How to Build the Future: Demis Hassabis

In this interview, Demis Hassabis outlines what he believes remains missing on the path to AGI—continual learning, long-term memory, and more sophisticated reasoning—predicting AGI around 2030 while warning it may arrive mid-journey for many deep-tech ventures. He reflects on his path from chess prodigy to co-founding DeepMind, highlighting AlphaGo, AlphaFold, and the Nobel Prize, and argues that current techniques like pre-training and chain-of-thought are likely part of the final architecture, though one or two big ideas are still missing. Hassabis stresses that continual learning is a key unsolved blocker for agents, noting that context windows are a brute-force form of working memory and that reasoning models still overthink and lack introspection. He sees AI agents as "just getting started" and believes near-term wins will come from humans amplified to "1000x" rather than fully autonomous systems. He also discusses DeepMind’s strategy of keeping edge models open, the multimodal foundation of Gemini, and the promise of smaller models compressing frontier capability. Finally, he shares the long-term mission of using AI as "the ultimate tool for science," including Isomorphic Labs’ push toward a full working simulation of a cell—a goal currently limited by the lack of live-cell nanometer-resolution imaging.

• ▶ 0:00 Hassabis lists continual learning, long-term reasoning, and aspects of memory as still-unsolved ingredients required for AGI, predicting AGI around 2030 and warning it may arrive mid-journey for new deep-tech ventures.
• ▶ 0:40 Hassabis's path: chess prodigy, Theme Park designer at 17, PhD in cognitive neuroscience, then co-founding DeepMind in 2010 with the mission to "solve intelligence"; highlights AlphaGo, AlphaFold (free to all scientists), and the Nobel Prize in Chemistry.
• ▶ 2:09 Hassabis says current techniques (large-scale pre-training, RLHF, chain-of-thought) are very likely part of the final AGI architecture, but one or two big ideas may still be missing—estimating the odds at roughly 50/50 and working both paths at DeepMind.
• ▶ 3:31 Continual learning remains an unsolved challenge, with current fixes like “dream cycles” feeling like duct-tape patches rather than principled solutions.
• ▶ 4:14 Brain-inspired experience replay—already used in DeepMind’s 2013 DQN Atari work—shows how replaying important episodes during sleep-like states can help integrate new knowledge.
• ▶ 5:20 Context windows act like working memory, but stuffing everything into them is brute force and unsatisfying; for example, a million tokens only covers ~20 minutes of live video, making long-term understanding impractical.
• ▶ 6:10 DeepMind’s RL heritage—from Atari and AlphaGo—now directly fuels modern frontier-model advances like chain-of-thought reasoning, tree search, and “thinking modes.”
• ▶ 8:00 Distillation is a core strength: Google compresses frontier-model power into small models like Flash, delivering ~95% capability at ~1/10 the cost to serve billions of users.
• ▶ 9:35 Demis sees no fundamental ceiling on small-model intelligence, expecting frontier capabilities to be compressible into tiny, near-edge models within six months to a year.
• ▶ 12:48 Lack of continual learning is a key blocker for agents; they need to learn the specific user context, and cracking this is needed for "full general intelligence."
• ▶ 13:38 Reasoning paradigms are still simplistic and brute force; models overthink and get stuck in loops, with major innovation left in monitoring and interjecting in chain-of-thought.
• ▶ 15:00 "Jagged intelligence" is seen in IMO gold-level success alongside elementary math errors; the missing capacity is introspection about the model's own thought process.

• ▶ 15:29 Hassabis says AI agents are "just getting started," not overhyped, and are the path to AGI, but we're at the very beginning of integrating them into real workflows.

• ▶ 16:50 No "vibe-coded" game has yet become a breakout hit; AI speeds up prototyping dramatically, but craft, taste, and "soul" are still missing — he expects this to be delivered within the next 6–12 months.

• ▶ 18:08 The first major successes will come from humans operating at "1000x" using AI tools, not fully autonomous systems; true AI creativity remains an open challenge, as no system can yet invent a game like Go from a high-level description.

