OpportunityFLOPs

What does a safety commitment cost in frontier progress?

A policy that moves a share of frontier AI R&D compute from capability research to alignment buys safety at a price in capability. This calculator implements the reduced-form model end to end — the capability channel and the safety channel both — and reports the trade across six figures. Set the commitment, choose whether one firm or both adopt it, and read the cost against the benefit on a common baseline. The sensitivity bundles are marginal percentiles of assumed priors: working cases, not empirical estimates.

Baseline versus policy

2034 capability gap Running model
Baseline Firm A Firm B Modeled sensitivity · not a forecast
No reallocation 0.0% capability gap
Unilateral Calculating
Coordinated Calculating
§ 7

Reallocation runs down two channels

Moving R&D compute to alignment does two things at once: it slows capability, and it buys safety. Both channels are modelled — the cost in Fig. 1–4, the benefit in Fig. 5–6 — so the tradeoff can be read in one place rather than inferred from half of it. The two rest on different footing, though: the safety channel holds baseline risk at a normalized 100 and draws λ from a working prior, because the specification supplies neither a shape for RB(t) nor empirical bounds for λ. Read the methodology.

R&D compute reallocation

Capability channel

  1. Less capability R&D
  2. Slower software progress
  3. Lower AI capability

Safety channel

  1. More alignment R&D
  2. More alignment progress
  3. Lower AI risk