kuchaev: ChatGPT, o1, and agents weren't scaling wins but training-stack innovations
kuchaev · x · 2026-11-08
kuchaev argues each capability jump came from training innovations, not scaling: ChatGPT was unlocked by alignment/RLHF over GPT-3, o1-style reasoning models required post-training innovation (RLVR), and agentic models needed new training-stack work beyond reasoning models. Innovation, he says, will continue.
More from AGI Musings
- Lobsters thread: how AI traffic is degrading the web for human developers — vboykis · 2026-11-08
- Does humanity have a moral obligation to build superintelligence? A future-lives argument — VraserX · 2026-11-08
- jun_song: 'Human slop'—people clueless about AI yelling for bans—annoys me more than AI slop — TinfoilTricorn · 2026-11-08
- Dev slams LLM consciousness essay as "zero-reasoning copium" that equates LLMs with calculators — mgostIH · 2026-11-08
- Should labs run adversarial evals to detect models deceiving them about AI R&D? — dfrsrchtwts · 2026-11-08
- Pliny's take: other species have coexisted with a recursively self-improving superintelligence — us — ZeroStateReflex · 2026-11-08