Google DeepMind CEO Demis Hassabis told the World Economic Forum in Davos on January 20 that today’s large language models, including systems like ChatGPT, are not capable of producing genuine scientific breakthroughs on their own, because they lack what researchers call world models.
Words without understanding
Hassabis’s argument is not that current models are weak. It is that they do not understand the world in terms of cause and effect, physical laws and the structure of reality. Without that grounding, he said, a model can assemble language convincingly but cannot discover anything new. Real discovery, in his view, requires a system that grasps why something happens, not only what tends to be said about it.
The position puts him at odds with the strategy most associated with OpenAI, which has emphasized making models larger and feeding them more data. Hassabis suggested that approach is running into a fundamental limit: scale alone is not producing the kind of intelligence that drives scientific progress.
DeepMind’s alternative
DeepMind is pursuing world models directly through projects such as Genie 3, a system that generates interactive three-dimensional environments. The idea is that a model which can simulate how an environment behaves learns something closer to physical and causal understanding than one trained only on text. Hassabis has argued that combining such capabilities with language models is a more promising route to systems that can reason about the real world.
AGI odds and the race with China
Asked about artificial general intelligence, Hassabis gave a measured estimate: roughly a 50 percent chance of AGI by around 2030, neither a certainty nor a distant dream. He also said AGI, when it arrives, should produce a fresh understanding of how the world works and could lead to major advances in science and medicine.
On competition, Hassabis said Chinese AI companies, including ByteDance, are technically perhaps only six months behind the leading Western labs. The open question, he said, is whether they can move beyond rapid replication to original innovation. Being six months behind on reproducing results, he noted, is different from being six months behind on discovery.
The remarks reflect a broader divide in the industry over where the next gains will come from. For Hassabis, the next phase of AI will be defined less by ever-larger language models and more by machines that understand the world they are describing.
