#416: The Relentless Missionary Creating AGI: Demis Hassabis
What this covers
David's solo episode traces Demis Hassabis from a childhood in competitive chess to the head of DeepMind. Early on, Hassabis took his father's advice to "try your best" literally, deciding the only proof was pushing to the edge of burnout. Later chapters cover his time at the Bullfrog studio under Peter Molyneux, where Gödel, Escher, Bach convinced him to dedicate his career to AI, and his first company, Elixir, which failed by attempting the most complex video game ever. David frames him as a missionary rather than a mercenary founder, and links this to founder-problem fit.
The technical arc is the middle of the episode. DeepMind began with Atari games because they were cheap to compute, numerous, and gave constant score feedback. It then moved to Go, where self-play and a reward signal produced AlphaGo Zero's strategies unknown to humans, and finally to AlphaFold and protein folding, a problem Hassabis had fixated on since college. David also covers the recruiting pitch, which offered an escape from academia's cautious, race-to-publish culture. About 80% of people at conferences rolled their eyes at the pitch.
The contested material comes last. David looks at the sale to Google, which spared Hassabis the endless fundraising, and at Peter Thiel's contrarian-investor logic. He argues Hassabis's charisma created an over-inspiration feedback trap, and that DeepMind was too absorbed in its own research to pivot to language models before OpenAI. The episode closes on Hassabis's forecast of long-horizon planning agents and his insistence that AGI builders disclose their motives.
David argues that Demis Hassabis's relentless missionary drive, formed in childhood competition and a conviction that information and intelligence are the fundamental units of reality, explains both DeepMind's scientific wins and its strategic choices, including the sale to Google.
- Hassabis's childhood chess training taught him to push himself to the edge of burnout, a pattern he carried into every later venture
- His early epiphany that information is the fundamental unit of reality drove him from games to AI to protein folding
- His willingness to sell DeepMind to Google, and his wartime response to ChatGPT, followed from treating money and power as means to scientific knowledge
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Hassabis's team believed that pattern-matching human play would not reach the level of the top human player, so the system needed to discover novel, non-human moves; the early version learned human strategies, then discarded some that could be counteracted, and became alien in style as it strengthened.
“if you pattern match what humans do, it's not going to take you all the way to beating the top human.”
AlphaGo Zero was trained exclusively through self-play using random moves and a reward signal rather than expert human games, and it outclassed its predecessor by a wide margin, discovering strategies unknown to human players.
“Learning only from self-play, the system outclassed its predecessor by a mile.”
DeepMind's AlphaFold project predicted protein structures and the company released its results free to the world, with the team of roughly 20 at its peak defeating decades of work by hundreds of academic scientists.
“Hundreds of academic scientists had spent decades on protein folding. How was it that the DeepMind team, numbering perhaps 20 at its peak, had defeated all of them?”
Hassabis's actions show his priorities: he had been fixated on protein folding since college, believing AI could finally solve it and unlock medical advances such as drug design and treatments for Parkinson's and Alzheimer's, which are linked to misfolded proteins.
“Knowing how those structures formed might unlock cures for Parkinson's or Alzheimer's, since both diseases were thought to be linked to incorrectly folded proteins.”
Peter Thiel believed that venture investments should be made by a single contrarian decision-maker rather than by committee votes, because consensus-chosen startups would produce middle-of-the-road portfolios with mediocre returns, since venture profits come from a few improbable moonshots.
“The way he saw things, if investments were chosen based on voting, the Founders Fund portfolio would consist of middle-of-the-road startups to which nobody objected.”
Solving protein folding matters because protein shape determines function and knowing structures helps design drug molecules, and understanding misfolding could unlock treatments for Parkinson's and Alzheimer's, which are linked to incorrectly folded proteins.
“Knowing how those structures formed might unlock cures for Parkinson's or Alzheimer's, since both diseases were thought to be linked to incorrectly folded proteins.”
DeepMind started with Atari video games as its testbed because the simple graphics kept computing needs affordable, the dozens of games offered many opportunities to prove generality, and constantly updating scores gave the agent the feedback it needed to learn.
“Given that most Atari games featured a constantly updating score, the agent would have the feedback it needed to learn how to play better.”
DeepMind's business plan argued that building general, proactive agents that learn broadly from their environment was more likely to succeed than narrow task-specific AI, because the narrow approach limits what systems can achieve.
