What does Claude know? Philosophy & LLMs
Tom Pashby, Visiting Assistant Professor of Philosophy
Nearly 50 years ago, Searle’s Chinese Room argument convinced philosophers to answer “no” to the question: Can computers think? In this talk I argue that the remarkable ability of recent LLMs to seemingly understand English (i.e., natural language) should lead us to doubt Searle’s conclusion. Rather than the rule-following, syntax-only machines Searle considered, LLMs operate as dynamical systems described by physics and are trained on vast natural language corpora. By learning patterns that reflect how language is actually used in the world, I argue that LLMs can be said to understand English.
As we’ve recently seen, systems like Claude Code can interpret and create computer programs with considerable skill, and I claim that these meta-linguistic abilities are the best evidence that computers can break the syntax-semantics barrier. Therefore, Searle’s argument fails, which opens the door to a Dennett-inspired view that the linguistic abilities of LLMs constitute real patterns with ontological weight. As such, it is correct to say that LLMs think but I reject the suggestion that such systems are conscious or have personal identity—-Descartes should have said “I think, therefore there is thinking.”
Twilight 201