Hassabis, Demis

CEO and co-founder of DeepMind

Demis Hassabis: DeepMind - AI, Superintelligence & the Future of Humanity

Lex Fridman Podcast · Jul 1, 2022

• Demis Hassabis
• DeepMind
• AlphaGo and AlphaZero lineage
• MuZero
• AlphaFold and protein folding
• Turing test critique
• AI and creativity levels
• nuclear fusion plasma control
• quantum chemistry functionals
• virtual cell simulation
• consciousness and AI
• Fermi paradox and great filter
• simulation hypothesis

Core ideas

Hassabis opens by critiquing the Turing test as a rigorous benchmark, arguing Turing's 1950 paper was more a philosophical thought experiment than a formal specification (it never fixed the judge's expertise or time allotted), and proposes that AI progress should instead be measured by generalization across a very large number of tasks spanning modalities — though he notes such multi-task capability likely still gets expressed and communicated through language, since language is the primary tool humans use to generalize and explain understanding. He traces his own path into AI from chess (competitive junior player, master-level by age 12, second-highest-rated player in the world in his age group after Judit Polgár) and early programming on a ZX Spectrum and Commodore Amiga, then into professional game design (notably Theme Park and Black & White, where he says a reinforcement-learning-driven virtual pet responding to player behavior first showed him how compelling adaptive AI could be), and finally into DeepMind's founding thesis: that games are an ideal, efficient testbed for general learning algorithms because they offer clear reward signals, strong human benchmarks, and parallelizable simulation.

He walks through the AlphaGo lineage as a progression toward generality: AlphaGo (trained partly on human game data) → AlphaGo Zero (pure self-play from random initialization, no human data) → AlphaZero (generalized to any two-player board game) → MuZero (learns the environment's rules itself, extending to video games). He frames chess computer Deep Blue as impressive but narrow — "it couldn't even play tic-tac-toe" — arguing true intelligence requires generality and learning, not just hand-coded domain mastery, and credits Kasparov's post-loss versatility (able to do everything else a human can do) as more impressive to him than Deep Blue's raw chess strength. On creativity, Hassabis distinguishes three levels: interpolation (averaging seen examples), extrapolation (AlphaGo's famous "Move 37," a genuinely novel strategic idea within Go's existing rules), and true invention (inventing a game like Go itself) — arguing current AI systems, including DeepMind's, are capable of the first two but not yet the third, since specifying abstract, compositional design goals ("a game learnable in minutes but taking lifetimes to master") remains beyond current systems' ability to be given as a task.

Discussing AlphaFold, he explains the 50-year-old protein folding problem (predicting a protein's 3D structure, which determines its function, from its one-dimensional amino acid sequence) as a "Levinthal's paradox" of astronomical search space (roughly 10^300 possible foldings) that physics somehow resolves in milliseconds. He describes AlphaFold 2's key innovations — over 30 combined algorithmic components, hard-coded physical/evolutionary constraints that don't block learning, self-distillation to expand a training set of only ~150,000 known structures, and making the system fully end-to-end (amino acid sequence directly to 3D structure, without an intermediate distance-matrix step) — as reflecting a broader DeepMind pattern ("Alpha-X projects") of starting with hand-crafted structure and progressively letting learned, end-to-end systems absorb more of the pipeline. He notes AlphaFold has been used by over 500,000 researchers within about a year of release (citing the nuclear pore complex structure as a landmark result combining AlphaFold predictions with experimental data), and explains DeepMind's decision to open-source both AlphaFold and the MuJoCo physics engine as case-by-case judgments about maximizing benefit to the scientific community, while noting future biology tools may be released commercially, non-profit, or open-source depending on what best serves impact — with dual-use safety and bioethics consultation (30 external bioethicists consulted before AlphaFold's release) as an explicit consideration.

Hassabis discusses ongoing DeepMind science applications beyond biology: reinforcement learning for real-time control of tokamak fusion plasma (published in Nature, using a test reactor at EPFL Switzerland to hold plasma in specific shapes) and learning approximate density functionals to simulate quantum-mechanical electron behavior at a scale beyond traditional simulation (aimed at accelerating materials science, e.g. batteries or superconductors). He frames these as steps toward a long-term goal of building a full "virtual cell" simulation (a decade-plus ambition discussed for 20 years with biologist Paul Nurse) and argues AI may become "the perfect description language for biology" the way mathematics is for physics, since biological systems are too emergent and dynamic for elegant closed-form laws.

On speculative and philosophical topics, Hassabis states he does not believe in a literal simulation-hypothesis scenario (our world deliberately constructed by descendants), but does think information, not matter or energy, may be the most fundamental description of physical reality. Regarding Roger Penrose's view that consciousness requires quantum effects in the brain, Hassabis says mainstream neuroscience finds no evidence of quantum processes there and that his own bet is that brain function can eventually be approximated on classical (Turing-machine) computation. On extraterrestrial life, he states his personal view — based on reasoning about von Neumann probes, Dyson spheres, and the "loud" radio/TV signature a spacefaring civilization would likely produce — is that humanity is probably alone, finding arguments for a deliberately hidden or silent civilization ("zoo hypothesis") less persuasive than the possibility that advanced life is simply exceedingly rare; he identifies the evolution of multicellular life (via mitochondrial capture) as, in his estimation, the most plausible "great filter" bottleneck, following ideas from biologist Nick Lane.

On consciousness and AI, Hassabis argues intelligence and consciousness are likely "doubly dissociable" (each can exist without the other, citing conscious-seeming but unintelligent animals versus intelligent but plainly non-conscious systems like AlphaGo), proposes that our confidence in other humans' sentience partly rests on shared biological "substrate" (a point from his debates with philosopher Daniel Dennett) that AI systems will never share, and states his current view that language models are "not ready" for wide deployment as anything beyond tools, given unresolved interpretability and safety questions — while advocating, as a matter of current DeepMind design philosophy, that AI systems should not be built to simulate or claim sentience prematurely, and should always disclose that they are AI. Asked about the risk that concentrated control over transformative AI could corrupt its builders, he says he tries to stay grounded via longstanding friendships, a deliberately multidisciplinary and humbling work life, and surrounding himself with ethically grounded colleagues, and states his view that AI's benefits should ultimately belong to humanity broadly rather than any single organization.

He closes with personal reflections: a night-owl routine built around collaborative meetings during the day and solitary deep-thinking/writing sessions from roughly 10pm to the early morning hours; advice to young people to identify both their genuine passions and their distinctive strengths through active exploration; and, asked for the meaning of life, he answers that gaining and understanding knowledge is the closest thing to an intrinsic virtue he can identify, expressing a sense that the universe's apparent structure (its comprehensibility, the surprising fact that computation is physically realizable at all) feels almost suspiciously well-suited to the pursuit of knowledge. Asked what single question he'd ask a superintelligent AGI, he says he would ask for the true nature of reality.

Excerpts