Reid Hoffman Pledged a Billion to Free DeepMind. Hassabis Never Played the Card

Jacob Effron, hosting Unsupervised Learning for Redpoint, sits down with Sebastian Mallaby a few weeks after the publication of *The Infinity Machine*, the biographer's sixth book and the product of three years and more than thirty hours with Demis Hassabis in the same British pub, coffee on the table rather than pints. Effron wants the reporting used to understand the people and decisions that shaped the race we are living through. He starts with the biggest structural question in the book: was this lab-versus-lab contest inevitable, or could different choices have produced a different world?

Mallaby's answer is blunt. It was inevitable. When a technology is this strong, multiple labs in multiple countries will be desperate to build it. China's stack is strong enough that a semiconductor shortage was never going to keep them out, and they are doing pretty well. France has Mistral, Canada has Cohere, and the United States has a crowd. The technology is too sweet to interest only one team. What is strange, he says, is that this is not how people saw it *ex ante*. When Hassabis was starting DeepMind, he hoped he could avoid the race. In retrospect it looks naive. It was still what he hoped.

The hope had a ritual. Interviewing candidates, Hassabis would describe a day when the group might be close to AGI, fly to a bunker, and finish the job as one team. Would you get on that flight? Anthropic's founding story, Effron notes, carried a cousin of the same idea: when you reach the precipice, our people should be the ones with their hands on the dial. Mallaby says Hassabis has since swung from one extreme to the other. He began in a singleton scenario that really meant DeepMind and himself. He now sees a crowded field in which it is almost pointless for one lab to pursue safety alone. If your house is safe and the others are not, the world is not safer. Safety is a collective-action problem that only a government can solve.

The 2015 lesson arrived at SpaceX. DeepMind hosted a safety summit with Elon Musk as host, the plan being to bring him inside the tent, even have him chair a safety oversight board, so that he would not start a competitor. By the end of 2015 he had started OpenAI. Mallaby says you still cannot trust the other labs, which is why future collaboration needs a government enforcer and a level playing field: pre-testing, slowdowns, the same rules for everyone. The next objection will be China, which is why the endgame is a US-China deal, remote as that sounds. He points to the FDA, and to Britain's AI Safety Institute, whose chief scientist Geoffrey Irving came out of DeepMind, also spent time at OpenAI, and is a close collaborator of Dario Amodei's. That public team has found vulnerabilities private labs missed and quietly shared them. The last line of Mallaby's book is Hassabis saying he is still optimistic. Crises produce action, Mallaby argues; DeepSeek in January 2025 did not kill that line. COVID showed governments doing things nobody expected. Protectionism, a slow poison, does not. An acute scare, an Anthropic model such as Mythos that suddenly worries Washington, can flip the United States from laissez-faire to control. The shock is what matters.

Mallaby started the project just before ChatGPT, after AlphaFold, and spent the research years watching the popular story fail to catch up. The thing that still astonishes him is how thoroughly people discounted Hassabis. Interviewers since the book came out often cannot pronounce the name. Penguin Press wanted a photograph on the cover and also feared that a face nobody recognized would not sell books. The compromise was a fuzzed image, an enigmatic geek. Sam Altman, Dario Amodei, and Elon Musk would have been fine as cover faces. Hassabis founded the original lab in 2010, years before the others, and built the model OpenAI later copied. When Mallaby went to see Amodei during the reporting, Amodei said yes: Demis was the original figure, and at the time he had the AI-for-science space to himself.

Google was discounted on the same script: innovator's dilemma, too slow, OpenAI went first with a chatbot and now owns a brand nobody can catch. By late 2025 that story was already wrong. Google DeepMind's Gemini 3.0 models were better on the leaderboards than the rivals, before a later Anthropic surge. People crowned OpenAI and Altman too quickly, Mallaby says, and underestimated both the person and the company. Effron pushes on a different scoreboard. The two monumental consumer moments were ChatGPT and then Claude Code. Google's models look good; the products have not owned either zeitgeist.

People use Gemini more than the chatter implies, Mallaby notes, because it is bundled into Google search's AI mode. He still takes the broader point. Neither seminal consumer experience came from Google DeepMind. Part of the reason, as he reads Hassabis, is formation. A neuroscience PhD produced a very broad picture of intelligence and therefore a try-everything research culture. Two paths appear, and they take both; a third appears, and they take that too. Anthropic got to coding because it was willing to make a concentrated bet. It never went into generative video and never had a Sora equivalent. OpenAI, as a startup, did not have Google's reputational baggage, so it could put out a chatbot that hallucinated a lot at the beginning.

