The Nobel Prize laureate in Chemistry has left DeepMind, predicting the realization of AGI within the next three years, to dedicate all his efforts to curing cancer using artificial intelligence.
Demis Hassabis often stays awake at night; not out of worry, but out of passion. Midnight is his favorite time to work; the world grows quiet, his mind is set free, and he can think without interruption. If work is going well, he glances at the schedule for the next morning and sometimes decides not to sleep at all. He knows this habit is harmful to the brain; he has studied the brain for years.
On August 5, 2026, Hassabis wrote to his employees that he had spent his entire life working to build Artificial General Intelligence (AGI) and now sees it within reach. Right in that same message, he announced he was stepping down as CEO of Google DeepMind, the lab he had built himself. Why would someone hand over the steering wheel on the verge of achieving a lifelong dream?
Two and a half months later, in an interview with the Wall Street Journal, he made his previous prediction regarding the arrival time of Artificial General Intelligence, or AGI, even bolder, stating that we are two to three years away from AGI rather than three to five years, and adding: "My estimate keeps getting closer."
Why did a scientist whose achievements are respected even by his fiercest rivals abandon the CEO position at one of the world's most advanced AI labs?
We might understand Hassabis's decision when we look back; to the boy who used to sit on a telephone directory to see a chessboard, the game designer who turned down one million pounds, the neuroscientist who researched imagination, and the man who had predicted his Nobel Prize years in advance. Hassabis's master plan was always clear: "Step one, solve intelligence. Step two, use it to solve everything else."
Checkmate in Liechtenstein
In 1980, a four-year-old boy in London—born to a Greek-Cypriot father and a Singaporean-Chinese mother—learned the rules of chess simply by watching adults play. He wasn't tall enough to reach the table, so he sat on a telephone directory to see the board. By age 13, he achieved a rating of 2300; in his age group globally, only Judit Polgár ranked ahead of him.
At a chess tournament in Liechtenstein, the 12-year-old Hassabis battled an experienced opponent for hours. The hall gradually emptied, and he eventually lost. But what stuck in his mind from that day wasn't the loss. He looked around the hall and thought that if all these brains were working on a real-world problem instead of moving pieces, cancer might have been cured.
From his teenage years, Hassabis was more fascinated by the mind that played chess than by chess itself.
He later said that at that moment, he felt life had to offer something more than just moving around "lifeless pieces of plastic."
The young chess player used his tournament prize money to buy his first computer and taught himself programming. From those early days, the ultimate puzzle for him wasn't chess itself; it was the mind that played chess. In his own words, chess led him to "think about thinking itself." Reading an excessive amount of science fiction also played a role; he laughs and says it was one of the reasons he entered this field.
One Million Pounds to Drop Out of University
Hassabis was as talented at programming games as he was at playing them. At sixteen, before entering university, he temporarily put his studies on hold and entered the game development industry. During this period, he joined Bullfrog Productions and worked on Theme Park; an amusement park simulation game that sold millions of copies and featured virtual visitors with spontaneous and unpredictable behaviors. His first serious experiments with elementary artificial intelligence took shape within these very games.
The head of the studio made him a tempting offer: one million pounds on the condition that he skip university. Hassabis declined and went to Cambridge; the university where Newton and Turing had studied. He studied computer science and, naturally, kept playing games as well. He taught himself Go, the Chinese board game that is vastly more complex than chess. He also conquered foosball in the city's pubs. Hassabis proudly recalls, "I was the champion of Cambridge."
Foosball is no minor joke in Demis Hassabis's biography; it even opens a window into his personality. Sebastian Mallaby recounts in his book The Infinity Machine that Hassabis later brought a foosball table to his workplace and challenged his colleagues. When he finally lost, he stood up and shouted, "It feels like my soul is on fire!"
Those close to him describe him with traits like brilliant and tireless, and certainly so competitive that he cannot stomach even a loss in foosball.
