DeepMind's Zahavy: LLMs can't jump, but world models might

google deepmind's tom zahavy published 'llms can't jump,' arguing that language models lack the cognitive mechanism for 'manipulative abduction' required for true scientific discovery. zahavy contends llms excel at deduction and induction but cannot perform the creative leaps needed to invent new axioms without pre-existing linguistic templates. he illustrates this with einstein, who derived relativity through embodied simulation rather than data optimization, implying that future progress requires integrating world models or embodied ai rather than scaling statistical pattern recognition.