Yann LeCun, a leading figure in artificial intelligence research, has declared that large language models are a dead end. He predicts the technology will be obsolete within five years.
LeCun, who serves as Meta's chief AI scientist, argues that current LLMs lack the fundamental components needed for true intelligence. Instead of relying solely on text, he is pushing the field toward systems that learn from sensory experiences — sight, sound, touch — the way humans and animals do.
Why LeCun sees a dead end
LeCun's criticism centers on the limitations of models trained exclusively on text. While LLMs like ChatGPT and GPT-4 can generate fluent language, they have no understanding of the physical world. They cannot plan, reason causally, or build mental models of reality. LeCun has long argued that such systems are inherently brittle and will hit a ceiling.
“Large language models are a dead end,” he stated, according to remarks reported from recent discussions. He believes that without grounding in sensory data, LLMs will never achieve the robustness or generality required for advanced AI.
The sensory shift LeCun wants
LeCun advocates for a shift in AI research. He envisions models that integrate sensory inputs — video, audio, touch — to build internal representations of the world. This approach, sometimes called “world modeling,” is inspired by how animals learn: by interacting with their environment, not just by reading text.
His own work at Meta focuses on developing these world models. The goal is to create systems that can learn from raw sensory data, predict outcomes, and act in the real world. LeCun has said that such models, not LLMs, will be the foundation of the next generation of AI.
What this means for the AI landscape
LeCun's prediction carries weight given his stature. He co-won the Turing Award in 2018 for his work on convolutional neural networks, a key technology behind modern computer vision. His view puts him at odds with the current boom in LLMs, which has driven massive investment and public excitement.
Other researchers disagree. Many believe LLMs will continue to improve with scale and new techniques. But LeCun's argument forces a question: can text-only systems ever bridge the gap to general intelligence? His answer is no.
For now, the AI field remains split. LeCun has not said when he expects sensory-integrated models to surpass LLMs, but he has set a five-year deadline for LLMs' relevance. Whether the industry pivots or stays the course remains an open debate.