Hassabis turned down the one-million-pound offer to study computer science at Cambridge
After Cambridge, Demis Hassabis founded his own game development studio, Elixir Studios. Despite its ambitions, Elixir did not survive, and Hassabis experienced the first major failure of his life. His next step surprised many: he returned to university, this time to pursue a PhD in Cognitive Neuroscience at University College London.
However, Demis had been dreaming of Cambridge for years, perhaps since his teenage years when he watched a film about the discovery of the structure of DNA and wonderingly asked himself whether scientists in Cambridge really sat in pubs and made such discoveries? During the university entrance interview, he even surprised the professor with the speed of his mental calculations.
The research conducted by Hassabis and his colleagues during this period showed that patients suffering from amnesia due to hippocampal damage also had difficulty imagining the future and experiencing new imaginary scenarios. The findings indicated that parts of the brain's mechanism responsible for recalling the past are also crucial for constructing imaginary scenes. This research later influenced Hassabis's perspective on artificial intelligence and machine learning.
Bell Labs at the Heart of the Manhattan Project; Google's $650 Million Bet
In 2010, Hassabis, along with Shane Legg and Mustafa Suleyman, founded DeepMind. The vision he had for the company in his mind was a combination of Bell Labs and the Manhattan Project: a laboratory for fundamental discoveries, with the urgency of a historical mission. And naturally, the starting point was gaming once again.
Hassabis founded DeepMind to solve the problem of intelligence
One of DeepMind's first testing grounds was Atari games. The artificial intelligence agent learned what to do simply by seeing the screen image and receiving a score, without knowing the rules of the game. In the game Breakout, it initially struggled to return the ball, but as practice continued, it discovered more effective strategies—including clearing a path through the brick wall and sending the ball into the space behind them.
In 2015, researchers demonstrated that this same approach could achieve human-level or superior performance across a range of games. When the Atari paper was published, Hassabis stood up at a company meeting and raised a glass to celebrate the achievement. Then he added a sentence that might have seemed like a joke at the time: "One day we'll win a Nobel with this technology."
In 2014, Google acquired DeepMind in a deal reported to be worth around $650 million. Google's financial backing allowed DeepMind's scientists to pursue ambitious research that did not necessarily generate short-term revenue. However, Hassabis's relationship with the new owner was not always tension-free.
In the very early days of joining Google, Hassabis had stated that building Google's computational and engineering infrastructure for DeepMind might take ten years. Now, with the help of that same infrastructure, they could run countless experiments simultaneously. Larry Page's personal interest in artificial intelligence also played a role in his decision; but the benefits of this partnership could not eliminate subsequent disagreements over the future of DeepMind.
DeepMind fought for independence from Google, but that very dependence later turned into one of its greatest advantages.
In his book, Mustafa Suleyman narrates the behind-the-scenes of one of the least-told chapters of this story. Hassabis and Suleyman secured a one-billion-dollar commitment from Reid Hoffman, co-founder of LinkedIn, to make DeepMind independent once again. Long legal and financial negotiations ensued, but ultimately DeepMind remained a subsidiary of Google.
According to Reuters, at a gathering in Scotland in 2017, DeepMind leaders even told employees of their intention to break away. Hassabis believed that only by freeing itself from Google's profitability pressures could DeepMind ensure the responsible development of artificial intelligence.
The independence plan failed, and according to accounts, these very battles transformed the idealistic Hassabis into a pragmatic scientist. Interestingly, the cage that Hassabis could not escape provided a massive advantage for DeepMind: access to Alphabet's financial resources, computing infrastructure, and global network, without the laboratory having to bear all the costs of competition on its own.
AlphaGo's Move 37; A Piece That Showcased Machine Creativity
In March 2016, at a hotel in Seoul, Lee Sedol, one of the greatest Go players in history, sat before the board to compete against his opponent, a computer program named AlphaGo.
Go had long been considered one of the most formidable challenges in artificial intelligence. The number of legal configurations on its 19x19 board far exceeds the estimated number of atoms in the observable universe, and the number of possible game paths goes beyond even that. No program can examine all these paths one by one; top players also rely more on experience and intuition than on mere calculation to choose their moves.
