Eli Lifland reasons through whether pausing training without pausing experiments helps AI safety
eli_lifland · x · 2026-09-30
AI forecaster Eli Lifland step-by-step reasons about the strategy of slowing frontier training while continuing experiments. He argues it may not matter much on the timescale of training the next model, with two exceptions: (a) you lose a better model for doing RSI, and (b) some failures only appear at frontier-scale training runs, so serial training experience is valuable.
He adds that training compute is generally small relative to experiment compute, so pausing training (redirecting its compute to safety) without pausing experiments likely has limited effect over months.
More from AGI Musings
- Verifying all of nixpkgs: 95% of machine code for ~$134M by end of 2027, study models — ctjlewis · 2026-09-30
- X User Likens FTX to EA/Rat Orgs' 'Hiring Your Friends' Problem — EigenGender · 2026-09-30
- kalomaze calls ES/genetic-algorithm training pointless unless backprop is truly intractable — kalomaze · 2026-09-30
- Edge Agent Village V2: hundreds of humans to live with personal AI agents for 3 weeks — sydneymlevine · 2026-09-30
- Peter Diamandis: We're Still Living in the Singularity — PeterDiamandis · 2026-09-30
- Judea Pearl points to Eric Horvitz's remarks on probabilistic reasoning's 42-year-old rise in AI — yudapearl · 2026-09-30