The substrate split with LeCun

On December 22, 2025, Demis Hassabis posted on X that Yann LeCun was “just plain incorrect, confusing general intelligence with universal intelligence.” The line was a four-clause rebuttal to a podcast riff LeCun had given on the Information Bottleneck a few days earlier, where Meta’s former chief AI scientist had called “general intelligence” the concept “complete BS.” Within twenty-four hours LeCun had replied that the disagreement was “largely one of vocabulary.” The exchange ran for a news cycle as a personality trade and then sank. A month later, on January 23, 2026, both men appeared on separate panels at Davos and the substance of the disagreement surfaced in primary-source form. They agreed current systems are nowhere near artificial general intelligence. They split, sharply, on whether language models are the right substrate to keep building on.

The row they share

Locating the two men on a common grid clarifies what is actually being argued. One axis asks whether scaling alone is sufficient; the other asks whether the language model is the right substrate. Altman and Amodei hold that scaling a language model is sufficient. Hassabis holds that scaling is insufficient but the language model is the right base. LeCun holds that scaling is insufficient and the language model is the wrong base. No major frontier voice holds that scaling is sufficient on some other substrate. Hassabis and LeCun share a row and disagree about the column.

The bottom row — scaling alone is insufficient — is also where most of the AAAI’s surveyed researcher community sits. The 2025 Presidential Panel on the Future of AI Research, chaired by outgoing AAAI President Francesca Rossi, found that 76 percent of 475 surveyed researchers think scaling up current approaches is unlikely or very unlikely to yield AGI. Hassabis and LeCun are not outliers in believing the recipe is incomplete. They are outliers in being unusually specific about what is missing, and what would replace it.

The columns are the disagreement. Hassabis treats the scaled language model as a necessary prior — the place where a great deal of human knowledge already sits, encoded in a form a sufficiently large neural network can compress and condition on — and proposes to bolt three things on top: AlphaZero-style planning and search, foundation “world models” that learn intuitive physics, and parallel-thinking inference. LeCun treats the scaled language model as the wrong substrate full stop, arguing that an agent that has to act in physical space needs a system that predicts the consequences of its actions in advance, which next-token prediction structurally does not. He proposes to train video-and-image models with a non-autoregressive objective he calls a “joint embedding predictive architecture,” or JEPA. Both men have a published 2025 paper trail behind their respective lines.

Where the disagreement starts and where it doesn’t

The shared part is worth dwelling on, because the X exchange disguised it. On the Google DeepMind podcast on December 16, 2025, Hassabis told its host, Hannah Fry, that the lab had moved from training answerers to training researchers — “In the past, we trained answerers; now, we need to train researchers” — and named two prerequisites for the next jump: “world models” that learn intuitive physics, and automated experimentation that closes the perception-action loop. On the Big Technology Podcast at Davos on January 21, 2026, he extended the framing: continual learning, persistent memory, “true world models,” and automated experimentation are the gating items, and “one or two more breakthroughs” at the level of the Transformer or AlphaGo are probably needed before AGI is in reach. None of that is an Altman or Amodei position. None of it is a LeCun position either. It sits between them, and it has been Hassabis’s stable framing across at least eight dated venues from February 2024 through May 2026.

LeCun’s diagnosis at AI House Davos on January 23, 2026 was structurally similar in what it ruled out. “We’re never going to get to human-level intelligence by training LLMs or by training on text only,” he told the Fortune-covered panel. “We need the real world.” On the same panel: “The AI industry is completely LLM-pilled… everybody is working on the same thing. They’re all digging the same trench.” The phrasing is different from Hassabis’s. The diagnostic content — that text-trained autoregressive systems are not going to reach AGI by themselves, however much compute is added — is recognizably the same.

Where the prescriptions split, both men have evidence to point at, and the evidence is strong enough that neither line dismisses easily.

LeCun’s argument, fairly

LeCun’s strongest single piece of empirical evidence is V-JEPA 2, released June 11, 2025, with 47 authors including LeCun himself. The paper, V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning (arXiv 2506.09985), trains a non-autoregressive video model on more than a million hours of internet video plus a million still images, with no language pretraining and no human labels in the encoder. The headline numbers: 77.3 top-1 on Something-Something v2; 39.7 recall-at-five on Epic-Kitchens-100 action anticipation, which is state-of-the-art; and, once the encoder is aligned with an 8-billion-parameter language model, 84.0 on PerceptionTest and 76.9 on TempCompass video question answering.

The follow-up is the heavier piece. The action-conditioned variant V-JEPA 2-AC was post-trained on under 62 hours of unlabeled robot video drawn from the DROID dataset — the Distributed Robot Interaction Dataset, an open multi-lab corpus of teleoperated manipulation — then deployed zero-shot, no per-environment finetuning, no per-task demonstrations, on Franka robot arms in two different laboratories. Meta reports pick-and-place success of 80 percent with a cup and 65 percent with a box, averaged across the two labs over ten trials per lab and object, with individual lab results from 50 to 80 percent, using only image goals as the planning specification. That is what “the wrong substrate” claim is meant to cash out as: a system that learned the dynamics of physical space from pixels can drive a robot it has never seen, in a lab the model never trained for, well enough to manipulate objects it has never met.