“DeepMind asserted that its ultra-ambitious conception of AI made progress more likely.”
Hassabis sold DeepMind to Google partly to escape the endless fundraising negotiations with venture investors that were draining his attention, and he judged that a deep-pocketed parent company was necessary for AGI research.
“A deep-pocketed parent company could free him from the endless fundraising negotiations that cluttered his life and pulled his attention away from DeepMind's research.”
Hassabis predicted that future AI systems would plan long-term objectives the way humans do, breaking goals into weekly and monthly steps, so that an AI tasked with solving energy scarcity by inventing a superconductor could draw up a reading list, run experiments, and pursue the goal over a year or more.
“Future AIs would behave in the same way. Tasked, for example, to help solve energy scarcity by inventing a superconductor.”
Hassabis's charisma over-inspired his team, so engineers told him things were possible only because he had inspired them, and as a result no one gave him real feedback, deluding both him and the team.
“you can get to the point where you're deluding your team and then they are deluding you also.”
AlphaGo Zero's results showed that humans had not understood how little they understood about Go, and that AI could evaluate centuries of human strategic wisdom and keep some verdicts while overturning others.
“AI stood in judgment over centuries of human wisdom, vindicating some verdicts and tossing out others.”
Hassabis's father told him to try his best, and Hassabis took this literally, concluding that the only way to know he had done his best was to push himself to the point just before burnout or death, a belief that shaped his lifelong approach to work.
“The only way I could know is if I basically push myself to the point just before death, because that is literally when you've done your best.”
The internet was for AI what coal and oil were for the Industrial Revolution, because the vast data resource was an unplanned byproduct of human activity that AI could later extract, much as fossil fuels were stored by chance.
“The analogy here is the internet has been for AI what coal and oil were for the Industrial Revolution.”
Hassabis argued that every builder of AGI should be clear about motives, since his own reasons are scientific while others may be building it for other purposes, and Steve Jobs's view that motives make a foundational difference supports this concern.
“I think there's a question for anyone trying to build AGI. What are your reasons for building it? My reasons are scientific. Some are definitely building it for other purposes.”
Daniel Ek's idea of founder-problem fit holds that geniuses are seldom brilliant in general but are brilliantly suited to a particular mission, which David finds applicable to Hassabis.
“geniuses are seldom brilliant in a general way. They tend to be brilliantly suited to a particular mission. My friend Daniel Ek, who's the founder of Spotify, he calls this founder problem fit.”
Hassabis came to believe that information is the fundamental unit of the universe, and that a machine that learns for itself how to induce nature's patterns is the most powerful tool for understanding reality, which he later described as a potential meta-solution to any problem.
“Number one, information was the fundamental unit of reality. Number two, a machine that learned for itself how to induce nature's patterns was the most powerful imaginable tool with which to apprehend reality.”
When DeepMind's founders pitched an AGI company at conferences, about 80% of people rolled their eyes and walked away, and the founders treated this as an efficient way to identify who to recruit and talk to.
“80% of the people would roll their eyes at us, literally roll their eyes at us and turn around and walk away. We figured that this was a very efficient way to discover who we should be talking to.”
DeepMind should have pivoted to language models as OpenAI did, but it was too absorbed in its own research, unable to imagine a copycat overtaking it, and Hassabis rebelled against following OpenAI's example, which let OpenAI get ahead despite being founded later.
“DeepMind was too excited by its own research. It was accustomed to being the world's top AI lab.”
Hassabis holds that intelligence is fundamental, the root of all else and the mechanism through which humans perceive reality, and that building an artificial analog is the way to understand human intelligence, following Feynman's dictum that what one cannot build one does not understand.
“Intelligence is fundamental. It is the root of all else.”
Working at the Bullfrog game studio, where the blurring of work and philosophy fed big ideas, and receiving Gödel, Escher, Bach from founder Peter Molyneux, convinced Hassabis that computers would soon do whatever the brain could do and that he should dedicate his career to AI.
“I decided then that I was going to dedicate my career to working on AI, Demis recalled.”
DeepMind pitched recruits on escaping academia, which it described as both boringly cautious because it pursued incremental advances and terrifyingly competitive because scientists rushed to publish first, by offering the pursuit of big leaps with few rivals.
