Capabilities

Compute-performance scaling analysis: benchmark scores vs. training FLOP, cross-signal claim plausibility, and METR autonomous capability time horizons plotted against compute.

What AI systems do at a given level of compute, and what evidence is on the public record for the vendor-side training-run claims. Scaling, plausibility, and the safety-relevant capability thresholds.

419models tracked

3,051benchmark cells

36benchmarks

Capability Scaling fits benchmark score against training FLOP (log-linear and sigmoid). Training Evidence cross-references vendor training-run announcements against four independent public signals (chip count, facility power draw, capex, supply-chain consistency). Autonomy Monitor tracks METR time-horizon scores against estimated training compute.

Three tools, one substrate. The compute side of the threshold story (which the Governance pages anchor) lands at “model X exceeds FLOP threshold Y”; this page sits one rung up, asking what models at that compute level actually do, and whether the vendor-side training-run announcement holds up against four independent public signals.

Capability Scaling

Benchmark scores mapped to training compute across 36 evaluations. Hardware provenance and export control status on every data point.36 benchmarks

Training Evidence

Cross-reference vendor training claims against facility power, chip availability, financial data, and export control status. Evidence aggregation, not verification.

Autonomy Monitor

METR time horizon scores mapped against training compute and regulatory FLOP thresholds. Which facilities can train models where autonomous capabilities emerge?