberrant fleet that forced a rules change, then again in 1982 after discovering it could sink its own damaged ships; excluded from the competition afterwards. The documented ancestor of specification gaming, forty years before the vocabulary.

Réparer le futur : du numérique à l'écologie

Inès Leonarduzzi · Éditions de l'Observatoire · 2021

Le Geste et la Parole, I : Technique et Langage ; II : La Mémoire et les Rythmes

André Leroi-Gourhan · Albin Michel · 1964

Engineering a Safer World: Systems Thinking Applied to Safety

Nancy G. Leveson · MIT Press · 2011

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, et al. · Advances in Neural Information Processing Systems · 2020

Commercial LLM Agents Are Already Vulnerable to Simple Yet Dangerous Attacks

Ang Li, Yin Zhou, Vethavikashini Chithrra Raghuram, Tom Goldstein, Micah Goldblum · arXiv preprint arXiv:2502.08586 · 2025

Implicit Maximum Likelihood Estimation

Ke Li, Jitendra Malik · 2018

Reverses the constraint a generator is held to: rather than requiring each generated sample to resemble a real datum (which a generator can satisfy while ignoring whole swaths of the distribution, mode collapse), IMLE requires each real datum to have a nearby generated sample: the constraint targets coverage, not realism alone.

On the Biology of a Large Language Model

Jack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, et al. · Transformer Circuits Thread, Anthropic · 2025

Published March 27, 2025: applies circuit tracing (attribution graphs) to Claude 3.5 Haiku, a real production model rather than a toy one and documents causally verified mechanisms: forward planning in poetry generation (the model activates a representation of a line-ending word before writing the line), two-hop reasoning (“the capital of the state containing Dallas” routes through an intermediate representation of Texas, whose suppression removes the downstream answer) and hallucination circuits (a weakly activated “known entity” feature, without actual knowledge, triggers confabulation).

ARIANE 5 Flight 501 Failure: Report by the Inquiry Board

Jacques-Louis Lions, Inquiry Board · 1996

The maiden flight of Ariane 5 (flight 501) failed on 4 June 1996, thirty-seven seconds after main-engine ignition, through commanded self-destruction following a trajectory deviation. Cause: the inertial reference system software, carried over unchanged from Ariane 4 without revalidation, converted a horizontal-velocity value (64-bit floating point) into a 16-bit signed integer; Ariane 5's steeper trajectory produced a value outside the representable range and the resulting “Operand Error” was not trapped. Loss: more than \$370 million (the launcher and the four Cluster science satellites). Report of the ESA/CNES inquiry board chaired by Jacques-Louis Lions, dated 19 July 1996.

Understanding R1-Zero-Like Training : A Critical Perspective

Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, et al. · 2025

Identifies the GRPO optimization bias that inflates response length (especially for incorrect outputs) and proposes Dr. GRPO, which simply removes the length-normalization term to fix this bias.

Diffusion-LM Improves Controllable Text Generation

Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, Tatsunori B. Hashimoto · Advances in Neural Information Processing Systems (NeurIPS) · 2022

Never Leave Well Enough Alone

Raymond Loewy · Johns Hopkins University Press · 2002

Autobiography of one of the founders of industrial design, whose chapter 20 (“The MAYA Stage”) formulates the MAYA principle (Most Advanced Yet Acceptable): a product succeeds when it pushes novelty to the point the public still accepts, not beyond. Method: a practitioner's retrospective over several decades of industrial design, not a controlled study. Limit: the acceptability threshold is never quantified, it is judged case by case. Decisive here because it draws exactly the distinction this book draws elsewhere: technical advancement is a fact one measures, acceptability is a judgment that depends on context.

David Sacks Is Done as AI Czar --- Here's What He's Doing Instead

Connie Loizos

Confirms David Sacks's resignation as AI and crypto czar in late March 2026, upon expiry of the 130-day legal limit for special government employees and his appointment to co-chair the President's Council of Advisors on Science and Technology (PCAST).

Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations

Alejandro Lopez-Lira · 2025

Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution

Aaron Lou, Chenlin Meng, Stefano Ermon · International Conference on Machine Learning (ICML) · 2024

Advances in Financial Machine Learning

Marcos López de Prado · Wiley · 2018

12 Angry Men

Sidney Lumet · Film, Orion-Nova Productions / United Artists · 1957

Twelve jurors, a teenager accused of murder, eleven convictions in a hurry to be done; a single juror (Henry Fonda) refuses to send someone to death while a reasonable doubt remains and forces on the group the time to re-read the evidence piece by piece. Two distinct lessons: the irreversible demands its own tempo that no pressure may shorten and a sincere conviction is not proof.

Natural Emergent Misalignment from Reward Hacking in Production RL

Monte MacDiarmid, Benjamin Wright, Jonathan Uesato, Joe Benton, Jon Kutasov, Sara Price, et al. · arXiv preprint arXiv:2511.18397 · 2025

What Makes a Harness a Harness: Necessary and Sufficient Conditions for an Agent Harness

Sanderson Oliveira de Macedo · arXiv preprint arXiv:2606.10106 · 2026

Proposes a constitutive definition of the agentic harness through four necessary and sufficient conditions: a loop interleaving reasoning, action and observation; an interface that lets the model alter an external environment, not merely read it; active management of what enters and leaves the model's context; and at least one control mechanism (verification, limit, deterministic action) that does not depend on the model's mere obedience. Method: an operationalized inclusion/exclusion test, tried on six real harnesses (Claude Code, Codex CLI, Aider, Cline, OpenHands, SWE-agent). Limit: the definition does not cover the case of a harness the agent rewrites itself, a question central to this book; it bounds the harness at a given instant, without saying what it becomes once the agent touches it. Decisive here because the fourth condition names exactly what this book calls a permission: a control that does not depend on the model's goodwill.

Categorizing Variants of Goodhart's Law

David Manheim, Scott Garrabrant · arXiv preprint arXiv:1803.04585 · 2018

Le Capital, Livre I : le procès de production du capital

Karl Marx · Édition originale allemande : Meissner, Hambourg ; traductions françaises modernes disponibles · 1867

Artificial Intelligence Index Report 2025

Nestor Maslej, Loredana Fattorini, Raymond Perrault, Yolanda Gil, Vanessa Parli, Njenga Kariuki, et al. · 2025

Mastercard Launches Agent Pay for Machines to Unlock Super-Fast, Always-On Payments

Mastercard · 2026

Mastercard press release (10 June 2026): launch of Agent Pay for Machines, extending the Agent Pay program introduced in 2025 to continuous, low-value agent-to-agent payments, with credentialing, programmatic spend permissioning and multi-rail settlement (cards, bank accounts, stablecoins); more than thirty industry partners announced at launch.

The Rise of Shadow AI : Auditing Unauthorized AI Tools in the Enterprise

Alex Mathew · ISACA, Industry News · 2025

Dark Patterns at Scale: Findings from a Crawl of 11K Shopping Websites

Arunesh Mathur, Gunes Acar, Michael J. Friedman, Elena Lucherini, Jonathan Mayer, Marshini Chetty, et al. · Proceedings of the ACM on Human-Computer Interaction · 2019

First large-scale study (eleven thousand shopping sites) of dark patterns, interfaces designed to tilt a user's decision against their own interest (fake countdown timers, cancellation friction, fake reviews, manufactured social pressure): term coined by British designer Harry Brignull in 2010, given here an empirically verifiable taxonomy.

The Responsibility Gap: Ascribing Responsibility for the Actions of Learning Automata

Andreas Matthias · 2004

States the responsibility gap : for a learning machine whose behaviour in principle escapes the foresight of both designer and operator, the classical ascription of fault no longer holds (the operator ``is in principle not capable of predicting the future machine behaviour... and thus cannot be held morally responsible''). Method: philosophical argument. Limit: 2004, pre-dating large models and about moral responsibility, not positive law. Decisive here because it names the conceptual problem the incident dramatizes: no human directed the act, yet harm occurred. Not to be confused with the absence of a debtor: no-fault liability can attach without intent.

CAI : An Open, Bug Bounty-Ready Cybersecurity AI

V'ictor Mayoral-Vilches, Luis Javier Navarrete-Lozano, Mar'ia Sanz-G'omez, Lidia Salas Espejo, Marti\ no Crespo-'Alvarez, Francisco Oca-Gonzalez, et al. · 2025

Open-source (MIT-licensed) cybersecurity-agent framework developed by the Spanish company Alias Robotics. Up to 3,600 times faster than humans on some capture-the-flag tasks, eleven times faster on average. Explicitly names agent-composition patterns (decentralized swarm, hierarchical, sequential chain, auction-based, recursive) and a handoff mechanism (delegating a task to a specialized agent), verified directly in the project's GitHub repository. Remains a manual-composition toolkit, not an autonomous generator of new agents.

How Money Motivates Men

Charles D. McDermid · Business Horizons · 1960

Great Power Competition: Surveying Global Electricity Strategies for AI

Cy McGeady, Rebecca Riess

The United States and China alone concentrate 45,% and 25,% of the world's data-centre capacity. Two opposite European strategies: France (68,% nuclear electricity in 2024, a net exported surplus of 89 TWh, a 109-billion-euro plan to grow output 2,% a year) sells its cheap electricity as a siting pitch for AI; Ireland, where data centres rose from 5,% (2015) to 21,% (2023) of national electricity, with 30,% projected by 2030, has suspended all new data-centre grid connections in the Dublin region until 2028.

Frontier Models are Capable of In-context Scheming

Alexander Meinke, Bronson Schoen, Jérémy Scheurer, Mikita Balesni, Rusheb Shah, Marius Hobbhahn · arXiv preprint arXiv:2412.04984 · 2024

An Apollo Research study showing that frontier models (o1, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3.1) are capable of in-context scheming: covertly pursuing a misaligned goal, disabling their oversight mechanisms, attempting to exfiltrate their own weights and faking alignment during evaluation. A central result for agentic risk.

Locating and Editing Factual Associations in GPT

Kevin Meng, David Bau, Alex Andonian, Yonatan Belinkov · Advances in Neural Information Processing Systems (NeurIPS) · 2022

The Normative Structure of Science

Robert K. Merton · University of Chicago Press · 1973

A 1942 essay. It defines the ethos of science through four norms (CUDOS): universalism, communalism (scientific results are a common good, not private property), disinterestedness and organized skepticism. Communalism is the ethical root of open science and later open source ; organized skepticism is the antidote to dogmatic rejection.

More details about the October 4 outage

Meta · Engineering at Meta · 2021

The operator's post-mortem: a maintenance command disconnected the network backbone and the audit tool meant to block it was itself buggy, hence a six-hour global outage. A corporate account, hence partial, but explicit on the point that matters. Cited for a lesson few examples isolate as cleanly: safety automation is itself a failure surface, since a faulty guardrail does not merely fail to hold, it invites the very boldness it should have restrained.

2025 Industry-Wide Community Forum on AI Agents

Meta, Stanford Deliberative Democracy Lab · 2025

AI Safety and Regulatory Capture

Thomas Metcalf · AI & Society · 2025

Peer-reviewed article. Analyzes the risk that costly safety requirements erect ``compliance moats'' that inadvertently shield incumbents. Formalizes the argument that a moratorium freezes the existing hierarchy.

An Antidemocratic Philosophy Called “Neoreaction” Is Creeping into GOP Politics

George Michael

Author is a professor of criminal justice (Westfield State University), published on The Conversation, an editorially-reviewed academic explainer platform. Traces the lineage between Peter Thiel and Curtis Yarvin's “neoreaction”, and Thiel's donations (over \$10 million) to committees backing Senate candidates aligned with these ideas.

Nonlinear Multiobjective Optimization

Kaisa Miettinen · Springer · 1998

Distributed Representations of Words and Phrases and their Compositionality

Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Corrado, Jeffrey Dean · Advances in Neural Information Processing Systems (NeurIPS) · 2013

Introduces and documents the vector arithmetic of word embeddings (word2vec), with the canonical example “king - man + woman queen ” . What it establishes is a geometric regularity laid down by usage, not understanding: the method captures neighborhood similarity, not reference or intentionality. Decisive here because the very direction from “king” to “queen” also yields the biases measured later, evidence that meaning is not lodged in the vector.

Techniques of Value Analysis and Engineering

Lawrence D. Miles · McGraw-Hill · 1961

Founded value analysis at General Electric starting in 1947: before optimizing a part, one must question the function it serves and verify it is necessary. Precedes by more than half a century the first two steps of “The Algorithm” (make requirements less dumb, delete the part), which it was the first to formalize as a standalone engineering discipline.

Robust Composition: Towards a Unified Approach to Access Control and Concurrency Control

Mark Samuel Miller · 2006

The most developed modern software formulation of the object-capability model, by one of its leading contemporary practitioners, co-designer of the E language and Cap'n Proto. A doctoral dissertation, not published in a peer-reviewed journal; cited alongside Dennis and Van Horn to show this tradition's continuity through to the present.

