Will a new kind of AI that understands physical reality change the world again?
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 based on his JEPA approach
- Fei-Fei Li launched World Labs roughly a year earlier with $230 million to develop AI with 'spatial intelligence'
- Google DeepMind released Dreamer 4 in September 2025, an agent trained on Minecraft that learns by 'dreaming' the consequences of possible actions; team lead Danijar Hafner says it can figure out diamond collection without being shown how to play
- DeepMind's Genie 3 generates interactive virtual environments from text prompts like 'snowy mountain at dusk' where users can move around and trigger events like rain
- World Labs' Marble generates coherent 3D scenes from text prompts, though the article notes it is essentially a 3D video generator rather than a true agent-internal world model
- A 2024 study found LLMs could provide reasonable NYC taxi routes but 'failed miserably' when asked to take odd detours, illustrating their lack of mental maps
- Melanie Mitchell at the Santa Fe Institute warns 'world model' has become 'a kind of shorthand for all the things that current AI systems can't do well', blurring substance and hype
Why it matters: Three of AI's most prominent figures — LeCun, Fei-Fei Li, and DeepMind — have redirected over $1.2 billion into world models, publicly breaking from the LLM-only path to AGI. DeepMind's Hafner frames the robotics breakthrough as recipe-complete, with only compute, data, and scaling standing between Dreamer 4's Minecraft success and real-world deployment.
Ask SkimNews
