Grammaire générative et intelligence artificielle générative - Luigi Rizzi

The presentation is a masterful synthesis of two research programs, delivered with clarity and depth. Rizzi’s expertise in generative grammar is evident, and he successfully bridges the gap between theoretical linguistics and AI. The argumentation is rigorous, building from historical foundations (Descartes, Turing) to contemporary issues. He carefully distinguishes the goals of generative grammar (explanatory, biological) and LLMs (technological, practical), avoiding reductionism. The core of the talk focuses on the abstract hierarchical nature of language, using the operation ‘merge’ as a unifying principle. He provides concrete linguistic examples, such as agreement and coreference, to demonstrate that these phenomena are governed by hierarchical structure, not linear order. This is a crucial point, as it challenges the assumption that LLMs, which process text sequentially, can fully capture these properties. Rizzi acknowledges the rapid progress of AI and the principle that what is impossible today may become possible tomorrow, showing intellectual honesty. The sources cited are primarily his own work and foundational texts (Chomsky, Turing), which are appropriate for the topic. The talk is well-structured, with clear transitions and a logical flow. The only minor weakness is that the discussion of LLM limitations is based on current models, which may evolve, but he explicitly acknowledges this. Overall, this is an excellent, thought-provoking presentation that offers valuable insights for both linguists and AI researchers. The title accurately reflects the content, and the talk delivers on its promise to explore the complementary nature of the two programs.

The title accurately reflects the content, which compares and contrasts generative grammar and generative AI, highlighting their complementary nature.

9/10

The speaker is a renowned linguist and professor at Collège de France, providing a rigorous and well-structured presentation. The content is based on established linguistic theory and recent AI developments, with clear arguments and references to foundational works. The institutional setting and lack of commercial bias enhance reliability.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction: comparison of generative grammar and generative AI, their distinct goals and common roots.
• 02:00Historical roots: Descartes' Discourse on Method, Turing's imitation game, and the shared term 'generative'.
• 05:00Galileo and Darwin on combinatorial language; Chomsky's contribution to formalizing linguistic combinatorics.
• 10:00The minimalist program and the operation 'merge' as a single rule for building hierarchical structures.
• 15:00Properties of merge: infinity, hierarchy, and displacement; hierarchical structure is invisible but determines linguistic properties.
• 20:00Agreement as an example of hierarchical locality; the verb agrees with the hierarchically closest NP, not the linearly closest.
• 25:00Coreference constraints: examples from French showing that coreference is possible when the pronoun precedes the name in certain hierarchical configurations.
• 30:00Discussion of challenges for LLMs: hierarchical properties remain problematic for current models.
• 35:00Potential contributions: generative grammar can inform AI architecture, and AI can provide computational models for linguistic theory.
• 40:00Conclusion: the two programs are complementary, and their interaction is fruitful for both fields.

• Collège de France - Colloque de rentrée 2025 — Official page for the symposium where this talk was given.
• Collège de France — Institutional website providing general information and resources.
• Fondation du Collège de France — Support page for the institution.
• Collège de France on Bluesky — Social media presence.
• Collège de France on LinkedIn — Social media presence.
• Collège de France on Threads — Social media presence.

• Chomsky, N. (1957). Syntactic Structures — Foundational work in generative grammar, referenced implicitly.
• Turing, A. (1950). Computing Machinery and Intelligence — The paper introducing the Turing test, discussed in the talk.

• Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data — This paper argues that LLMs lack true understanding, which contrasts with the optimistic view of AI progress presented in the talk.