How much slower does takeoff go with 10× less compute?

According to LessWrong, a new analysis finds that slashing AI research and development compute by ninety percent—a ten-fold reduction—would slow AI capability improvements by a factor of about six. The study models how reductions in experiment compute, automated research agents, and training runs would compound to degrade progress rates. The researchers built this on a semi-endogenous growth model accounting for how software efficiency improvements become progressively harder to achieve as capability rises. The eighty percent confidence interval spans three point five to eight times slower, suggesting the relationship is fairly robust. The implication: compute is not just an accelerator for AI progress—it's a multiplier on research productivity itself, so constraints on it scale nonlinearly with slowdown.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton