The bet against the language model

In November 2025, after twelve years at the company, Yann LeCun announced that he was walking out of Meta; he left at the end of the year. The phrasing he used was unusually plain: he called it a battle he had lost. The battle in question was internal, and it was about architecture. LeCun founded Facebook AI Research in 2013 and directed it until 2018, when he became Meta’s chief AI scientist. He shaped it as a place where the dominant approach of the moment — first deep learning, later transformers — was treated as one option among several. By 2025 the lab he founded was, like every other frontier lab, scaling large language models. He left to start one that would not.

The new outfit is AMI Labs, short for Advanced Machine Intelligence, and LeCun took the role of executive chairman in January 2026. The name is deliberate. It is the framing he has used in technical writing since at least 2022, when he posted a position paper called A Path Towards Autonomous Machine Intelligence and proposed a concrete alternative to the architecture that everyone else was about to spend three years scaling. The paper is the reason the departure mattered. With it, the move was the test case for a four-year argument: that the frontier labs are climbing the wrong mountain, and that someone with no commercial book to defend should go climb a different one to prove it. In March 2026, AMI raised a $1.03 billion seed round at a $3.5 billion pre-money valuation.

LeCun is one of three co-laureates of the 2018 Turing Award. His specific contribution was the convolutional neural network, the architecture he developed at Bell Labs in the late 1980s that made image recognition work and that, by the mid-2010s, was running inside every camera-equipped phone in the world. What is in dispute is whether the architecture currently running inside every chatbot — the autoregressive transformer trained on internet-scale text — is on a path to anything that could reasonably be called intelligence. LeCun’s answer, given to The Next Web in April 2024, is that it is “an off-ramp, a distraction, a dead end on the path to human-level intelligence.”

The bet against language models is, technically, a bet for something. The something is JEPA — Joint Embedding Predictive Architecture. It is more LeCun’s bet than his company’s: in July 2026 AMI’s chief executive, Alexandre LeBrun, called world models and LLMs “complementary, not replaceable” in physical-world systems, and AMI had not yet shipped a product.

JEPA inverts the geometry of prediction. Instead of predicting the next token in a sequence of symbols, a JEPA network predicts the embedding (a learned, abstract numerical representation) of an unseen part of its sensory input. Show it a video with a patch of a frame missing, and the system learns to predict what that patch’s representation would look like in the abstract space the network has built up, not the pixel values themselves. The argument LeCun makes for the inversion is that prediction in a learned latent space, applied to video and physical-world data, forces the system to acquire the kind of grounded model of objects, of their persistence, of cause and effect, that text prediction does not. LeCun’s post-Meta work is not confined to JEPA: in January 2026 he became founding chair of the technical research board at Logical Intelligence, a startup building energy-based reasoning models.

He put the argument plainly at Davos in January 2026, in remarks reported by Fortune:

I cannot imagine building agentic systems without an ability to predict the consequences of their actions.

If the only mechanism a system has for figuring out what comes next is to ask what word a hypothetical author would write next, the system does not, in any architectural sense, have a model of the world. It has a model of a corpus.

Meta’s JEPA releases include image and video encoders, a vision-language model, and an action-conditioned model for robot planning. I-JEPA, published in 2023, is a vision model. V-JEPA 2, released with code and checkpoints in June 2025, is a video model, and its action-conditioned variant, V-JEPA 2-AC, plans robot actions. VL-JEPA, the vision-language variant, reports zero-shot classification and retrieval results against CLIP, OpenAI’s 2021 model for matching text descriptions to images, and, in a separate controlled comparison, roughly half the trainable parameters of a token-generative vision-language model. That is a credible architectural argument that latent prediction is not just a metaphor. It is not a demonstration that JEPA can do what LeCun says LLMs cannot.

The shipped surface area, four years on from the position paper, is competitive encoders plus one narrow robot-planning result. LeVJEPA, an open video encoder released on August 27, 2026 by AMI-affiliated researchers including LeCun, reported needing 5.6 to 20.8 times less matched-epoch pretraining compute than V-JEPA 2. That, too, is an encoder result.

The gap between encoder results and a product was visible while LeCun was still at Meta. At CES in January 2025, according to Rajeev Chand of the venture firm Wing VC, who led the conversation, LeCun predicted that if everything went right, JEPA would make its way into Llama 5. Ten months later he announced that he was leaving Meta.

