Fireside Chat, Interview
Demis Hassabis: We're Three Quarters of the Way to AGI
- Demis Hassabis identified AI as his primary career goal from age 15–16, deliberately structuring his education and early ventures to build toward this objective.
- His 1990s detour into game development served as a strategic pathway to master cutting-edge hardware (GPUs) and apply AI to complex economic simulations, exemplified by Theme Park, which sold over 10 million copies.
- Hassabis founded Elixir Studios immediately after college, aiming to "fund AI through the back door," though the company learned that timing is critical: technology must be five years ahead of the market, not 50.
- In 2009, DeepMind was established based on the conviction that combining deep learning (then niche) with reinforcement learning could scale to Artificial General Intelligence (AGI), despite academic skepticism regarding "strong AI."
- The original DeepMind mission statement defined a two-step roadmap: first, solve intelligence to build AGI; second, apply this capability to solve other global challenges.
- Hassabis predicts AGI will be achieved around 2030, aligning with the 20-year timeline the field originally projected in 2010.
- The company formally established an AI for Science division nearly a decade ago, shortly after the AlphaGo victory, waiting for algorithms to reach a level of generality suitable for real-world scientific problems.
- AlphaFold is cited as the pivotal moment that revolutionized biology by solving the 50-year-old protein folding challenge, allowing the industry to know protein structures accurately.
- Isomorphic Labs, a DeepMind spinout, aims to automate the subsequent drug discovery phase by designing compounds to bind to specific protein targets, aiming to reduce the typical 10-year drug development cycle to months or weeks.
- Hassabis envisions "world models" and AI-driven simulations as the foundation for new sciences in social domains (e.g., economics), enabling controlled experiments that are currently impossible in the physical world.
- He proposes that machine learning, not traditional mathematics, is the ideal descriptive language for complex, emergent biological systems characterized by weak signals and massive data volumes.
- Hassabis posits that information is the fundamental building block of the universe, potentially equivalent to matter and energy, and that AI represents the ultimate tool for organizing and understanding this information.
- He argues that classical Turing machines, specifically modern neural networks, are sufficient to model systems traditionally thought to require quantum computing, such as protein folding.
- Regarding consciousness, Hassabis suggests a two-step approach: first build a highly intelligent tool (AGI), then use it to analyze the nature of consciousness, noting that behavioral equivalence and substrate equivalence are key criteria.
- He believes a classical computer cannot compute everything, though he views the brain as an approximate Turing machine and remains skeptical about the immediate necessity of quantum hardware for modeling complex biological systems.
- Hassabis cites Kant's concept of the mind creating reality and Spinoza's deterministic view of the universe as philosophical pillars that inform his scientific inquiry into the nature of reality.
- His rapid-fire responses included selecting The Fabric of Reality by David Deutsch as the definitive book to read post-AGI and von Neumann as the ideal historical figure to partner with on a strategy game.