A DeepMind position paper argues that language models are not enough to produce the kind of conceptual leap behind major scientific revolutions. The Decoder reports that Tom Zahavy frames the gap through Einstein’s account of discovery: experience, an intuitive jump to axioms, and then logical deduction.
The distinction matters because today’s models are already strong at induction, or finding patterns, and increasingly strong at deduction, or deriving results from rules. Zahavy’s claimed bottleneck is “manipulative abduction,” the act of inventing a new cause or framework when no known linguistic template already exists.
The article uses Einstein’s general relativity as the example. An optimization-driven system might have explained Mercury’s orbital anomaly by adding another planet, as many people did at the time, rather than rethinking space and gravity before decisive data arrived.
The paper does not dismiss AI for science. It suggests that richer world models and embodied simulation may be needed if machines are to move beyond recombining known explanations toward generating new foundations.