In the second game, AlphaGo placed its stone in a spot where a human master was highly unlikely to play. DeepMind itself later described the probability of a human choosing such a move as about one in ten thousand. Commentators initially thought the machine had made a mistake, and Lee Sedol left the room, taking a few minutes to return.
AlphaGo's move 37 demonstrated that the machine could discover strategies capable of surprising even grandmasters.
Several dozen moves later, the significance of that strange decision became clear. Move 37 helped AlphaGo win the second game and showed that the program could devise strategies unexpected even for top-tier players.
AlphaGo won the match 4 to 1, and move 37 turned into one of the most controversial moments in the history of artificial intelligence—a moment where the machine did not merely calculate, but seemed to exhibit creativity. For Hassabis, the boy who once sat on a phone book to see the chessboard, this victory held a meaning far beyond defeating a champion.
Yet, Hassabis never saw games as the ultimate goal; he viewed them as a training ground.
Moving Beyond Games and Attacking the Protein Puzzle
Hassabis had found his next challenge years earlier, beside that very foosball table in Cambridge. A biologist friend had told him about the protein folding problem; proteins are chains of amino acids that fold into complex three-dimensional shapes, and it is this exact shape that determines their function.
Protein structure prediction from sequence was a long-standing problem in biology that could help researchers better understand protein function, investigate the mechanisms of certain diseases, and advance the discovery of new drugs. Hassabis saw this problem as an opportunity to transfer the power of artificial intelligence from games to scientific research.
Pushmeet Kohli, DeepMind's current VP of Research, discussed joining the company a decade ago during a long walk with Hassabis. He later said: "I expected to talk to a CEO, but I came away with the feeling that I had spoken to a scientist."
According to Reuters, Hassabis had told his employees to tackle problems big enough that solving them would win a Nobel Prize. In a meeting with executives from BlackRock, the world's largest asset management company, he compared DeepMind to NASA’s Apollo program; an ambitious project aimed first at solving the problem of intelligence and then moving on to other grand challenges.
AlphaFold enabled the structure prediction of millions of proteins
Against such a vision, using AI to turn a profit in financial markets was not a high priority, at least for Hassabis. The DeepTick project, despite its promising performance in trading simulations, was eventually quietly shut down.
The first signs of their success appeared in 2018. DeepMind surprised long-standing competitors at the CASP competition, where research groups test their ability to predict protein structures. Among the 43 proteins whose structures had to be predicted without a template, AlphaFold provided the most accurate answers in 25 cases; the closest competitor succeeded in only three.
In 2020, AlphaFold 2 achieved a breakthrough in predicting the 3D structures of proteins that astonished researchers. A problem that biologists had struggled to solve for over half a century was now largely solvable with the help of artificial intelligence; but Demis Hassabis was not satisfied.
In a meeting about the next step, his colleague John Jumper said they could predict the structures of millions of proteins in a month. Hassabis interrupted him, "Why not do it? We have to do it." And that is exactly what happened. DeepMind predicted the structures of over 200 million proteins and made the results freely available to researchers worldwide.
Hassabis and John Jumper received the 2024 Nobel Prize in Chemistry for predicting protein structures.
In 2024, Hassabis's old prediction also came true. He and John Jumper jointly received half of the Nobel Prize in Chemistry for developing the protein structure prediction method. The other half of the prize went to David Baker, a researcher at the University of Washington, for the computational design of novel proteins.
Hassabis signed his name in a book previously signed by Feynman, Curie, Bohr, Planck, Crick, and Einstein.
Fundamental Research Sacrificed in the Chatbot Wars
The Nobel Prize arrived right as Hassabis's world was changing. ChatGPT had disrupted everything, and Google, once considered the undisputed frontrunner in artificial intelligence, now had to keep pace with OpenAI and Anthropic.
In 2023, DeepMind merged with Google Brain, and Hassabis became Google's commander-in-chief in the war of language models: Gemini, chatbots, and product-by-product competition—a world that the scientist within him was never built for.
