AI can ace the International Mathematical Olympiad and spot patterns in oceans of data. But according to a new position paper from Google DeepMind’s Tom Zahavy, it lacks the fundamental spark that made Einstein Einstein. The paper, titled “LLMs can’t jump,” argues that current language models are brilliant deductive and inductive reasoners but are structurally incapable of the creative, intuitive leap—what the philosopher Charles Sanders Peirce called ‘abduction’—that births truly revolutionary theories.
Zahavy builds his case on a framework Einstein himself described in a letter to a friend: a cycle where sensory experience fuels an intuitive jump toward fundamental axioms, from which logical deduction then flows. Language models, Zahavy argues, can handle the deduction and induction parts beautifully. He even concedes that a modern AI could probably derive general relativity if you handed it Einstein’s assumptions first. The roadblock is formulating those assumptions in the first place, a process he calls “manipulative abduction.” This isn’t matching symptoms to a known disease; it’s inventing a disease no one has named.
To illustrate the blind spot, Zahavy points to the physics of Einstein’s era. Newtonian mechanics was a triumph, its only blemish a tiny wobble in Mercury’s orbit. An optimization-driven AI, trained to minimize error against existing data, would have logically “fixed” the problem by inventing another planet—which is exactly what astronomers of the time did, hypothesizing a world called Vulcan. Without a data crisis, there was no error signal to push a system toward rethinking space and time itself. The leap came from a different source: Einstein’s embodied imagination, his “happiest thought” of a falling observer feeling no gravity.
This is the chasm. Zahavy compares language models to John Searle’s Chinese Room, fluently shuffling symbols of physics with zero sensory grounding. Systems like Sakana’s AI Scientist or DeepMind’s AlphaEvolve are impressive optimizers, but they only recombine existing concepts or need a clear error signal to climb step by step. Zahavy’s proposed path forward isn’t a bigger text model; it’s building physically consistent world models that can learn the intuitive physics a baby picks up by knocking blocks over. The question he leaves us with isn’t whether AI can solve harder problems, but whether it can figure out which problems are worth solving without first being told.