• ▶ 21:47 DeepMind's strategy is to keep edge models (for Android, glasses, robotics) fully open, like Gemma, because they run on user surfaces and are "vulnerable anyway" — this supports a competitive Western open-source stack and has driven ~40 million downloads in weeks.
• ▶ 22:47 Gemini was built multimodal from the start, which was harder initially but pays off long-term: it powers world models (Genie), robotics, Waymo, and future assistants that must understand physical context and intuitive physics.
• ▶ 24:11 Demis doubts inference will ever be truly free, citing Jevons paradox — demand (millions of agents in swarms, ensembled reasoning) will absorb all available compute; near-zero cost would require physical breakthroughs like fusion or better batteries.
• ▶ 25:13 Inference remains constrained and organizations will still face "rationing" on the inference side, even as models become more capable.
• ▶ 25:15 Inference must be used efficiently, making resource-conscious deployment an ongoing priority alongside raw performance.
• ▶ 25:18 Smaller models are getting "smarter and smarter," offering a promising trend to relieve inference burdens and improve accessibility.

• ▶ 25:24 A question asks about AlphaFold 3's expansion beyond proteins to a broad spectrum of biomolecules, and how close we are to modeling full cellular systems.

• ▶ 25:39 The speaker highlights Isomorphic Labs, spun out of DeepMind after AlphaFold 2, saying it is "going amazingly well" and is focused on more than just protein structure prediction.

• ▶ 25:47 Isomorphic's broader mission is to build out adjacent biochemistry and chemistry to design the right compounds with the right properties, with "big announcements" promised very soon.

• ▶ 26:07 The long-term goal is a full working simulation of a cell that can be perturbed and produce scientifically useful predictions, potentially skipping many experimental search steps and generating synthetic data.
• ▶ 26:40 DeepMind is starting with a virtual nucleus because it is relatively self-contained, though the key challenge is choosing the right subsystem and modeling the outside world through approximate inputs/outputs.
• ▶ 27:17 A major bottleneck is insufficient data: truly transformative progress would require imaging a live cell without killing it at nanometer resolution, turning the problem into a vision problem, but that technology does not yet exist.

• ▶ 28:31 Demis explains that all scientific domains are exciting, and his lifelong motivation for AI is to use it as "the ultimate tool for science" to advance understanding, discovery, medicine, and our understanding of the universe.

• ▶ 28:53 He shares his original two-step mission statement: first, solve intelligence by building AGI; second, use that intelligence to solve everything else.

• ▶ 29:06 The wording of the mission statement had to be changed over time because people would question the ambitious framing—"do you really..."—though the thought is cut off.

• ▶ 29:17 Hassabis describes "root node problems": fundamental scientific challenges that, once solved, unlock new avenues across many fields.
• ▶ 29:25 AlphaFold exemplifies this, with over 3 million researchers using it and pharma reporting that almost every future drug will involve it.
• ▶ 29:59 Other domains are at an "AlphaFold 1" moment—promising results but grand challenges unsolved—and major progress is expected in the next few years, especially in materials science and mathematics.
• ▶ 30:18 The "Promethean" feeling: AI's power in scientific discovery is compared to stealing fire from the gods—a transformative gift to humanity.
• ▶ 30:24 Hassabis adds a caution: with such Promethean power comes the obligation to be careful about how it is used, what it is used for, and especially to guard against misuse.

• ▶ 31:02 Demis advises startups to "intercept where the AI tech is going" and combine it with another deep technology area, calling the sweet spot fields like materials, medicine, and other hard sciences.

• ▶ 31:22 Working in the "world of atoms" with interdisciplinary teams is highly defensible because these areas are "pretty safe from just getting swarmed by whatever the next update is to the foundation models."

• ▶ 32:14 Founders need deep belief and conviction, understanding why their approach is different this time, plus expertise in both machine learning and the application domain — where huge impact can be made.

