Training AI Researchers on Small Compute Risks Learning 'Anti-Bitter-Lesson' Habits
1a3orn · x · 2026-11-22
1a3orn discusses the risk of training research-capable AIs at small compute scale: models might learn a bunch of "anti-bitter-lesson" adaptations that transfer awfully to larger scales.
He adds a cost argument: even a minimal experimental run costs $1k–$10k (MATS-fellow scale), suggesting a 10^5-order floor on training compute cost — assuming models must learn to infer facts about large runs from much smaller ones. A possible sidestep: models getting good via theoretical math and insight rather than experiments.
More from AGI Musings
- Will everyone publish a 'values constitution' for their AI agents? — round · 2026-11-22
- The underrated detail in 'Claude solved it': Fields-medal-caliber minds with AI tools — avlok · 2026-11-22
- Rob Miles to Anthropic: 'You're good people, but you might kill us all' — zetalyrae · 2026-11-22
- McLuhan's 1970 note resurfaces: obsolescence is the matrix of innovation — round · 2026-11-22
- Anil Seth vs Rohit Krishnan to debate AI consciousness and interiority again — anilkseth · 2026-11-22
- AI reshapes how proofs get made: slop first, abstract later — and a BCI twist — doomslide · 2026-11-22