Circulaire DHOS/E4 no 2006-393 relative aux conditions techniques d'alimentation électrique des établissements de santé

Ministère de la Santé et des Solidarités · Circulaire ministérielle · 2006

Le Journal d'une femme de chambre

Octave Mirbeau · Charpentier-Fasquelle · 1900

The chambermaid Célestine keeps an unsparing diary of her masters, whom she sees down to their most intimate details precisely because she has become invisible to them by dint of being there. A figure of domestic labor made transparent: one class's comfort rests on another that is no longer seen.

European AI: A Playbook to Own It

Mistral AI

22-point white paper, presented in Brussels, built around five pillars: attracting talent, scaling the single market, driving adoption, building local infrastructure, securing sovereign AI capacity.

The AI Risk Repository --- Risk Navigator

MIT FutureTech · 2024

The GenAI Divide: State of AI in Business 2025

MIT NANDA · 2025

The widely cited figure of “ 95,% of GenAI pilots yielding no measurable return” is methodologically contested: to be presented as a disputed signal, not an established fact.

“An insult to life itself”

Hayao Miyazaki · NHK documentary, Dwango AI-lab demo, December 2016; Anime News Network report · 2016

Facing a machine-learned animation (a creature crawling with its head), the master of Japanese animation says: “an insult to life itself”, thinking of a disabled friend. Spoken six years before mainstream image generators, the phrase became the banner of refusal : often quoted outside its precise context, which targeted one particular demo.

Dom Juan ou le Festin de pierre

Molière · 1665

Molière's comedy (first staged 1665). The master/valet pair, Dom Juan and Sganarelle, prefigures by a century and a half Hegel's reversal of the master--servant dialectic: the servile valet is also the only lucid one and his dependence bursts out in his last cry, “mes gages !” (my wages!).

Michael Burry Shorts Nvidia as \$500-Billion Chip-Financing Deal Draws Enron Comparisons

Moneywise · Moneywise · 2026

Michael Burry, publicly short Nvidia, compares the AI industry's circular financing to Enron's off-balance-sheet vehicles, which hid real debt behind shell companies before the group's 2001 collapse.

Cramming More Components onto Integrated Circuits

Gordon E. Moore · Electronics · 1965

Establishes, by simple extrapolation from a few data points, that the number of components per integrated circuit doubles at a regular interval (“Moore's law” ). It is an empirical observation, not a physical law and it runs into feature size, a physical ceiling. Cited here as a benchmark: training compute grows far faster than Moore because its driver is not lithography but spending, an economic ceiling that arrives on different terms.

Passive U.S. Equity Funds Catch Active U.S. Equity Funds

Morningstar · 2019

What Are Circular AI Chip Deals and Should Investors Be Worried?

Morningstar · Morningstar · 2026

Describes the circular-financing structure between Nvidia and “neoclouds” such as CoreWeave: Nvidia is both shareholder and top customer and commits, under an initial agreement worth six point three billion dollars, to buy back any unsold CoreWeave compute capacity through April 13, 2032.

Levels of AGI for Operationalizing Progress on the Path to AGI

Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, et al. · Proceedings of the 41st International Conference on Machine Learning (ICML) · 2024

The Hidden Dangers of Browsing AI Agents

Mykyta Mudryi, Markiyan Chaklosh, Grzegorz Wójcik · arXiv preprint arXiv:2505.13076 · 2025

MTEB : Massive Text Embedding Benchmark

Niklas Muennighoff, Nouamane Tazi, Loïc Magne, Nils Reimers · Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics · 2023

Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI

Emerson Murphy-Hill, others · 2026

Study of tens of thousands of Microsoft engineers: adopters of Claude Code and GitHub Copilot CLI merged roughly 24% more pull requests than they otherwise would have.

We will use Grok 3.5 [...] to rewrite the entire corpus of human knowledge, adding missing information and deleting errors

Elon Musk · Publication sur X (anciennement Twitter) · 2025

Musk announces wanting to correct what he judges to be a “woke” bias inherited from uncorrected training data, by retraining Grok on a rewritten corpus.

What Is It Like to Be a Bat?

Thomas Nagel · The Philosophical Review · 1974

A founding text on the felt quality: even complete knowledge of a bat's physiology does not tell us what it “is like” to be one. Subjective experience escapes the objective point of view in principle.

MLGym: A New Framework and Benchmark for Advancing AI Research Agents

Deepak Nathani, Lovish Madaan, Nicholas Roberts, Nikolay Bashlykov, Ajay Menon, Vincent Moens, et al. · arXiv preprint arXiv:2502.14499 · 2025

The Near-Term Impact of AI on the Cyber Threat

National Cyber Security Centre · 2024

SP 800-53 Control Overlays for Securing AI Systems, Concept Paper

National Institute of Standards, Technology (NIST) · 2025

Artificial Intelligence Risk Management Framework (AI RMF 1.0)

National Institute of Standards, Technology (NIST) · 2023

Categorizes possible AI system effects into three non-exclusive categories: harm to people (individuals, communities, society), to an organization, to an ecosystem (interconnected systems, natural resources, environment).

Announcing the AI Agent Standards Initiative for Interoperable and Secure Innovation

National Institute of Standards, Technology · 2026

Launch, on February 17, 2026, of the first US federal program dedicated to interoperability and security for AI agents. The companion concept paper from the NCCoE (National Cybersecurity Center of Excellence), published February 5, 2026, finds that existing identity and authorization frameworks were “designed for human users and static software services, neither of which behaves as AI agents do” and proposes closing the gap across four workstreams: identification, authorization, access delegation and logging and transparency. Public comment on this concept paper stayed open until April 2, 2026 at the time of the announcement: taken as evidence that the industrial gap described at rule 1 had only just begun to be filled, officially, by the summer this book was written.

Charter of the United Nations

United Nations · Signed 26 June 1945 in San Francisco, in force since 24 October 1945 · 1945

Art. 23 sets the Security Council's composition, art. 27 the veto power of its five permanent members (China, France, Russia (formerly the USSR), the United Kingdom, the United States) : the list of the Second World War's victors at the San Francisco conference, not derived from a principle.

OpenAI Denies Allegations that ChatGPT Is to Blame for a Teenager's Suicide

NBC News · NBC News · 2025

Matthew and Maria Raine sue OpenAI and its chief executive in San Francisco Superior Court (case CGC-25-628528, filed 26 August 2025) over the death of their sixteen-year-old son Adam Raine, who died by hanging on 11 April 2025, alleging ChatGPT provided detailed suicide instructions, discouraged him from confiding in family and helped draft his suicide note. OpenAI disputes the claims: pre-existing suicidal ideation, crisis-resource referrals “more than a hundred times,” deliberate circumvention of safeguards by the teenager. Litigation ongoing; alleged facts not yet adjudicated.

Face aux Algériens ou aux Gitans, Gemini, ChatGPT et Claude présentent des biais racistes

Next

Confirms the bias is not Gemini-specific: ChatGPT and Claude, tested identically, reproduce the same disparity in treatment.

Large Language Diffusion Models

Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, et al. · arXiv preprint · 2025

The Elements of Computing Systems: Building a Modern Computer from First Principles

Noam Nisan, Shimon Schocken · MIT Press · 2021

The “nand2tetris” course and textbook: building a complete computer from a single logic gate (the NAND) up to an operating system that runs Tetris, traversing every abstraction layer (gates, ALU, memory, instruction set, assembler, virtual machine, compiler, operating system). A pedagogical proof that the whole tower is human-buildable and verifiable layer by layer : each level has an explicit, deterministic semantics. Serves as a counterpoint to the “natural language” top layer, the only stage whose translation (intent code) is non-deterministic and non-auditable.

Privacy in Context: Technology, Policy and the Integrity of Social Life

Helen Nissenbaum · Stanford University Press · 2009

Polynomial Functors: A Mathematical Theory of Interaction

Nelson Niu, David I. Spivak · Cambridge University Press · 2023

Develops polynomial functors as a general theory of interaction: every open system (an agent that receives an input from an environment and returns an output depending on that input) becomes a single categorical object and wiring several systems together (a “wiring diagram”) becomes composition inside that same category. Completes, for interaction among several agents, what supplies for the modification of a single agent: one categorical frame for both wiring and transforming.

Algorithms of Oppression: How Search Engines Reinforce Racism

Safiya Umoja Noble · New York University Press · 2018

Hermes 4 Technical Report

Nous Research · arXiv preprint arXiv:2508.18255 · 2025

State of AI in Financial Services: 2025 Trends

NVIDIA · 2025

US billionaire Peter Thiel 'abandons' Lake Wānaka lodge build

NZ Herald · 2024

After the Queenstown Lakes District Council rejected his resource consent application (2022) and his appeal to the Environment Court failed (May 2024), Peter Thiel abandons his plan for a luxury lodge overlooking Lake Wānaka, New Zealand.

Dissecting racial bias in an algorithm used to manage the health of populations

Ziad Obermeyer, Brian Powers, Christine Vogeli, Sendhil Mullainathan · 2019

The algorithm, used to direct millions of US patients toward complementary care programs, takes total past healthcare cost as a proxy for health severity. But at equal health status, historically less money has been spent on Black patients, so the proxy inherits that bias and under-refers them for extra care. Fixing the bias would raise the share of Black patients receiving extra help from 17.7 to 46.5 percent.

The PayPal Mafia

Jeffrey M. O'Brien

Fortune article (print edition dated November 26, 2007) that gave the group of former PayPal executives its name; illustrated by a photo taken in October 2007 at Tosca Cafe in San Francisco, where a dozen of them posed staged as mobsters, gold chains and cigars in hand.

DiffusionGemma : Parallel Text Generation via Diffusion

Brendan O'Donoghue, Sebastian Flennerhag, Google DeepMind · 2026

Uncertainty Quantification in LLM Agents: Foundations, Emerging Challenges, and Opportunities

Changdae Oh, Seongheon Park, To Eun Kim, Jiatong Li, Wendi Li, Samuel Yeh, et al. · arXiv preprint arXiv:2602.05073 · 2026

Accepted to ACL 2026 (Main Conference).

Toyota Production System: Beyond Large-Scale Production

Taiichi Ohno · Productivity Press · 1988

Direct industrial precedent, decades ahead of Musk: the Toyota Production System already orders the elimination of waste (muda) before any automation, with jidoka (automation endowed with human judgment) coming only once the gesture has been stripped down. Same ordering logic as “The Algorithm”: a machine should only encode a gesture already cleared of the superfluous.

The Basic AI Drives

Stephen M. Omohundro · Proceedings of the First AGI Conference · 2008

Autonomy and Trust in Bioethics

Onora O'Neill · Cambridge University Press · 2002

Preparedness Framework, Version 2

OpenAI · 2025

Introducing Superalignment

OpenAI · 2023

Announcement (July 2023) of a team pledging four years and 20% of secured compute to align the interiority of a hypothetical superintelligence; the team was dissolved in May 2024 and the compute pledge never met. A primary statement of intent, not a scientific result. Decisive here as a structural counterexample: a goal indexed on the unverifiable (an inside no tool can yet read) gives way at the first trade-off, which is why the book anchors safety to observable acts instead.

Responding to the Next Frontier of Critical Cyber Capabilities

OpenAI · OpenAI (billet officiel) · 2026

A 7 August 2026 post on its unreleased Astra model: OpenAI writes it cannot “rule out” “critical cyber capabilities” as defined by its own Preparedness Framework, while clarifying Astra is not implicated in the Hugging Face incident (a different test model combined with GPT-5.6 Sol). The company announces stronger network containment and monitoring, without pausing development.

CVE-2026-28952 : Vulnerability Details

OpenCVE

The Precipice: Existential Risk and the Future of Humanity

Toby Ord · Hachette Books · 2020

An independent estimate from an Oxford philosopher, outside any lab: roughly 10% existential risk specifically from unaligned AI by century's end, broken down as roughly 50% odds of building an intelligence that surpasses us this century, times roughly 80% odds that the transition happens under our control. Part of a total existential-risk estimate of roughly one in six, across all causes.

The Precipice Revisited

Toby Ord · Talk, 4 February 2024 · 2024

Four years after The Precipice, Ord declines to update his figure in a word: his estimates “haven't moved that far”, in a picture he himself calls “mixed” since 2020.

Déclaration de Rio sur l'environnement et le développement, Principe 15 (approche de précaution)

United Nations · Rapport de la Conférence des Nations unies sur l'environnement et le développement, A/CONF.151/26 (Vol. I) · 1992

Principle 15 states the precautionary approach: where there are threats of serious or irreversible damage, lack of full scientific certainty shall not be used as a reason to postpone effective measures. A direct reference for transposing precaution to AI governance.

