Yann LeCun Calls LLMs ‘Token Generators’ While Llama Hits a Billion Downloads

At Nvidia’s GTC 2025, Meta’s chief AI scientist, Yann LeCun, dismissed large language models (LLMs) as mere “token generators,” suggesting that their limitations—stemming from their reliance on discrete tokenized spaces—render them ineffective in the long run. LeCun didn’t just criticize LLMs; he predicted their obsolescence within the next five years before as well saying that nobody in their right mind would use them anymore.

LeCun’s skepticism is well-documented. He argues that modern AI systems fail in four key areas: they lack awareness of the physical world, have limited memory and no continuous recall, are incapable of reasoning, and struggle with complex planning. For him, these are fundamental weaknesses that cannot be solved by simply increasing scale. Reinforcement learning, according to LeCun, has been a “huge failure,” and generative AI is incapable of establishing real truth or understanding the world.

He contrasts AI’s learning process with human cognition. While LLMs ingest vast datasets, a four-month-old infant, simply by observing its environment, processes 450 times more information than the largest LLM. LeCun sees this as a clear indication that current AI models fall drastically short of real intelligence.

However, his remarks come just as Meta reported a major milestone for its Llama models. Meta announced that Llama, its open-source series of large language models, has now been downloaded a billion times.

Ahmad Al-Dahle, Meta’s VP and Head of GenAI, celebrated this achievement, calling it “a vote of confidence in open-source AI.” This makes Meta’s position contradictory—on one hand, its chief AI scientist dismisses LLMs as a dead-end technology, while on the other, the company continues to invest heavily in their development and distribution.

LeCun’s critiques are not new. In his position paper, “A Path Towards Autonomous Machine Intelligence,” he has previously argued that current machine learning approaches are fundamentally flawed. He asserts that LLMs are doomed due to their reliance on autoregressive generation, which inevitably leads to compounding errors. He also rejects contrastive training techniques like GANs and BERT, instead favoring regularized training methods such as Principal Component Analysis (PCA) and Sparse Autoencoders (AE).

His broader argument is that modeling probabilities in continuous domains leads to infinite gradients, making true reasoning impossible. He dismisses generative modeling as misguided, arguing that much of reality is unpredictable and should not be artificially modeled. Instead, he believes AI should learn passively from visual observation, similar to how humans acquire knowledge. This claim raises questions, as it does not account for the intelligence of those born blind, who develop reasoning abilities without passive visual intake.

Meanwhile, OpenAI is advancing a different vision. Their newly launched Deep Research model, built on OpenAI’s most advanced reasoning system (o3), is gaining traction among professionals. Unlike traditional LLMs that merely generate text, Deep Research autonomously explores the web, selects relevant sources, clicks links, and compiles research into detailed reports. The tool, available through OpenAI’s $200-per-month ChatGPT Pro plan, has been widely used since its February 2 launch, earning praise from figures like Stripe CEO Patrick Collison and Wharton professor Ethan Mollick. OpenAI’s approach challenges LeCun’s view, demonstrating that LLMs can evolve beyond simple “token generators” when paired with new capabilities.

LeCun also questions the necessity of massive models, pointing out that a mouse’s brain containing only tens of millions of neurons outperforms current robotic AI. He suggests that intelligence is about efficiency, not scale, and that current AI models are bloated and inefficient.

Meta’s simultaneous criticism and promotion of LLMs highlight the current divide in AI research. While some, like OpenAI, are doubling down on improving generative AI, others, like LeCun, believe that a new approach is needed. Meta’s billion Llama downloads indicate strong demand for LLMs, contradicting the idea that they are on the verge of obsolescence.