Noam Chomsky: Language, Cognition, and Deep Learning | Lex Fridman Podcast #53

Chomsky argues that what we call language is chiefly a mental faculty localized in the brain, which generates structured thoughts; speech and writing are secondary ‘externalizations’ of this inner system.

Referencing Marvin Minsky’s work with simple Turing machines, Chomsky suggests that any sufficiently intelligent species would likely have arithmetic, and that human language’s core operations approximate arithmetic, offering a possible shared basis for communication.

Just as our genes allow complex mammalian vision but preclude insect eyes or wings, our cognitive endowment enables powerful reasoning but likely blocks certain forms of understanding, challenging the assumption that humans can in principle comprehend everything.

Interpretation often relies on hierarchical structure rather than linear proximity (e.g., where “carefully” attaches in a sentence), showing that speakers unconsciously use complex structural computations on representations they never directly hear.

Chomsky sees systems like Google Translate as valuable tools that find patterns in huge corpora, but notes they are optimized for coverage of random ‘experiments’ (sentences), not for testing theoretical questions, and they can perform just as well on languages that violate known principles of human grammar.