Summary

Hassabis’s core bet is not a product, but a single AGI machine that will “solve AI first, then use AI to solve everything.” He sees science as “a spiritual pursuit” and wants to build a machine capable of making discoveries greater than those of Einstein and Newton and exploring the universe; the financing story shows investors were betting on general-purpose R&D spanning games, life sciences, and foundation models, while accepting the tail risk that the technology could run out of control.
Google acquired DeepMind for roughly $500M-$650M, with talent valuation clearly anchoring the negotiations. Top researchers were valued at about $15M each, while Hassabis himself was reportedly valued at around $140M; 泓君’s retrospective is that the deal later produced AlphaGo, AlphaFold, and Gemini, as well as the Nobel Prize in Chemistry for Hassabis and others—“worth far more than a few hundred million dollars.”
周健工 believes DeepMind’s early lag behind OpenAI was largely driven by Hassabis and David Silver’s fixation on reinforcement learning. Hassabis long argued that language was “ungrounded,” placing large language models on the third branch of his research portfolio; but the program also notes that o1, DeepSeek R1, and other reasoning models are borrowing from reinforcement learning, showing that the competition is not about one route replacing another, but about the continued fusion of deep learning, Transformers, and reinforcement learning in post-training.
AlphaFold proved that DeepMind’s value lies not only in model leaderboards, but in applying AI to high-value scientific problems. The team moved from sequence modeling and image recognition to deep learning and Transformers, combining roughly 100,000 existing protein structures with self-generated distance data; it beat 97 teams in 2018, AlphaFold 2 scored 92.4 in 2020, and the work ultimately led to the 2024 Nobel Prize in Chemistry.
ChatGPT forced Google to recognize that research leadership cannot deliver “general” without engineering, products, and distribution. After GPT-4 arrived and Bard stumbled at launch, Google merged DeepMind and Google Brain in April 2023, named Hassabis CEO of Google DeepMind, concentrated compute, cut some blue-sky research, reduced the immediate release of frontier results, and revived skunkworks-style collaboration; Gemini, Flash Thinking, native multimodality, and long context later narrowed the gap, underscoring Google’s organizational leverage.
DeepMind’s governance history shows how difficult it is for AI-safety principles to withstand control, commercialization, and internal politics. The “three-three-three” ethics committee envisioned at the time of the acquisition never truly operated, and an independence plan collapsed for lack of funding; 泓君 notes that a Suleyman-led NHS project used 1.6M patient records to train a kidney-failure detection model, triggering a privacy controversy, while 周健工 adds that the media amplified the crisis and that Suleyman also did substantial data-governance work. The co-founder eventually left to become Microsoft’s AI chief, turning internal fractures into external competition.
The program does not cast Hassabis simply as a savior, preserving the tension between the “good man,” his appetite for victory, and an AI race slipping out of control. He says he does not want to control other people, yet also says that in any intellectual game, “once I play, I have to beat the other person”; a postscript at the end has him publicly warning that AI is becoming increasingly autonomous and that the consequences could be severe if it departs from its intended path—the unresolved question remains: “He wants to do the right thing, but can he?”

Deep dive

1. Hassabis Built AGI to Understand the Universe, Not to Launch a Product

泓君 puts the central contradiction at the opening: Hassabis runs Google’s AI research while spending his life trying to build something smarter than humans and understanding better than anyone that it could run out of control. Every day, he passes the intersection where Szilard conceived the idea of a nuclear chain reaction and thinks that what he is doing may be “as important as the atomic bomb.”
Sebastian Mallaby uses The God Machine as the book’s original title, presenting AGI as a “god machine” that could improve human welfare or do great harm. 周健工 believes the biography’s real question is not about technical specifications, but what “the small group of people” inventing the ultimate technology are actually like, because their motives may directly shape civilization’s fate.
泓君 says the shock from the recent arrival of DeepSeek was that AI may already be better than she is at developing podcast topics and interview outlines. Hassabis is therefore no longer merely a biographical subject, but a live question: how does a scientist think about a machine that is accelerating the replacement of human work?

2. Chess Gave the Prodigy Two Epiphanies: Intelligence Should Not Be Used Only to Win Games

Hassabis watched chess at 4, began playing at 5, joined Britain’s under-11 team at 9, became a master at 13, and ranked second globally among his peers. As a child, he needed two thick books to see over the chessboard, routinely played games lasting 8 to 10 hours, and had a father who drove him around Britain and Europe in a battered car for tournaments.
At around 11, he fought the Danish champion for 10 hours. He could have continued playing for a draw, but looked around at the roomful of intelligent people and suddenly asked: “Was I really going to spend my whole life playing chess?” He resigned voluntarily and became less fixated on the outcome of individual games, caring more about whether intelligence could solve problems larger than games.
A short book on computer chess then took him from the board to Shannon. In 1950, Shannon had already argued that if a computer could play chess, it might also perform many other cognitive tasks. For the first time, Hassabis saw that games could be an entry point to general-purpose computation rather than the endpoint of a life.

