LLMs are not like you and me—and never will be.

The post argues that claims about LLMs converging on human inner lives are wrong. It says they share some surface similarities with humans, but they do not build reliable world models and instead work as autocomplete systems that often miss obvious constraints in time, history, and common sense.

Reading notes

#

• LLMs may share some computational features with humans, but that does not make them similar in how they reason.
• The post says LLMs never induce proper world models, which helps explain chess failures and recurring obvious mistakes.
• Examples include errors about inflation and pricing, nonexistent legal cases, biographical mistakes, and broken lists of Canadian and US leaders.
• The post treats these failures as evidence that LLMs do not reliably process time, history, biology, or other everyday relations.
• It says even “thinking mode” does not fix the problem because the systems still lack robust models.
• The argument is that LLMs are better at mimicking language than at doing what a human would do in the same situation.
• The post warns that using LLMs as agents is risky because their unreliability persists in multi-step tasks.
• It closes by saying they should be treated as function approximators rather than trusted as minds.