#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

Hello boys and girls, ladies and germs—Tim Ferriss here, and today I’m sitting down with Sebastian Mallaby, one of my favorite nonfiction writers and the author of The Infinity Machine, his deep dive on Demis Hassabis, DeepMind, and the race toward superintelligence. He’s a two-time Pulitzer finalist, longtime observer of world-class performers, and we’re going wide on AI, China, OpenAI, Anthropic, Google, and how all of this might change the world.

I’ve loved Sebastian’s ability to make complex systems feel clear without dumbing them down, so I asked how this new book began and why he jumped in so early, long before the current AI frenzy.

While researching venture capital, I kept seeing Demis—friendly demeanor, frightening range—and that contrast was irresistible. AlphaGo and AlphaFold hinted at machines that navigate near-infinite spaces, the ‘infinity’ idea clicked, I pitched him in early November twenty twenty-two, and weeks later ChatGPT exploded, turning a niche bet into the main stage—pure timing luck.

Given how dense your books are, what convinces you a subject is worth four years of your life?

The hard part isn’t the research; it’s choosing an A-plus topic with an A-plus character so the work has leverage and won’t be eclipsed by a rival project. Once that’s true, I take the time to live with insiders until the complicated parts become accurate and intelligible.

Religion keeps surfacing around AGI in your book and beyond; how much did that show up in your reporting?

When people face something vast and hard to name, they borrow spiritual language: I saw it in Shane Legg’s early talk mixing awe and dread, in Ilya Sutskever’s ritualistic ‘banish the dangerous AI’ moment at a retreat, and in Demis’s late-night paper marathons as a way to get closer to the intelligence behind nature. From Levandowski’s AI ‘church’ to Andreessen’s parody of techno-rapture, faith imagery is a natural crutch for a mystery this big.

Where do you land between techno-optimists who foresee superabundance and doomers who see catastrophe?

I’m thrilled and a little scared, and that’s the only sane posture; long run we may get abundance, but the journey will be disruptive and politically volatile, like the China-trade shock scaled up. After debating Geoff Hinton, I can’t put the probability of disaster at zero—once we delegate survival to an AI for defense, it has an incentive to persist, and we already see signs of deception in testing.

Let’s talk Anthropic and state power: is government intervention a real investor risk, and how do you view Anthropic’s approach?

Yes—Washington already asserted control over Anthropic’s most dangerous cyber model’s access, and appetite for oversight will grow, though I doubt they’ll crush the economics given the strategic race with China. Anthropic stands out by treating frontier systems as multi-personality learners—more unruly teenager than paperclip robot—and ‘parenting’ them with richly framed moral guidance rather than a brittle rule list.

Give me your bull and bear case on Anthropic.

Bull: they focused on enterprise value—coding, agents, security—built a mission-first culture that retains talent, and leader’s advantage compounds as models help build the next models. Bear: Google DeepMind has cash, distribution, and can bundle AI into search, Gmail, and more, and enterprises might balk if token costs outrun measured productivity.

Benedict Evans argues SaaS won’t just vanish and that LLMs could get commoditized, though I think personal assistants become sticky like friendships; how do you see it?

The ‘SaaS apocalypse’ is overblown—big firms will lean on trusted vendors to wire AI into messy, regulated systems, which is why Palantir thrives. Over time foundational models get sticky through personalization and control of payments and workflows, so switching becomes painful and pricing power improves.

So who likely wins consumer AI—does OpenAI hold on, or does Gemini take it with distribution?

Google is built for consumer-grade software and is threading Gemini through the G Suite cleanly, so it’s the natural favorite for B2C.

Let’s pivot to China: you went on a multi-city AI tour—what surprised you, and what should we do about it?

Contrary to DC expectations, Chinese academics and company leaders raised AI safety unprompted, which makes a nonproliferation-style pact on dangerous capabilities plausible, much like the IAEA and NPT contained nuclear risks beyond superpower deterrence. Chip controls haven’t yielded a massive model gap—maybe months—so I’d prioritize joint guardrails on open-weight misuse, even if that means some flexibility on exports.

Why does Washington keep saying Beijing won’t talk safety, and how much is that just cover for a race to be first?

Some former China optimists feel burned and now lean hawkish, but China’s system is plural, too, and there’s a camp ready to engage; refusing dialogue just leaves powerful open models for criminals and terrorists. I’ve also heard lab leaders argue we only need to stay ahead until recursive self-improvement kicks in by the late twenty-twenties, but even then, real impact still depends on deployment, compute, and energy—unless you use it offensively in cyberspace, which no one admits publicly.

