A DeepMind Researcher Says LLMs Structurally Cannot Make Einstein's Leap

A DeepMind Researcher Says LLMs Structurally Cannot Make Einstein's Leap

Google DeepMind researcher Tom Zahavy argues in a new position paper, "LLMs Can't Jump," that large language models can prove existing theorems but structurally cannot originate new scientific axioms the way Einstein did with his 1907 elevator thought experiment. The paper lands days after OpenAI declared an "AGI era" with GPT-6 Astra, offering a sharp internal counterpoint from inside one of the field's top labs.