Google’s Demis Hassabis says Meta’s Yann LeCun is ‘plain incorrect’ in fresh AI intelligence debate

Artificial Intelligence

A renewed debate over the nature of intelligence has erupted among some of the most influential figures in artificial intelligence, after comments made by Meta’s chief AI scientist Yann LeCun were publicly challenged by Google DeepMind CEO Demis Hassabis.

The disagreement began after Yann LeCun appeared on a podcast and argued that the idea of general intelligence is fundamentally flawed. According to LeCun, human intelligence itself is not truly general but instead highly specialised for interacting with the physical world. He suggested that what people often label as general intelligence is largely an illusion created by the limits of human imagination.

LeCun argued that humans appear versatile only because they are unaware of the vast range of problems they are incapable of solving. In his view, intelligence is always constrained by context, environment and evolutionary pressures, making the concept of a broadly general intelligence misleading.

A short clip from the podcast was later shared on X, where it quickly drew the attention of Demis Hassabis, who responded with a detailed and strongly worded rebuttal. Hassabis, who leads Google’s AI research arm and is also a Nobel laureate, publicly disagreed with LeCun’s position, describing it as “plain incorrect”.

In a lengthy post, Hassabis said LeCun was conflating two very different ideas: general intelligence and universal intelligence. He argued that while no system can be universally optimal across all possible tasks, that limitation does not negate the existence of general intelligence.

Hassabis emphasised that the human brain remains the most complex and capable system known, and that it is, by any reasonable definition, highly general. He acknowledged that practical systems must always show some degree of specialisation due to finite time, memory and data, referencing the well-known no free lunch theorem. However, he argued that this does not undermine the theoretical generality of intelligence.

According to Hassabis, general intelligence should be understood in a computational sense. He explained that architectures capable of learning any computable function, given sufficient resources, qualify as general. In this context, he said both the human brain and modern AI foundation models can be viewed as approximate Turing machines, systems that are theoretically capable of learning a wide range of tasks rather than being locked into narrow domains.

Hassabis also pushed back on LeCun’s use of chess as an example of specialised skill. While LeCun has previously argued that excellence in games like chess does not imply general intelligence, Hassabis countered that the very invention of chess, along with broader human achievements, is evidence of general cognitive capability.

He pointed out that humans have created not just abstract games but entire fields of science, engineering and technology, from modern physics to commercial aviation. Using chess grandmasters as an example, Hassabis noted that while humans may not achieve perfect optimisation due to biological constraints, the level of reasoning and creativity displayed is remarkable given that human brains evolved for survival tasks like hunting and gathering.

The exchange highlights a deeper philosophical divide within the AI research community. LeCun has long argued for more grounded, world-model-based approaches to intelligence and has been sceptical of claims surrounding artificial general intelligence. Hassabis, by contrast, has consistently maintained that general intelligence is both a valid concept and an achievable long-term goal for AI systems.