Gary Marcus, professor emeritus at NYU, explains the differences between large language models and "world models" — and why he thinks the latter are key to achieving artificial general intelligence.