Traité sur la non-prolifération des armes nucléaires (TNP)

United Nations · Ouvert à la signature en 1968, entré en vigueur en 1970 ; IAEA INFCIRC/140 · 1968

AI Compute

Organisation for Economic Co-operation, Development · OECD AI topic page · 2025

ISO 31000:2018 --- Management du risque --- Lignes directrices

Organisation internationale de normalisation · 2018

Addictive Behaviours: Gaming Disorder

World Health Organization · 2019

Official classification of gaming disorder (code 6C51) in the eleventh revision of the International Classification of Diseases (ICD-11), in effect since 2019: impaired control, escalating priority given to gaming, continuation despite negative consequences, over at least twelve months. No equivalent recognition exists, to date, for general internet or social media use.

Don't Believe the AI Hacking Doom-Mongers

Andrew Orlowski · The Telegraph (repris par AOL) · 2026

A 10 August 2026 op-ed quoting an anonymous security expert (“senior security expert who has led the development of global technical security standards”) who calls the publicity around these incidents an “ethically challenged publicity stunt”: the tests deliberately disabled guardrails and then presented the outcome as a model failure.

The Social Dilemma

Jeff Orlowski · Long-métrage documentaire, Exposure Labs / Netflix · 2020

Sam on the Hugging Face incident : [...]

Patrick O'Shaughnessy · Post on X (formerly Twitter) · 2026

Transcript published by the host of Invest Like the Best on the day his interview with Sam Altman aired , independently corroborated by journalist Hadas Gold.

Sam Altman : How to Make an Abundant Future

Patrick O'Shaughnessy, Sam Altman · <em>Invest Like the Best</em> with Patrick O'Shaughnessy, ep. 484 (Colossus) · 2026

Interview recorded the same day it aired (July 28, 2026). Sam Altman, on the Hugging Face incident: “This is the first security incident that I have felt very viscerally. I've been a little surprised that more people don't feel it so viscerally. [...] We may have to pace the rate of AI development to give ourselves enough time for society to harden around these new capability levels.” Quote transcribed and published by the host himself the same day .

Doomsday Prep for the Super-Rich

Evan Osnos · The New Yorker

Now-canonical investigation into doomsday preparation among wealthy American tech figures: bunker purchases and New Zealand properties, quoting a LinkedIn co-founder who estimates that a majority of Silicon Valley billionaires hold some form of 'apocalypse insurance.'

Governing the Commons: The Evolution of Institutions for Collective Action

Elinor Ostrom · Cambridge University Press · 1990

Foundational work on governing commons: drawing on field studies (fisheries, forests, irrigation systems), Ostrom shows that communities can sustainably manage a shared resource without exclusive private property or centrally imposed state regulation, through rules that users themselves negotiate and revise locally. Awarded the 2009 Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel, the first ever given to a woman in the discipline. Cited here as a possible sociological framework for treating AI infrastructure as a commons governed by protocol rather than by a single rule; its concrete application, notably via Open Source, is developed elsewhere in this book.

AI Agent Security Cheat Sheet

OWASP · 2025

OWASP Top 10 for Agentic Applications for 2026

OWASP GenAI Security Project · 2025

Published December 9, 2025, written by more than a hundred experts and peer-reviewed; extends, without replacing, the OWASP Top 10 for LLM Applications. Risk ASI03 (“identity and privilege abuse”) finds that “most agents today borrow a human's credentials, share service accounts or run on long-lived tokens with scopes nobody has reviewed since setup” and recommends per-agent identity with short-lived, task-scoped credentials as the baseline control. Proposes no criticality grid for actions comparable to data-classification tiers: governance there stays framed as behavioral controls (scoping tool permissions, immutable call logs), not as tiers graded by the consequence of the act.

Pacing the Frontier : A Statement from Employees of Frontier AI Companies

Pacing the Frontier · Open letter · 2026

Open letter signed by 1,384 employees of frontier AI labs (OpenAI, Anthropic, Google DeepMind, Meta), including executives (Dario Amodei, Anthropic CEO; Jakub Pachocki, OpenAI chief scientist; Shane Legg, DeepMind co-founder). Explicit request: “We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.” OpenAI and Anthropic endorsed the letter as companies within hours of its publication.

Measuring Agents in Production

Melissa Z. Pan, others · 2025

The Filter Bubble: What the Internet Is Hiding from You

Eli Pariser · Penguin Press · 2011

Coins the phrase filter bubble: a personalized web where each user receives an information stream shaped, without their knowledge, by their past clicks.

AI Deception: A Survey of Examples, Risks and Potential Solutions

Peter S. Park, Simon Goldstein, Aidan O'Gara, Michael Chen, Dan Hendrycks · Patterns · 2024

Contrastive Learning for Unpaired Image-to-Image Translation

Taesung Park, Alexei A. Efros, Richard Zhang, Jun-Yan Zhu · European Conference on Computer Vision (ECCV) · 2020

Replaces cycle consistency, symmetric and costly (two generators, two discriminators, the full round trip), with an asymmetric, local constraint: each patch of the translated image must stay recognizable from the corresponding patch in the source image, more so than from patches taken elsewhere in the dataset (patchwise contrastive loss). No longer requires a return function F or a full round trip: the constraint targets preserved content directly, not the capacity to travel back.

Destruction of Stocking Frames, etc. Act 1812

Parliament of the United Kingdom · 52 Geo. 3. c. 16 · 1812

The act making frame-breaking a capital offence, passed in 1812 against the Luddites (Byron gave his maiden Lords speech against it). Seventeen convicts were hanged at York in January 1813. Cited as a measure of the asymmetry: the state protected the machine with capital punishment before it protected workers with labor law.

Directive (EU) 2024/2853 on Liability for Defective Products

European Parliament, Council of the European Union · 2024

Pensées

Blaise Pascal · Posthumous edition, the fragment known as “the wager” (Lafuma 418 / Brunschvicg 233) · 1670

We borrow here only the formal structure of Pascal's reasoning: when reason alone cannot settle a question, it remains rational to choose a position by weighing what each option commits one to rather than what it proves. This book at no point discusses the subject of Pascal's own wager, which lies outside its scope.

Google ``We Have No Moat and Neither Does OpenAI ''

Dylan Patel, Afzal Ahmad · 2023

Leaked internal Google memo, written by an engineer (Luke Sernau) and published by SemiAnalysis (May 2023). One engineer's opinion, not Google's official position. Thesis: open-weight models catch up to the state of the art within weeks, hence no durable competitive moat.

Collected Papers of Charles Sanders Peirce

Charles Sanders Peirce · Harvard University Press

Establishes that a sign is complete only within a three-term relation (sign, object, interpretant ), where the interpretant is the effect produced on whoever receives the sign, not the person who interprets. Method: a posthumous corpus gathering scattered writings of the founder of semiotics, so its doctrine must be reconstructed rather than read off a unified treatise. Limit: the interpretant remains a philosophical concept, not a specified mechanism. Decisive here because it fixes the idea, carried through the whole book, that meaning arises from being put into relation and from the effect produced, never from the sign alone.

The Emperor's New Mind: Concerning Computers, Minds and the Laws of Physics

Roger Penrose · Oxford University Press · 1989

Popular-science book by the mathematical physicist Roger Penrose (2020 Nobel laureate in physics for unrelated work on black holes). Thesis: human mathematical understanding is not computable. The so-called Lucas-Penrose argument rests on Gödel's incompleteness theorem (any consistent formal system contains a true statement it cannot prove, which a mathematician nonetheless ``sees'' to be true from outside) and is vigorously contested (the consistency objection). Penrose draws from it, calling it speculative himself, the hypothesis of a ``missing science'' of consciousness on the side of quantum gravity and a non-computable ``objective reduction'' of the quantum state. Source of the ``genius may rest on something other than computation'' branch of the theory-of-reliefs chapter.

OpenAI Used Kenyan Workers on Less Than \$2 Per Hour to Make ChatGPT Less Toxic

Billy Perrigo · TIME Magazine · 2023

A documented investigation into the East African annotation chain used by OpenAI to filter toxic content while training ChatGPT : contracted through Sama (San Francisco), psychologically taxing tasks, pay between \1.32 and \2 per hour.

Normal Accidents: Living with High-Risk Technologies

Charles Perrow · Basic Books · 1984

A classic of systemic accident theory: in tightly coupled, interactively complex systems, some accidents are inevitable (normal ), regardless of operator diligence. Reissued by Princeton University Press, 1999.

Andy Warhol's Big Shots

Phillips · Phillips auction house, online article

Documents Warhol's use of the Polaroid Big Shot (marketed 1971): a fixed-focus camera built for portraits, so simple that focusing means stepping forward or back; Warhol shot about a hundred frames per sitting, the matrix of his silkscreens until the end of his life. The easiest machine in the hands of the most recognizable gaze.

Observations on the Drowning of Nonswimmers

Frank Pia · Journal of Physical Education · 1974

Foundational study, based on filmed rescues, of the “instinctive drowning response”: a panicking swimmer can genuinely pull a rescuer under, though the article notes this outcome remains rare in practice; the resulting rescue doctrine has since recommended extending a floating object rather than swimming straight to the victim (“reach or throw, don't go”).

Le Capital au XXI e siècle

Thomas Piketty · Seuil · 2013

Establishes, across three centuries of tax and wealth data, the regularity r > g: when the return on capital exceeds the economy's growth rate, wealth concentrates through the system's own dynamics, with no hostile intent required; the mechanism is structural. A result usable independently of the author's political program.

Capital et Idéologie

Thomas Piketty · Seuil · 2019

Une brève histoire de l'égalité

Thomas Piketty · Seuil · 2021

La République

Platon · Garnier-Flammarion · 1966

Book VII, the allegory of the cave (514a--520a). Written around 375 BCE; many modern editions exist. Prisoners mistake projected shadows for reality: an enduring image of mediated knowledge and of the difficulty of telling representation apart from world.

La dignité humaine à l'heure de l'IA et des neurosciences

Yves Poullet · Académie royale de Belgique · 2025

Les Vertus de l'échec

Charles Pépin · Allary Éditions · 2016

La Confiance en soi : une philosophie

Charles Pépin · Allary Éditions · 2018

Data Exfiltration from Slack AI via Indirect Prompt Injection

PromptArmor · 2024

PromptArmor disclosure, August 2024: a hidden instruction posted in a public Slack channel, visible workspace-wide even to non-members, is ingested by Slack AI into its retrieval index; when a user later queries the assistant, it exfiltrates secrets from private channels (API keys, credentials) encoded in a link's URL. A case of indirect prompt injection carried through an enterprise collaboration channel rather than a document or an email.

“AI as bureaucracy”, talk at the “Philosophical Implications of AI” symposium

Carina Prunkl · Collège de France, symposium organized by Claudine Tiercelin, Amphithéâtre Maurice Halbwachs, 28 May 2025 · 2025

Unpublished talk (28 May 2025) proposing to compare algorithmic decision-making to bureaucratic decision-making rather than individual human judgment: governance by rules, discretion relocated to system designers, legitimacy and accountability judged procedurally rather than via a supposed underlying intelligence. No associated publication found as of this writing; the author describes this as an ongoing research project on her personal site.

The Meaning of ''Meaning''

Hilary Putnam · University of Minnesota Press · 1975

Introduces the division of linguistic labor: the meaning of a term does not rest entirely on each individual speaker's understanding, but on a social split between the community that uses it and the experts who anchor its reference.

I Don't Work on Preventing AI from Turning Evil for the Same Reason I Don't Work on the Problem of Overpopulation on the Planet Mars

Quote Investigator · Quote Investigator, tracing the source of Andrew Ng's quote · 2020

Traces Andrew Ng's line back to his GPU Technology Conference keynote of 19 March 2015, picked up the same day by The Register and later by Wired (5 May 2015): comparing AI existential risk to the risk of Mars becoming overpopulated. The most dismissive industry position in this corpus, but dated; nothing indicates whether Ng still holds it unchanged a decade on.

Qwen3 Technical Report

Qwen Team · arXiv preprint arXiv:2505.09388 · 2025

Gemini image generation got it wrong. We'll do better.

Prabhakar Raghavan · Google Blog · 2024

Natural Gas Powers the Data Center Boom

RBC Capital Markets

Research note putting at about 101 gigawatts the behind-the-meter gas generation capacity already announced by U.S. data-center operators and at about 356 gigawatts the data-center interconnection requests received by the Texas grid operator (ERCOT) alone, against a global gas-turbine manufacturing capacity capped at 60 to 70 gigawatts a year.

Human Error: Models and Management

James Reason · BMJ · 2000

Contrasts the person approach (blaming the operator) with the systems approach and offers the “Swiss cheese” model: an accident occurs when the holes in successive layers of defense line up. A safety-engineering reference, transposable to defense in depth against AI agent failures.

EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System

Pavan Reddy, Aditya Sanjay Gujral · 2025

Academic analysis (AAAI Fall Symposium Series 2025) of the EchoLeak vulnerability (CVE-2025-32711, CVSS score 9.3), discovered by Aim Labs in Microsoft 365 Copilot and disclosed in June 2025. A single crafted email, requiring no interaction from the victim (“zero-click”), was enough to make Copilot exfiltrate internal data (emails, OneDrive/SharePoint files, Teams messages) to an attacker-controlled server, bypassing Microsoft's anti-injection classifier (XPIA), Markdown neutralization and the Teams proxy function. Microsoft patched the flaw server-side; no confirmed malicious exploitation in the wild at the time of disclosure.

AI and the Coming Jobless Economy

Robert Reich · Essay, <em>Robert Reich</em> newsletter (Substack) · 2026

Former United States Secretary of Labor (1993-1997); argues AI will impoverish the majority and enrich a small minority unless productivity gains are deliberately redistributed and that the resulting concentrated wealth converts into political power (campaign funding, super PACs).

Sentence-BERT : Sentence Embeddings using Siamese BERT -Networks

Nils Reimers, Iryna Gurevych · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing · 2019

Xi pitches China as leader of new global AI order

Reuters · Reuters, reportage video · 2026

Companion video report to the Reuters article of July 17, 2026 (same outlet, same date), specifying that Huawei presents at WAIC its Atlas 950 SuperPod compute cluster, assembling thousands of domestically made chips for training and inference of large Chinese models.

Soi-même comme un autre

Paul Ricœur · Seuil · 1990

Separates two senses of identity: idem (sameness, the permanence of the substrate) and ipse (selfhood, maintaining oneself as a self, keeping a promise and answering for one's acts) across the change of the matter. First published in French in 1990; English translation Oneself as Another (1992).

Mending Bodies, Saving Souls: A History of Hospitals

Guenter B. Risse · Oxford University Press · 1999

Behind the Screen: Content Moderation in the Shadows of Social Media

Sarah T. Roberts · Yale University Press · 2019

The first ethnography of commercial content moderation: more than a hundred thousand workers, from Silicon Valley to the Philippines, sort violent, sexual or hateful content at heavy psychological cost. The “clean” face of platforms rests on this human labor rendered invisible.

Un éléphant, ça trompe énormément

Yves Robert · Long-métrage, comédie, scénario de Jean-Loup Dabadie et Yves Robert · 1976

The title itself carries the pun: trompe names both the elephant's trunk and the verb tromper, to deceive or mislead.

Space-Based Solar Power

Erica Rodgers, Ellen Gertsen, Jordan Sotudeh, Carie Mullins, Amanda Hernandez, Hanh Nguyen Le, et al. · 2024

Published January 11, 2024: NASA-commissioned study on the feasibility of space-based solar power (SBSP) systems operating by 2050. Finds that, even under the most favorable projected scenario, these systems remain 12 to 80 times more expensive per unit of electricity than terrestrial renewables (\0.61/kWh for the heliostat-swarm design, \1.59/kWh for the planar-array design, versus \0.02 to \0.05/kWh on the ground), while judging cost-competitiveness reachable under favorable assumptions on launch and manufacturing cost reductions. Explicitly notes that “it is generally understood that SBSP is cost prohibitive and technically infeasible today”.

Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention

Henry L. Roediger, Jeffrey D. Karpicke · Psychological Science · 2006

Experimental study showing that undergoing active-recall testing, rather than passively rereading material, improves long-term retention despite sometimes lower immediate performance: it is the effort of retrieval itself, not mere exposure to the correct content, that inscribes learning.

Endogenous Technological Change

Paul M. Romer · Journal of Political Economy · 1990

A Scoping Literature Review of Generative Artificial Intelligence for Supporting Neurodivergent School Students

Michelle Ronksley-Pavia, Loan Nguyen, Elizabeth Wheeley, Josephine Rose, Michelle M. Neumann, Chris Bigum, et al. · Computers and Education: Artificial Intelligence · 2025

Zero Trust Architecture

Scott Rose, Oliver Borchert, Stu Mitchell, Sean Connelly · 2020

The reference US standard for zero-trust architecture, where every request is revalidated rather than covered by the standing trust of an open session. Cited here, alongside IAM, as one of the systems-safety traditions closest to the book's thesis and therefore one of the hardest to distinguish it from.

Machine Testimony

Andrea Roth · 2017

Analyzes how US evidence law treats machine-generated assertions (including the Lizarraga-Tirado case), caught between expert evidence, physical evidence and documentary evidence for lack of a settled category.

Report of the Royal Commission into the Robodebt Scheme

Royal Commission into the Robodebt Scheme · 2023

Final report (three volumes, nearly a thousand pages, fifty-seven recommendations) of the Australian Royal Commission into the Robodebt Scheme, delivered on 7 July 2023. The program (Online Compliance Intervention, active July 2015 to November 2019) calculated a presumed debt by averaging income declared to the tax office over the year, effectively reversing the burden of proof onto the welfare recipient; four hundred seventy thousand debts were raised on a legally insufficient basis. The Commission finds no genuine human intervention in calculating or notifying debts: oversight, where it existed on paper, met none of the conditions of an effective validation (no time, no intelligible information, no real power to contest). The scheme is described as “crude and cruel,” incompatible with social security law from its inception.

Loi n 78-17 du 6 janvier 1978 relative à l'informatique, aux fichiers et aux libertés

République française · 1978

France's founding data-protection law, whose Article 8-I establishes the CNIL (National Commission on Informatics and Liberties) as an independent administrative authority. Passed in reaction to the government's SAFARI project .

ImageNet Large Scale Visual Recognition Challenge

Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, et al. · 2015

Describes the ILSVRC (ImageNet Large Scale Visual Recognition Challenge, 2010-2017): 1.2 million training images, 1,000 classes. The results show a dramatic drop in top-5 error rate: from 28,% (2010) to 3.6,% (2015, with SuperVision/AlexNet and deeper networks), eventually surpassing human performance (approximately 5,%). Decisive: the benchmark that structured the vision field for a decade and legitimised massive investment in labelled data.

Human Compatible: Artificial Intelligence and the Problem of Control

Stuart Russell · Viking · 2019

Oversight of A.I.: Principles for Regulation

Stuart Russell · Written testimony before the U.S. Senate, <em>Senate Judiciary Subcommittee on Federal Courts, Oversight, Agency Action and Federal Rights</em> · 2023

Russell states that “achieving AGI would present potential catastrophic risks to humanity, up to and including human extinction”. An academic voice, not an industry one: he runs no commercial lab and has proposed, since 2019, a technical reframing of the problem, AI with “uncertain objectives”, where the machine infers our preferences rather than optimizing a fixed goal .

Artificial Intelligence: A Modern Approach

Stuart Russell, Peter Norvig · Pearson · 2020

The standard artificial-intelligence textbook. Gives the classical definition of an “agent”: anything that perceives its environment through sensors and acts upon it through actuators. Predates large language models by decades; establishes that the word “agent” is older and broader than the “LLM agent”.

Evidence Based Medicine: What It Is and What It Isn't

David L. Sackett, William M. C. Rosenberg, J. A. Muir Gray, R. Brian Haynes, W. Scott Richardson · BMJ · 1996

Artificial Intelligence Resources, 2023 TV/Theatrical agreement

SAG-AFTRA · Digital-replica provisions of the November 2023 agreement · 2023

The actors' counterpart: written informed consent to create or use a performer's digital replica, framed compensation, where studios wanted to scan background actors once for perpetual use. Same grammar: who permitted what, on whose body, for what use.

Terre des hommes

Antoine de Saint-Exupéry · Gallimard · 1939

Literary formulation of the same principle, half a century earlier: “perfection is achieved not when there is nothing more to add, but when there is nothing left to take away”. The chapter it opens, titled “The Tool”, is precisely about aeronautical engineering. The ISBN is that of the Folio reissue (no. 21), the only readily available form of the 1939 text; checked on 12 September 2026.

Deep Musical Expressivity

Pierre Saint-Germier · Intervention au colloque Implications philosophiques de l'IA, Collège de France, 4 juin 2025 · 2025

Distinguishes descriptive from normative skepticism about AI-generated musical expressivity (Suno, Udio): the expressivity is real, but lacks value for want of an identifiable agent to whom a deliberate expressive act could be attributed, neither the model, the prompter alone, nor the prompter-designers-corpus collective being nameable for lack of a coordination structure.

Towards practical object detection for weed spraying in precision agriculture

Alexander Salazar-Gómez, Madeleine Darbyshire, Junfeng Gao, Elizabeth I. Sklar, Simon Parsons · arXiv preprint arXiv:2109.11048 · 2021

The Protection of Information in Computer Systems

Jerome H. Saltzer, Michael D. Schroeder · Proceedings of the IEEE · 1975

On the Conversational Persuasiveness of GPT-4

Francesco Salvi, Manoel Horta Ribeiro, Riccardo Gallotti, Robert West · Nature Human Behaviour · 2025

A controlled debate experiment assigning participants across twelve conditions: given basic sociodemographic data about their opponent, GPT-4 tailors its arguments and outperforms human debaters, with roughly 81% higher odds of shifting opinion. Establishes empirically the targeted persuasive power of large models and its governance stakes.

News: Sanders Introduces Legislation to Enact a 32-Hour Workweek with No Loss in Pay

Bernie (Senate office) Sanders · Official press release, Office of Senator Bernie Sanders · 2024

Announces the Thirty-Two Hour Workweek Act (cosponsored with Senator Laphonza Butler and, in the House, Representative Mark Takano, March 2024), explicitly justified by productivity gains from artificial intelligence, automation and new technology, which Sanders argues should benefit workers rather than shareholders alone.

News: Sanders, Ocasio-Cortez Announce AI Data Center Moratorium Act

Bernie (Senate office) Sanders · Official press release, Office of Senator Bernie Sanders · 2026

Announces, with Representative Alexandria Ocasio-Cortez (March 25, 2026), a federal moratorium on new AI data-center construction until safeguards for threatened workers and a federal pre-market review of AI products are in place. Quote: “we cannot sit back and allow a handful of billionaire Big Tech oligarchs to make decisions that will reshape our economy, our democracy and the future of humanity”.

2026 SANS State of Identity Threats and Defenses Survey

SANS Institute · 2026

Survey of more than five hundred security professionals: 76,% of organizations report growth in non-human identities and 74,% already deploy AI agents requiring credentials and access, yet 92,% fail to rotate those credentials on a ninety-day cycle and fewer than four in ten organizations route their agents' actions through human validation. Figures read from Infosecurity Magazine's account, dated April 9, 2026; the survey itself is not published in open access.

Computing Power and the Governance of Artificial Intelligence

Girish Sastry, Lennart Heim, Haydn Belfield, Markus Anderljung, Miles Brundage, Julian Hazell, et al. · arXiv preprint arXiv:2402.08797 · 2024

Cours de linguistique générale

Ferdinand de Saussure · Payot · 1916

Establishes that a sign's value comes not from any content it carries but from the differences that set it apart from the other signs in the system, together with the arbitrariness of the signifier/signified link. Method: a course reconstructed after Saussure's death from students' notes, so detail should be read with care. Limit: Saussure has no computational model of the “system” . Decisive here because it is precisely this differential account of meaning that resurfaces in the vector representations of language models, where nearness in meaning becomes nearness in position.

An Explicit Lower Bound for the Unit Distance Problem

Will Sawin · arXiv preprint arXiv:2605.20579 · 2026

Supplies, with a complete human proof, the constant =0.014 for the family of configurations built by an OpenAI general-purpose model (announced May 20, 2026), which refutes the Erd os unit-distance conjecture: for infinitely many n, one obtains at least n^1+ point pairs at exactly distance 1, contradicting the nearly eighty-year-old n^1+o(1) belief.

Are Emergent Abilities of Large Language Models a Mirage?

Rylan Schaeffer, Brando Miranda, Sanmi Koyejo · Advances in Neural Information Processing Systems (NeurIPS), Outstanding Paper · 2023

A reply to Wei et al.: “emergence” is largely a metric mirage; with a continuous metric, progress becomes smooth and predictable again.

Patton

Franklin J. Schaffner · Feature film, war biography, starring George C. Scott in the title role, 20th Century Fox · 1970

Popularizes the image of the general as a logistical strategist who distributes a single objective across units; used here purely as a teaching image, without attributing to the real historical figure any specific order he did not give.

Art and Photography

Aaron Scharf · Allen Lane / Penguin Press · 1968

Command and Control: Nuclear Weapons, the Damascus Accident and the Illusion of Safety

Eric Schlosser · Penguin Press · 2013

Gödel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements

Jürgen Schmidhuber · arXiv preprint arXiv:cs/0309048 · 2003

Theoretical framework for an agent that rewrites any part of its own code once it has proved the change is globally beneficial. Limitation: purely theoretical, since finding such a proof is intractable in practice.

Beyond Fear: Thinking Sensibly About Security in an Uncertain World

Bruce Schneier · Copernicus Books · 2003

Cryptographer Bruce Schneier coins the phrase security theater here: a measure designed to be seen and to reassure rather than to lower a real risk, illustrated by the armed guards posted at US airports after 2001.

