AI safety allocation · compiled 26 September 2026

Where AI safety's people, money and compute go

Every number has an 80% interval and a source grade, from C1 (counted from a primary source) to C5 (essentially a guess): what the grades mean. Built by parallel research agents, rechecked twice, then audited independently by a second model family (GPT-6 via Codex). Use the controls at the top to switch every chart between absolute values and shares, and between all, inside-lab and outside-lab work.

Same field, five very different shapes

What the nine categories mean

Evidence, confidence and method

Confidence grades

C1Direct primary source, counted by us (e.g. mentor lists, grant databases, 990 lines). C2Primary source but partial or sampled, or a credible secondary source reporting the number. C3Inferred from real data with the working shown. C4Weakly grounded: few examples, or heavy modelling assumptions. C5Essentially a guess, shown so the structure is visible.

A headline takes the grade of its weakest load-bearing input. Intervals are 80% ranges. They combine source noise with a model of classification ambiguity: on recheck, about 10% of items were firmly miscategorised and 20–35% were genuinely ambiguous. For compute they also include shared errors: annualising a one-week measurement, other labs' unknown safety shares, and scope.

Load-bearing claims, rechecked

Each was re-fetched from its source by an independent verifier. Filter by grade or by check result.

Caveats

Full methods