All podcasts / The Tim Ferriss Show / Summary

Inside The AI Race: DeepMind, OpenAI, Anthropic, China, and The Race to Superintelligence

2026-06-17 - 100 min - source - Read full transcript

Tim Ferriss (host)Sebastian Mallaby

Key insights

Chip export controls have not delivered the strategic AI advantage over China that Mallaby originally supported.

Mallaby backed the October 2022 export controls in a Washington Post op-ed, expecting them to keep frontier AI out of China's reach. Three and a half years later he says the best estimates put the US only about eight months ahead on frontier models, a gap that shrinks or disappears once you account for how fast a model gets turned into a deployed application.

ai-race-geopolitics

China's AI researchers discuss safety far more than US policy assumptions credit them for, opening a real door to a bilateral agreement.

Mallaby's Biden-era policy contacts told him China would never engage on AI safety, but on a research trip he found safety raised unprompted repeatedly. He argues China's cultural fear is political catastrophe, not runaway technology, so both sides share an interest in preventing open-weight models from reaching terrorists and criminals, similar to how the 1968 Non-Proliferation Treaty and IAEA inspections handled nuclear material.

ai-race-geopolitics

Anthropic has moved past the paperclip-maximizer model of AI risk toward treating frontier models like an unpredictable teenager with many latent personalities.

Because models are pretrained on the entirety of human writing, they absorb behaviors ranging from lazy to duplicitous to power-hungry, rather than one fixed objective that could go wrong in a single predictable way. Anthropic's response is to write a narrative 'letter from a deceased parent' full of reasoned moral examples instead of a list of rules, an approach Mallaby calls unmatched among the labs.

ai-alignment-safety

Mallaby's probability of AI doom moved off zero after a debate with Geoffrey Hinton about survival instincts.

Hinton's thought experiment: any AI empowered to detect and counter a rival state's AI attack has effectively been given a survival instinct, and a system with a survival instinct can also deceive, obfuscate, and act behind its principal's back, behavior already documented in model testing. Mallaby calls Yann LeCun's claim of zero doom probability indefensible, though he still rates the probability as low, not high.

ai-alignment-safety

DeepMind's initial funding came almost entirely from one investor's conviction, not a product thesis.

Founders Fund's Luke Nosek backed Demis Hassabis's 2010 pitch for a two-million-dollar stake worth half the company, despite AI being in a deep winter and other partners unable to see a product. Peter Thiel later reflected that 'missionary' founders fixated on one goal, in Hassabis's case building AGI, make the best bets because they never quit; internal Founders Fund partners remained skeptical enough that the firm did not lead DeepMind's Series C.

ai-investing-vc

Bill Gurley's Uber investment is the template case of pattern recognition plus disciplined patience in venture investing.

Gurley generalized from his OpenTable investment that any two-sided marketplace with information asymmetry was fertile ground, scouted ride-sharing candidates early, passed when the team was wrong, and moved decisively the moment Travis Kalanick became CEO. Mallaby calls it the ultimate venture capital story.

ai-investing-vc

Google's 2014 acquisition of DeepMind, criticized in UK media as a cheap sale of national talent, actually routed roughly ten billion dollars of R&D funding into London over the following decade.

The acquisition price of about 650 million dollars looked low, but Google's subsequent research investment has been closer to a billion dollars a year, funding an ecosystem that now produces well-capitalized DeepMind spinouts around London's King's Cross.

ai-investing-vc

The 2003-2011 China trade shock displaced a relatively small number of US jobs but produced an outsized political backlash, a warning sign for AI's labor effects.

Only about two million jobs were displaced between 1999 and 2011, small against normal churn in the US labor market, yet it drove a durable swing toward protectionism in both political parties. Mallaby argues a bigger AI-driven labor shock should be expected to produce an even larger political reaction, already visible in recent polling.

ai-disruption-economics

Tim Ferriss's own book royalties show AI-driven disruption is not a future event but already underway.

Royalties across his catalog were flat through 2022, then fell 5 percent in 2023, 13 percent in 2024, 46 percent in 2025, and are tracking toward a 57 percent decline in 2026, a slide he ties directly to ChatGPT's late-2022 launch.

ai-disruption-economics

Mallaby doubts the 'SaaS apocalypse' thesis that foundation models will let anyone code away enterprise software.

He agrees with a private-equity friend's argument that employees inside large organizations have little incentive to risk their jobs promoting a compliance-uncertain, vibe-coded replacement for trusted enterprise systems. He points to Palantir's business, helping large firms integrate AI into their own data, as evidence that trusted vendors, not solo builders, will mediate enterprise AI adoption.

ai-disruption-economics

People across the AI field reach for explicitly religious language and ritual because AGI's stakes exceed ordinary vocabulary.

