World Models: The Next AI Frontier Beyond LLMs
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Yann LeCun left Meta in December 2025 to found AMI Labs, raising just over $1 billion to build world-model systems, arguing current LLMs can't reason about physical reality because they learn only from text.
- Fei-Fei Li started World Labs in 2024 with $230 million to develop AI with "spatial intelligence"; its product Marble generates coherent 3D scenes from text prompts.
- Google DeepMind released Dreamer 4 in September 2025 — an agent trained mostly on Minecraft data that "dreams" consequences of actions and learned to collect diamonds without being shown how.
- A fault line has emerged between approaches: most world models try to faithfully reconstruct future observations, but LeCun's JEPA learns in latent space, predicting only what matters for reasoning — mirroring how humans don't mentally simulate pixel-by-pixel.
- A 2024 study found that LLMs trained on NYC taxi directions could give reasonable routes but "failed miserably" on odd detours, because they lack the mental maps needed to imagine scenarios they haven't seen described.
- Melanie Mitchell at the Santa Fe Institute notes that "world model" has become "a kind of shorthand for all the things that current AI systems can't do well."
Why it matters: Over $1.2 billion from LeCun's AMI Labs and Fei-Fei Li's World Labs is now chasing world models on the thesis that LLMs hit a wall without physical-world experience — a 2024 study cited in the article showed language models failing on NYC taxi detours they hadn't seen described, and DeepMind's Dreamer 4 is the first agent proof-of-concept that plans by simulating imagined consequences.
