There's a question circulating through AI labs that few dare to raise openly: does ChatGPT actually understand anything? The short answer is no. Yann LeCun, one of the pioneers of deep learning and a Turing Award laureate, has repeated this with a bluntness that stirs debate across Silicon Valley.

To understand why LeCun is charting a different course than OpenAI, Anthropic, or Google, we need to examine what LLMs actually are and why their apparent success masks a fundamental limitation. A large language model, like GPT-4 or Claude, functions as a statistical tool for predicting text. During training, it analyzes enormous volumes of textual data and answers a basic query: given this sequence of words, which one most likely comes next? It uses transformer architectures with billions of parameters, but at its core it's generating a map of associations between linguistic tokens.

This matters because language represents the world—it isn't the world. It's an imperfect version of reality, filtered through cultural context. When an LLM states that "ice is cold," it isn't deducing this from any sensory experience, but from how frequently those words co-occur in its training data. Ask it about something underrepresented in the data, and it stumbles. Not for lack of information, but for lack of any causal understanding of the physical environment. LLMs capture linguistic structures through statistics, but they don't model the world beyond those textual correlations.

LeCun sums it up as "glorified text autocomplete." It sounds provocative, and it is, but it points to a real technical truth. A three-year-old intuits that tilting a glass spills the water, even without ever having seen it happen before. No LLM has that physical intuition unless it's been described in text somewhere. This gap separates language modeling from modeling reality itself.

LeCun didn't arrive at this idea overnight. While serving as Chief AI Scientist at Meta, he developed an alternative architecture called JEPA (Joint Embedding Predictive Architecture). Its core differs radically from LLMs: instead of anticipating the next word within a textual domain, JEPA generates abstract representations of the world, predicting future states in a high-level space without needing exact sensory detail.

In practical terms, a JEPA system would observe its environment, build an internal causal model of how it works, and apply that model to plan actions. It resembles the way a mammal learns in its habitat far more than it resembles optimizing a text compressor. LeCun argues this approach opens a path toward human-like intelligence, though he's careful to avoid the term AGI. He isn't just criticizing the status quo—he's proposing an alternative grounded in concrete technical foundations.

LeCun has now left Meta to found a startup that will push these ideas forward in an independent setting. Operational details remain scarce, but the goal is clear: build AI systems that construct models of the world, reason causally, and operate without depending on masses of labeled text. They would learn from video, physical interaction, and observing consequences—much like animals do. It's the same principle that led Judea Pearl to reformulate statistics around causality, a minority current that coexisted for decades alongside the dominance of correlational learning.

Would this AI be more explainable? There are signs that it might. LLMs are opaque: their decisions emerge from interactions across trillions of parameters that are impossible to trace. Models built on world representations could offer greater interpretability, since their internal structures would reflect causal concepts rather than mere textual regularities. Still, "more interpretable" doesn't mean "fully transparent." How much of that clarity is real versus aspirational at this current stage remains an open question.

Would it lead to AGI? LeCun remains skeptical of the concept itself. He believes LLMs won't reach general intelligence no matter how much data or parameters get thrown at them. His proposal emphasizes learning about the physical world, causal reasoning, and hierarchical planning as the essential path forward. If JEPA delivers on its promise, it would yield systems far more capable at tasks requiring genuine environmental understanding. Calling it AGI, without solid evidence yet, would be premature.

The philosophical contrast is stark. OpenAI, Anthropic, and Google are betting that scaling LLMs—more data, more compute—will eventually produce something close to understanding. LeCun sees that path as a sophisticated dead end: impressive, but running into limits that don't disappear with more compute. It's like climbing a ladder to reach the Moon. Other researchers argue that LLMs could develop causal models through sheer scale alone. The hypothesis remains open, though the theoretical grounding for it is still thin.

LLMs attract massive investment because they shine in demos, generate persuasive text, and enable fast product launches. LeCun's vision is harder to commercialize in the short term: learning from the physical world demands time, cost, and complex testing, far removed from an instant chatbot. His track record carries weight, though. The convolutional networks he developed built the foundation of modern computer vision, and for years they were considered a marginal bet.

LeCun's central question is an uncomfortable one: are we trying to build systems that mimic human language, or ones that model the world the way we do? These are different goals, with different consequences. A causal AI would be more robust, safer in critical applications, and better aligned with human values by reasoning about real-world impact rather than just textual ethical regularities.

This bet carries real risk and a distant horizon. But throughout the history of technology, visions built on solid foundations tend to eventually prevail—even if the people who started them don't always live to see them bear fruit.

Stones don't lie, but historians sometimes do.

Sources:

1. LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Meta AI Research. (Public whitepaper)

2. LeCun, Y. Presentations at NeurIPS 2023 and interviews on the Lex Fridman Podcast (episodes 258 and 416)

3. Bengio, Y., Hinton, G., LeCun, Y. (2018). Deep Learning. ACM Turing Award.

4. Marcus, G. & Davis, E. (2019). Rebooting AI: Building Artificial Intelligence We Can Trust. Pantheon Books.

5. Pearl, J. & Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books.