Yann LeCun Warns Large Language Models Hit Limits, Launches AMI Labs Shift

Yann LeCun, one of the “godfathers of AI,” has warned that Silicon Valley’s multibillion‑dollar bet on large language models is “marching into a dead end,” arguing that current chatbots cannot be scaled into human‑level intelligence because they lack any real understanding of the physical world nytimes +1. His critique has sharpened since he left Meta in November 2025 to launch AMI Labs, a Paris-based startup pursuing a rival approach built on so‑called “world models” observer +1.

LeCun, a Turing Award winner who helped invent the neural networks behind today’s AI boom, said companies have already poured “hundreds of billions of dollars” into LLM‑driven systems and data centers, driven by a “herd effect” that funnels talent and capital into a single paradigm nytimes. “They’re all digging the same trench,” he said, warning that ever‑larger text models will hit hard limits on reasoning, planning and reliability businessinsider.

Why LeCun Says LLMs Are a Dead End

LeCun’s technical objection is that models trained mainly to predict the next word in text “cannot truly reason or plan, because they lack a model of the world” technologyreview. He argues that language alone doesn’t contain enough continuous, causal information for an AI system to predict the consequences of actions in the real world, a capability he sees as central to any human‑level intelligence technologyreview.

Instead, AMI Labs is betting on architectures like his Joint Embedding Predictive Architecture (JEPA), which learn abstract representations from video, audio and sensor streams and make predictions in a compact “latent space” rather than generating every pixel of future frames businessinsider +1. Such “world models,” LeCun says, are needed for robots and digital agents that can reliably navigate, manipulate objects and plan several steps ahead. By contrast, LLMs remain prone to hallucinations and brittle reasoning, weaknesses he portrays as structural rather than merely engineering bugs businessinsider +1.

Industry Pushback and a Split Over the Path to AGI

LeCun’s broadside has intensified an already public rift over how—and how soon—artificial general intelligence might emerge. At this year’s World Economic Forum in Davos, Google DeepMind chief Demis Hassabis agreed that today’s models are “nowhere near” human‑level intelligence and suggested “one or two more breakthroughs” are still needed, though he gave a 50% chance that AGI could arrive within a decade ynetnews. Hassabis did not, however, endorse LeCun’s description of LLMs as a dead end, instead casting them as powerful but incomplete building blocks ynetnews.

Others have been more openly bullish on the current LLM trajectory. Anthropic CEO Dario Amodei told Davos attendees that AI systems could effectively replace the work of all software developers within a year and eliminate half of white‑collar jobs within five years, implying that scaling and integrating today’s models may be enough to trigger rapid economic upheaval ynetnews +1. Elon Musk has publicly sided with Hassabis in his feud with LeCun over AGI timelines, underscoring a divide between those who see current architectures as fundamentally limited and those who expect transformative gains from further scaling businessinsider +1. Major firms including OpenAI, Anthropic and Meta continue to invest heavily in LLM‑centered products even as they quietly fund research into multimodal and world‑model variants ynetnews +1.

The Bigger Picture

The clash over LeCun’s “dead end” warning is less about whether AI is advancing—few dispute that—and more about what kind of systems will ultimately matter. If he is right, the current LLM boom could prove a costly detour that delays investment in architectures capable of robust reasoning and action in the physical world. If his critics are closer to the mark, world‑model work like AMI Labs’ may end up as one ingredient in a broader LLM‑based stack rather than a wholesale replacement. Either way, the debate is already reshaping research roadmaps, investor bets and regulatory discussions, as governments and companies decide whether to double down on language‑first AI or hedge toward a more embodied, multimodal future.