Hassabis disputes LeCun on LLMs
DeepMind CEO Demis Hassabis publicly challenged Yann LeCun’s claim that large language models are a ‘dead end,’ arguing that scaling laws still apply and that a small number of breakthroughs — for example in memory — could keep LLMs central to AGI progress. The exchange was posted on social media as part of an ongoing public debate about research direction in AI. (x.com)
Large language models are systems that learn by predicting the next word, and Demis Hassabis said that approach is still improving fast enough to remain central to artificial general intelligence. (x.com) Hassabis made the case in a public exchange with Yann LeCun, the Meta chief artificial intelligence scientist who has argued that scaling up large language models will not reach human-level intelligence. A recent Big Technology interview clip summarized LeCun’s position as: “We Won’t Reach AGI By Scaling Up LLMs.” (x.com) (youtube.com) The technical split is straightforward. LeCun says next-word prediction cannot by itself deliver planning, persistent memory, common-sense physics, and reasoning, while Hassabis said a small number of additions, including better memory, could extend the same model family much further. (youtube.com) (x.com) That argument lands in a field that has spent six years measuring “scaling laws,” the regular pattern that language-model performance improves as model size, data, and computing power rise. OpenAI’s 2020 paper reported power-law gains across more than seven orders of magnitude, and that result became a core justification for building ever-larger models. (arxiv.org) (openai.com) Memory is the concrete bottleneck both sides keep returning to. Google DeepMind has published work on long-range memory in the Compressive Transformer and later work on memory consolidation for long-context video understanding, both aimed at helping transformer systems keep useful information over longer spans. (deepmind.google 1) (deepmind.google 2) LeCun’s side of the debate is tied to “world models,” systems meant to learn how the world changes, more like an internal simulator than a chatbot. Brown University described his April 1, 2026 lecture this way: world models would learn from video, audio, and other sensory data that large language model developers often leave aside, and LeCun sees the current large language model path as a dead end. (brown.edu) Hassabis has not argued that text alone is enough. Google DeepMind’s own research agenda spans multimodal systems, agents, planning, video models, and memory-heavy architectures, which fits his narrower claim that language models may remain the core scaffold if a few missing pieces are solved. (deepmind.google 1) (deepmind.google 2) The dispute is also about where companies place billions of dollars in computing and research talent. If LeCun is right, the next leap comes from architectures built around perception and world simulation; if Hassabis is right, the current large language model stack still has room to run. (openai.com) (brown.edu) (x.com) For now, the public fight is less about whether large language models work than about what they are missing. Hassabis said those gaps may be narrow enough to close; LeCun said they point to a different road. (x.com) (youtube.com)
