Pause Training but Not Experiments? Compute Reallocation May Not Matter Much

eli_lifland · x · 2026-09-30

Forecaster elilifland, discussing rumors of a lab pausing training, said he puts only 25% on an actual substantial slowdown and that the duration of any pause is unclear.

He added a compute perspective: training compute is generally small relative to experiment compute, so pausing training while experiments continue may not matter much on the timescale of training the next model, even if the freed training compute were redirected to safety work.

Related event: Forecaster Eli Lifland doubts training pause will meaningfully slow frontier AI, puts ~25% odds on real slowdown(6 posts)→

Original post →

More from AGI Musings

AGI Musings channel →