Demis Hassabis on Building DeepMind, AlphaFold, and the Final Stretch to AGI
In a Nutshell
Demis Hassabis recounts his path from chess prodigy and game developer (Theme Park, Elixir Studios) to founding DeepMind in 2009, leveraging games, neuroscience, and deep reinforcement learning to pioneer AI breakthroughs like AlphaGo and AlphaFold, which solved the 50-year protein folding challenge and now fuel Isomorphic Labs' drug discovery revolution. DeepMind's mission—solve AGI first, then apply it to science, simulations (e.g., GraphCast weather, virtual cells), and new fields—views information as the universe's fundamental building block, enabling accurate world models and potentially extracting equations from complex systems like biology. Hassabis predicts AGI by 2030, names Von Neumann as top historical scientist ally in strategy games, and cites Kant, Spinoza, and David Deutsch's *The Fabric of Reality* as key inspirations for understanding reality through AI.
These notes were generated by AI and may contain inaccuracies.
Host introduces Demis Hassabis as an original thinker, founder, visionary in AI, true believer, and scientist. Conversation begins with early days, DeepMind, then science and questions.
Demis was a chess prodigy, founder of a gaming company, neuroscientist, founder of DeepMind, now leader of a major company. These seem unrelated but share a common thread: wanted to do AI since teenage years, decided it was the most important and interesting thing. From age 15-16, planned to build a company like DeepMind, chose studies and activities to help. Detour into games in the 90s because cutting-edge technology was there, especially graphics and hardware like first GPUs designed for graphics engines, used them in late 90s. All games for Bullfrog and own company Elixir Studios involved AI as main gameplay component.
Most well-known game: Theme Park at age 17, simulation of amusement park with thousands of people entering, playing rides, buying from shops; economics AI model underneath, one of first like Sim City. Sold 10 million+ copies; delighted people interacting with AI reinforced commitment to career in AI. Neuroscience for inspiration from brain: algorithmic ideas. Brought all together for DeepMind when timing right; used games as early proving ground for AI ideas.
Room full of founders; Demis founded twice. Started Elixir Studios straight out of college after working at Bullfrog Productions, legendary UK/Europe game studio. Wanted to push AI, funded AI through games, combine with cutting-edge creativity, still relevant for blue-sky research today.
Biggest lesson: be 5 years ahead of time, not 50. Tried Republic at Elixir: simulated whole country, overthrow dictator in various ways, living breathing cities on late 90s Pentium PC; graphics and AI for million people on home PC, too ambitious, caused issues. Lesson: ahead of time but not too far; if 50 years ahead, can't succeed; when obvious to everyone, too late.
In 2009 decided there would be AGI, maybe 10 years ahead. Convinced first brilliant talents (high-caliber early team) despite seeming sci-fi.
Thought 5 years ahead, maybe 10. Deep learning just invented by Jeff Hinton et al. in academia, few realized importance. Knew reinforcement learning; huge progress combining with deep learning (rarely mixed outside toy problems, siloed). GPUs useful (now use TPUs). From PhD/postdoc and team (computational neuroscientists): brain principles useful, including reinforcement learning scaling to AGI. Felt like keepers of secret; academia/industry didn't believe big progress possible, rolled eyes at AGI/strong AI (tried in 90s, failed). Postdoc at MIT center for expert systems/logic languages (seemed archaic); convinced even more because would fail differently/originally if unsuccessful.
Would spend life on AI regardless; turned out better than optimistic predictions, still within 2010 forecast (20-year mission, on track). Original DeepMind mission: Step 1 solve intelligence (build AGI). Step 2 use to solve everything else. Most important technology ever, interesting artifact, tool for science, understand own minds (consciousness, dreaming, creativity). Missing analysis tool like AI for controlled experiments comparing systems.
Personal passion: advance science, medicine, world understanding via meta approach—build ultimate tool, then use for breakthroughs like AlphaFold.
Always at heart of DeepMind; AI for science group/division led by Pushmeet Kohli, started nearly decade ago, day after AlphaGo match in Seoul (10 years ago). Waited for algorithms powerful/general enough; cracking Go was signal to apply to real-world scientific challenges (cure diseases, healthier lifespans, medicine; then materials science, environment, energy).
Already had moment with AlphaFold: 50-year grand challenge, protein folding/3D structures key for medicines/understanding biology (one part of drug discovery).