Of course, Hassabis did not deny the importance of product development. In an interview with Wired in June 2025, he admitted that following the introduction of the Transformer architecture in 2017, Google had focused too heavily on fundamental research and had missed the opportunity to turn this breakthrough into products more quickly.
In his view, Google needed to learn from this experience. The paradox of Hassabis's professional life became even more apparent here: a scientist who deeply valued long-term research now had to lead a laboratory where the speed of product releases also determined the fate of the competition—and it soon became clear that this competition came with a heavy price.
In June 2026, John Jumper, Hassabis's Nobel Prize co-recipient, left Google for Anthropic, and two of his colleagues followed him. Several other members of the AlphaFold core research team started their own startup, and another researcher joined OpenAI.
In 2026, Hassabis relinquished the day-to-day management of DeepMind to focus on research and the future of artificial intelligence.
Google reassigned the remaining team members to activities closer to Gemini, and in July, dissolved what was left of AlphaFold. Two years after the Nobel Prize, the award-winning team no longer existed.
On August 5, 2026, Google announced that Demis Hassabis would become Chairman of DeepMind and Chief Scientist of Alphabet, while Koray Kavukcuoglu, the Chief Technology Officer, would take over the day-to-day management of the lab. Sundar Pichai, Alphabet's CEO, stated that the new role allows Hassabis to focus all his attention on shaping the future of AGI. Jeff Dean, Google's legendary engineer, also left the company after nearly three decades in the same major shakeup.
According to Semafor, the groundwork for DeepMind's major changes began at least a year ago. Hassabis had gradually handed over the management of Gemini and consumer product strategy to Kavukcuoglu. He was not pushed out against his will; rather, he found less satisfaction in playing the role of a tech executive compared to working as a visionary scientist.
Hassabis said at the time that he wanted to focus on long-term strategy, accelerating scientific discoveries, and his work at Isomorphic; as if the scientist wanted to be a scientist once again.
A New Target; Drug Design Without Empty Promises
Hassabis knows that the AI industry's grand promises are not convincing to everyone. Industry executives have been promising a cure for cancer for years, and the public has grown tired. Concerns about the risks of AI and doubts regarding its tangible benefits have diminished the technology's popularity in the American public sphere.
However, Hassabis has another way in mind to win back the public's trust: instead of repeating grand promises, it must be demonstrated that AI can truly help cure diseases like cancer. In his own words, people will then be able to feel the benefits of AI directly in their lives.
Hassabis wants artificial intelligence to play a role in the discovery of real-world drugs
His path to achieving this goal goes through Isomorphic Labs; a company that spun out of DeepMind in 2021. Its name comes from the mathematical concept of isomorphism—a relationship between two structures that, despite potential differences in appearance or constituent elements, are equivalent in terms of their fundamental relations.
His and the lab's goal is not just to find a single drug. Hassabis says they want to build a suite of models and an entire pipeline that revolutionizes drug discovery itself. It feels like we are seeing his signature approach all over again, much like when DeepMind entered the protein structure prediction competition; he wasn't satisfied with just winning—he wanted to make the competition itself obsolete. He wanted to beat the game entirely.
Investors believe in this vision too. In May 2026, Isomorphic managed to raise $2.1 billion in funding—one of the largest private investments in the history of AI-driven drug discovery. However, the transition from in-silico molecule design to human clinical trials is progressing slower than initially planned.
Hassabis had previously promised that the first AI-designed drugs would enter clinical trials by the end of 2025, but in January 2026, he announced that this goal had been postponed to the end of that year. Naturally, he is not surprised by the slow pace.
When asked if he expected it to take this long, he said of course. Biology is perhaps one of the most complex systems we know, and Hassabis finds the mere functioning of AI in this field to be astonishing. He views the obstacles as "difficult, but not impossible" and echoes a pattern he has seen time and again: "Progress is slow, slow, slow; but then suddenly it accelerates and everything changes."
The Einstein Test
Hassabis has a strict definition for AGI, describing artificial general intelligence or human-like intelligence as a system possessing the full range of cognitive abilities of the human brain—not just a part of it. For him, the G in AGI, meaning general, is as important as the intelligence itself.
