World Models Could Change Everything

If World Models work properly, AI could move beyond text patterns to a stage where it learns real-world interaction. Applications like household robots, automated kitchens, and manufacturing lines become possible, but the timeline is not yet fixed.

Investment is already pouring in. AMI Labs is Yann LeCun's $1.03 billion seed round bet, and World Labs, Skild AI, and Physical Intelligence also pulled in hundreds of millions of dollars in 2024. By early 2026, the stakes grew even larger: World Labs at $1 billion, AMI Labs at $1.03 billion, and Skild AI at $1.4 billion.

The core bottleneck is data friction. LLMs only read the traces of language left on the internet — they don't learn by seeing, touching, and failing with objects. Reading about physical laws and mastering them in reality are entirely different things.

This is also why LeCun goes so far as to criticize current AI as being worse than a cat. Without sufficient memory, reasoning, and planning, a system remains merely good at predicting language rather than a model that understands the real world.

ARC-AGI-3 tests exactly this kind of ability: exploring an unfamiliar environment, building an internal model, adapting on the fly, and planning multi-step actions. Humans can generally solve it on the first try, but frontier models still lag far behind.

• o3 scored 87.5% on the original ARC-AGI.
• On ARC-AGI-2, it collapsed to under 3% at launch, but Gemini 3 Deep Think later recovered to 84.6%.
• On ARC-AGI-3, the frontier score at launch was 0.26%, and on the public leaderboard, Gemini 3.1 Pro scored 0.37% while Grok scored 0%.

In the end, the piece concludes that while world models could be the breakthrough for real-world AI, success will hinge less on capital and more on how cheaply and quickly real interaction data can be gathered. For now, this means application-specific progress is more likely to appear before any general-purpose revolution.