The same pattern, Mallaby says, is the transformer story in miniature. Google invented the architecture. OpenAI went all in on scaling it. DeepMind pursued several paths, arrived late to language models after 2017, arrived late again to productizing behind ChatGPT, and then caught up by 2025. With Google's pockets, bench, talent, and compute, you can afford not to be Apple, which has taken the extreme of not doing the AI thing and charging to put it on the iPhone. Google is in the fight. It does not mind being a couple of years behind, because it figures it can catch up. Mallaby does not think the multiple-bets habit has been cured. A young scientist from the shop told him as much the week before the taping: they always make multiple bets, they never really go hard down one avenue.

The most explosive passage in the book is the secret spinout plan, Project Mario. Hassabis was not happy when Mallaby learned of it from other sources. Google DeepMind's general counsel tried to talk him out of writing it. Other people leaked the documents. Mallaby had them, they were true, and he wrote it. The fight underneath was safety oversight. Hassabis wanted independent safety control over the models. Mountain View would not give it. So he needed a credible threat of leaving, and he went to Reid Hoffman, who pledged a billion dollars to finance a spinout.

Hassabis used that pledge to pressure Google without, Mallaby thinks, ever mentioning that Hoffman had made it. He pushed knowing the option was in his back pocket. Advisers in the room split. One camp said spin out: independent-startup incentives, alertness, flexibility, high-powered equity for the team. Hassabis's view, in the end, was that there might be a court battle over whether they even had the right, and he wanted to do science, not live in legal fights. He wanted tons of compute. He stayed.

Was it right? A year after he gave up the fight with the parent company, DeepMind shipped AlphaFold, the protein-folding model that led to his Nobel Prize. Mallaby tells that punchline with a laugh and then, when Effron asks how serious the team really was, turns colder. They wanted a plan B because why not have one. They were never quite sure how to use it. They determined never to tell Google explicitly, only to hint that if safety oversight did not improve, something unexpected might happen. They never waved the threat in the open. They never exercised it, even though Google still did not give them the oversight they wanted.

In that sense it is a story of naive founders. Hassabis and Mustafa Suleyman, the co-founder now at Microsoft, went in not quite sure what the endgame was. They spent three years going back and forth to Mountain View with a card they did not play. Mallaby calls some of that a waste. The other reading, which he also allows, is that staying is what made AlphaFold and the Nobel possible. The leverage existed. The safety concession did not arrive. The science did.

What does Hassabis regret about the last decade? Not the energy poured into AI for science. Effron loves the AlphaGo detail: the day they beat the best player in the world, someone caught the next sentence on a microphone. Hassabis rested on his laurels for about ten seconds and then said the next thing they were going to do was solve protein folding. Mallaby thinks the serious point is larger than a Nobel. Hassabis sees visible scientific benefit as central to whether society will accept AI at all. If the technology is mostly job disruption, politically useful for productivity and painful for the people on the receiving end, a huge backlash becomes easy to imagine. Science is how you earn the rollout.

Does he regret being late to chatbots? Of course. The contrast is Ilya Sutskever on the day the transformer paper dropped in 2017: out of his chair, down the corridor, finding Alec Radford, announcing that they were going to build a language model on this architecture. Sutskever had a prepared mind. His PhD had been about sequential data such as text, so he saw the significance immediately. DeepMind took two or three years to get there. When Mallaby tells Hassabis they were about three years late, Hassabis still says no, they had Chinchilla, they had other models. Mallaby's reply is that those models did not come out until the end of 2020, by which time Sutskever had been on the problem for three and a half years. You were late. Hassabis resists the point, Mallaby thinks, because he does regret it.

Effron connects the resistance to neuroscience. Hassabis believed the path to AGI had to look more like human learning, the reinforcement-learning world, and less like consuming the internet. Scaling showed you can understand more of being human from that corpus than anyone had anticipated. Mallaby wonders whether robotics will stage the same dilemma. Training robots in simulation needs real-world simulations, and video may be a key tool, which means DeepMind has cards. The old habit remains: put a finger in a dozen approaches while a startup goes all in on one. If the recipe is unclear, keeping options open is not irrational. It also leaves you open to whoever goes all in on the next transformer.

