AlphaZero Sparks Debate Over LLM Creativity and Extrapolation

On October 9, prominent AI researcher François Fleuret posed a sharp question on X to those convinced that "LLMs can't truly invent anything, only interpolate rather than extrapolate, and lack creativity": what about AlphaZero? AlphaZero surpassed human chess knowledge through self-play and is often cited as an example of machines going beyond existing data.

Confirmed

• Guillaume Lample (syhw, a researcher at Google DeepMind) responded that he doesn't hold the "LLMs only interpolate" view, but noted that Fleuret's argument is too easy to refute—because AlphaZero simply isn't an LLM.
• Lample argued the core of the disagreement lies in the difference between "closed-world vs. open-world": a sufficiently large lookup table (as in AlphaZero's case) can beat the strongest humans at Go, and "we know how to fill it"; but for open problems like room-temperature superconductivity or cancer cures, "we don't yet know how to fill the table."
• Fleuret then continued sparring with Lample over how the "a lookup table can beat humans at Go" analogy relates to the capability boundaries of large models.

Why it matters

• The debate touches on one of the most divisive questions in today's AI community: whether LLM achievements are mere interpolation over massive data, or reflect some capacity for extrapolation and creativity.
• Lample's "closed vs. open world" and "lookup table" analogies offer a more precise framework for discussing model capability limits: the key isn't whether a model can surpass human data, but whether the target domain has a structured space that can be effectively searched/filled.
• Neither party is a lightweight—Fleuret is a well-known AI researcher, and Lample has worked at DeepMind—so the disagreement itself reflects a genuine split in academia over the nature of LLMs.

2026-10-09 ~ 2026-10-10 · 5 related posts

• [source] François Fleuret challenges LLM skeptics: what does AlphaZero say about the interpolation argument? — francoisfleuret · 2026-10-09
• giffmana on the LLM creativity debate: AlphaZero isn't an LLM — giffmana · 2026-10-09
• [source] DeepMind researcher: AlphaZero fills a Go lookup table, but not one for cancer cures — syhw · 2026-10-10
• [source] Fleuret Clashes Over 'Lookup Table Can Beat Humans at Go' Analogy — francoisfleuret · 2026-10-10

1 near-duplicate retellings: syhw