Demis Hassabis has had one of the most extraordinary careers in technology. A chess prodigy who designed his first hit video game at 17, he went on to earn a PhD in cognitive neuroscience before co-founding DeepMind in 2010 with a single stated mission: solve intelligence.
His lab did things most researchers thought were decades away. AlphaGo beat a world champion at Go. AlphaFold cracked protein structure prediction, a 50-year grand challenge in biology. He gave AlphaFold away free to every scientist on earth. The work won him the 2024 Nobel Prize in Chemistry.
Today, Hassabis leads Google DeepMind, where he is building Gemini and pushing toward the same goal he set as a teenager: artificial general intelligence. In a live episode of How to Build the Future, he sat down with Y Combinator's Garry Tan to explain what is still missing, where agents are headed, and what founders should be building right now.
The missing pieces
Hassabis believes the current paradigm of large-scale pre-training, RLHF and chain-of-thought reasoning will form part of the final AGI architecture. But he identified specific gaps that remain. Continual learning, long-term reasoning and certain aspects of memory are still unsolved. So is consistency. Models can solve gold-medal problems from the International Mathematical Olympiad and then make elementary arithmetic errors in the next breath.
He put the odds at roughly 50/50 that one or two fundamental new ideas will be needed to close those gaps, rather than incremental scaling of existing techniques. Researchers at Google DeepMind are pursuing both paths simultaneously.
Memory is harder than it looks
Context windows have grown to a million tokens, but Hassabis argued this is still a brute-force approach. A million tokens store about 20 minutes of live video. To understand a person's day, let alone a month, something more selective is required. The brain manages this through the hippocampus, consolidating episodic memories during sleep and replaying the ones that matter. DeepMind borrowed this idea early on. Its first Atari-playing program, DQN, used experience replay drawn directly from neuroscience research.
The challenge is not storage. It is retrieval. Even with enormous context windows, the cost of finding the right piece of information at the right moment is non-trivial.
Agents are just getting started
Hassabis pushed back on the idea that agents are overhyped. He argued that an active system capable of solving problems on its own is a necessary step toward AGI, and that agent capabilities are still in their infancy. People are experimenting, but few have yet produced outputs that justify the effort. Nobody has used agents to create a AAA game that tops the app store charts, for example.
He predicted the full value of agents would emerge in the next six to 12 months, with individual operators achieving 1,000x efficiency before automation follows. The philosophy behind DeepMind's agent work, built on reinforcement learning and search stretching back to AlphaGo, remains embedded in Gemini today. Ideas such as Monte Carlo search and augmented reinforcement learning are now being re-examined at the scale of modern foundation models.
The jagged intelligence problem
Hassabis described a phenomenon he called jagged intelligence. Current systems can solve complex competition-level mathematics but stumble on questions a smart undergraduate would handle without difficulty. When he plays chess against these systems, they sometimes consider a move, recognise it as a blunder, and play it anyway because they cannot find an alternative. The thinking traces reveal basic reasoning errors sitting alongside sophisticated problem-solving.
He suggested the issue may relate to a lack of introspection about their own thought processes. The systems are still fairly simplistic in how they monitor and correct their own chains of reasoning. One or two key innovations in this area could make a significant difference.
Smaller models, bigger reach
Distillation is making smaller models far more capable than their size suggests. Hassabis said models a fraction of the size of frontier systems can now deliver 95% of the performance at a tenth of the cost. Google needs this. It has to serve search, YouTube, Maps and the Gemini app with low latency across billions of users. Flash-sized models make that possible.
He said no known limit exists to the distillation process. Models running on the edge offer additional benefits in privacy, security and speed, with local processing keeping personal data on-device and only delegating to the cloud when necessary.
What happens when inference gets cheap
Even if energy costs fall toward zero through breakthroughs in fusion or superconductors, Hassabis argued that inference will never truly be free. Physical chip production creates its own bottleneck. And whatever capacity becomes available will be consumed by swarms of agents working in parallel, or single agents thinking in multiple directions and ensembling their results. He invoked Jevons' paradox: as inference becomes more efficient and widely available, demand will expand to absorb it.
Gemma and the open model strategy
Google released Gemma, its open-weights model, which recorded 40 million downloads in two and a half weeks. Hassabis said the strategic logic is straightforward. Edge models deployed on Android devices, smart glasses and robotics hardware are vulnerable once shipped. Making them open allows a broader developer community to build on and improve them. The aim is to unify models at nano scale and make them fully accessible.
The AlphaFold pattern and what comes next
Hassabis outlined a specific pattern behind the AlphaFold breakthrough that he believes can be applied elsewhere. It requires three things: a massive combinatorial search space that cannot be solved by brute force, a clear objective function that enables hill climbing, and enough data or a simulator capable of generating in-distribution synthetic data. Drug discovery, he said, fits this pattern precisely. The goal is to find a compound that treats a disease without side effects, and AlphaFold has already demonstrated the ability to find needles in that kind of haystack.
He described a 10-year ambition to build a full working simulation of a cell, a virtual system that can be perturbed and produce outputs close enough to experimental results to be scientifically useful. Isomorphic Labs, his drug discovery company, is building out the adjacent biochemistry. The biggest obstacle is data. Imaging a live cell without killing it would convert the problem into a vision challenge and transform the field.
The Einstein test for AI discovery
Hassabis proposed a benchmark for genuine AI creativity that he called the Einstein test. Train a system with all the knowledge available up to 1901 and see whether it can independently arrive at the discoveries Einstein made in 1905, including special relativity. No AI system has yet passed this test, or anything close to it. He argued that true scientific discovery requires the ability to generate interesting hypotheses, not just solve existing problems, and to use analogical reasoning to move beyond the boundaries of current knowledge.
Advice for founders
Hassabis told founders to intercept where AI technology is heading and combine it with a deep technology area such as materials science or medicine. The combination, he said, creates a defensible position that pure software plays cannot match. Interdisciplinary founding teams with genuine expertise in both AI and the target domain will create the most value.
Related reading
- Google and Kaggle relaunch AI Agents Vibe course
- Google Maps and the Gemini chatbot that can plan your day. The results are good, the privacy element less so
- Google DeepMind releases AI model that understands text, images, video and audio together
He urged founders starting 10-year deep tech journeys to plan for AGI arriving in the middle of the process, currently estimated around 2030. That means building systems flexible enough to integrate with general-purpose AI when it arrives, rather than being made redundant by it.
His final point was personal. He said he would have worked on AI even if it had not succeeded, because it was the most consequential and interesting problem he could find. Founders, he argued, should feel the same way about whatever they choose to build.