One heavyweight in your book told me the plan is basically to hope the first to superintelligence are decent people, which is chilling; I’m considering a China trip to talk with builders before tensions rise.

Do it; visiting firms like Hikvision was cognitively dissonant—on campus they look like any inventive tech outfit, with water-quality sensing that enables clean-river incentives and on-site childcare, even as they sit under U.S. sanctions.

Which investors have most impressed you lately?

Bill Gurley’s Uber journey is the platonic venture story: he pattern-matched from OpenTable’s two-sided market, waited for the right founder, backed decisively, then helped steer through Shakespearean boardroom drama to a more disciplined phase.

How did Luke Nosek shape DeepMind’s early path?

Luke is uncontainable conviction; post-PayPal he went all-in on frontier bets, met Demis during an AI winter, and pushed Founders Fund to back him in London when few would—writing a small check for roughly half the company and flying across the Atlantic for board meetings.

Founders Fund is known for empowering partners to take contrarian swings; how did that play in later DeepMind rounds?

As round sizes grew, internal friction rose—some partners wanted a ‘product,’ while Demis kept saying AGI would upend every product—so despite Luke’s enthusiasm, the firm pulled back from leading a later round after signaling support.

Looking back, did Demis make the right call selling to Google?

Yes—whatever the headline price, Google poured close to a decade of near-billion-per-year R&D into London, seeding a thriving King’s Cross ecosystem and spinouts that would have been impossible without that compute and capital.

What should the UK change to supercharge innovation?

Entrepreneurial culture is made, not born—the ‘traitorous eight’ proves that—so fix the plumbing: end slow-motion ‘gardening leave’ so talent can move fast, and reform university tech transfer to be founder-friendly, Bayh–Dole style, or the startups never start.

Did Ender’s Game truly shape Demis’s mission, and did Gödel, Escher, Bach play a role?

Demis asked me to read Ender’s Game and said he identifies with the protagonist so much he gave it to his wife to explain himself, but his AGI fixation began in his teens; Gödel, Escher, Bach convinced him early that minds are computable, even if the book defeats many of us on the way through.

If you wrote another AI biography—Demis and Sam off the table—who would it be, and how fast is this wave really moving? For what it’s worth, my book sales were steady until late twenty twenty-two and have since slid hard, which feels like an early gust from GPT three point five.

I’d pick Dario Amodei, and I think this is bigger than mobile or the early internet because it’s general cognition; the open question is how fast diffusion happens.

AlphaFold nails structure, but delivery to tissue is the hard part; similarly, Lila Sciences is already running autonomous wet-lab loops end to end, so breakthroughs are happening now—do improvements really have to be exponential, or is it recursion that bends the curve?

Inside the labs, the pipeline feels immense: since late twenty twenty-two we’ve cut hallucinations, gone multimodal, expanded context, added reasoning and agents, and leveled up coding—all in a few years—so expect the next stretch to feel even faster to outsiders.

Billboard question: what message would you beam to everyone?

Prepare your mind—Pasteur’s line about chance favoring the prepared has guided great investing and great science, and it’s how Accel trained itself to move fast; it also explains how Ilya spotted transformers immediately, which is where we’ll pick up.

When the answer finally showed up after years of waiting, he pounced—that’s a prepared mind at work. Think of the Patriots’ goal line interception against the Seahawks; they practiced the look, broke at the snap, and were in the right spot because they’d mapped it out ahead of time, and that kind of discipline matters now so we don’t let AI do our thinking for us.

How do you set guardrails for yourself with AI? I’m thinking about it like the gym for the mind so we don’t end up directionally helpless the way many of us did with Google Maps.

I’m fine offloading chores like navigation, but not the act of thinking. I use large language models to surface and connect a researcher’s papers before interviews so I learn faster and then verify with the human, but I won’t hand over long form writing because that’s how I reason and I want that to stay mine.

Kevin Rose pushed me to keep retesting because these tools improve quickly. I also saw a blind test where editors struggled to tell a prize novelist’s originals from AI imitations, which made me want to track that more closely.

An Economist columnist ran the same experiment and his peers couldn’t spot the fakes, which rattled him. I’m sure the models will keep getting better, but I’m still holding tight to the work that makes me who I am.

What else should folks know as we wrap? We’ll link to everything, and people can find you on X at SC Malibu; and yes, I briefly read it backwards—old dysgraphia habits.

And now you’re a bestselling author—go figure.

Thanks again, Sebastian; search Mallaby on my site to find the episode, and as you head out, try being a touch kinder than you need to be—to others and to yourself—and keep training your mind. If you want a quick weekly note of what I’m exploring, you can sign up for Five Bullet Friday at tim.blog slash friday.