Compute Thresholds: Regulating by FLOPs

Regulators want to catch frontier risk without freezing the whole field. They need a trigger that's measurable, hard to game, and correlated with capability. Training compute — floating-point operations used during model training — is the best proxy currently available.

The main thresholds (as of ~2025)

Why this approach has critics

• Algorithmic efficiency means same capability at lower compute over time — thresholds drift
• Small specialized models can be dangerous without being compute-heavy
• Inference compute (test-time reasoning like o1) is not captured by training FLOPs
• Distillation can transfer capability from a big model to a small one
• Open-source releases create compute-independent diffusion

The big idea: compute is the only thing regulators can easily count. That makes it the default regulatory hook, for better and worse.