2026: AI newcomers collectively bet on world models

Summary

The article examines the investment wave directed at 'world models' in 2026. These are AI systems that predict physical states rather than simply forecasting the next token. The text references massive funding rounds, including Yann LeCun's AMI Labs raising $1.03 billion in a seed round, World Labs securing $1 billion, and numerous Chinese startups collecting billions of RMB. It also maps the technical camps competing in video generation, spatial reconstruction, embodied simulation, and autonomous driving.

Why it matters

The piece documents a genuine inflection point, indicating that the industry is seeking a post-LLM paradigm. Scaling law gains are flattening; GPT-4.5 performs only 2-7% better than GPT-4o on SWE-bench. Additionally, the public text corpus is estimated at approximately 300 trillion tokens. The article provides a useful competitive map showing how players in video, 3D, robotics, and automotive are approaching physical world modeling from different entry points, offering a practical reference for where capital and expertise are shifting.

Highlights

• Comparisons between BLINK and PhysBench reveal a persistent gap in physical reasoning. GPT-4V scores approximately 51% on spatial tasks. Physical understanding does not reliably scale with increased model size or data volume.
• There is no unified definition for a 'world model.' The Zhiyuan Institute proposes four categories, while World Labs suggests three functional types. Technical approaches (pixel-based, 3D structure, latent space, lightweight) remain fragmented.
• China's competitive advantage is framed not as computational superiority, but as scenario density. Manufacturing floors, complex urban traffic, and robotic supply chains are highlighted as elements that generate the physical feedback loops world models require.

Critical take

The article reads more like an industry support document than an independent analysis. It repeatedly labels world models as a 'battlefield.' It quotes LeCun's claim that 'LLMs are completely wrong' without presenting counterarguments. It presents this still fragmented and undefined field as an established consensus. The emphasis on Chinese startups' funding rounds and the closing argument that 'the next AI company may be born in a workshop, not a lab' replace critical scrutiny with narrative promotion. This approach sidelines the question of whether any of these approaches have yet produced verifiable and generalizable physical reasoning capabilities.