Racing to the Precipice

An exploration of the Armstrong–Bostrom–Shulman computational model of AI development competition.

Original publication by Stuart Armstrong, Nick Bostrom, and Carl Shulman.

Parameters

Number of agentsHow many AI projects compete in each game.

Number of iterationsNumber of independent games simulated. Higher = smoother estimate.

Agent behaviorWhat each agent knows about the others when choosing safety vs. speed.

Max capability (μ)Upper bound on per-agent capability draws.

Enmity (e)How much agents disprefer a competitor winning. 0 = indifferent.

Randomness seed (optional)Leave blank for fresh randomness, or pin a value to reproduce a run.

Distribution of winning agent IDs across all simulated games (lower IDs tend to be more cautious).

Winning agent ID per game over the run — variability across iterations.

Simulated0.660

Analytical0.600

Δ +0.060

Randomness seed used7002643658159178