Mallaby has written both about venture capital and, now, about a hyperscaler, so Effron teases him for the matchup. The contest is interestingly balanced. Generative AI is capital intensive; deep pockets matter. In January Mallaby wrote in the *New York Times* that OpenAI had a 50% chance of going bust by the following summer, meaning selling itself at a discount to some hyperscaler. Asked on the podcast whether he is still at 50/50, he is. He gives the company credit for cost cutting, including canceling Sora, and he does not doubt the tech. The business model is the problem, because the opponent is Google, which can spend you into the ground. Rumors of Sam Altman being replaced by Bret Taylor are widespread. Leadership looks like damaged goods. In one sense this platform shift helps incumbents. In another, unlimited talent, compute, and money make concentrated strategic bets less likely, and a cluster of venture-backed startups each going all in on a different thesis can get a hit.

Anthropic is the thin-margin version of that story. There was always a question whether it could raise enough to stay in the race or would end up deeply embedded with a hyperscaler, Google or Amazon. Then the coding models hit, hugely, and Anthropic became the hottest company in the Bay. If they had hit six or twelve months later, Mallaby and Effron agree, the clean retrospective narrative would read completely differently. Google at least stays in the arena. Apple has taken the other extreme. Researcher Jack Rae left DeepMind for OpenAI because OpenAI had put all its chips on language models, which was the thing he did. Drop a robotics researcher into a careful, diversified lab that believes in you and also in five other bets, and the passion piece goes missing.

Isomorphic Labs, the AlphaFold spinout that Alphabet still owns a large share of, had practical reasons: partnerships with big pharma, health's own politics and long lead times, a sense that the effort needed a different home. Effron's extra reason is recruiting. You give people a place where they *are* the bet. Mallaby thinks that tool matters. He also notes that Anthropic's churn is very low relative to everyone else. Amodei's unfiltered safety identity, the long essays, the extreme concern, means some people will never join because they think he is wacky. The ones who do join love him. Loyalty is part of the product.

David Silver, the reinforcement-learning true believer behind the Atari agent, AlphaGo, and AlphaZero, left DeepMind after the book came out to start his own company. Mallaby thinks it fits. Colleagues came to find the intensity too much. Silver and Rich Sutton, his PhD supervisor, wrote the paper Mallaby half-recalls as arguing that you should learn from experience, not from data, because data includes mistakes. Train on even expert human Go games and you inherit the human ceiling; superintelligence has to come from the machine's own play. Hassabis told Mallaby more than once that this approach may win after AGI, when you are polishing toward superintelligence, but the road to AGI has to bootstrap on existing data, which is what the language-model revolution showed. Silver left just as RL came back into fashion, a human lag between feeling swamped and walking out. He wanted a small organization where his vision is the vision.

The personal histories among lab leaders are less documented on Hassabis's side than the Sam-Dario or Elon-Sam files. Mallaby starts with Musk. Both men were Founders Fund companies. Around 2012 they presented at the same LP offsite, then got talking. Musk, being Musk, said he had the most important technology in the world: even if Hassabis's AI screwed up Earth, SpaceX could make us multiplanetary. Hassabis answered that the AI would conquer spaceflight and follow him to Mars, so Mars would not be safe either. Silence. Then Musk said he would like to invest in the Series B, and wrote a five-million-dollar check.

At the start of 2014, with Google circling, Musk wanted to buy DeepMind to stop the sale. Luke Nosek, the Founders Fund partner on SpaceX's board, found Musk at a party in Los Angeles. They decided Larry Page was a transhumanist you could not trust. They went upstairs into a closet and Skyped Hassabis in the middle of the London night: sell to us, to SpaceX, to Tesla, anything but Google, they are evil. Hassabis said Google had the compute, he was selling to them, goodbye, good night, and put the phone down. Musk went nuts, started calling Hassabis an evil genius, a nod to a game Hassabis had worked on as a designer, and stayed obsessed. Mallaby says the recent Musk-versus-Altman trial brought some of that back. Hassabis is keen to say they get on fine now. Mallaby has not heard it from Musk, but guesses the water is under the bridge and that Musk has moved on to fighting Altman.

