On the last day of October, in the auditorium of Westlake University's Yungu campus, roughly ninety industry R&D chiefs and investors gathered to hear about world models — and the first order of business was explaining what the term means.
The occasion was the founding of the Westlake World Model Center, led by Yu Kaicheng, a 33-year-old professor who has been working on this since 2023, when the phrase was still cold. His own lab, AutoLab, is where the definitional problem becomes concrete.
In 2024, AutoLab won the ECCV autonomous driving track by a wide margin by generating synthetic extreme scenarios — rain, sudden accidents, the long-tail events you cannot ethically stage on a real road. Then, after a deep deployment with a major Chinese automaker, the team reversed course.
If the model could already produce photorealistic driving data, it had, in some sense, understood the world. So what was pixel rendering for?
The reasoning was almost philosophical: if the model could already produce photorealistic driving data, it had, in some sense, understood the world. So what was pixel rendering for? At the end of 2024, AutoLab pivoted to what it calls concept world models — no pixels, just reusable physical concepts like spatial relations, materials, and contact states, decomposed and recombined in an abstract latent space.
That pivot, more than any benchmark win, is the tell.
Because the money and the science are now pulling hard in opposite directions. Venture investors committed over $3 billion to world-model startups in the first half of 2026 alone. Yann LeCun's AMI raised a record $1.03 billion seed; Fei-Fei Li's World Labs took $1 billion. Anthropic is reportedly in talks to acquire Decart for about $6 billion — framed as buying into model infrastructure, world models, and compute optimization. In the same stretch, AMD agreed to acquire World Labs for roughly $8.2 billion in stock.
That is a lot of capital converging on a term nobody can define consistently.
The Forbes coverage of the funding wave put it plainly: the field lacks a unified definition, with systems categorized variously as renderers, simulators, or planners. These are not three flavors of the same thing. They are three incompatible research programs.
The first camp builds pixel renderers — video world models like Tencent's WorldCrafter, which pursues consistency through implicit 3D-aware memory. The second builds latent concept reasoners, the Westlake bet. The third builds formal provers. And the gap between them is not semantic. It is measurable, and Tsinghua just measured it.
Yao Shunyu's team — with Andrew Yao, the 2000 Turing laureate, and Yuan Yang — built the Algebraic World Model, or AWM, and tested it on ARC-AGI-3, a game set that tells you nothing: no rules, no goals, just environments to probe.
The comparison is devastating. Across twelve tests on unsolvable levels, GPT-6 Astra and Claude Opus 5.5 together correctly judged "impossible" only five times — and in four of those five, they gave the wrong minimum step count anyway. In one case, Astra tried eight rounds, grew frustrated, and declared that 18 steps could never win, estimating it needed at least 22. The true answer was one step more.
AWM, running on Opus 5.5, proved the same level unsolvable in two attempts. Across all 183 ARC-AGI-3 levels, it cleared everything at 99.9%, using an average of 43% of the human action baseline — the most efficient system known.
The distinction the Tsinghua team draws is precise. Astra's symbolic rule dumps are Tycho Brahe's notebook: voluminous, accurate, and merely descriptive of phenomena. AWM tries to be Kepler — extracting the laws that generate the phenomena, then formalizing them in Lean 4 so a machine can check the proof.
Note what that means. AWM's output is falsifiable by construction. A video renderer's output is judged by whether it looks right. A latent planner's output is judged by whether the task succeeds. Only one of these three admits the word "impossible" as a checkable claim.
This is why the debate over what a world model "is" is not academic throat-clearing. It determines what counts as evidence.
The structural problem runs deeper still. Fusion Fund's Charlotte Xia and Matt Wong argue that the LLM parallel — the "Great Parallel" Jim Fan presented at Sequoia's AI Ascent in April — is more wish than roadmap, and the reason is not missing data. Text has a universal training unit: the token. The physical world has none. Joint angles, point clouds, force-torque readings, and pixels are not the same kind of thing and cannot be mixed without conflict.
Without a universal unit, there is no consolidation. Without consolidation, billions in competing architectures do not converge the way the transformer bet did. The $6 billion that flowed into embodied world models in Q1 2026 was not buying a single paradigm. It was buying lottery tickets in three different lotteries.
The counterargument is that falsifiable wins are starting to exist. NVIDIA's DreamZero, a 14-billion-parameter world-model system, hit 62% average task progress on zero-shot generalization where pretrained VLA baselines reached 27% and from-scratch VLAs near zero. That is the kind of specific, published demonstration the thesis needs.
But one result on one benchmark family is not a paradigm. And the research warnings are piling up on the other side: a vision-only Genie Envisioner baseline scores 27.3% on contact-rich manipulation against 70.6% for a tactile-augmented system, and separate work finds video world models losing physical and spatial consistency over long horizons. Demis Hassabis and others are calling for third-party evaluators, which is what people say when the benchmarks are immature.
So read the capital carefully. AMD's stated rationale for the World Labs deal is understanding how AI workloads are evolving as compute demands diversify. Anthropic's reported interest in Decart is bundled with compute optimization. These are chip and infrastructure buyers hedging on a workload. That is not the same as validating a science.
Which brings the story back to Westlake, and to the sentence that should be underlined. Yu's lab abandoned the approach that won it a world championship, at the moment it was winning. The four robotics benchmark suites Awomo claims to top — environmental disturbance, fine manipulation, long-horizon tasks — are vendor-reported, with no third-party reproduction published. Treat them as a starting line, not a verdict.
The economic wager underneath is the same one driving the whole camp: if a general physical foundation model exists, downstream developers need not collect millions of hours of task data — 20 to 30 hours of post-training could specialize a model for a new domain. That is the open thread from our 10/26 piece on whether specialist layers build durable moats when recipes are published openly, now restated in physical form. It remains unproven at scale.
Yu, for his part, has been studying ancient Chinese calligraphy. Oracle bone script, he says, is the earliest latent space — the compression of real-world entities into two-dimensional symbols, the same mathematical act as converting intent into brushstrokes, or intent into robot trajectories. Thousands of years later, we are still abstracting the laws of the world, just with different tools.
That is a lovely thought, and it is also a warning. Abstraction without a falsifiable prediction is poetry.
The post-LLM paradigm will not be settled by who raises the most or scales the hardest. It will be settled by which camp can state, in advance and in checkable form, what the world will do next — and be proven wrong when it doesn't. Westlake's self-reversal is the tell because it is the rare case of a lab changing its definition before the market forced it to. The labs that win will be the ones willing to do that on purpose.