LeCun left Meta at the end of 2025, after twelve years, five of them as the founding director of FAIR — the Fundamental AI Research lab Meta still runs. AMI Labs, the Paris startup he chairs and Alex LeBrun runs as chief executive, raised $1.03 billion at a $3.5 billion pre-money valuation on March 10, 2026, the largest seed round in European startup history. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with NVIDIA, Samsung, and Temasek participating. He has framed the move as following a project Meta would not back at frontier scale, an argument deslop.media has covered separately. An early test of whether the JEPA-class line scales beyond the V-JEPA 2 record without Meta’s compute behind it came on August 27, 2026, when AMI-affiliated researchers, LeCun among the authors, released LeVJEPA, an open video encoder reporting 5.6 to 20.8 times less matched-epoch pretraining compute than V-JEPA 2. It is an encoder result, not the planning system the line promises.

The substrate-is-wrong claim, in other words, is not rhetoric. It has a published 2025 paper behind it, two laboratories’ worth of zero-shot transfer behind it, and a billion-dollar lab now dedicated to extending it.

Hassabis’s argument, fairly

The DeepMind line has its own 2025 evidence, structured around the same year.

Genie 3, announced August 5, 2025, is the foundation “world model” that Hassabis has consistently framed as a “key stepping stone on the path to AGI.” The DeepMind blog reports that Genie 3 generates interactive 3D environments that retain consistency “for a few minutes at a resolution of 720p” from a text prompt, producing playable worlds that an agent can act inside. The DeepMind framing is that “world models” of this class enable unlimited training curricula for agents, which is the bottleneck Hassabis names whenever he talks about closing the perception-action loop on the LLM substrate. On the Hannah Fry podcast he made the architectural claim as a single sentence: AGI requires “world models” and automated experimentation joined into a research closed loop.

Gemini Deep Think is the second 2025 datapoint. On July 21, 2025, DeepMind announced that an advanced version of Gemini with Deep Think had achieved gold-medal standard at the International Mathematical Olympiad — solving five of the six problems perfectly, end-to-end in natural language, within the 4.5-hour competition limit. The score, 35 of 42, was officially certified by IMO President Gregor Dolinar. Of 630 human contestants, 72 received gold; the cutoff was the same 35-of-42 the system reached. The system is parallel-thinking search built on a pretrained LLM. It is a fully bolted-on architecture in the Hassabis sense — a “thinking system on top of a model,” in the phrasing he used on the Big Technology Podcast at Google I/O on May 21, 2025.

What does Hassabis’s line require to hold? “World model” fidelity has to keep climbing. If Genie’s successors stall at minutes-of-720p-consistency through 2027, the “language alone is not the world” objection becomes the more parsimonious read of what the LLM substrate cannot do. Conversely, ARC-AGI-3 — the third iteration of the Abstraction and Reasoning Corpus benchmark, launched March 25, 2026 by the ARC Prize Foundation that François Chollet, the Keras creator, chairs — opened with frontier scores below one percent against humans at one hundred percent; on September 2, 2026, ARC Prize reported OpenAI’s GPT-6 Astra, not a Gemini release, at 62.7 percent on its Semi-Private set under the Standard harness at max reasoning effort. ARC-AGI-3 is the interactive-environment benchmark that Hassabis’s bolted-on architecture has to close on. If multiple Gemini releases through Q4 2026 leave ARC-AGI-3 near zero, the Genie-and-Deep-Think line resembles what LeCun says it is: a delaying tactic, not an answer.

There is an empirical record beyond Genie and Deep Think. AlphaProof, published online in Nature in November 2025, formalizes mathematical proofs by running an AlphaZero-style RL agent in Lean — pretrained language model seeded, tree search on top — and reached IMO silver in July 2024. AlphaEvolve, the LLM-driven evolutionary code-search system DeepMind disclosed in May 2025, has been in production inside Google for over a year recovering 0.7 percent of worldwide compute, and ships new matrix-multiplication algorithms and improved best-known solutions on twenty percent of fifty open math problems. These are bolted-on architectures, not pure-scale Transformer outputs, and they are running.

The hybrid-architecture trend at production scale corroborates that the bolt-on path is broader than DeepMind’s Alpha line. NVIDIA’s Nemotron-H (April 2025), the Nemotron 3 family, announced in December 2025, IBM Granite 4.0 (October 2025) with its 9:1 Mamba-to-attention ratio and over 70 percent reduction in long-context memory, and AI21’s Jamba 1.5 (ICLR 2025) all interleave non-attention layers with Transformer attention at frontier scale. None of them are LeCun’s JEPA. None of them are pure-scale Transformers either. The architectural pluralism Hassabis describes is, at the production layer, where most of the field already is.

