Language models can't spark scientific revolutions, but world models might

Zahavy argues that today’s AI can handle deduction, induction, and increasingly mathematics, but still cannot make the abductive leap needed for new scientific theories.

He says breakthroughs like Einstein’s relativity and Archimedes’ buoyancy insight came from embodied, physical intuition that machines lack.

Language models are useful for recombining known ideas, but they do not yet have the grounded world models needed for genuinely novel principles.

Tools such as AI Scientist and AlphaEvolve may speed up research workflows, but they remain limited compared with human paradigm-shifting discovery.