The strongest empirical pressure on LeCun’s position does not come from the labs he names. It comes from a benchmark designed by someone who agrees with him about what is missing. François Chollet, then at Google and now running his own outfit Ndea, published On the Measure of Intelligence in 2019 and built ARC-AGI, the Abstraction and Reasoning Corpus, to operationalize what he and LeCun both said the field was failing to test. The training set teaches no rule that appears in the test set.

The benchmark was supposed to be the test that exposed pattern matching dressed up as reasoning. For most of its life it did exactly that: large language models scored in the single digits, well below the human baseline. Then, in 2024, frontier models began chipping away at it. In December 2024 ARC Prize reported that a preview version of OpenAI’s o3 scored 75.7% on the Semi-Private evaluation set of ARC-AGI-1 within the prize’s compute limits, and 87.5% in a high-compute configuration. ARC Prize later priced that high-compute run, at o3-pro rates, at about $4,560 per task, a cost that raised its own question about whether such a solve counts. Chollet has been candid in interviews that partial solutions on ARC-AGI-1 do not amount to general intelligence. He released the harder ARC-AGI-2 in March 2025 and ARC-AGI-3 in March 2026. By February 2026 ARC Prize had verified Gemini 3 Deep Think at 84.6% on ARC-AGI-2’s Semi-Private set, and in September 2026 it reported GPT-6 Astra at 62.7% on ARC-AGI-3’s Semi-Private set under its Standard harness, a score of action efficiency against human testers.

But the simplest version of LeCun’s empirical claim has gotten harder to defend. Every benchmark LeCun has cited as a wall LLMs cannot climb has eventually had a route up it found, even if the route is expensive and partial.

The empirical record also runs inside the models themselves. Anthropic has spent the last two years publishing on mechanistic interpretability. Will Douglas Heaven covered the most striking results for MIT Technology Review (March 2025). The findings cut both ways. In one experiment, the team found that when the model was asked to write a rhyming couplet, it picked the rhyme word for the second line before it began writing the line, committing in its activations to “rabbit” and then constructing a sentence to land on it. That finding is awkward for the simplest version of LeCun’s critique. Something that looks structurally like planning, on a small scale, is happening in the latent space.

In the same body of work, the team found cases that cut the other way. The model’s commitment to grammatical fluency could override its safety training. Joshua Batson, the Anthropic researcher Heaven quotes, called the findings “tip-of-the-iceberg stuff” (MIT Technology Review, March 2025), and noted that the analysis runs on what Anthropic calls a replacement model. In the published study of Claude 3.5 Haiku, learned cross-layer transcoders stand in for the model’s MLP layers, and Anthropic says the substitute can use different mechanisms from the model it approximates.

JEPA has produced vision and vision-language encoders and a robot-planning system. In June 2025, Meta reported that V-JEPA 2-AC used image goals to pick and place objects with Franka arms in two unfamiliar lab environments, after post-training on fewer than 62 hours of robot video. In October 2026, LeCun and coauthors posted H-JEPA, a hierarchical latent world model evaluated on simulated visual-planning tasks and on offline robot video. Language-model-derived robotics systems report physical results too: Google DeepMind’s Gemini Robotics, built on Gemini 2.0 and announced in March 2025, controls robots and reports handling tasks it had not seen in training. Whether either line generalizes broadly across physical tasks is open. Large language models have shipped four years of capability gains, partial solutions on benchmarks they were not supposed to touch, and interpretability findings that show both more internal structure than the harshest critics expected and more brittleness than the most enthusiastic defenders admit. Neither side has the result that would settle the argument.

The result that would settle it is roughly knowable. The unresolved comparison is whether JEPA-based systems generalize across physical tasks more reliably and efficiently than competing architectures. A JEPA system that matched human efficiency on ARC-AGI-3 under ARC Prize’s Standard harness, at a fraction of a language model’s compute, would also count for LeCun. A large language model that did the same, or a language-model-based robot system that planned reliably across physical tasks it had not seen, would vindicate the labs LeCun keeps naming.

LeCun’s bet is not that LLMs will fail to improve. They have improved, and he has not been right about the ceiling so far. The bet is that those gains are on the wrong axis, and that a system that learns by predicting in latent space from video will, when it arrives, do what the largest language model cannot.