“At DeepMind, we're promising the opposite experience, the thrilling pursuit of the big leap and the near absence of rivals.”
Hassabis's core theme is that money and power were not ends in themselves but means to scientific knowledge, and he describes his mission as a lifelong work that is infused in him and cannot be separated from him.
“Demis's core theme is that money and power were not ends in themselves. They were a means to scientific knowledge.”
Hassabis's first venture capital experience convinced him that taking outside money meant giving up control, as investors wanted more than half his equity in exchange for funding, which he saw as selling his soul and which later motivated his sale of DeepMind to avoid endless fundraising.
“They wanted our souls in exchange for the money.”
Under Hassabis, DeepMind embraced strict unity around one model and meritocracy in which any member could propose and test an upgrade, and only measured performance gains determined what entered the master code, regardless of seniority, personality, or theoretical claims.
“Only measurement mattered.”
Demis Hassabis is an authentic missionary entrepreneur who feels compelled to solve a specific challenge and starts a company as a means to tackle it, rather than a mercenary who chooses a startup idea to get rich, and missionaries keep working on the problem even without pay.
“Demis was an extreme case of an authentic entrepreneur, not a mercenary who starts with a desire to get rich from a startup, then casts around for a plausible idea, but rather a missionary who feels compelled to work on a particular challenge”
After Musk and Hassabis had lunch in 2015 and Musk accused DeepMind and Google of irresponsibility, Sam Altman wrote Musk an email arguing that stopping AI development was almost impossible and that someone other than Google should build it first, which led to the creation of OpenAI as a rival.
“If it's going to happen anyway, it seems like it would be good for someone other than Google to do it first.”
Hassabis's first company, Elixir, failed because he plunged it straight into building the most complex video game ever, and that overreach doomed the project.
“At Elixir, he had plunged his company straight into making the most complex video game ever, and that overreach had doomed the project.”
Hassabis argued that DeepMind's sale was the right call because there was no time to waste, and that he would not trade a large amount of money for extra years of life unless those years let him accomplish his goals.
“There are so many things that have to be done while I'm still alive and my brain is still in gear.”
Hassabis argued that belief comes before ability, because a team cannot crack a hard problem if its leader believes it is impossible, and this conviction drove his expectation that AI could produce scientific breakthroughs.
“you definitely can't crack a hard problem if the person leading the team thinks it's not possible.”
DeepMind nearly ran out of money during fundraising because investors kept reneging on commitments, and this experience convinced Hassabis that blue-sky research was a poor fit for venture capital and that he needed a deep-pocketed backer.
“At one point they almost ran out of money.”
At a young age, after a tournament full of brilliant players exhausting themselves over a board game, Hassabis concluded that this collective mental effort was wasted and should instead be applied to science or medicine, which led him to resolve that his life needed a mission.
“I thought we were wasting our minds.”
DeepMind's approach set a grand ultimate goal but let people tinker while the scientific team was assembled, then set demanding intermediate targets once the team was in place, which the author presents as Hassabis's improved judgment.
“Demis had let people tinker while he was building out the scientific team, not setting a demanding goal for them.”
Hassabis's father drove him to chess tournaments every weekend for about six years in a van, sleeping in sleeping bags or cheap hostels, which strained the family's finances and crowded out a normal childhood.
“Chess consumed every weekend and every day of school vacation, squeezing out the easy recreation of a normal childhood.”
DeepMind's business plan argued that human brains have limited storage, human lifespans are limited, and large organizations show diminishing returns, so the complexity of society's problems exceeds human capacity and AGI is the solution.
“AGI is the solution to this problem.”
In response to ChatGPT, DeepMind pared back its broad blue sky research bets, stopped publishing mission-critical research that competitors could copy, shifted toward engineering over science, and merged Google Brain with DeepMind to end duplicate teams, beginning work on the Gemini model.
“The company would stop publishing mission-critical research that competitors could copy.”
Google bought DeepMind in January 2014 for $650 million, and Hassabis netted $136 million, and DeepMind's staff costs soon reached $260 million a year, six times its total spending over its first three years.
“Google bought DeepMind for $650 million. Demis netted $136 million.”
Steve Jobs argued that the motives behind a company matter a great deal, citing HP's early focus on making great products rather than being the biggest or richest, and David uses this to frame Hassabis's question of why people build AGI.
