Code execution and the debate over language models’ reasoning

A post shares remarks attributed to Richard Socher from the Latent Space podcast arguing that reinforcement learning, tool use, and the ability to write and run code are making large language models more sophisticated. His argument challenges the criticism that these models lack neuro-symbolic reasoning: coding, he says, is itself structured, symbolic reasoning. As models improve at generating and executing code, he argues, the boundary between language models and more general reasoning systems becomes increasingly blurred.