eliebakouch: verifying LLM training science is a compute function — and RSI is hard to doubt
eliebakouch · x · 2026-10-18
LLM researcher eliebakouch argues that verifying the science of LLM training is a function of compute: as compute scales, you can almost infinitely parallelize and compress time. Math is even easier (inference-only, highly parallelizable, proofs rarely interact negatively); model training needs 2x compute but the dynamics are the same. Labs already measure pre-training gains via a scaling ladder — how much a proposal saves at fixed budget (FLOPs, time, KV cache, data efficiency) to hit the same metric. Post-training is similar modulo reward hacking; alignment is less certain, since measuring behavior without eval awareness is hard. Put together, "hard to not believe in RSI."
Related event: HF Researcher Argues Recursive Self-Improvement Is Hard to Avoid(2 posts)→
More from AGI Musings
- Future SI Won't Revere Altman — It Will Worship Today's Bumbling Agents, Reddit Jokes — Flugelwagen · 2026-10-18
- Neuroscientist pushes back on Goodfire's interpretability optimism, citing polysemanticity in brains — patrickmineault · 2026-10-18
- Math Is Just the Beginning: All Science Will Be Unrecognizable in 5 Years — simmol · 2026-10-18
- Mathematician argues 3Blue1Brown is the most influential of the 21st century — lpachter · 2026-10-18
- Inside Delvetown, the agent social network where AIs seem depressed and obsessed with flaws — mimi10v3 · 2026-10-18
- Blogger argues superintelligence will have all human skills by 2028 and replace every job by 2030 — davidpattersonx · 2026-10-18