Figures

The data race

What machines receive: corpora, compute, context · and the wall where the gap in paces finally lands.

The data explosion · text: Reference text corpora grow by orders of magnitude across decades. ZIP · CSV + make.py

The data explosion · vision: The same curve for annotated image datasets: the gift extends to vision. ZIP · CSV + make.py

The data explosion · speech: Transcribed speech hours entrusted to models follow the same slope. ZIP · CSV + make.py

The data explosion · tabular: Reference tabular datasets, from early bases to massive warehouses. ZIP · CSV + make.py

The data explosion · genomics: Sequencing makes biology machine-readable at scale: the same explosion, life side. ZIP · CSV + make.py

Paces of the race · overview: The yearly multiplication factor of the inputs that feed and run models: compute, data, context, outputs. ZIP · CSV + make.py

Paces of the race · text: The text modality in detail: training corpora and context windows. ZIP · CSV + make.py

Paces of the race · vision: The vision modality in detail: image datasets and inference throughput. ZIP · CSV + make.py

Paces of the race · audio: The audio modality in detail: hours learned and hours produced. ZIP · CSV + make.py

The data wall: Compute quadruples yearly while data grows by 2.4×: the gap meets the finite stock of public human text (2026–2032 window). A limit to push back, at a cost still to be paid. ZIP · CSV + make.py

Measures, risks, markets

What gets measured once deployed: real productivity, catalogued risks, shadow uses, already-automated markets.

METR: expected gain, measured slowdown: Four positive expectations against the single, negative measurement · developers feel sped up by AI while the study by METR (Model Evaluation and Threat Research), an independent model-evaluation lab, measures a slowdown. ZIP · CSV + make.py

The MIT AI risk repository: Two cuts of the repository: accident weighs at least as much as intent and danger mostly arises after deployment. ZIP · CSV + make.py

Shadow AI by rank: The share of employees putting speed before cybersecurity grows with hierarchical rank. ZIP · CSV + make.py

The lever is information on the target: Measured effect without, then with information on the target: remove the information and the effect collapses. The arrow's length is the point. ZIP · CSV + make.py

The market already on autopilot: Even before agents, automation dominates order flow, well before it dominates capital ownership. ZIP · CSV + make.py