Chollet: The jagged frontier may be math+code via RLVR while everything else plateaus on human data
fchollet · x · 2026-10-07
François Chollet poses a key open question: the jagged AI capability frontier may be mainly math + code, which RLVR can push arbitrarily far, while other domains plateau because they remain bottlenecked by human-generated data.
Non-verifiable areas keep improving steadily, but far slower than math and code. The crucial uncertainty: is that steady improvement a side effect of higher general capability driven by RLVR, or merely a function of the ongoing massive injection of new human data into training? He argues a lot depends on the answer.
More from AGI Musings
- Richard Ngo's hard sci-fi Tinker sketches how AI designs chips for its successors — teortaxesTex · 2026-10-07
- tszzl: Don't be an uncomprehending audience clapping at the AI magic show — Afinetheorem · 2026-10-07
- VraserX: Personal AI agents like Grok Bot, Muse and Dot will soon be everywhere — VraserX · 2026-10-07
- Predicting a 'third thing' RSI: capability leap in ~9 months via automated R&D, short of full FOOM — teortaxesTex · 2026-10-07
- Hinton, Bengio and 20+ experts urge governments to track how much AI research is done by AI — rohanpaul_ai · 2026-10-07
- For Which Generation Will Death Become Optional? — thesaraharminta · 2026-10-07