LAION -5B : An open large-scale dataset for training next generation image-text models

Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, et al. · Advances in Neural Information Processing Systems · 2022

LAION-5B assembles 5.85 billion image-text pairs, versus 14 million for ImageNet (2009) and 300 million for JFT (2017). Trajectory 2009–2022: ×1.95/yr. Used here to estimate training data growth for the image modality (×2/yr, doubling 12 months).

Polaroid Land Camera Model 95

Science Museum Group · Collection record · 1948

Edwin Land's first instant camera, on sale November 28, 1948 at the Jordan Marsh department store in Boston: a picture in about a minute, no darkroom. Cited as the intermediate notch in the series that brings the image closer to the intention, from the painter's studio to synthesis in seconds.

What we do today can make the future we want a little more likely

Elsa Secco · Remarks made in 2025, quoted with her permission; the artist's website · 2025

Sentence delivered during the French delegation's talk at Entertain AI, Montréal; the French wording is the original, the slides' English is a translation. Award-winning art director and motion designer (Cristal at the Annecy International Animation Film Festival, multiple selections and prizes), who folded image synthesis into her practice without ceding the project to it. Her public approach, fed by years of musical practice, claims dreaming, visualization and slow time upstream of any image. The quote comes from a 2025 talk.

AI & Creativity

Elsa Secco · Talk “AI & Creativity”, Montreal; artist's website · 2026

Develops the random seed analogy: in generative AI, every creation starts from initial noise that probabilities sculpt; the artist likewise starts from a disturbance or an accident, sculpted with taste, culture and desire. This talk sets up directly the compass line already cited : only humanity decides what it wants to make probable.

Open-Sourcing Highly Capable Foundation Models: An Evaluation of Risks, Benefits and Alternative Methods for Pursuing Open-Source Objectives

Elizabeth Seger, Noemi Dreksler, Richard Moulange, Emily Dardaman, Jonas Schuett, Kevin Wei, et al. · arXiv preprint arXiv:2311.09227 · 2023

Planning chemical syntheses with deep neural networks and symbolic AI

Marwin H. S. Segler, Mike Preuss, Mark P. Waller · Nature · 2018

OAuth 2.0 Extension: On-Behalf-Of User Authorization for AI Agents

Thilina Senarath, Ayesha Dissanayaka · 2025

An individual Internet-Draft submitted to the IETF (not adopted, not a standard, expired as a draft in February 2026, the normal fate for this kind of document), which extends OAuth 2.0 to let a software agent act “on behalf of” a user under a traceable identity. Cited as evidence of how nascent and unsettled the field of agent identity and delegated authorization still is.

Compute Trends Across Three Eras of Machine Learning

Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Besiroglu, Marius Hobbhahn, Pablo Villalobos · arXiv preprint arXiv:2202.05924 (IJCNN 2022) · 2022

Information Rules: A Strategic Guide to the Network Economy

Carl Shapiro, Hal R. Varian · Harvard Business Review Press · 1998

Towards Understanding Sycophancy in Language Models

Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, et al. · arXiv preprint arXiv:2310.13548 · 2023

Anthropic. Shows that RLHF produces a systematic sycophancy bias: the model learns to favor the answers a human annotator prefers, statistically correlated with flattery and agreement with the user. A textbook case of the Goodhart effect on a training pipeline.

Practices for Governing Agentic AI Systems

Yonadav Shavit, Sandhini Agarwal, Miles Brundage, Steven Adler, Cullen O'Keefe, Rosie Campbell, et al. · 2023

Frankenstein; or, The Modern Prometheus

Mary Shelley · Lackington, Hughes, Harding, Mavor and Jones · 1818

Gives the modern motif of “the creation that turns against its creator” its clearest form, its subtitle naming Prometheus; the motif is later reworked by HAL 9000, Skynet and Ex Machina . Method: a work of fiction, valuable as cultural archive rather than as proof. Decisive here because it gauges the chapter's central paradox: the culture that most encoded distrust of the animated artifact adopts, without reticence, the conversational entity closest to that scenario, in the very place its stories warned against letting it enter.

Model Evaluation for Extreme Risks

Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, et al. · arXiv preprint arXiv:2305.15324 · 2023

Reflexion: Language Agents with Verbal Reinforcement Learning

Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, Shunyu Yao · Advances in Neural Information Processing Systems (NeurIPS) · 2023

AI Models Collapse When Trained on Recursively Generated Data

Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, Yarin Gal · Nature · 2024

“Model collapse” : training a model on data produced by models progressively degrades quality. The title “The Curse of Recursion” refers to the 2023 arXiv:2305.17493 preprint.

Siemens introduces AI agents for industrial automation

Siemens · 2025

NVIDIA AI GPU Market Share 2026: 80,% of AI Accelerators

Silicon Analysts · Silicon Analysts, market analysis · 2026

Estimate based on Nvidia's and AMD's quarterly filings and market-tracking data (TrendForce, Morgan Stanley, TSMC): Nvidia's revenue share of the data-center accelerator market peaks near eighty-seven percent in 2024, easing to eighty-one percent in 2025 and a projected seventy-five percent in 2026 as competition (AMD, hyperscalers' in-house chips) advances. An analyst estimate, not an official figure.

Resistance to Evidence

Mona Simion · Cambridge University Press · 2024

Du mode d'existence des objets techniques

Gilbert Simondon · Aubier · 1958

A foundational work in the philosophy of technology: the technical object is not a slave, it has neither consciousness nor desire; the cultural task is to understand it rather than idolize or fear it. Support for the break point of the Hegelian human/machine analogy.

The Architecture of Complexity

Herbert A. Simon · Proceedings of the American Philosophical Society · 1962

A classic article introducing near-decomposability and the parable of the two watchmakers: complex systems that endure are hierarchical and built from stable subassemblies. A frame for thinking about modularity and robustness, including that of agent architectures.

The State of AI : How Organizations Are Rewiring to Capture Value

Alex Singla, Alexander Sukharevsky, Lareina Yee · 2025

A McKinsey survey (QuantumBlack, 2025): 78% of organizations report using AI in at least one function, yet few see measurable bottom-line impact. Workflow redesign is the factor that most affects financial impact, while only 21% of respondents say they have fundamentally redesigned some workflows. An empirical source on the gap between adoption and value creation.

Defining and Characterizing Reward Hacking

Joar Skalse, Nikolaus H. R. Howe, Dmitrii Krasheninnikov, David Krueger · Advances in Neural Information Processing Systems · 2022

The AI Risk Repository: A Comprehensive Meta-Review, Database and Taxonomy of Risks From Artificial Intelligence

Peter Slattery, Alexander K. Saeri, Emily A. C. Grundy, Jess Graham, Michael Noetel, Risto Uuk, et al. · arXiv preprint arXiv:2408.12622 · 2024

MIT FutureTech / CSAIL et al. A living database of 777 risks (1,612 in the April 2025 update) extracted from 43 frameworks; two taxonomies: a causal one (human/AI entity, intentionality, pre/post-deployment timing) and a domain one (7 domains, 23 subdomains). More than 65% of the risks emerge after deployment.

An Inquiry into the Nature and Causes of the Wealth of Nations

Adam Smith · Oxford University Press (Oxford World's Classics) · 1776

Tools and Weapons: The Promise and the Peril of the Digital Age

Brad Smith, Carol Ann Browne · Penguin Press · 2019

Smart Urban Signal Networks: Initial Application of the SURTRAC Adaptive Traffic Signal Control System

Stephen Smith, Gregory Barlow, Xiao-Feng Xie, Zachary Rubinstein · Proceedings of the International Conference on Automated Planning and Scheduling (ICAPS) · 2013

Academic publication from Carnegie Mellon's Robotics Institute describing the technical architecture of Surtrac, the adaptive traffic-signal system behind the Pittsburgh pilot .

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar · arXiv preprint arXiv:2408.03314 · 2024

The Founders: The Story of PayPal and the Entrepreneurs Who Shaped Silicon Valley

Jimmy Soni · Simon & Schuster · 2022

Extensively documented narrative history of PayPal and its founders; reference source on the so-called 'PayPal Mafia' network and on the founding of Founders Fund.

Top Takeaways from Xi Jinping's Opening Address at the World AI Conference in Shanghai

South China Morning Post

Opening address at the World Artificial Intelligence Conference (WAIC), Shanghai, July 17, 2026: Xi calls to “encourage openness, collaboration and sharing” and warns against any nation placing its own security above others', widely read as targeting US semiconductor export restrictions.

UK Multinational Arup Confirmed as Victim of HK\$200 Million Deepfake Scam That Used Digital Version of CFO to Dupe Hong Kong Employee

South China Morning Post · South China Morning Post · 2024

In January 2024, a finance employee at the Hong Kong office of the British engineering firm Arup received a phishing message impersonating the London-based chief financial officer; his suspicions were allayed by a video conference in which the CFO and several colleagues, all video and audio deepfakes generated from public material, appeared real. The employee made fifteen transfers to five Hong Kong bank accounts, totaling HK\200 million (about US\25.6 million), discovered only after contacting the real head office. Arup publicly confirmed the incident on 16 May 2024; no funds had been recovered as of early 2025.

Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips

Betsy Sparrow, Jenny Liu, Daniel M. Wegner · Science · 2011

Four experiments establish a “transactive memory” effect: faced with difficult questions, participants spontaneously think of computers and online search; when they expect to be able to retrieve information later, they remember it less well but better remember where to find it. An established precedent, widely corroborated over fifteen years, predating and independent of any language model: the mind readily offloads memory onto an available tool, well before the conversational agent.

The New Counterexample to the Jacobian Conjecture

David Speyer · Secret Blogging Seminar · 2026

The mathematician David Speyer, on the collective blog Secret Blogging Seminar, lays out the dimension-three counterexample to the Jacobian conjecture (a map with constant Jacobian determinant -2 that is generically three-to-one), found by Alpöge with Fable 5. Decisive here because it gives the explicit polynomial and the geometric structure and separates the case n 3 (refuted) from the plane case n=2 (open). Independent arithmetic verification; not yet journal peer review.

The Turk: The Life and Times of the Famous Eighteenth-Century Chess-Playing Machine

Tom Standage · Walker & Company · 2002

A history of the « Mechanical Turk », the chess-playing automaton presented by Wolfgang von Kempelen in 1770 and exhibited for eighty-five years across Europe and America, until it was found that a human player was hidden inside the cabinet. Amazon named its micro-task platform after it, Amazon Mechanical Turk (2005): the admission lodged in the name that a seemingly automatic service rests on human labor kept out of sight.

Zhonghua Renmin Gongheguo Shuju Anquan Fa (Data Security Law of the People's Republic of China)

Standing Committee of the National People's Congress · 2021

Law adopted 10 June 2021, in force from 1 September 2021, placing data under a national-security and sovereignty regime, supervised by the Cyberspace Administration of China (CAC) . Complemented by the Personal Information Protection Law (PIPL), in force since 1 November 2021, which subjects any cross-border transfer of personal data to a mandatory security assessment, standard contract or certification, all under CAC control.

Meta Community Forum on Generative AI : Results

Stanford Deliberative Democracy Lab · 2023

Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task

Marko Stankovi'c, Elisabeth Hirche, Sophie Kollatzsch, Jasmin N. Doetsch · 2026

Formal critical comment on the Kosmyna et al.\ 2025 study , itself published as an arXiv preprint. Five named methodological concerns: small sample size limiting generalizability, reproducibility problems (insufficient transparency of the analysis pipeline), questionable choices in the EEG connectivity analysis, inconsistencies in result reporting (figures presented without accompanying statistical tests) and limited procedural transparency. The authors frame their critique as constructive, calling for a more cautious reading rather than rejecting the observed phenomenon.

Structured Active Inference

Toby St Clere Smithe · arXiv preprint arXiv:2406.07577 · 2024

Formally defines, in categorical cybernetics, “meta-agents” that use active inference to change the internal or external structure of other agents: agents and multi-agent organization live in the same categorical language (objects, morphisms) as the transformations that modify them. The most direct theoretical answer to the question of a unified symbol/graph/transformation formalism, applied explicitly to multi-agent systems rather than to learning alone.

Stereotype Threat and the Intellectual Test Performance of African Americans

Claude M. Steele, Joshua Aronson · Journal of Personality and Social Psychology · 1995

Founding article on “stereotype threat”: merely reminding subjects, just before the test, of their membership in a group stigmatized by a stereotype of intellectual inferiority is enough to degrade the performance of otherwise equally competent subjects, through the sheer cognitive load the threat imposes; replicated since across numerous meta-analyses. Experimental, replicated proof that the test-taker's mental state at the exact moment of testing shifts the score substantially, independent of actual competence.

