Gary Marcus pushes back on Chollet: symbolic tool use still critical to LRM progress
GaryMarcus · x · 2026-10-02
François Chollet argued the key difference between pre-2024 base LLMs and modern LRMs isn't symbolic tool use, but the shift from a transductive paradigm (intuit the answer) to an inductive one (intuit the program producing the answer, i.e., test-time induction of NL programs).
Gary Marcus disagrees on several points: the switch would be useless without a symbolic code interpreter to execute the generated code; the role of formal verification tools like Lean is critical; and presumably symbolic search is also a major contributor. He suggests rephrasing it as "a critical distinction" rather than "the critical distinction," rejecting the claim that symbolic tool use doesn't matter.
Related event: Chollet and Gary Marcus Debate AI Paradigm Shifts and Symbolic Tools(4 posts)→
More from AGI Musings
- Duvenaud's "Permanent Periphery": nations without frontier AI get risks but no benefits — teortaxesTex · 2026-10-03
- Calling It "Superintelligence" Risks Cognitive Surrender, Psychology Today Essay Argues — AnnaCiaunica · 2026-10-03
- "We all become meat proxies": why adding value beats avoiding AI dependence at work — brandon_galang · 2026-10-03
- D-Day's chief forecaster refused to "trust your gut": a masterclass in data-driven go/no-go decisions — shashib · 2026-10-03
- Joel Grus: I read every popular consciousness book (Dennett, Chalmers) until I realized no one had anything useful to say — ZeroStateReflex · 2026-10-03
- Merzmensch: AI itself can't make art, but people can create art with it — Merzmensch · 2026-10-03