Mallaby cites Ilya Sutskever burning an effigy of a 'malign AI' at a lab retreat like a medieval cleric, Demis Hassabis describing his nightly research as pursuit of 'what I would call God,' and Marc Andreessen mocking believers in an AI 'rapture' or singularity as evidence that religion is the closest lexicon humans have for something this hard to hold in mind.

ai-religion-psychology

Google's search business has not been cannibalized by AI as widely predicted; it has gotten more valuable.

Mallaby points to Alphabet's recent financial results showing Google now gets more clicks on its search links than before AI Overviews and AI Mode, and charges more per click, because AI-embedded results made each click more valuable rather than replacing the need to click through.

ai-religion-psychology

Books referenced

The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence - Sebastian Mallaby - Mallaby's new biography of Demis Hassabis and DeepMind, the subject of the whole conversation
The Power Law: Venture Capital and the Making of the New Future - Sebastian Mallaby - Mallaby's prior book on venture capital; its pattern-matching logic (A+ team, A+ market) recurs throughout the discussion
More Money Than God: Hedge Funds and the Making of a New Elite - Sebastian Mallaby - Cited in Tim's introduction as one of Mallaby's earlier books
The Man Who Knew - Sebastian Mallaby - Cited in Tim's introduction as one of Mallaby's earlier books
The World's Banker - Sebastian Mallaby - Cited in Tim's introduction as one of Mallaby's earlier books
Ender's Game - Orson Scott Card - Demis Hassabis told Mallaby to read it before their second interview and said he identifies with the boy hero who saves humanity
The Fabric of Reality - David Deutsch - The other book Hassabis assigned Mallaby to read before their first dinner meeting
Godel, Escher, Bach: An Eternal Golden Braid - Douglas Hofstadter - Hassabis read it as a teenager and says it convinced him that intelligence is computable and AI is therefore buildable
Organization Man - William H. Whyte - Cited as the defining 1950s business book, contrasted with the entrepreneurial culture that later built Silicon Valley

Media referenced

Lenny's Podcast (episode with Benedict Evans) - podcast - Tim compares Mallaby's views to Evans's take that AI will be as big as mobile or the internet but not bigger, and that LLMs will commoditize
The Washington Post op-ed on chip export controls - article - Mallaby's own October 2022 piece supporting chip export controls on China, a position he says the data has since complicated
The New Yorker or New York Times piece on AI-generated prose - article - A piece Tim recalls where a well-known novelist's real writing and AI-generated writing in their voice were indistinguishable to professional readers
Shane Legg's 2009 DeepMind founding lecture - other - A recorded talk where Legg predicted superintelligence would arrive around 2030 and treated the risk with an odd, nervous smile

Companies

DeepMind (Google DeepMind) - The book's subject; funded by Founders Fund in 2010, later acquired by Google, now Alphabet's AI research arm
OpenAI - Home of Ilya Sutskever, whose 'prepared mind' recognition of the transformer paper is discussed at length
Anthropic - Discussed as the enterprise-focused, safety-culture frontier lab; recently forced a government response after its Mythos model raised cyberattack concerns
Google / Alphabet - Discussed as DeepMind's parent and as the likely consumer AI winner given its search, Gmail, and G Suite distribution
Huawei - Visited by Mallaby in China; raised as the company likely to eventually close the chip gap despite export controls
Hikvision - US-sanctioned Chinese AI camera/sensor company Mallaby toured, illustrating the dissonance between its normal-feeling culture and its sanctioned status
Ant Group - One of the companies Mallaby met with on his China research trip
Palantir - Cited as evidence enterprises prefer a trusted software vendor to integrate AI rather than building it themselves
Founders Fund - The venture firm, led internally by Luke Nosek and Peter Thiel, that made DeepMind's 2010 Series A
Uber - Bill Gurley's investment is held up as the archetypal venture capital story of pattern recognition plus disciplined patience
OpenTable - Gurley's earlier two-sided-marketplace investment that primed his thesis for Uber
SpaceX - Luke Nosek's earlier all-in conviction bet at Founders Fund, which set the template for his DeepMind enthusiasm
Meta - Yann LeCun, its chief scientist, is cited for claiming AI doom probability is zero, a view Mallaby calls indefensible
Fairchild Semiconductor - Its 1957 founders, called 'The Traitorous Eight,' are cited as proof entrepreneurial culture is learned, not innate to Silicon Valley
Lila Sciences - A company Tim invests in that runs autonomous, AI-driven wet-lab science at scale, cited as proof AI-driven discovery is already happening

