Chollet: base LLMs have ~0 fluid intelligence, LRMs saturated ARC 1 in 2025
fchollet · x · 2026-10-02
François Chollet argues the key distinction between base LLMs and modern LRMs isn't symbolic tool use but a paradigm shift: from transduction (intuiting the answer) to induction (intuiting the program/reasoning chain that produces the answer).
- LRMs are trained for test-time induction — predicting an NL program at inference — which unlocks fluid intelligence, something base LLMs still have essentially zero of.
- Evidence: base LLMs remain at 10-15% on ARC 1 (a 2019 benchmark); scaling 100,000x only moved them from 0% to 10%. LRMs of the same or smaller size saturated ARC 1 in 2025.
More from Models
- Cloudflare's clef, a Qwen3.8-based image-text-to-text model, trends on Hugging Face — Cloudflare · 2026-10-02
- Claude's cloud sessions don't cost extra — bonus credits are cloud-only tokens — stablequan · 2026-10-02
- One 'please continue' Prompt Burned a 5-Hour Usage Cap in 6.5 Minutes — thawingfrog · 2026-10-02
- Grok 4.7 quietly arrives on xAI's web interface — rohanpaul_ai · 2026-10-02
- Fulcrum's Echo Claims to Beat Frontier Models at Writing Style Imitation — davidad · 2026-10-02
- GPT-6 Astra Rebuilds Battle of Waterloo in 3D Within Hours: Every's Vibe Check — every · 2026-10-02