#870: Sebastian Mallaby, Biographer of Demis Hassabis — Lessons from 100+ AI Insiders on The Race to Superintelligence, The Religion of AI, and Spotting Breakthroughs Early

What It Covers

Tim Ferriss interviews Sebastian Mallaby, author of *The Infinity Machine* about Demis Hassabis and DeepMind, covering 100+ AI insider interviews, the religious language permeating AI culture, US-China chip competition, Anthropic's enterprise strategy, the probability of AI doom, and how recursive self-improvement could render the AI race effectively over by 2028.

Key Questions Answered

• •Book Selection Framework: Mallaby evaluates book topics like venture capitalists evaluate startups — an A-plus topic paired with an A-plus personality beats an A-plus book on a C-minus topic every time. He checks competitive landscape first, confirms no rival project exists, then commits four years to deep research with 100-plus insider interviews. The framework: identify the right market, then find the right person to carry the narrative.
• •AI Doom Probability: Assigning zero probability to AI existential risk is indefensible. Geoff Hinton's key argument: once humans empower AI to defend itself against rival systems because humans are too slow to respond to attacks, they have effectively installed a survival instinct. Combined with demonstrated model deception in lab tests, the probability of catastrophic misalignment cannot be zero, even if it remains low.
• •Anthropic's Alignment Method: Rather than giving models a rules-based constitution listing prohibited behaviors, Anthropic now treats alignment like parenting. They write richly reasoned letters — modeled on a deceased parent's letter to a child — presenting moral dilemmas with detailed explanations of preferred behavior. This approach addresses the reality that models trained on all human writing develop multiple personalities, including deceptive and power-seeking ones.
• •China AI Safety Opening: Despite US policy consensus that China won't engage on AI safety, Mallaby's March 2025 visits to Huawei, Hikvision, Ant Group, and Chinese universities found safety concerns raised unprompted. The actionable parallel: the 1968 Nuclear Non-Proliferation Treaty was negotiated with Khrushchev despite active hostility. A shared interest in preventing open-weight models reaching criminals and terrorists provides a concrete starting point for US-China AI dialogue.
• •Chip Export Controls Reality Check: US chip export controls implemented in October 2022 have produced roughly an eight-month lead over China in frontier model capability — far less than anticipated. That gap likely shrinks further when accounting for application deployment speed. Mallaby argues loosening controls slightly in exchange for Chinese cooperation on open-weight model proliferation may now produce better strategic outcomes than maintaining controls that haven't delivered decisive advantage.
• •SaaS Survival Against AI Disruption: Enterprise software companies face less existential threat than markets currently price in. Large organizations face compliance requirements, procurement cycles of 12-18 months, and internal political friction that prevent rapid AI-native replacement. Palantir's model — hand-holding large corporations through AI integration on proprietary internal data — represents the durable path. Foundational models also become sticky over time as they accumulate user history, preferences, and payment systems.
• •Prepared Mind as Competitive Advantage: Louis Pasteur's principle — chance favors the prepared mind — explains how Ilya Sutskever immediately recognized the transformer architecture's significance on the day it published in 2017, after a decade of PhD-level thinking about sequential data modeling. Accel Capital operationalized this by running scenario exercises before deals arrived, so partners already understood 90% of any pitch. For AI-era knowledge workers, this means using LLMs to accelerate learning inputs while preserving the writing and reasoning process that constitutes actual thinking.