“the older I get, the more I'm convinced that motives make so much difference.”
On a flight back to California, Luke Nosek and Larry Page talked AI with Elon Musk, who had concluded after debate with Hassabis that powerful AI might matter more than spaceflight, and Musk then promised to invest in DeepMind, calling it the only AI company that would work.
“Elon says there's only one AI company that I think is going to work. I'm an investor in that company, DeepMind.”
Reading David Levy's The Chess Computer Handbook introduced Hassabis to the themes of computing and chess, and at age 12 he built an Othello program applying those principles that was good enough to beat his younger brother.
“The program proved intelligent enough to beat Demis's little brother.”
Bullfrog's founder offered Hassabis £500,000 to stay and work on the studio's next game, roughly $1.7 million in today's money, and Hassabis refused in order to attend Cambridge.
“He writes out a check for £500,000 to get him to work on Bullfrog's next game.”
Past innovations from gunpowder to nuclear fission made wars more terrifying and accidents more lethal, but their general effect on technological change has been to amplify human experience and extend lifespans, which supports an optimistic view of AI discovery.
“Past innovations from gunpowder to nuclear fission have made wars more terrifying and accidents more lethal, but the general effect on technological change has been to amplify our experiences and extend our lifespans.”
Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for protein structure prediction, the breakthrough that resulted from DeepMind's 2020 solution of nearly all protein shapes in nature by stitching together 32 algorithms.
“in 2024, Demis and John Jumper were awarded the Nobel Prize in Chemistry for protein structure prediction.”
Hassabis began chess at about age four, beat adults within weeks, and started competing in tournaments by age five, sitting on a telephone book to reach the board.
“Within a few weeks, he had mastered the game well enough to defeat adults.”
After ChatGPT reached 100 million users within two months, Hassabis saw the moment as a fight for survival and described it as wartime, with OpenAI and Microsoft having 'parked the tanks on the lawn.'
“This is wartime, he answered. OpenAI and Microsoft have literally parked the tanks on the lawn.”
In December 2010 Founders Fund wired $2.3 million to DeepMind for a bit less than half the company, and no other capital was available, showing how hard it was to fund the AGI idea.
“Founders Fund wired $2.3 million to DeepMind, and they assumed ownership of a bit less than half the company. There was no other capital available.”
Hassabis argued that when powerful people understand a transformative technology, they will not stay on the sidelines as advisors, so the supposed advisors were in fact rivals.
“They won't be content to just be your advisors.”
Hassabis's childhood was shaped by chess tournaments that consumed weekends and school holidays, financial strain at home, and a father who erupted in anger after losses, which left Hassabis with a strong sense of control and competition.
“When Demis had a bad game, his father would erupt. There was one time I lost horribly, my dad went mental.”
Larry Page pitched Hassabis that he could use Google's accumulated resources to pursue AGI instead of spending a career building a company like Google, and Hassabis found this argument persuasive because it aligned with his mission.
“Why don't you take advantage of what I've already created, Larry Page? Asked Demis.”
Near the end of his Cambridge time, Hassabis planned to found a company to pursue AI, an idea that was shocking in a British academic setting that had no Silicon Valley equivalent and few students thinking about starting companies.
“Entrepreneurship was a foreign concept on the Cambridge campus. Britain had no equivalent to Silicon Valley.”
David Gammon would invest only if DeepMind pitched to his home, with his wife and three teenage sons each having an equal vote, illustrating how bizarre and difficult early fundraising was.
“Each family member would get an equal say on whether to invest.”
Elon Musk argued humans needed to colonize Mars in case disaster struck Earth, and Hassabis countered that killer AI could be such a disaster and could follow humans to Mars, after which Musk concluded powerful AI might be more consequential than spaceflight and promised to invest.
“Elon had declared that humans needed to colonize Mars in case disaster struck Earth.”
Hassabis combined intense competitiveness with unusual kindness and approachability, and he said he avoided manipulating people because he wanted to help them.
“I want to help people and I feel very strongly that it's really bad to manipulate or control people.”
Hassabis won Peter Thiel's interest by talking chess at a party, and Thiel judged DeepMind to be an A+ on the science and possibly an F on the business model, but still backed the project as a missionary venture.
“began to think this project was an A+ on the science and maybe an F on the business model.”