The two lines, side by side

The tests are concrete on both sides. JEPA weakens if it cannot scale to long-horizon manipulation, or if no comparative scaling-law plot beats Transformer at matched compute when an independent third party finally runs one — Meta’s V-JEPA 2 paper deliberately does not present such a plot. The Hassabis line weakens if Genie-class “world model” fidelity stalls at minutes through 2027, or if ARC-AGI-3 stays near zero through Q4 2026 across multiple Gemini releases.

That is what makes the substrate split trackable rather than a personality clash. Neither man has set these deadlines himself; they are the results that, on this reading, would count against each line.

The “vocabulary” defense

The most generous reading of LeCun’s December 22 follow-up — that the disagreement is “largely one of vocabulary” — is that he and Hassabis have been arguing past each other on the meaning of “general.” Hassabis’s framing, made explicit in the same X thread, is that brains and large foundation models are “approximate Turing Machines,” with generality understood as a spectrum rather than as optimality on every possible task. “Universal intelligence,” in his framing, is the philosophical maximum no biological system reaches; “general intelligence” is the operationally useful generalization across the kinds of tasks humans handle. LeCun’s dismissal was aimed at the philosophical maximum. Hassabis was defending the operational version. On that reading, the two men converge on more than the X feed suggested: both want a system that learns continuously, predicts the consequences of actions, and acts in the real world.

Gary Marcus, the cognitive scientist and longtime LLM critic, has read this overlap from the outside, from a third architectural direction. In an October 2024 essay he framed the AlphaFold-Nobel as “the first Nobel Prize for Neurosymbolic AI” — substantial agreement with his own position that pure end-to-end neural is insufficient. His January 2026 read of Hassabis’s late-2025 messaging called him “the latest to say that ChatGPT is a dead-end and that we must turn our focus to world models.” Marcus’s framing flatters his own camp; it also corroborates that the LLM-only road is contested across more than two voices. The disagreement remains real, however. Hassabis is not adopting LeCun’s substrate rejection; he is widening the architectural envelope around the LLM core, which is what he has been doing since at least Wired’s February 2024 interview, where he articulated “AlphaGo capabilities built on top of these large models.”

The “vocabulary” claim holds as far as it goes. It does not resolve which substrate the architecture rests on.

What the next two years actually test

The disagreement is a forecast, not just a debate. Both lines are now downstream of empirical work that will be visible to outsiders by the end of 2027.

For LeCun’s line, three observable items: whether AMI Labs follows its August 2026 LeVJEPA encoder with a working planning system, or an independent group reproduces LeVJEPA’s compute savings; whether any independent third-party scaling-law plot — academic or otherwise — finds JEPA-class systems beating Transformer at matched compute, in either direction; whether V-JEPA 2’s successor lands long-horizon dexterous manipulation comparable to what RT-2-class systems — the Google DeepMind robotic-transformer family that converts vision and language into low-level motor commands — target. The first depends on AMI’s own research schedule. The second is a 2026–2027 academic-paper question that Meta, by the structure of its V-JEPA 2 paper, has visibly chosen not to settle internally. The third is the test the substrate-is-wrong claim has built up to since 2018.

For Hassabis’s line, three observable items: Genie-and-SIMA successors, where “world model” fidelity either keeps climbing past minutes-at-720p or stalls; ARC-AGI-3 frontier scores through Q4 2026, where the bolted-on architecture either closes the gap or does not; the DeepMind UK government partnership announced December 10, 2025, which commits to an automated DeepMind materials laboratory in 2026, with researchers overseeing robotics-driven experiments, and is the closest thing to a shipping test of the “world models plus automated experimentation” framing.

Each line has a specific public experiment pending. That is the empirical version of the disagreement.

The harder question — the question neither the X exchange nor the Davos panels resolved — is whether the substrate question will resolve on the experiments at all, or whether it will resolve only by which line ships AGI first. Hassabis’s own framing, in the Axios interview on December 3, 2025, allows for the latter. “The scaling of the current systems, we must push that to the maximum,” he said, and “one or two more breakthroughs” are still likely needed at the level of the Transformer or AlphaGo. If the breakthrough comes from the JEPA line, LeCun’s substrate rejection is vindicated. If it comes from a Genie-class “world model” docked to a parallel-thinking system on top of Gemini, Hassabis’s bolted-on architecture is vindicated. Either way, the field updates on a result, not on a panel.

The deeper part of the disagreement, which neither man has named directly, is whether “the substrate” is even the right unit of analysis at frontier scale. Production systems shipping in 2025 and 2026 — the Nemotron family, Granite 4.0, Jamba — are already hybrid in ways the X exchange did not contemplate. Mamba-3, a linear-time state-space architecture that researchers at Cartesia, CMU, Princeton, and Together AI announced on March 11, 2026 and described in a preprint five days later, came with the prediction that linear layers will be predominantly used in conjunction with global self-attention. If that is the production future, the question of which substrate is the substrate may already be the wrong question; the empirical answer might be that no one substrate carries the load and that the architecture is the combination.

LeCun, on the X thread, left it at “largely one of vocabulary.” Hassabis has not yet said whether he agrees.