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

Google DeepMind's Tom Zahavy argues in a position paper that language models cannot spark scientific revolutions because they lack the cognitive mechanism for true novelty. World models might enable such breakthroughs.