Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough

Hassabis expects today’s large models, reinforcement learning, and reasoning methods to remain part of AGI, but says continual learning, long-term reasoning, memory, and consistency remain open problems. He estimates his AGI timeline at around 2030 and says one or two major ideas may still be missing.

Hassabis traces today’s agent work to DeepMind’s game-playing research, including AlphaGo, and says search and reinforcement learning could improve foundation models. He sees continual learning as a key barrier to agents that can complete tasks independently. Smaller distilled models can make iteration faster and support private, local uses such as robotics, while frontier models remain important for capability.

Hassabis says a full virtual cell may be about 10 years away; DeepMind is starting with a virtual nucleus, while better live-cell imaging and learned simulators could help. He describes promising work in areas such as materials and mathematics, and says breakthroughs like AlphaFold often involve a huge search space, a clear objective, and enough data or simulation.

For science startups, Hassabis recommends combining AI with deep expertise in fields such as materials or medicine, where work involving the physical world may be more defensible. He advises founders to consider how AGI could arrive during a long research journey and build systems that can use specialized tools such as AlphaFold rather than putting every capability into one model.