Demis Hassabis and the Race to Build AI with Sebastian Mallaby

What It Covers

Sebastian Mallaby, author of *The Infinity Machine*, traces Demis Hassabis's founding of DeepMind in 2010 through the current multi-lab AI race. The conversation covers Hassabis's early conviction about AGI, DeepMind's Google acquisition for $650M, AlphaFold's protein-folding breakthrough, and why the safety-first vision has collapsed into a fragmented global competition.

Key Questions Answered

• •Early conviction as competitive advantage: Hassabis formed his AGI conviction at age 18 after reading Gödel, Escher, Bach — seventeen years before DeepMind's founding and nineteen years before convolutional neural networks reliably recognized images. His 2010 business plan predicted AI surpassing human intelligence around 2030. Mallaby argues this extreme early belief, not capital or timing, was the primary accelerant that pushed the field forward faster than it would have moved otherwise.
• •Acquisition strategy over independence: Hassabis sold DeepMind to Google in January 2014 for $650M, but the sale price was secondary — he negotiated nearly $1B annually in research budget with operational independence in London. For capital-intensive deep research, acquisition from a patient, wealthy acquirer can outperform venture funding. Google treated DeepMind as an insurance policy against AI disrupting search, funding it for years without demanding near-term commercial returns.
• •Race dynamics undermine safety intentions: Every major AI lab publicly endorses safety while privately reasoning that winning the race first is the safest outcome — because they trust themselves more than rivals. This mutual logic produces maximum acceleration with minimum coordination. Mallaby concludes that private-sector safety commitments are structurally ineffective when six-plus US labs, multiple Chinese labs, and European players like Mistral are all competing simultaneously with deep personal animosities between founders.
• •Government regulation as the only viable safety mechanism: With too many competing labs across too many jurisdictions for voluntary coordination, Mallaby argues only two governments — the US and China — have the leverage to impose binding safety standards, vetting processes, and deployment definitions. Industry self-regulation fails game-theoretically when any single defector gains competitive advantage. Hassabis's pivot from leading Google DeepMind's research to serving as Alphabet's chief scientist signals a personal shift toward influencing governance rather than building capability.
• •Institutional bottlenecks cap individual productivity gains: Even when AI doubles individual developer output, organizational throughput increases far less — perhaps 10% — because product, legal, and compliance functions remain unchanged bottlenecks. Mallaby draws a parallel to electrification: factories took 30-40 years to redesign layouts after electricity arrived because early adopters simply swapped power sources without restructuring workflows. Meaningful institutional productivity gains from AI likely require several years of deliberate bottleneck elimination, not just tool adoption.
• •Open-weights models differ fundamentally from open-source code: The traditional open-source safety argument — many eyes make shallow bugs — breaks down with trillion-parameter models whose weights are mathematically opaque, not human-readable. Additionally, the iterative attacker-defender dynamic that hardened traditional software assumes small, contained exploits. A criminal group with a capable open-weights cyber model can simultaneously attack thousands of targets, making the cost of each iteration catastrophically higher than in conventional software security cycles.