Altman is a different mismatch. Hassabis has a Nobel Prize. Altman did not finish his first degree. Therefore, Mallaby says in the line the podcast uses as a cold open, Hassabis does not take him very seriously. To Hassabis, Altman is the skillful ultimate embodiment of the Silicon Valley network: fake it till you make it, tell the premature truth, raise money, leverage connections. That is all real, and it is not the same as being a serious scientist, and you cannot trust somebody like that. Mallaby thinks the public has come around toward Hassabis's view of Altman to quite some extent. Hassabis always felt it. Mallaby does not think he ever liked him.

Amodei took the celebrity path Hassabis will not. Pick a public fight with the Pentagon, handle a Mythos release in a way that manufactures attention, and you become famous immediately. Hassabis is, Mallaby says, in some ways too sensible to pick a public fight with the government, and the byproduct of that strategic caution is that he is less notorious. He is good at retrospective story: thirty hours with a biographer, *The Thinking Game*, the AlphaGo documentary team brought to Seoul on purpose. He is not good at the prospective trick. In his heyday Altman could kill attention on a DeepMind release by trailing two posts on X the day before, not even shipping anything, because his following was five or six times larger. That hurts product adoption and recruiting. Google DeepMind knows the problem. Whether it can fight it is another matter.

*The Economist*, talking about the book, treated it as a test of great-man history. Mallaby once wrote *The Man Who Knew*, about Alan Greenspan, who understood that bubbles were dangerous, had written his PhD about them, and still could not stop 2008. He says he could almost have given Hassabis the same title. Hassabis has known from the beginning that this thing is dangerous. As the leader of one lab, even a rich one, even as a Nobel winner, making his own house safe does not stop the others. The race has a kind of inevitability. Personalities still matter. Amodei changed the safety narrative by fighting the Pentagon in public and, more importantly, by how he did the Mythos release. Altman chose to release ChatGPT while it was hallucinating; he did not have to, and that colored the whole race. Hassabis, more quietly, told Rishi Sunak the UK should host a global AI safety summit. Bletchley Park happened, the Chinese came, and an international conversation started. Leaders count. Underlying forces count too.

Asked which moment most shifted his view of Hassabis, Mallaby talks about the depth of a conviction he had not suspected. In two-hour pub conversations Hassabis would erupt, banging the table: this table is a mystery, atoms with gaps, yet it is solid; a laptop is sand and copper, so why can it think; the world is set up to function, there must be some intelligence behind it, this cannot be coincidence. Maybe, if we approach science the right way and understand more of nature, we get closer to something we could perhaps call God. Mallaby had no idea Hassabis felt that way. It explains, for him, why a man who knows the AI is dangerous still pushes to build it. It is a quasi-spiritual quest.

Hassabis drew lines. Family was out; Mallaby left it out, beyond the passing facts of a Cambridge girlfriend who became his wife, two children, a normal life. Fights with Sundar Pichai and Mountain View were not supposed to be in the book; other people talked, and Mallaby wrote them anyway. The firing of Suleyman was another sore spot. Hassabis would say he did not really fire him, that there was a bullying inquiry run by the general counsel. Mallaby's version is that the inquiry was the trigger, not the deep reason. Nothing happened at DeepMind unless Hassabis wanted it. He decided it was time for Suleyman to go because they disagreed on too many issues. Some preferences Hassabis won. Some he did not.

What Mallaby wishes were more discussed is scientific taste inside a deep-tech company. Elixir Studios, Hassabis's first game startup, blew up by being too ambitious on graphics and early reinforcement learning; the technical team went over the brink and could not ship. In 2018, two years into AlphaFold, the team had the best protein-prediction system in the world. Andrew Senior wanted to declare victory. Precise prediction, he said, was impossible. Hassabis sat in the meetings and listened for what he calls fluidity: if researchers bounce possibilities off one another, add chips; if there is silence, stop. The room was fluent. He moved Senior aside and put in John Jumper, later co-winner of the Nobel. That, for Mallaby, is the lesson of frontier science inside a firm. As for Hassabis becoming Google CEO, it is possible and he is conflicted. He still does a double shift, dinner with family, then the desk until 4 a.m. for science, a window already being eaten by Mountain View calls. CEO would leave nothing. If Pichai left and the alternative were someone Hassabis trusted, he would probably stay a scientist. If not, he might put his hat in the ring rather than work for someone he did not agree with.