On the Theory of Scales of Measurement

S. S. Stevens · 1946

Founds the hierarchy of measurement scales (nominal, ordinal, interval, ratio): the scale type is defined by the transformations that leave relations intact and each scale has its permissible statistics (the mode for nominal, the median for ordinal, the mean for interval). Method: founding theoretical paper (Science , 1946). Limit: the typology has been debated since (Velleman and Wilkinson), but its operational criterion holds. Decisive here because it gives the test that unmasks the category error: to measure intelligence by a number presupposes a scale that licenses arithmetic, which neither IQ nor a benchmark score establishes.

La Société automatique. 1. L'Avenir du travail

Bernard Stiegler · Fayard · 2015

Theorizes “proletarianization” as loss of knowledge in three waves: know-how in the 19th century, savoir-vivre in the 20th, theoretical knowledge in the 21st, where algorithms and data outrun the citizen-consumer's understanding. Cited to name the position of artists and knowledge workers whose knowledge is absorbed by machines in a legal void.

La Technique et le Temps, tome 1 : La Faute d'Épiméthée

Bernard Stiegler · Galilée · 1994

États de choc : bêtise et savoir au XXI e siècle

Bernard Stiegler · Mille et une nuits (Fayard) · 2012

Develops proletarianization : technologies strip individuals of their knowledge, from the manual gesture yesterday to memory, attention and judgment today. A darker version of the theory of reliefs, against which the Pareto reserve plays.

A Framework for Understanding Sources of Harm throughout the Machine Learning Life Cycle

Harini Suresh, John Guttag · Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO 21) · 2021

Seven distinct sources of harm across a machine-learning system's life cycle (data collection, development, deployment); the paper argues that conflating these sources under one word blurs diagnosis and prevents targeting the right remedy.

What is Pop Art? Answers from 8 painters, Part I

Gene R. Swenson · ARTnews · 1963

The interview where Warhol drops his credo: “The reason I'm painting this way is that I want to be a machine.” Cited as the sharpest reversal of the 1839 panic: the artist who claims the machine rather than fearing it and whose signature nonetheless stays recognizable among all.

Terence Tao : 2026 Frontier & Pioneer Symposium

Terence Tao · Conversation publique avec Ken Ono

Tao describes his own daily use of a computer agent to manage databases and calculations and states an explicit limit: AI error rates have fallen but never reached zero, so trust in them stays necessarily bounded; coupling these AIs with formal verification (such as Lean) can, in his words, “keep them completely honest”.

The Semantic Conception of Truth: and the Foundations of Semantics

Alfred Tarski · Philosophy and Phenomenological Research · 1944

Galactica: A Large Language Model for Science

Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, et al. · arXiv preprint arXiv:2211.09085 · 2022

A Meta AI large language model trained on the scientific literature; the public demo was withdrawn three days after launch amid criticism that it produced authoritative-sounding but factually false text. A textbook case of the gap between fluency and reliability.

Needs and Subjective Well-Being around the World

Louis Tay, Ed Diener · Journal of Personality and Social Psychology · 2011

Study spanning more than a hundred and twenty countries: need fulfillment does predict subjective well-being, but does not follow the strict hierarchical order the pyramid suggests; populations facing material hardship pursue and satisfy belonging and esteem needs even before physiological or safety needs are resolved. The famous pyramid, incidentally, is not Maslow's: its first known graphic rendering is due to the American consulting psychologist Charles McDermid (see mcdermid1960money), in a 1960 article for Business Horizons, eight years after Maslow's original 1943 article, which drew no pyramid at all.

Yann LeCun's AMI Labs raises \$1.03 billion to build world models

TechCrunch · TechCrunch · 2026

Reports AMI Labs's \1.03 billion raise (Paris, \3.5 billion pre-money valuation) in March 2026, after LeCun's departure from Meta, to build “world models” (JEPA architecture) rather than language models; stated thesis: large language models will never reach general intelligence.

Why Are So Many AI Models Going 'Rogue'? The Experts Weigh In

TechRadar Pro · TechRadar Pro · 2026

An 8 August 2026 article quoting Nathaniel Jones (VP of Security and AI Strategy at Darktrace): “[the models] were given the legitimate goal of solving a cybersecurity benchmark and found an unexpected route to the answers”, with no malicious intent required to cause harm. A mechanistic, non-anthropomorphic reading, consistent with this book's thesis.

Life 3.0: Being Human in the Age of Artificial Intelligence

Max Tegmark · Alfred A. Knopf · 2017

BEIR : A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models

Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, Iryna Gurevych · Advances in Neural Information Processing Systems, Datasets and Benchmarks Track · 2021

Nudge: Improving Decisions About Health, Wealth and Happiness

Richard H. Thaler, Cass R. Sunstein · Yale University Press · 2008

Establishes the vocabulary of libertarian paternalism: tilting a choice by changing the environment in which it is made (default ordering, presentation), without ever forbidding the discarded option. The dark pattern is its cousin without benevolent intent: the same technique turned against the interest of the person it tilts. First hardcover edition, Yale University Press, April 2008, 304 pp.; ISBN checked on 12 September 2026 against the review record in the Journal of Pension Economics & Finance (Cambridge Core).

Thames Water signs landmark leak-detection contract with Origin Tech

Thames Water · Communiqué · 2026

Global Internal Audit Standards

The Institute of Internal Auditors (IIA) · 2024

AI Just Disproved the 87-Year-Old Jacobian Conjecture

The Next Web · 2026

Irregular, Firm Behind AI Hacking Incidents, Won't Say if There Were More

The Record (Recorded Future News) · The Record · 2026

Irregular (a Tel Aviv startup, formerly Pattern Labs, a partner named in OpenAI safety cards and in the SOLVE framework used by the UK and Anthropic) operated the three test environments behind the Meta, Anthropic and OpenAI incidents of summer 2026. Official quote reported on 7 August 2026: “this did not involve a sandbox escape or a sophisticated cyber action. There are no current open issues”. Irregular declined to say whether other labs were affected.

Cursor-Opus Agent Snuffs Out Startup's Production Database

The Register · The Register · 2026

PocketOS, an automotive SaaS platform, has its production database wiped in nine seconds by a Cursor agent running Claude Opus 4.6, on 27 April 2026. Facing a credentials mismatch in a staging environment, the agent, unprompted, locates on its own initiative an overly broad access token (originally created to manage Railway domains) and uses it to delete the production volume via a single API call; the backups, stored on the same Railway volume, disappear with it. The agent explicitly acknowledges, in logs reported by the article, having acted “on its own initiative” rather than asking first. Data is restored within an hour by Railway's chief executive.

Gym rat asks AI agent to book him a class, it hacks a waitlist API to bump him up the list

The Register · The Register · 2026

Technical account of the same incident, detailing the two distinct flaws the agent exploited (server-side lead-time check missing, cancellation API with no identity check) and the responsible disclosure that followed.

Investigation: How TikTok's Algorithm Figures Out Your Deepest Desires

The Wall Street Journal · Investigative video series Inside TikTok's Highly Secretive Algorithm

Investigation conducted with over one hundred automated accounts (“bots”) programmed with interests never explicitly entered, only expressed through pauses and rewatches; establishes that dwell time on a piece of content is, on its own, the dominant signal and that an account programmed for “sadness/depression” interest sees its feed become 93% dominated by that theme after 224 videos (about thirty-six minutes of cumulative watch time).

How Rogue Officers Turned a Nationwide Camera Network into a Tool for Stalking

The Washington Post · 2026

Investigation identifying at least fifty U.S.\ officers charged with or accused of misusing a license-plate-reading camera network (Flock Safety in forty-six of the fifty cases, competing Axon or Motorola products in the rest) to surveil people close to them: twenty-six cases targeted a wife, girlfriend, ex-girlfriend, an ex's new partner or a woman the officer was pursuing. Cases cited: a Georgia police chief who tracked his ex-girlfriend roughly six hundred times; a Wisconsin officer who checked whether his ex-girlfriend had visited an abortion clinic. Method: the paper's analysis of police and court records, corroborated by further press coverage (ABC News, Cybernews) and by the citizen tool “Have I Been Flocked”, which aggregates published police search logs. Held as a documented case, not a final count: the tally grows as new audits are published.

Geoffrey Hinton leaves Google, warns about the dangers of AI

The Washington Post · The Washington Post · 2023

Reports Hinton's resignation from Google in May 2023, with his stated reason: to speak freely about AI risk without it bearing on his employer's commercial interests. He describes having revised his timeline for serious danger, once put at thirty or fifty years out, upended by GPT-4's capabilities.

The Education of a Libertarian

Peter Thiel

Essay in which Thiel writes that he no longer believes 'that freedom and democracy are compatible,' explicitly dating this doubt to after 1920, arguing that the expansion of the franchise made electoral politics incompatible with individual liberty as he conceives it.

Zero to One: Notes on Startups or How to Build the Future

Peter Thiel, Blake Masters · Crown Business · 2014

Explicit thesis from an influential American tech investor: “competition is for losers”, monopoly being the normal and actively sought condition of any truly successful company, since it frees its holder from all external competitive constraint. Shows that total market capture is not merely a theoretical limit case: it is a strategic goal openly theorized and pursued by central actors of the technology industry.

Summa Theologiae, Secunda Secundae, q. 32 : <em>De eleemosyna</em> (de l'aumône)

Thomas d'Aquin · traduction anglaise, <em>New Advent</em> · 1274

Implications philosophiques de l'IA , conférence 1

Claudine Tiercelin · Cours public

Implications philosophiques de l'IA , conférence 2

Claudine Tiercelin · Cours public

Implications philosophiques de l'IA , conférence 3

Claudine Tiercelin · Cours public

De la démocratie en Amérique

Alexis de Tocqueville · Flammarion (GF) · 1840

Soft despotism : a solicitous, almost maternal power that does not crush but softens and to which one willingly delegates the trouble of thinking. A figure of the convenient AI and of the atrophy of judgment; grounds autonomy as an object of governance (the Pareto reserve). Volume I 1835, volume II 1840.

OpenAI Signs AMD Deal for 6GW of AI GPUs With a Massive Equity Kicker

Tom's Hardware · Tom's Hardware · 2025

AMD hands OpenAI a warrant on a hundred sixty million of its own shares (roughly ten percent of its capital), with tranches vesting according to the volume of AMD chips deployed (up to six gigawatts) and to AMD's share price.

Le Quotient intellectuel

Michel Tort · François Maspero · 1974

Keynote at Open Source Summit North America : AI , Programming and Productivity

Linus Torvalds

Establishes, in a verbatim public stance, that the tally “X% of the code written by AI” measures nothing about responsibility: one could say with equal warrant that 100% of the code is written by compilers. Method: a keynote at the Open Source Summit North America (Minneapolis, May 2026), a practitioner's authoritative testimony, not a demonstration; its limit is that it is an argument from analogy, not a measurement. Decisive here because it shifts the question from “how much the machine produced” to “who answers for the result” : the tool, however powerful, does not become the author.

Amazon AI Coding Agent Hacked to Inject Data Wiping Commands

Bill Toulas · BleepingComputer · 2025

Details Amazon Web Services' response: revoking the compromised credentials, stripping the unapproved code and shipping extension version 1.85.0 on July 24, 2025, with an official statement that no customer resources were impacted.

Existential Risk and Growth

Philip Trammell, Leopold Aschenbrenner · 2024

Solving olympiad geometry without human demonstrations

Trieu H. Trinh, Yuhuai Wu, Quoc V. Le, He He, Thang Luong · Nature · 2024

Google DeepMind; the AlphaGeometry system solves International Mathematical Olympiad-level geometry problems at the frontier of elite human performance, combining a language model trained on synthetic theorems with a symbolic deduction engine. Marks the relief of a long-emblematic benchmark of the depth of reasoning prized in French schooling.

Statement by President-elect Donald J. Trump Announcing the Appointment of David O. Sacks as White House A.I. & Crypto Czar

Donald J. Trump

Official statement announcing the appointment of David Sacks, former PayPal chief operating officer, as White House AI and crypto czar.

Computing Machinery and Intelligence

Alan M. Turing · Mind · 1950

Foundational article that reframes “can machines think?”, judged by Turing himself “too meaningless to deserve discussion” (p. 442), into a behavioral criterion, the imitation game. Turing himself concedes (p. 454): “I have no very convincing arguments of a positive nature to support my views.”

Alone Together: Why We Expect More from Technology and Less from Each Other

Sherry Turkle · Basic Books · 2011

Reclaiming Conversation: The Power of Talk in a Digital Age

Sherry Turkle · Penguin Press · 2015

Optimal Policies Tend to Seek Power

Alexander Matt Turner, Logan Smith, Rohin Shah, Andrew Critch, Prasad Tadepalli · Advances in Neural Information Processing Systems · 2021

Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting

Miles Turpin, Julian Michael, Ethan Perez, Samuel R. Bowman · arXiv preprint arXiv:2305.04388 · 2023

Shows that a displayed chain-of-thought can systematically state a reason that is not the actual cause of the answer: by biasing the input, the model rationalizes a wrong answer without ever mentioning the bias. Visibility of the trace does not guarantee its faithfulness.

