Yann LeCun argues LLMs will drive real-world applications, but not human-level thinking

Yann LeCun, Meta’s chief AI scientist and one of the godfathers of deep learning, is making a nuanced argument that cuts against both AI hype and AI doomerism simultaneously. Large language models are commercially useful, he says. They’ll justify the billions being poured into GPU clusters and data centers. But the bubble isn’t in the infrastructure spending. It’s in the belief that these models can think like humans.

The case for LLMs as utility, not oracle

LeCun’s argument is straightforward once you strip away the academic jargon. LLMs are good at a growing list of practical tasks: coding assistance, enterprise search, document summarization, customer service automation. These applications generate real revenue and solve real problems. That makes the massive infrastructure buildout, the GPU farms and the power plants, a defensible investment.

LeCun draws an aggressive line in the sand about what these models fundamentally cannot do. He’s been arguing for years that next-token prediction, the core mechanism behind every major LLM from GPT-4 to Claude to Llama, is a “dead end” for achieving anything resembling genuine intelligence.

LLMs learn by consuming trillions of tokens of text. A child learns to understand the physical world from far fewer hours of visual experience. The model can generate a convincing paragraph about how gravity works. The toddler can actually catch a ball. These are fundamentally different kinds of understanding.

World models and a billion-dollar bet

Advanced Machine Intelligence, or AMI Labs, is a new venture focused on building AI systems that learn from raw sensory inputs rather than text. The company raised $1.03 billion in seed funding, reportedly the largest seed round ever for a European startup.

LeCun’s lab at Meta has also been working on this front directly. They introduced a model called Leham, built on what’s known as a JEPA-based architecture, short for Joint Embedding Predictive Architecture. Rather than predicting the next word in a sequence, JEPA-based systems learn by predicting abstract representations of future states. The goal is improved planning efficiency and a more grounded understanding of how things actually work in the physical world.

What this means for investors and the AI market

The companies that may be most exposed are those whose valuations depend on LLMs eventually achieving something close to artificial general intelligence. If LeCun is right that the current architecture has a hard ceiling, then the premium baked into certain AI stocks for “the AGI upside” is mispriced.

The $1.03 billion flowing into AMI Labs suggests that at least some sophisticated investors agree. When capital at that scale moves toward an alternative research paradigm, it’s worth paying attention to what’s being implicitly abandoned, not just what’s being funded.

The risk, of course, is that LeCun is wrong and scaling laws continue to surprise. OpenAI, Anthropic, Google, and others are betting that bigger models with more data and more compute will continue to unlock new capabilities. But if you’re allocating capital based on that assumption, you should at least be aware that one of the most credentialed researchers in the field — one of the three researchers who won the Turing Award for foundational work in deep learning — thinks you’re making a category error.