Isomorphic Labs (latest spin-out): build adjacent tech in biochemistry/chemistry to design compounds fitting/binding right protein part. Know protein shape/surface/target; need compound that binds strongly there, not elsewhere (avoid toxicity). Dream: 99% exploration in silicon, wet lab only for validation. Reduce drug discovery from average 10 years to months/weeks/days; all diseases in reach, personalized medicine (variations off base medicines). Whole area revolutionized in next few years.
AI will create new sciences:
- Understanding/analysis of AI systems: engineering science; complex as human mind/brain, need study (mech interp part, more needed).
- AI unlocks new sciences via simulations (loves simulations, all games were). Addresses social sciences (economics, humanistic) hard for controlled studies (emergent like biology). Can't repeat experiments (e.g., raise interest rates 0.5%, real-world only). Accurate simulators enable rigorous sampling, better decisions in uncertain domains.
Work on learning simulators where math unknown/too complex (can't write direct simulator). Done with weather: GraphCast/WeatherNext, most accurate weather simulator, faster than meteorologists'.
Biology: virtual cell (dynamical emergent). ML perfect for biology like math for physics: loads of weak signals/correlations, tons of data humans can't analyze, causalities in mass data. Until now, math couldn't (too complex for top mathematicians, lacks expressive power for emergent dynamical systems). Also messiness/stochastic nature.
Once learn simulators, extract explicit equations (sample as many times as want); maybe as fundamental as Maxwell's for emergent systems.
Theory: basic building block of universe is information (E=mc²: energy/matter equivalent; info has equivalency). Organization of matter/structure (biology resisting entropy) = information processing systems. Convert matter/energy/info; info most fundamental (opposite classic 1920s physics view). Evidence supports; AI profound: organizes/understands info, constructs informational objects; deeply connected via info processing lens.
Classical Turing machine computes everything? Refer to as Turing's champions; Alan Turing hero, profound result: everything computable by simple machine description. Brains likely approximate Turing machines. Connection to quantum interesting.
AlphaGo/AlphaFold show classical (modern neural net) models quantum systems (protein folding: small particles, quantum effects like water bonds)—approximate optimal on classical.
Many things thought to need quantum might be modelable classically if right approach.
AI as tool (telescope/microscope/astrolabe). When modeling anything (even if not quantum), stops being tool?
Build incredibly intelligent/precise tool first (already profound), then next rubric (autonomous/agentic—now in agent era). Further: agency? Conscious? Address as second step, use tool to help; understand own brains/minds, define consciousness precisely.
No addition beyond thousands of years philosophy. Necessary (not sufficient): self-awareness, self/other, continuity over time.
Open question; discussed with philosophers like Daniel Dennett. Behavioral: does it behave consciously? (Near-AGI AI might.) Why think each other conscious: behave like it + same substrate. No substrate equivalence with artificial; hard to close gap (behavioral vs. experiential). Post-AGI ways possible, out of scope.
Brilliant. Uh so we're going to open the room to questions in just a moment. Get your questions ready. Uh but you brought
You've mentioned Kant and Spinoza as two of your favorite philosophers. Kant is deontological, highly duty-driven. Spinoza has a deterministic view of the universe. How do you connect those two beliefs and your thinking on how the world works?
The reason I like those two is that Kant's statements about the mind creates reality—that's basically true. That's another reason to study the mind and how the brain works. I'm interested ultimately in the nature of reality, so we have to understand how the mind is interpreting that. That's what I took from Kant.
Spinoza is more about the spiritual dimension: if you're trying to understand the universe using science as the tool, you're understanding some deep mystery about how the universe works in a deep way. That's what we're doing when we work on AI and build these tools—we're reading the language of the universe.
Beautiful way to say what you do every day. Demis: scientist or philosopher?
AGI Timeline
Over/under on year of AGI distribution, or reject premise?
- I've been pretty consistent about that.
Must-Read for AGI Achievement
Must-read book, poem, or paper for when we achieve AGI?
My favorite book is The Fabric of Reality by David Deutsch. I'd hope to answer the questions in that book with the AGI. That's my post-AGI work.
Proudest DeepMind Moment
Proudest moments so far in DeepMind?
AlphaFold.
If you were engaged in a high-stakes turn-based strategy game like Civ or Polytopia, and you could select one scientist from history—Einstein, Turing, Newton—who would you select to be on your team?
Von Neumann. He's a game theorist, and the best.
Well, Demis, you do it all. Thank you so much for being with us. Please join me in thanking.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