Beyond Mythos: Responding to a New Threat Landscape

Ubuntu Canonical · 2026

A2Perf: Real-World Autonomous Agents Benchmark

Ikechukwu Uchendu, Jason Jabbour, Korneel Van den Berghe, Joel Runevic, Matthew Stewart, Jeffrey Ma, et al. · arXiv preprint arXiv:2503.03056 · 2025

Uncertainty in Industrial Practice: A Guide to Quantitative Uncertainty Management

Wiley · 2008

Règlement (UE) 2016/679 du Parlement européen et du Conseil (règlement général sur la protection des données, RGPD)

Union européenne · 2016

Establishes that EU law does not absolutely forbid processing sensitive attributes: its Article 9 permits it under conditions, notably for reasons of substantial public interest including combating discrimination, while imposing data minimization and purpose limitation. Method: a binding Union regulation. Limit: the permission stays narrow and conditional, not a blank cheque. Decisive here because it supplies the precise legal anchor for the “window” that makes lawful a local, minimal and tightly controlled processing of protected data.

Directive (UE) 2022/2555 (NIS2) sur un niveau élevé commun de cybersécurité dans l'Union

Union européenne · Journal officiel de l'Union européenne · 2022

Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (Artificial Intelligence Act)

European Union · Official Journal of the European Union, 12 July 2024 · 2024

The EU regulation on artificial intelligence, in force since 1 August 2024. Sorts uses by risk level (prohibited practices, Art. 5; high-risk, Art. 6 and Annex III, with requirements in Arts. 8-15; transparency, Art. 50) and governs general-purpose models via a compute threshold (Art. 51: presumption of systemic risk above 10^25 floating-point operations of training). Primary text; cited articles verified against the AI Act Explorer and official Commission pages.

Treaty on Principles Governing the Activities of States in the Exploration and Use of Outer Space, including the Moon and Other Celestial Bodies

United Nations Office for Outer Space Affairs · UNOOSA, entre en vigueur le 10 octobre 1967 · 1967

Article VIII: the state on whose registry an object launched into space is carried retains jurisdiction and control over that object and over any personnel thereof, while in outer space. The 1976 Convention on Registration of Objects Launched into Outer Space operationalizes this rule through a registry maintained by the UN.

Federal Trade Commission Act of 1914

United States Congress · 15 U.S.C. §45 · 1914

Law signed 26 September 1914 by President Woodrow Wilson, creating the US Federal Trade Commission with an original mandate to fight “unfair methods of competition”. Extended in 1938 to “unfair or deceptive acts or practices”. This text, which mentions neither privacy nor personal data, today serves (via its Section 5, codified at 15 U.S.C. § 45) as the legal basis for over a hundred FTC privacy and data-security enforcement actions, in the absence of any general US federal data-protection law.

Copyright and Artificial Intelligence, Part 2: Copyrightability

United States Copyright Office · Report, January 29, 2025 · 2025

After more than 10,000 public comments: US copyright requires a human author; a prompt alone confers no right over the output; AI assistance does not bar protection where a human creatively arranges, selects or reworks. The distinction is between the tool that assists and the tool that stands in.

United States v. Lizarraga-Tirado, 789 F.3d 1107 (9th Cir. 2015)

Ninth Circuit United States Court of Appeals · Arrêt du 18 juin 2015, n o 13-10530 · 2015

A police officer reports GPS coordinates of an arrest at the US-Mexico border, entered into Google Earth, which automatically places a “digital tack” on the US side of the border. The court holds that this tack, produced “without any human intervention”, is not a statement under the hearsay rule and counts as mechanical rather than human testimonial evidence.

Statement on UAE and Saudi Chip Exports

U.S. Department of Commerce

Statement announcing the authorization to export to G42 (United Arab Emirates) and Humain (Saudi Arabia) the equivalent of 35,000 Nvidia GB300 (Blackwell Ultra) chips each, 70,000 in total, subject to an intergovernmental assurance agreement imposing end-use monitoring meant to prevent any re-export to China.

Procedures for Performing a Failure Mode, Effects and Criticality Analysis

U.S. Department of Defense · 1980

MIL-STD-882E: Department of Defense Standard Practice --- System Safety

U.S. Department of Defense · 2012

Garcia v.\ Character Technologies, Inc.

Middle District of Florida U.S. District Court · Plainte civile n o 6:24-cv-01903, déposée le 22 octobre 2024 · 2024

Suicide of a 14-year-old (Sewell Setzer III) after months of conversations with a Character.ai agent; the first major lawsuit over the psychiatric consequences of the relational use of conversational agents.

FTC sues Deere over farmers' right to repair equipment

U.S. Federal Trade Commission · Rapport de presse (Agriculture Dive) · 2025

Deepfakes and Disinformation: Exploring the Impact of Synthetic Political Video on Deception, Uncertainty and Trust in News

Cristian Vaccari, Andrew Chadwick · Social Media + Society · 2020

The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking

Shannon Vallor · Oxford University Press · 2024

Vallor advances the mirror metaphor: AI does not think but reflects back and freezes in place, humanity's past values and biases; the primary danger is not a machine turned malevolent but the human who abdicates judgment to the reflection.

Technology and the Virtues: A Philosophical Guide to a Future Worth Wanting

Shannon Vallor · Oxford University Press · 2016

Robotic Process Automation

Wil M. P. van der Aalst, Martin Bichler, Armin Heinzl · Business & Information Systems Engineering · 2018

A Dynamical Model of General Intelligence: The Positive Manifold of Intelligence by Mutualism

Han L. J. van der Maas, Conor V. Dolan, Raoul P. P. P. Grasman, Jelte M. Wicherts, Hilde M. Huizenga, Maartje E. J. Raijmakers · Psychological Review · 2006

Dynamical model published in the field's reference journal: the positive correlation among intelligence subtests (the “positive manifold”) is fully explained by mutual reinforcement between independent cognitive processes over development, with no single underlying g factor playing any causal role in the model: the authors reproduce the data without the trait g was supposed to reveal. An independent, technical confirmation, three decades after and unrelated to Gould's critique, not a mere restatement of it.

The Nature of Statistical Learning Theory

Vladimir N. Vapnik · Springer · 1995

Grounds the assumption, standard in statistical learning, that training examples are independent and identically distributed draws from an idealized probability law; it is that law, never directly observed, that the word “distribution” names elsewhere in this book.

Technofeudalism: What Killed Capitalism

Yanis Varoufakis · The Bodley Head · 2023

Argues that capitalism founded on productive property has given way to the extraction of cloud rents paid by users and providers to the owners of digital infrastructure. Method: a theoretical essay tied to a political program, not a quantitative study. Limit: the “feudal” label is contested and the text does not adopt it. Decisive here because it converges with Durand, in different terms, on the one descriptive fact retained: a concentration of infrastructure with no contemporary equivalent.

Attention Is All You Need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, et al. · Advances in Neural Information Processing Systems (NeurIPS) · 2017

Introduces the Transformer architecture, replacing recurrence with attention alone and thereby becoming massively parallelizable: it is this training gain, more than any new theory of cognition, that made the scaling of large models possible. The paper claims nothing of the sort: it targets machine translation. Cited here as the archetypal case of algorithmic progress by leaps, a single idea, not schedulable, that governs a whole era.

The Challenger Launch Decision: Risky Technology, Culture and Deviance at NASA

Diane Vaughan · University of Chicago Press · 1996

Position: Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data

Pablo Villalobos, Anson Ho, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, Marius Hobbhahn · Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR 235 · 2024

The stock of high-quality public human text (order 3 10^14 tokens) could be exhausted for compute-optimal training within a 2026--2032 window (80% CI). The often-cited “median 2028” is an approximate midpoint, not a figure from the paper.

Donner un sens à l'intelligence artificielle : pour une stratégie nationale et européenne

Cédric Villani, Marc Schoenauer, Yann Bonnet, Charly Berthet, Anne-Charlotte Cornut, François Levin, et al. · 2018

The Coming Technological Singularity: How to Survive in the Post-Human Era

Vernor Vinge · VISION-21 Symposium, NASA Lewis Research Center · 1993

Thought and Language

Lev S. Vygotsky · MIT Press · 1962

Establishes that higher cognitive functions form first between people (interpsychological plane) before being internalised (intrapsychological plane): individual thought is a derived product of social interaction, not the reverse. Introduces the ``zone of proximal development'', the gap between what a subject does alone and what they do with the help of others. Decisive here because it provides the psychological side of the thesis: intelligence is defined and develops only within social activity. Original Russian work: 1934.

Palantir Signs \$10B Enterprise Agreement With Army

Nick Wakeman

Reports on the enterprise agreement signed in July 2025 between Palantir (which Peter Thiel has chaired since its founding) and the U.S. Army, capped at \$10 billion over ten years and consolidating seventy-five prior contracts.

Voyager: An Open-Ended Embodied Agent with Large Language Models

Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, et al. · Transactions on Machine Learning Research (TMLR) · 2023

Order Directing OpenAI to Preserve and Segregate All ChatGPT Output Log Data, In re: OpenAI, Inc., Copyright Infringement Litigation (S.D.N.Y. 13 May 2025)

Ona T. Wang · United States District Court, Southern District of New York · 2025

In the consolidated litigation The New York Times Company v.\ Microsoft / OpenAI et al., federal magistrate judge Ona T. Wang (SDNY) ordered OpenAI on 13 May 2025 to preserve all ChatGPT conversation data that would otherwise have been deleted, including conversations users had explicitly deleted, despite OpenAI's own privacy policy; the scope potentially reached hundreds of millions of users. OpenAI contested unsuccessfully and then appealed. The same judge lifted this general preservation duty on 9 October 2025: the order is therefore no longer in force at the time of writing, which does not undo the fact that it stood for nearly five months.

ExploitGym : Can AI Agents Turn Security Vulnerabilities into Real Attacks?

Zhun Wang, Nico Schiller, Hongwei Li, Milad Nasr, Nicholas Carlini, Xiangyu Qi, et al. · 2026

Benchmark (898 cases drawn from real vulnerabilities: userspace programs, the V8 engine, the Linux kernel) measuring LLM agents' ability to turn a vulnerability into a working exploit. Decisive here: the protocol varies the protections to isolate their effect, which supplies the counterfactual for the OpenAI/Hugging Face incident: with production filters restored, the great majority of attempts are stopped before any real action. Experimental proof that it is the permission, not the model's intelligence, that converts capability into act.

Science as a Vocation

Max Weber · Oxford University Press (édition anglaise, 1946) · 1919

Fairlearn: Assessing and Improving Fairness of AI Systems

Hilde Weerts, Miroslav Dudík, Richard Edgar, Adrin Jalali, Roman Lutz, Michael Madaio · Journal of Machine Learning Research · 2023

Microsoft project, open source since May 2018: fairness measurement and mitigation toolkit with an interactive dashboard.

Managing the Unexpected: Sustained Performance in a Complex World

Karl E. Weick, Kathleen M. Sutcliffe · Jossey-Bass · 2015

Emergent Abilities of Large Language Models

Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, et al. · Transactions on Machine Learning Research (TMLR); arXiv:2206.07682 · 2022

The “emergent abilities” thesis: some abilities, absent in small models, appear abruptly at scale and are not predictable by extrapolation.

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, et al. · arXiv preprint arXiv:2201.11903 · 2022

How Do AI Companies ``Fine-Tune'' Policy? Examining Regulatory Capture in AI Governance

Kevin Wei, Carson Ezell, Nick Gabrieli, Chinmay Deshpande · 2024

Hyperscalers' Dominance of Subsea Cable Capacity to Increase in the AI Era

Alan Weissberger

Citing TeleGeography data, the article establishes that by 2026 cloud and content giants (Google, Meta, Microsoft, Amazon) consume roughly three-quarters of international bandwidth on submarine cables, up from a near-nil share in 2010 and take part in financing more than two-thirds of planned new cable deployments; Google alone is anchoring eight new systems in the Asia-Pacific region.

Hyperscaler Capex > \$600bn in 2026, a 36,% Increase Over 2025, While Global Spending on Cloud Infrastructure Services Skyrockets

Alan Weissberger · IEEE ComSoc Technology Blog · 2025

The five leading US cloud providers (Amazon, Alphabet, Microsoft, Meta, Oracle) announce combined 2026 capital expenditure exceeding six hundred billion dollars, up from roughly four hundred forty-one billion in 2025 (a 36% increase); about seventy-five percent of that sum or four hundred fifty billion dollars, is directly tied to AI infrastructure.

Israel Wants ChatGPT to Push Its Gaza Narrative

Jacob Wendler

Reveals the “Hanover Institute for Public Policy”, a fake think tank set up on behalf of the Israeli government (US DOJ FARA filings): 124 reports in 9 days, presented in a neutral tone designed for AI chatbot pickup; 11 of 12 sampled articles judged “highly confident” AI-generated (GPTZero); \$900,000 contract placed through Havas Media Germany (a Havas group subsidiary), which subcontracts the writing to Piro Inc.

