In April, I published The Evidence Sharpens, a Q1 update on what AMI Labs’ $1.03 billion seed round is buying. The scoreboard then: the Architect’s Road had built a research ecosystem, the generative camp had built further into robot control, and the defining experiment comparing the two still did not exist. The tension was speed versus complexity.
The second quarter changed the axis. In late May, two preprints from LeCun’s research circle appeared within days of each other. One is a theorem. One is a stress test. Read together, they define exactly what the bet must prove and measure exactly how far current systems are from proving it. Then, in June, the generative camp stopped publishing evidence and started shipping infrastructure. The new tension is provable structure versus shipped scale.
The bet acquired mathematics
When Does LeJEPA Learn a World Model? by David Klindt, Yann LeCun, and Randall Balestriero is the most important single paper the Architect’s Road has produced. It proves that LeJEPA, the alignment-plus-Gaussian-regularization recipe behind LeWorldModel, linearly recovers the world’s true latent variables from nonlinear observations. The property is called linear identifiability. Feed the model raw pixels generated by hidden physical causes, and it untangles those causes up to a rotation.
The conditions matter as much as the guarantee. The proof holds when the latent variables are Gaussian and evolve under stationary, additive-noise dynamics. The paper’s sharpest result is a uniqueness claim: among all such worlds, the Gaussian is the only latent distribution for which the guarantee holds. The authors also prove that the guarantee degrades gracefully near the boundary, and that this form of identifiability enables optimal planning in latent space.
This is the first genuine evidence on the question the original article called “just a foundation model.” JEPA’s training objective now has a provable property that generative objectives have not demonstrated. A recipe variation does not come with a theorem. A moat might.
The boundary immediately became the battle line. Real-world dynamics are frequently non-Gaussian: fat tails, phase transitions, nonlinear feedback. Within weeks, a team at ARYA Labs published Identifiability Without Gaussianity, arguing that the Gaussian limit is an artifact of statistical alignment itself and that symbolic world models escape it. The theorem strengthened the bet and constrained it in the same motion. It proved the destination exists. It also drew a fence around the terrain where the proof applies.
The bet audited itself
Five days before the theorem, the same research circle posted stable-worldmodel, a reproducible evaluation platform for world models, and used it to stress-test the field’s current systems, their own included. The results are unsparing. Planning success rates drop sharply under mild visual perturbations. Changing background colors cuts performance. Adding small visual distractors produces a quadratic collapse in success rate, and the pattern holds across every baseline tested. The conclusion in the paper’s own language: current world models exhibit limited zero-shot generalization.
The companion theory explains why the two papers must be read together. The identifiability guarantee assumes the training data explores the state space broadly. Goal-directed robot data does not. The theorem defines the target. The benchmark measures the distance. And the distance is large.
The scientific posture here deserves notice. LeCun’s circle did not publish a demo reel. It published a formal statement of the conditions under which its architecture wins, then published the instrument showing that no current system, LeWorldModel included, meets those conditions. This is what winning on paper looks like: the proof is real, and so is the gap between the proof and the machines.
The generative camp became infrastructure
While the Architect’s Road was proving theorems, NVIDIA industrialized the opposing thesis. On June 1 at GTC Taipei, the company launched Cosmos 3, a fully open omnimodel that unifies vision reasoning, world generation, and action prediction in a single mixture-of-transformers system spanning text, image, video, ambient sound, and action. Two open sizes on Hugging Face, a permissive license, and a new Cosmos Coalition with Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI.
The competitive meaning is larger than the model. In Q1, NVIDIA was publishing the strongest papers against the JEPA bet. In Q2, it open-sourced the entire generative stack and organized an alliance around it. The company that invested in AMI Labs’ seed round now distributes the free alternative to AMI Labs’ thesis. The hedge has become a platform strategy.
The other audit
The generative camp received its own examination this quarter, and from third parties rather than from itself. A cluster of June benchmarks stress-tested video world models on state consistency: whether the model remembers what it can no longer see. Mbench measures memory capability directly. Out of Sight, Out of Mind? evaluates state evolution when objects leave the frame. A third paper states its finding in its title: Current World Models Lack a Persistent State Core.
The pattern across this literature is consistent. Video world models generate convincing futures and forget occluded objects. This is LeCun’s original critique of pixel prediction, that modeling appearance is not modeling state, returning as independent measurement. Each camp’s deepest weakness was quantified this quarter, and each camp’s weakness is the other camp’s founding argument.
What the five questions look like now
The controlled experiment. Still does not exist. But stable-worldmodel is the first reproducible platform capable of hosting it. The missing piece is no longer infrastructure. It is a lab willing to run JEPA and generative representations through the same gauntlet at matched scale.
The blueprint gap. Redefined. The gap is no longer only architectural. The theorem specifies what a completed JEPA world model must achieve, and the benchmark shows current implementations do not achieve it. The target is clearer than it has ever been. So is the distance.
The “good enough” question. Heavier. Cosmos 3 makes the generative baseline free, open, and unified. Any qualitative advantage JEPA claims must now beat a platform that costs nothing and ships today.
Competitor velocity. Transformed in kind. The metric is no longer papers per quarter. It is licenses, coalition members, and downloads. Research velocity became distribution velocity.
The “just a foundation model” question. First movement in four quarters. The identifiability theorem is evidence that JEPA’s objective has provable structure that generative objectives lack. The ARYA rebuttal is evidence that the structure’s domain is contested. The question is no longer unanswered. It is disputed, which is progress.
The updated scoreboard
The Architect’s Road holds the mathematics: a uniqueness theorem, an identifiability guarantee, and a proof that its representations support optimal planning. It also holds the measured brittleness: its systems, like everyone’s, collapse outside the training distribution.
The generative camp holds the distribution: an open unified model, an industrial coalition, and state-of-the-art robot control. It also holds the measured amnesia: its systems generate the world and forget its state.
The honest summary: JEPA is provably right and measurably fragile. The generative camp is empirically dominant and structurally forgetful. Each side spent the quarter accumulating exactly the strength the other side’s weakness demands.
The clock started in March 2026. This quarter, both hands got calibrated.
This article is part of Robonaissance’s coverage of the world model research frontier. For the original analysis, see The LeCun Bet. For the Q1 update, see The Evidence Sharpens. For the full landscape across all five research traditions, see the six-part series Roads to a Universal World Model.
It’s usually much easier to prove a theorem with necessary conditions (a => b) but good on them to be able to provide necessary and sufficient conditions (a <=> b). It just shows how difficult it is to get theoretical results to fully explain today’s very complex large deep neural networks.