Chollet: Is RLVR just boosting math and code skill, or raising general intelligence G?
fchollet · x · 2026-10-07
François Chollet poses a key question: does RLVR-driven improvement in math and code reflect a rising general intelligence factor "G", or is it just an isolated skill?
He hypothesizes that the "jagged frontier" may mainly be math + code—domains RLVR can push arbitrarily far—while everything else plateaus because it remains bottlenecked by human-generated data. The open question: is the steady progress in non-verifiable domains a side effect of higher G, or only of more data? This bears directly on forecasts of model capability ceilings and AGI timelines.
More from AGI Musings
- Three types of people vibe coding SaaS clones: nerds, broke devs, and paid shills — kylegawley · 2026-10-07
- Neuroscientist Zador: Give AI Agents Real Fears Instead of Just Telling Them Not to Be Naughty — TonyZador · 2026-10-07
- If Users Can Tell You're Cheating with AI, Maybe You Don't Love the Domain — brandon_xyzw · 2026-10-07
- Photoshop decompiled and reimplemented in Rust via LLM, sparking closed-source moat debate — JosephJacks_ · 2026-10-07
- Beff Jezos quips: human civilization is a hierarchy of meat proxies for the Kardashev scale — beffjezos · 2026-10-07
- BLUECOW009's chain of optimism: solve math, then energy, then the environment, then poverty — BLUECOW009 · 2026-10-07