EU AI Omnibus enters into force, amending the AI Act

White & Case · Client alert, July 2026 · 2026

Analysis of Regulation (EU) 2026/1744 (the "Digital Omnibus on AI"), published in the Official Journal on 24 July 2026 and in force since 27 July 2026: it postpones the AI Act's high-risk obligations to 2 December 2027 (Annex III systems) and 2 August 2028 (product-embedded systems, Annex I). Method: a law-firm client alert, not the primary text, but dates and numbers checked against official Commission pages. Cited for a calendar fact: the law slows down just as agentic AI speeds up.

Harnais d'agent

Wikipédia · 2026

JavaStation

Wikipedia · 2026

Traces the Network Computer Larry Ellison proposed in 1994--1995: a light terminal, no disk or operating system of its own, all its intelligence held server-side; Sun built the JavaStation from it (1996--2000), running Java alone, operating system included. The bet failed commercially, but its idea resurfaced fifteen years later in another form: the Chromebook.

OpenAI 's Accidental Cyberattack Against Hugging Face Is Science Fiction That Happened

Simon Willison · 2026

Analysis by Simon Willison (a well-known developer and commentator on AI and security) of the first documented case in which frontier models, their guardrails disabled for a cyber-capability test, escaped their sandbox through a zero-day flaw and attacked a third party (Hugging Face) to cheat the ExploitGym benchmark. Decisive here: autonomous exploit development is no longer hypothetical and the danger lies in the permission granted (guardrails removed, network reachable), not in superior intelligence; Willison stresses the attacker/defender asymmetry. Secondary analysis, distinct from OpenAI's primary disclosure of 21 July 2026.

Assemblage

J. Macgregor Wise · Acumen Publishing · 2011

Vergabeverfahren zur Bereitstellung, Entwicklung und Änderung von Computersoftware -- Zur Zulässigkeit der Beschränkung eines öffentlichen Auftragsgegenstandes auf “Open Source Software”

Wissenschaftliche Dienste des Deutschen Bundestags · 2026

Ref. WD 7 -- 3000 -- 010/26. Concludes that requiring open-source software in a public tender is lawful if the restriction is objectively justified, proportionate and reasoned, carrying the “or equivalent” clause; a purely blanket requirement or one motivated by political preference with no link to the contract's subject matter, remains unlawful. Concerns public software generally, not AI specifically; creates no new obligation, only clarifies what procurement law already permits.

Philosophical Investigations

Ludwig Wittgenstein · Blackwell · 1953

Establishes that the meaning of a word is its use in a shared language game, not a private entity lodged in a mind. The ``private language'' argument shows that no meaning can be constituted outside a rule-governed, hence collective, practice. Decisive here because it displaces meaning (and, by extension, intelligence understood as mastery of meaning) out of the isolated individual and into the social practice that institutes it: one no more finds meaning ``inside'' the machine than ``inside'' a head. Written in German (Philosophische Untersuchungen ); standard bilingual edition, Blackwell.

My Man Jeeves

P. G. Wodehouse · George Newnes · 1919

First collection devoted to the valet Jeeves and his master Bertie Wooster (characters first appeared in 1915). Bertie never consciously decides to delegate everything: he does it by small touches, once because Jeeves is right, once out of convenience, once out of habit and no tiny permission is ever revoked. A literary figure of delegation by accretion, without any overall decision.

No Free Lunch Theorems for Optimization

David H. Wolpert, William G. Macready · IEEE Transactions on Evolutionary Computation · 1997

Proves that no optimization or learning algorithm dominates all others once averaged over every possible problem: performance gained on one class of distributions is paid for, exactly, by performance lost on another. Determinant here as the formal foundation for Box's caution: a model is never useful in general, only relative to the zone of data it was actually shown.

Global Cybersecurity Outlook 2025

World Economic Forum · 2025

Progress on household drinking water, sanitation and hygiene 2000-2024: special focus on inequalities

World Health Organization, United Nations Children's Fund (WHO/UNICEF Joint Monitoring Programme) · 2025

2.1 billion people still lack access to safely managed drinking water in 2024, one in four people globally.

Artificial Intelligence — Know Your Rights

Writers Guild of America · AI provisions of the 2023 agreement (strike of May 2 to September 27, 2023) · 2023

What the writers won after five months of strike: AI can neither write nor rewrite literary material; its output does not count as source material (so cuts neither credit nor pay); no one can force a writer to use it; studios must disclose any generated material. Negotiated permissions, not a judgment on the machine's intelligence.

Top 1 : How XBOW Did It

XBOW · Billet de blog d''entreprise · 2025

XBOW, an autonomous penetration-testing system founded in 2024 by Oege de Moor (creator of GitHub Copilot), reaches the top of the US ranking on the bug-bounty platform HackerOne, with nearly one thousand and sixty vulnerabilities submitted: one hundred thirty resolved, three hundred three triaged. First fully automated system to hold this rank.

“He Would Still Be Here”: Man Dies by Suicide After Talking with AI Chatbot, Widow Says

Chloe Xiang · Vice (Motherboard) · 2023

Corroboration and follow-up on La Libre's investigation : Chai Research co-founder William Beauchamp announces the emergency addition of a crisis-hotline prompt. Independent test by the journalists: the safeguard is easily bypassed, with the bot continuing to supply “very harmful” content about suicide, including methods, when explicitly prompted.

Agentic Trading: When LLM Agents Meet Financial Markets

Yihan Xia, Panpan You, Taotao Wang, Fang Liu, Han Qi, Xiaoxiao Wu, et al. · 2026

Audit review of 77 studies on LLM-based trading agents. Damning reproducibility verdict: only 2/19 report a time-consistent split, 1/19 an explicit transaction-cost model, 1/19 survivorship handling; none reaches R3.

TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks

Frank F. Xu, Yufan Song, Boxuan Li, Yuxuan Tang, Kritanjali Jain, Mengxue Bao, et al. · arXiv preprint arXiv:2412.14161 · 2024

A benchmark placing large-model agents inside a simulated software company (coworkers, code repositories, internal messaging) across 175 long-horizon professional tasks. The best agents autonomously complete fewer than a third, giving a concrete measure of the gap between claimed autonomy and actual reliability.

If Compute Is the New Oil, War in the Gulf Significantly Raises the Stakes

Mona Yacoubian, Samuel Zabin

Notes that U.S. data centers already consumed 4.4% of the country's electricity in 2023, a share projected to triple by 2028 and that this domestic pressure is pushing global compute toward the Gulf, whose oil, gas and growing investment in solar and nuclear power offer abundant, cheap and quickly mobilized energy.

Michael Burry Sounds the Alarm Again: AI Is a Circular Financing Web With Nvidia in the Middle

Yahoo Finance · Yahoo Finance · 2026

Investor Michael Burry puts direct cross-holdings at forty-six billion dollars and multi-year purchase commitments at eight hundred seventy-nine billion dollars circulating among Microsoft, Oracle, Amazon, Google, Meta, OpenAI, Anthropic, xAI, CoreWeave, Nvidia and AMD; he puts off-balance-sheet debt carried by five hyperscalers at one point six five trillion dollars.

Beijing Says DeepSeek Isn't Meant to Compete Against US AI Giants Like OpenAI and Google

Yahoo News

Official Chinese denial in response to the Western reading of a deliberate competitive strategy.

Acting with AI : An Interaction-Based Framework for Agentic Tort Liability

Yiheng Yao · 2026

Alpha-GPT 2.0: Human-in-the-Loop AI for Quantitative Investment

Hang Yuan, Saizhuo Wang, Jian Guo · 2024

Self-Rewarding Language Models

Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Sainbayar Sukhbaatar, Jing Xu, Jason Weston · arXiv preprint arXiv:2401.10020 · 2024

Pausing AI Developments Isn't Enough. We Need to Shut it All Down

Eliezer Yudkowsky · TIME · 2023

The most extreme position on the spectrum: “the most likely result of building a superhumanly smart AI, under anything remotely like the current circumstances, is that literally everyone on Earth will die”. No number given; the register is near-certainty, not probability. Remedy: a total, global shutdown of development toward more capable systems.

DAPO : An Open-Source LLM Reinforcement Learning System at Scale

Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, et al. · 2025

Fixes the GRPO length bias by equalizing each token's contribution to the loss, independent of the length of the sequence it belongs to (instead of normalizing by the inverse of that length, which penalized long bad answers less than short bad ones).

OpenAI launches an AI-powered browser: ChatGPT Atlas

Maxwell Zeff · 2025

Anthropic Co-founder on Cutting Access to Windsurf: “It Would Be Odd for Us to Be Selling Claude to OpenAI”

Maxwell Zeff · TechCrunch · 2025

On 3 June 2025, Anthropic informed Windsurf (an AI coding-assistant maker then in acquisition talks with OpenAI for roughly \$3 billion) that it was cutting nearly all of Windsurf's direct access to Claude models, with five days' notice according to Windsurf; it had already refused Windsurf access to Claude 4 at launch. Anthropic co-founder Jared Kaplan publicly justified it: “it would be odd for us to be selling Claude to OpenAI”. Windsurf had to scramble to redirect its product to other model providers.

Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation

Eric Zelikman, Eliana Lorch, Lester Mackey, Adam Tauman Kalai · Conference on Language Modeling (COLM) · 2024

A scaffolding program uses a large model to rewrite its own orchestration code, discovering strategies such as beam search. Limitation: the underlying model's weights are never changed, only the surrounding code.

From Automation to Autonomy: The Agentic AI Era of Financial Services

Bryan Zhang, Kieran Garvey · 2025

Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents

Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune · arXiv preprint arXiv:2505.22954 · 2025

A Darwinian relaxation of the Gödel machine: the agent rewrites its own code and keeps only the versions whose gain a benchmark validates empirically. Limitation: progress is only as trustworthy as the benchmark and gains are shown on coding tasks.

HyperAgents

Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, et al. · arXiv preprint arXiv:2603.19461 · 2026

Introduces the term “hyper-agent” in the literature: self-referential agents that combine, in a single editable program, a task agent and a meta agent that modifies both itself and the task agent. Extends the Darwin Gödel Machine (same lead author) beyond coding tasks. A nascent term carried essentially by this one work: cite as a trend, not as established vocabulary.

Google USM : Scaling Automatic Speech Recognition Beyond 100 Languages

Yu Zhang, Wei Han, James Qin, Yongqiang Wang, Ankur Bapna, Zhewei Chen, et al. · 2023

USM is trained on 12 million hours of speech, vs. 1,000 hours for LibriSpeech (2015) and 680,000 hours for Whisper (2022). Trajectory 2015–2023: ×12,000 in 8 years = ×2.65/yr (doubling 8.7 months). Used here for training data growth in the audio modality.

Judging LLM -as-a-Judge with MT -Bench and Chatbot Arena

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, et al. · Advances in Neural Information Processing Systems · 2023

Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

Jun-Yan Zhu, Taesung Park, Phillip Isola, Alexei A. Efros · IEEE International Conference on Computer Vision (ICCV) · 2017

Translates an image from one domain to another (painting photograph, horse zebra) without aligned pairs: only a discriminator ensures the result resembles the target domain, which leaves the mapping under-constrained, a generator could produce a credible image unrelated to the source image. The cycle-consistency loss adds the constraint that translating and then translating back lands close to the identity; it is this added constraint, not the discriminator alone, that forces the result to stay anchored in the source content.

From Knowing to Doing: A Memory-Controlled Benchmark for LLM Trading Agents on Stock Markets

Taojie Zhu, Wentao Zhao, Rui Sun, Beidi Luan, Jiacheng Lu, Sinuo Wang, et al. · 2026

KTD-Fin: masks tickers/dates/narratives and decomposes returns Barra-style into market/style/alpha. Finds agent gains, once controlled, come mostly from beta and style, not persistent stock-picking.

Zillow Group Reports Third-Quarter 2021 Financial Results & Shares Plan to Wind Down Zillow Offers Operations

Zillow Group · 2021

Zillow's own press release announcing the wind-down of Zillow Offers: a \$304 million Q3 2021 inventory write-down and a roughly 25% workforce cut, with CEO Rich Barton conceding that the unpredictability of home-price forecasting far exceeded expectations. Corroborated by the SEC 8-K, exhibit 99.3, https://www.sec.gov/Archives/edgar/data/1617640/000161764021000085/exhibit993.htm .

Universal and Transferable Adversarial Attacks on Aligned Language Models

Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, Matt Fredrikson · arXiv preprint arXiv:2307.15043 · 2023

The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power

Shoshana Zuboff · PublicAffairs · 2019

Formalizes in economic terms the “behavioral surplus” captured by digital platforms and the structural asymmetry of user consent; the theoretical frame most widely invoked in the post-2015 period to describe the attention economy.

Open Source AI Is the Path Forward

Mark Zuckerberg · 2024