When Elon Musk learned that Google was about to buy DeepMind, he tried to acquire DeepMind himself, arguing that a giant corporation should not control AGI, and offered Tesla and SpaceX as alternative acquirers, though Hassabis rejected both as lacking the cash or computing power.
“Demis pointed out that Tesla was not generating enough cash to support DeepMind's research.”
Paul Graham wrote that Sam Altman is extremely good at becoming powerful and could be parachuted onto an island of cannibals and return in five years as king, and Hassabis cited this as a reason to scrutinize the motives of AGI builders.
“Sam is extremely good at becoming powerful.”
DeepMind staff called Hassabis's management style 'Demis-driven development,' in which review meetings set deadlines for upgrades and Hassabis always wanted more, so that no amount of progress was sufficient.
“If a review meeting with Demis has been scheduled for Tuesday, you were urged to complete the next round of upgrades by Monday.”
Hassabis identified strongly with the protagonist of Ender's Game, a boy genius manipulated and tested for a mission to save humanity, and recommended the book to Mallaby as a way to understand his endurance and sense of responsibility.
“Demis identified powerfully with Ender, and he suggested that I read Ender's Game”
David argues that Hassabis is a missionary who has been on the same mission since childhood, with his life experiences fitting together like a puzzle, and that his clarity about motives explains his decisions.
“all these experiences that he has, all the things that he's learning, they, they fit together almost like a puzzle.”
When Hassabis told Sergey Brin he wanted to build a computer to beat the world Go champion, Brin doubted it was possible, and Hassabis took that skepticism as a sign that success would be especially impressive.
“Brin seemed incredulous. Wouldn't that be impossible? Great, Demis thought to himself.”
A DeepMind researcher argued that the company probably needed to be second for a while to light a fire under its own team, since public humiliation is a powerful galvanizer of action.
“There's nothing like public humiliation for galvanizing action.”
Co-founder Shane Legg described Hassabis's determination as astonishing and unlike anyone else, saying he works, sleeps, eats, and breathes the mission 24 hours a day to a degree Legg had not seen in others, with football as his only hobby beyond the mission.
“He works, sleeps, eats, breathes the mission 24 hours a day to a degree that I haven't seen with other people.”
Hassabis's ideas owe their coherence to how his experiences, from physics to AI to neuroscience, fit together like a puzzle, which Mallaby saw as the reason his path looked deliberate.
“I marveled how Demis's experience and ideas appeared to slot together.”
Peter Thiel was instinctively suspicious of Hassabis because he believed a former chess prodigy who had spent his formative years crushing opponents should be treated with caution.
“Peter Thiel barely saw the DeepMind team, and he felt instinctively suspicious of a fellow chess player.”
Colleagues described Hassabis's persuasive power as uncanny, saying he could tell people they would believe what he said and they did, which made his early fundraising and recruiting effective but also contributed to over-inspiration.
“His powers of persuasion were uncanny.”
Hassabis says he is doing AI for knowledge and science, that the mission is inside him and cannot be separated from him, and that a leader who is a reed in the wind would fail at the job.
“I'm doing this for knowledge and science. This is my whole life's work.”
Hassabis said that because Google's AI effort was so large he no longer coded or designed things directly, and his job was holding about 100 projects in mind, setting the vision, picking intermediate targets, and nurturing people, with relentless shipping as the goal.
“I don't code anymore. I don't design things directly, so my skills are more about holding 100 different projects in my mind”
Less than two years after the Google DeepMind merger, the combined team had closed the technical gap with competitors, which David describes as a considerable achievement.
“Less than 2 years after the messy shotgun marriage that created Google DeepMind, Demis's team had closed the technical gap.”
David suggests the fact that coal and oil happened to be underground, and that the internet existed for AI to use, is improbable enough to raise a deeper question about what is really going on in reality.
“Why would we have this coincidence?”
Hassabis chose to remain in London, far from Silicon Valley's hype and commotion, which the book presents as evidence that he is a figure apart.
“It is not by coincidence that he has chosen to remain in London, far from Silicon Valley's hype and commotion.”
The DeepMind deck quoted Bill Gates saying that a breakthrough in artificial intelligence that lets machines learn would be worth 10 Microsofts.
“if you invent a breakthrough in artificial intelligence so machines can learn, that would be worth 10 Microsofts.”