Three years ago, Yann LeCun wrote a single sentence that's become one of the most cited and contested claims in AI: "auto-regressive LLMs, in and of themselves, will not lead to human-level AI." On September 20, 2026, he reposted it and defended it point by point against the years of LLM progress since — and within hours, Jürgen Schmidhuber used the reply thread to revive a separate, sharper dispute: that LeCun's signature alternative architecture, JEPA, isn't actually his own idea.
This lands in the middle of a year explainx.ai has tracked closely on the world-model side of the AI race — LeCun's Moravec's-paradox argument about physical agents in July, his revival of the 2019 GPT-2 mockery earlier this month, and a steady stream of JEPA-adjacent research including LeVJEPA's 20x compute reduction claim in August.
TL;DR: What people are asking
| Question | Direct answer |
|---|---|
| What's LeCun's actual claim? | Autoregressive LLMs alone won't reach human-level AI — a 2023 statement he says current evidence still supports, three years and multiple frontier model generations later. |
| What's his evidence? | Five points: current "reasoning" is inefficient token-space search, not the continuous-space search he advocates; self-improvement (RL) only works where outputs are automatically scorable (math, code); multimodal systems already use separate non-LLM encoders, validating his JEPA approach; there's still no consumer Level-4/5 self-driving car; and intelligence is "what you do when you don't know," per Piaget, which current systems still lack. |
| What did Schmidhuber say? | That JEPA, which LeCun introduced in 2022, is "essentially identical" to Schmidhuber's own 1992 Predictability Maximization technique — a priority dispute he's raised before, including in March 2026. |
| Did LeCun respond to Schmidhuber? | Not in the thread this post covers — no reply from LeCun to the JEPA-priority claim is present as of this writing. |
| Is this new? | The core 2023 claim isn't new — this is LeCun defending it three years later, in a thread that then branched into the separate, older JEPA-authorship dispute with Schmidhuber. |
| Where does LeCun work now? | He founded and chairs AMI Labs, alongside roles at NYU and 224 Ventures, after leaving his position as Meta's Chief AI Scientist. |
The five-point defense of a three-year-old claim
What makes this thread worth reading in full rather than skimming as a repost is that LeCun doesn't just restate his 2023 line — he lays out, point by point, why he thinks 2026's frontier models haven't disproven it:
- Current reasoning is search, but in the wrong space. LeCun concedes that today's models do use non-autoregressive search — which he says is exactly what he's "always advocated for" — but argues they perform it in discrete token space, which he calls "limited and inefficient." His long-standing position is that human-like reasoning requires search in continuous representation space instead, and he reads the industry's direction as slowly moving that way.
- Self-improvement is domain-limited. Reinforcement-learning-based self-improvement, as currently practiced, only works where outputs can be automatically scored without a human in the loop — math, code, and simulatable scenarios. LeCun's point: humans and animals learn new skills far more sample-efficiently than current RL methods require, and that gap hasn't closed.
- Multimodality already validates his approach, in his telling. Today's multimodal AI assistants generally use separately trained encoders rather than a single end-to-end LLM handling every modality — which LeCun frames as evidence for his own JEPA-style architecture, self-supervised learning over continuous embeddings rather than token prediction. He cites roughly 3,000 JEPA-related papers published in four years as a sign the research community is moving in that direction.
- The absent Level-4/5 car is his sharpest concrete test. If LLMs were a genuine path to human-level AI, LeCun argues, "we would have domestic robots and Level-4 or Level-5 self-driving cars for consumers by now" — and don't. He specifically notes the absence of cars that can "learn to drive in 20 hours or practice like any teenager," framing sample-efficient physical learning as a capability current systems still conspicuously lack, a continuation of the Moravec's-paradox argument explainx.ai covered from him in July.
- A definitional argument, via Piaget. LeCun closes by arguing that intelligence isn't "the mere accumulation and regurgitation of existing declarative knowledge" — which he characterizes as what LLMs do — and quotes Jean Piaget: "intelligence is not what you know, it is what you do when you don't know." His claim is that solving genuinely novel problems without prior training, and adapting quickly to unfamiliar situations, remains the gap current systems haven't closed.
