LeCun's New Lab Just Published the Math Behind His Bet Against LLMs

On October 29, Randall Balestriero and Yann LeCun posted a paper to arXiv titled "Proving and Improving the Platonic Representation Hypothesis with JEPAs." The affiliation line lists AMI Labs alongside Brown and NYU — the startup LeCun founded after leaving his twelve-year post as Meta's Chief AI Scientist in late 2025, and which raised a $1.03 billion seed round in March at a $3.5 billion valuation, built on the wager that joint-embedding predictive architectures, not autoregressive language models, are the right foundation for machine intelligence. It is AMI's second attempt this year to turn LeCun's long-running argument against LLMs into an actual theorem rather than a conference-keynote assertion.

A two-year-old hunch, finally proved (for one architecture)

The Platonic Representation Hypothesis has circulated informally since 2024: models trained on different data, even different modalities, tend to converge toward a shared underlying representation of reality as they scale. It became a popular talking point in interpretability circles without ever being proven for any specific training method. The new paper tries to close that gap, but only for JEPAs — and in doing so, it draws a sharp line between two kinds of self-supervised training that the authors say explains why that convergence happens at all.

The distinction is between "good views" and "bad views." A good view is two separate sensory captures of the literal same underlying scene — two camera angles of a room, or an image and a caption describing it. The paper's central theorem says JEPAs trained on good views converge to one identical representation, up to a single rotation, no matter what modality they were trained on. Next-token prediction — the objective behind every major LLM, from GPT-6 to Gemini — only ever sees bad views: one span of a single document predicting the span that follows it. In the authors' framing, that's not a smaller version of the same signal; it's a structurally worse one, because a prefix and its continuation aren't two independent glimpses of a shared reality, they're just two pieces of the same artifact.

The numbers AMI is using to make the case

The paper backs the theorem with sample-efficiency claims that are clearly aimed at an industry obsessed with compute budgets: a JEPA trained on good views reportedly reaches the same cross-modal alignment as next-token prediction on the same corpus using about 42 times fewer samples. A 109-million-parameter, text-only JEPA scores higher mutual-neighbor alignment with the image encoders DINOv2 and CLIP than language models up to 72 billion parameters, and beats every zero-shot model up to 8 billion parameters on an image-classification proxy task. As a demonstration, the authors take two encoders trained separately, on different modalities, with no paired data, and fit a single rotation between them that turns the pair into a working image captioner — scoring a respectable 0.838 CIDEr on COCO.

Those are striking numbers for a 109-million-parameter model to put up against 72-billion-parameter competition, but they're alignment-with-an-image-encoder numbers, not translation, coding, or conversation benchmarks — the tasks that actually generate revenue for the trillion-parameter models the paper is implicitly arguing against. The theorem tells you which architecture gets to a particular kind of cross-modal agreement more cheaply. It doesn't tell you that cross-modal agreement is the bottleneck standing between current AI and whatever LeCun means by "advanced machine intelligence."

The proof arrives the same month its premise got contested

The theorem itself is also narrower than the headline suggests. It holds at population optima, under idealized assumptions: that each modality injectively renders the same finite set of underlying world states, and that positive training pairs are genuinely two views of one state rather than noisy approximations of it. The authors concede real sensors are lossy and real training runs are finite-sample, which means actual results depend on inductive biases the theorem doesn't cover — an admission that echoes AMI Labs' own earlier paper this year on the identifiability of world models, which found that current JEPA architectures were brittle under minor visual perturbations once you left the clean benchmark setting.

The harder problem for the timing is that the empirical case for the Platonic Representation Hypothesis — the thing this paper just claims to have proven for JEPAs — is itself under fresh attack. A separate paper accepted to ICML 2026, "Revisiting the Platonic Representation Hypothesis: An Aristotelian View," argues that the reported evidence for global cross-modal convergence in trained models is largely an artifact of dimension and depth confounders in how researchers measure representational similarity; once those are calibrated out, only a weaker claim about local neighborhood structure survives. AMI's new theorem proves the strong, global version of the hypothesis is true for an idealized JEPA. Other researchers, working from the empirical side, are arguing the strong version doesn't actually show up in real trained models. Both things can be true at once — proofs about idealized objectives and measurements of real networks are different exercises — but it means AMI is publishing a mathematical defense of a claim that the field's own benchmarking is currently chipping away at from the other direction, eight months into a billion-dollar bet that rests on the claim being right.