He calls his benchmark for achieving human-like artificial intelligence the Einstein Test. Can a system trained only on human knowledge up to 1901 reproduce Einstein's work from his miracle year of 1905? That is, can it independently discover the photoelectric effect, Brownian motion, special relativity, and mass-energy equivalence?
In Hassabis's test, AI must be able to recreate Einstein's discoveries using the knowledge of 1901.
Hassabis also explained in 2025 that solving a difficult problem and creating a novel one are two different things. An AI might one day solve one of the Millennium Prize Problems, but can it formulate a question as significant as the Riemann hypothesis on its own? In his view, machines do not yet possess such capability.
His ultimate ambition is not merely to build an intelligence that answers human questions; he wants to build an intelligence that formulates entirely new questions about the nature of reality.
According to Hassabis, we are still a few years away from an AI capable of inventing modern physics; however, his prediction is now closer than what he imagined when receiving the Nobel Prize. In early 2026, he said three to five years, and now he says we will reach AGI in two to three years.
At the same time, he cannot ignore the risks of AI. In the summer of 2026, Hassabis spent his nights focusing on every single word of an article in which he outlined his proposal for the responsible development of AI. In this article, he wrote that AI resembles a miracle; we have practically found a way to "make sand think."
He views AI more like the discovery of fire and electricity than the invention of the internet, and suggests that an institution funded by the industry and supervised by the federal government should test the capabilities and risks of leading models prior to release. However, he does not support hasty legislation, warning that improper regulations could harm the progress and safety of this technology.
His proposal was praised by both Sam Altman and Elon Musk, two executives whose conflicts are well-known to the world. But not everyone agrees with Hassabis. Critics say that successive predictions about the proximity of AGI merely intensify market competition and hype.
Hassabis believes that Artificial General Intelligence must possess the ability to discover new knowledge.
Mallaby, who generally holds an admiring view of Hassabis, critiques him by saying that Hassabis is so engrossed in winning the race and bringing AI into the world that his approach has taken on a messianic tone, diverting DeepMind's mission.
As the book recounts, Demis Hassabis's dream of an international institution akin to CERN to monitor the final steps before AGI also failed to materialize, much like other governance experiments, and he now relies more on his own presence in decision-making rooms. Regarding the progress and dangers of artificial intelligence, Hassabis says: "It’s not all doom, and it’s not all utopia."
Standing in the Foothills of the Singularity
In an article published in the summer of 2026, Hassabis wrote a memorable sentence: "When we look back at this era, I think we will realize that we were standing in the foothills of the singularity. At the dawn of a new era where artificial intelligence can transform many of our current conceptions of science, economics, and human life."
He now has a new preoccupation and believes the AI industry does not spend enough time figuring out what should happen after AGI. What should we value? What will this technology do to the human condition and our institutions? What will the purpose of life be? Hassabis says he believes in a magnificent future, provided we consciously build toward it.
Hassabis dreams of a day when artificial intelligence, alongside scientists, helps discover the unknown laws of physics.
Hassabis admits in his interviews that he does not have much time for other activities. He reads a lot, but almost exclusively work-related and scientific books, and it has been a long time since he last read a book purely for pleasure. Among television series, he liked the first season of True Detective and Chernobyl, both of which came out years ago.
Even gaming has become like work to him; for example, he tests AI agents by remaking games he had designed in his adolescence.
When asked if he ever imagines a day when all these sleepless nights pay off, he gave an answer reminiscent of a mountaineer. According to him, halfway through the work, progress becomes so difficult that you don't think about the summit; you only look at your next foothold and make sure you don't fall. But when he envisions the peak, he sees a clear picture.
The boy in Liechtenstein who was disheartened by the wasted talent of hundreds of chess players in a tournament hall now wants to build an intelligence that, alongside human scientists, tackles problems that no single mind can solve alone. He says, "Even if I could take back all my nights just to think, it still wouldn't be enough."
Comments