The pushback LeCun didn't answer
The most substantive reply in the thread doesn't engage LeCun's five points directly — it goes after the foundation underneath point three. Jürgen Schmidhuber, whose own priority disputes with the field's dominant narratives are a running theme in AI research discourse, wrote: "What a misleading post. First of all, the so-called 'JEPA' family of techniques (2022) is actually the 1992 Predictability Maximization family of techniques." He linked back to a March 2026 post making the identical claim — "Dr. LeCun's heavily promoted Joint Embedding Predictive Architecture (JEPA, 2022) is essentially identical to our 1992 Predictability Maximization system (PMAX)" — meaning this isn't a new accusation prompted by this specific thread, but a standing dispute Schmidhuber has pressed for at least six months and chose to resurface here.
This matters for how to read LeCun's argument as a whole: his defense leans on JEPA as his own original contribution, cited as vindication that the field is validating his specific architectural bet. If Schmidhuber's priority claim is accurate — a question this post doesn't resolve, since it requires a technical comparison of the two techniques' actual mechanics that neither side's tweet-length exchange settles — then the "3,000 papers building on my idea" framing is at minimum contestable on authorship grounds, even if JEPA's practical approach and PMAX's original 1992 formulation share meaningful conceptual DNA. LeCun did not reply to Schmidhuber's specific claim in the thread as covered here.
Other replies pushed on narrower points: one commenter (musedivision) argued LeCun's actual underlying motivation is skepticism that LLMs lead to an AGI breakthrough at all, and that he "has not been vindicated" on that broader opinion regardless of the five-point technical defense; another (Andrei Stan) asked pointedly what AMI Labs, LeCun's well-funded post-Meta venture, has actually released since its founding — a fair question given LeCun's argument rests partly on what the rest of the field hasn't shipped yet (Level-4/5 cars, domestic robots), while his own lab's public output is not itself detailed in this thread.
Where this sits relative to LeCun's other 2026 disputes
This is a different kind of disagreement than the one explainx.ai covered when LeCun revived the 2019 GPT-2 mockery earlier this month — that thread was about open-weight release policy and safety-fear framing, a governance and access question. This one is narrower and more architectural: not whether LLMs are dangerous or should be open-sourced, but whether the autoregressive-token-prediction approach that underlies essentially every frontier chat model released in 2026 is, on its own, structurally sufficient to reach human-level intelligence. It's also distinct from — though adjacent to — his 2023 exchange with Geoffrey Hinton about existential AI risk, which was a safety and governance dispute rather than a claim about architectural sufficiency.
Honest limitations
- Whether Schmidhuber's priority claim about JEPA is technically accurate is not resolved by this post — it requires a detailed comparison of PMAX's 1992 formulation against JEPA's actual mechanics, which neither side has laid out in the exchange this post covers.
- LeCun's five-point defense is his own framing of the evidence, not an independently verified technical assessment — points like "reasoning abilities... are based on non-autoregressive search" describe a contested interpretation of how current frontier models actually work internally, which their creators haven't confirmed in these specific terms.
- The "3,000 papers on JEPA" figure is LeCun's own stated count, with no cited source or methodology in the thread this post is sourced to.
- This post covers a single X thread and its replies as of September 21, 2026 — no independent verification of AMI Labs' current output, or of the specific self-driving and robotics capability claims, was performed for this post.
What this means for builders
The concrete, checkable part of this exchange — separate from who said what first about JEPA — is LeCun's underlying architectural bet: that continuous-representation self-supervised learning, not further autoregressive scaling, is the more promising direction for tasks requiring sample-efficient physical or novel-situation learning. If your own work sits closer to robotics, physical agents, or world models than to chat-style LLM applications, that's a genuine, substantive research direction worth tracking directly — via the growing JEPA literature LeCun cites, and via world-model coverage explainx.ai has followed all year — independent of how the priority dispute with Schmidhuber eventually settles.
Related on explainx.ai
- Yann LeCun on LLMs, physical agents, and Moravec's paradox
- Yann LeCun revives the 2019 GPT-2 mockery — and the staged-release nuance
- LeVJEPA: video pretraining at 20x less compute
- What are world models? Starchild-1, Odyssey, complete guide
- Sakana's predictive-coding alternative to backpropagation
- Richard Sutton on Oak, AGI, and algorithms
- World Labs' Atlas: a multimodal 3D world model
This post is sourced to a public X thread from Yann LeCun dated September 20-21, 2026, and its replies, including a cross-reference to Jürgen Schmidhuber's own March 2026 post making the same JEPA-priority claim. No independent technical verification of the JEPA-versus-PMAX architectural comparison, AMI Labs' output, or LeCun's specific capability claims was performed for this post; read the framing above as a report of the public exchange, not an adjudication of its underlying technical dispute.