• ▶ 32:36 Before success, your idea looks impossible and people are arrayed against you; after success, it seems inevitable.
• ▶ 32:44 Nobody believes in a contrarian vision at first, so you must work on something you are genuinely passionate about.
• ▶ 32:52 Hassabis would have worked on AI no matter what, having decided from a very young age that it was his path.
• ▶ 32:57 Hassabis committed to AI young because it was the most consequential and interesting problem imaginable, a bet that could have failed if they were "50 years too early."
• ▶ 33:11 His commitment was unconditional: he would still work on AI even from a garage with no success, pivoting perhaps to academia rather than abandoning the field.
• ▶ 33:23 AlphaFold is presented as an example of a "spike" he pursued—a single, high-ambition bet within his lifelong AI dedication.
• ▶ 33:36 Demis outlines a repeatable pattern for AlphaFold-style breakthroughs: massive combinatorial search space, a clear objective function, and enough data or a simulator.
• ▶ 34:12 A well-defined objective (e.g., minimizing free energy or winning a game) is essential for hill-climbing toward rare solutions in the vast space.
• ▶ 34:45 Drug discovery fits the same pattern: the challenge is efficiently finding an existing, viable compound in an astronomically large chemical space.
• ▶ 35:33 Demis says AI systems capable of genuine scientific reasoning are close, citing efforts like Co-scientist and AlphaEvolve, but admits at ▶ 35:52 he has not yet seen a truly massive discovery made by AI.
• ▶ 36:17 True discovery requires creativity and analogical reasoning, not just pattern matching or extrapolation; the key benchmark, at ▶ 36:33, is whether an AI can generate an interesting new hypothesis rather than merely solve a posed one.
• ▶ 36:33 He sees AI solving a known Millennium Prize problem within a couple of years, but notes the far harder challenge is creating entirely new frontiers of science, like inventing a new set of Millennium problems.
• ▶ 36:57 Hassabis describes the hardest level of science—problems top mathematicians would deem worthy of a lifetime—and admits at ▶ 37:08 this is "another level harder" than anything AI has done, adding "I still don't think we know how to do that."
• ▶ 37:15 He insists this capability is not magical, and expresses confidence that "these systems will eventually be able to do that," suggesting we may be "missing one or two things."
• ▶ 37:22 He proposes an "Einstein test": give an AI the full physics knowledge of 1901 and see if it can independently derive Einstein's 1905 breakthroughs, including special relativity—concluding at ▶ 37:47 that passing this test would mean AI is on the verge of inventing truly novel science.
• ▶ 37:51 The final question targets technical audience members aspiring to build novel projects at the scale of DeepMind.
• ▶ 38:08 The interviewer expresses heartfelt gratitude to Demis Hassabis and the DeepMind team for their contributions.
• ▶ 38:13 The core question asks what he now knows about building at the frontier that he wishes he had known earlier.
• ▶ 38:21 Hassabis advises founders to tackle genuinely hard, meaningful problems: shallow and deep problems are "differently difficult," so you should put your "life force" into something that would truly be missing if you weren't there to push it.
• [38:19–39:22] Factor AGI timelines into long-term strategy: expect AGI around 2030, and since deep tech is a 10-year journey, AGI will likely arrive mid-project—so imagine that future world and build something AGI can leverage.
• ▶ 40:04 Future AI will be modular, not one giant brain: general systems like Gemini or Claude will orchestrate specialized tools like AlphaFold, because absorbing dense specialized knowledge into a generalist model causes regression and is information-inefficient.

• ▶ 0:00 AGI Gaps and Demis Hassabis's Background (0:00 - 3:31) - - Hassabis's unusual career and the current paradigm's unsolved pieces on the path to AGI.
• ▶ 3:31 Memory, Context, and Continual Learning (3:31 - 6:10) - - Continual learning, dream replay, context-window limits, working memory, and video token costs.
• ▶ 6:10 RL Heritage, Distillation, and Small Models (6:10 - 12:34) - - DeepMind's RL legacy, distillation's promise and limits, small-model applications, and a brief ad break.
• ▶ 12:34 Agent Limits, Reasoning, and Jagged Intelligence (12:34 - 15:27) - - Stateless agents, steering challenges, and gaps in chain-of-thought reasoning.
• ▶ 15:27 Agents, Creativity, and the 1000x Developer (15:27 - 20:18) - - Agent hype, vibe-coded games, human-AI collaboration, and AlphaGo-style creativity.
• ▶ 20:18 Open Models, Gemini Multimodality, and Inference Costs (20:18 - 25:18) - - Open-weight strategy, Gemini's multimodal future, and falling inference costs with energy bottlenecks.
• ▶ 25:18 AI for Science and Founder Advice (25:18 - 40:57) - - AlphaFold-style root-node science, Promethean risks, and advice for AI-for-science startups.