Post-training may exceed 75% of FLOPs, undercutting perplexity-based safety bars
stochasticchasm · x · 2026-11-23
In a meme-topped thread, the author welcomes OpenAI's new pretraining transparency but pushes back on the core argument: with post-training compute likely exceeding 75% (perhaps far more) of total FLOPs, arguing for "common limits" based on pretraining perplexity seems misguided. The target is OpenAI's "Perplexity and Intelligence" post by merettm, which proposes gating further progress on understanding AI behavior once perplexity-based limits are hit. The author closes with a tongue-in-cheek cheer for pretraining and a future 100T-parameter model.
More from AGI Musings
- Uber burned its annual AI budget in 4 months with no visible improvement — yaroslavvb · 2026-11-23
- FDA greenlights first human trial of age-reversal gene therapy — Dr_Singularity · 2026-11-23
- Gary Marcus doubles down: calling out overhyped LLM capabilities is engineering rigor — GaryMarcus · 2026-11-23
- AI safety researcher slams doxxing in the community: don't burn the commons on turn one — sjgadler · 2026-11-23
- AI proofs are a different kind of brute force: fast enough to skip inventing new math — moultano · 2026-11-23
- Google researcher: AI proofs skip human-style cleverness via speed and knowing all of math — moultano · 2026-11-23