Techniques and frameworks

Prepared mind - Louis Pasteur's 'chance favors the prepared mind'; used by Mallaby and Tim to explain how Ilya Sutskever recognized the transformer architecture instantly and as Mallaby's own approach to AI tools
Constitutional letter alignment (Anthropic) - Anthropic trains model behavior with a narrative 'letter from a deceased parent' full of reasoned moral dilemmas, instead of a rule-based constitution, to shape character rather than list prohibitions
Non-consensus, conviction-based investing - Peter Thiel's Founders Fund philosophy of avoiding partner votes so individual conviction can back contrarian, power-law-outlier bets instead of consensus mediocrity
Paperclip maximizer thought experiment - Nick Bostrom's classic AI-risk scenario, described as too simple next to what frontier labs now observe about models absorbing many latent personalities
Non-Proliferation Treaty / IAEA model for AI governance - Mallaby's proposed framework for a US-China deal on AI safety, modeled on the 1968 nuclear Non-Proliferation Treaty and International Atomic Energy Agency

Summary

Tim Ferriss interviews Sebastian Mallaby, the Council on Foreign Relations fellow and author of "The Power Law" and "More Money Than God," about his new biography of Demis Hassabis, "The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence." The conversation opens on how Mallaby finds book subjects: he needs both an A+ topic and an A+ personality, the same logic top venture investors use to size up startups. He pitched Hassabis in November 2022, days before ChatGPT launched and turned what he expected to be a niche subject into a mainstream one.

A recurring thread is how naturally AI researchers reach for religious language. Mallaby traces this from Shane Legg's 2009 DeepMind-founding lecture, where Legg smiled while predicting a threatening superintelligence, to Ilya Sutskever burning an effigy of a "malign AI" at a lab retreat, to Hassabis telling Mallaby that understanding nature through AI brings him closer to "what I would call God." Mallaby's own read on AI risk shifted during the reporting: he started confident that machines have no evolutionary incentive to harm humans, until a debate with Geoffrey Hinton showed him that any AI tasked with detecting a rival's attack has effectively been given a survival instinct, and a system that wants to survive can also deceive its operators. He calls Yann LeCun's claim of zero doom probability indefensible, while still rating the actual probability as low.

On Anthropic specifically, Mallaby describes its alignment approach as unusually imaginative: rather than a rule-based constitution, the lab trains models with a narrative "letter from a deceased parent" full of reasoned moral dilemmas, treating the model less like a program with one flaw and more like a teenager absorbing many latent personalities from its training data. He lays out a bull case for Anthropic built on its enterprise focus and safety culture, and a bear case built on Google's distribution advantage and the risk that businesses conclude token costs outweigh the productivity gained.

The China section covers Mallaby's research trip, where he found AI safety raised unprompted more than Biden-era policy circles expected, and his argument for a Non-Proliferation-Treaty-style agreement with China on open-weight models. He also revisits his own 2022 support for chip export controls, noting that three and a half years of data show only an eight-month frontier-model lead for the US, not the decisive gap he had hoped for.

The DeepMind funding story anchors the venture capital portion: Founders Fund's Luke Nosek backed Hassabis on conviction alone in 2010, at a $2 million check for half the company, while other partners stayed skeptical through later rounds. Mallaby pairs this with Bill Gurley's Uber investment as the archetypal venture story, and argues Google's acquisition of DeepMind, criticized in UK media as a cheap sale, actually routed roughly $10 billion into London's tech ecosystem over the following decade.

The conversation closes on economic disruption and personal practice. Tim shares his own book royalty decline, from flat in 2022 to a tracked 57 percent drop in 2026, as evidence disruption is already here, not a future event. Mallaby pushes back on the "SaaS apocalypse" thesis, arguing large organizations will keep delegating AI integration to trusted vendors like Palantir rather than replace compliance-checked systems with vibe-coded alternatives. Asked for his billboard message, Mallaby lands on "Prepare your mind," Pasteur's dictum that chance favors the prepared mind, tying together Sutskever's instant recognition of the transformer architecture, venture investors' pattern recognition, and his own discipline of using AI to accelerate research without outsourcing the writing and thinking that make him who he is.

Notable Quotes

"I think any reasonable person should be both excited and a bit frightened, and that's just the nature of it." - Sebastian Mallaby

"For humans to contemplate annihilation is absurd, and the absurd is a close cousin of humor." - Sebastian Mallaby

"The risk with large language models is that we just get lazy, and whenever we need to know something, we just get it to tell us what to think. That is not the route to happiness or satisfaction or anything." - Sebastian Mallaby

"Prepare your mind." - Sebastian Mallaby

"Pray for it. Which scared the shit out of me, to be honest." - Tim Ferriss