Forecaster Eli Lifland doubts training pause will meaningfully slow frontier AI, puts ~25% odds on real slowdown

In late September 2024, AI safety forecaster Eli Lifland (in discussion with researcher gleech) offered a systematic analysis of reports that a frontier lab (rumored to be OpenAI) had paused training its next-generation model. As of Saturday, training appeared to still be paused, but Lifland said this is not strong evidence — he puts only about 25% probability on a substantive slowdown and doesn't know how long the officially claimed pause will last.

Confirmed

• As of the discussion (Saturday), training appeared to still be paused, but both Lifland and gleech agreed this isn't strong evidence of a substantive slowdown.
• Lifland gave his explicit subjective probability: only about a 25% chance of a substantive slowdown.

Core arguments

• Lifland's key compute perspective: at frontier labs, training compute is typically far lower than experimental compute, so even pausing training without pausing experiments may have little impact on the months-long timescale of "training the next model" — even if the freed-up training compute is redirected to safety research.
• More generally, he expects OpenAI to overstate its slowdown in public statements, i.e., the advertised slowdown may be larger than the actual one.

Exceptions

Lifland acknowledged two exceptions to the "pause training, don't pause experiments" strategy: first, losing the opportunity to run recursive self-improvement (RSI) with stronger models; second, some risks and research can only be explored near next-generation model capability levels, and pausing training closes that window.

Why it matters

This analysis offers a quantified skeptical lens on the "training pause" rumors: even if true, its actual braking effect on frontier AI progress may be far smaller than it appears, since compute can be reallocated to experiments; but the RSI and next-model-dependent experiment exceptions keep the issue substantively relevant to AI safety debates.

2026-09-30 ~ 2026-09-30 · 6 related posts

• [source] Eli Lifland: OpenAI's training pause likely isn't a substantial slowdown, compute shifts to experiments — eli_lifland · 2026-09-30
• Eli Lifland: lab 'training pauses' likely oversold, only ~25% chance of real slowdown — eli_lifland · 2026-09-30
• Training compute is small vs experiments, so a training pause may barely matter — eli_lifland · 2026-09-30
• [source] Researchers put only ~25% odds on a real frontier AI training slowdown — gleech · 2026-09-30
• Pause Training but Not Experiments? Compute Reallocation May Not Matter Much — eli_lifland · 2026-09-30
• Eli Lifland reasons through whether pausing training without pausing experiments helps AI safety — eli_lifland · 2026-09-30