Paul Graham on LLMs 'Thinking'

Paul Graham compares calling large language models "thinking" to 16th-century astronomers who used the heliocentric model as a calculation tool before endorsing it as reality. He claims people will first adopt "think" as a convenient shorthand because it simplifies explanation and practice, and that over time that shorthand will be treated as literal - illustrating how instrumental language can morph into ontological acceptance.

Responses pick up the philosophical implications: one commenter invokes Ernst Mach's idea of atoms as useful fictions, others argue that whether LLMs "think" depends on definitions - if thought is confined to biological brains then LLMs cannot think, but if defined behaviorally the line blurs. Another note observes that anthropomorphic terms like "memory" and "recognize" began as metaphors and became standard. The exchange centers on whether descriptive convenience should reframe reality and what criteria - causal mechanisms, observable behavior, or explanatory power - should determine when a system counts as thinking.