3. A Failed Game Startup Taught Him About Agents, Compute, and Execution

While working at Bullfrog Productions, Hassabis was already experimenting with characters that acted on environmental feedback: eating salty food would make them thirsty, and drinking water would change their state again. 周健工 sees this as a hazy early prototype of the agent concept—the characters were not simply running scripts, but forming feedback loops inside an environment.
After graduating from Cambridge, he and David Silver founded Elixir Studios and tried to simulate a city and thousands of independently minded characters in Republic: The Revolution. The graphics cards, CPUs, and memory available in 2003 could not support the ambition. The trade-show demo failed, and the team eventually had to cut large parts of the feature set just to ship, with the market response fading into indifference.
泓君’s takeaway is that even the most advanced algorithm remains trapped on paper without enough compute. 周健工 explains Hassabis’s repeated insistence on building things himself with Feynman’s line: “If you can’t build something, you don’t really understand it.” He was not a scientist who merely talked about AGI, but a practitioner trained by failure.

4. Reinforcement Learning Built DeepMind—and Later Helped Slow It Down

DeepMind was founded in 2010 with a mission compressed into “solve AI first, then use AI to solve everything.” Hassabis and David Silver initially believed the road to AGI had to run through reinforcement learning, and that future AGI should be a single model capable of handling every task.
The route also reflected a clash of schools. Geoffrey Hinton represented deep learning, while Richard Sutton represented reinforcement learning, and the two Canadian camps were once openly hostile to each other. Reinforcement learning is especially powerful in games with clear rules, boundaries, and rewards, where trial and error, feedback, and rapid optimization can be pushed to the limit.
Practice kept forcing the approaches together. DeepMind recruited from Hinton’s camp and from OpenAI, while OpenAI also recruited from DeepMind. Deep learning played the larger role in AlphaFold, while OpenAI’s later reasoning models and human-feedback training used reinforcement learning. The program notes that o1 and DeepSeek R1 are borrowing from AlphaGo’s playbook: reinforcement learning and deep learning are not winner-take-all alternatives, but are converging again in the reasoning-model era.

5. The Three Founders Complemented One Another, but Power Was Unequal From the Start

Hassabis met Shane Legg while pursuing a PhD in computational neuroscience at University College London. During a chance elevator encounter, Legg said he was heading to Silicon Valley for a “Singularity Summit,” connecting Hassabis to the network around Peter Thiel, Elon Musk, and others. 周健工 believes that connection itself was a key contribution to the founding team.
Mustafa Suleyman came from a struggling immigrant family, dropped out of Oxford after 2 years studying theology, and began running small businesses. His conversations with Hassabis focused on “social justice” and “how to do business.” He also proposed ventures such as partnering on apartment rentals, giving him a role in fundraising, operations, and commercialization.
The ownership structure was visibly asymmetric: Hassabis held roughly 9x Shane Legg’s stake and 14x Suleyman’s. That gap also planted the seed for Suleyman’s later push for an independent domain.

6. DeepMind’s Fundraising History Was a Stress Test for Contrarian Capital

London’s financial community could not understand AGI and even advised the team to join a financial institution and do quantitative trading. The first British investor, however, concluded from religious conviction that it was “the god machine” and invested several hundred thousand pounds. Hassabis’s adviser at MIT put up the only $100,000 he had because he believed Hassabis would eventually win a Nobel Prize on the level of Feynman or Crick.
The Founders Fund partner who truly pushed the investment compared Hassabis to Musk, while Peter Thiel himself remained skeptical of the business model. The fund bypassed its usual investment-committee process; its first check was about $2.3M, which according to 泓君’s recollection came close to buying half the company, followed by another check of roughly $9M. After 3 rounds, its stake exceeded the combined holdings of the 3 founders.
Thiel stopped adding capital in 2013 not because he had decided AI was a failure, but because the AlexNet breakthrough had sent talent prices soaring and the industry had begun to reach consensus. In his framework, “consensus means a bubble”; continuing to invest would mean committing more capital for a less certain return.
The pivotal Series B raised about $25M, with the group led by 周凯旋 contributing roughly $10M to keep the company afloat when Founders Fund was unwilling to lead. The program also says Musk had promised to invest $5M. The money was not buying mature revenue, but keeping alive a scientific ambition that was almost impossible to commercialize at the time.

7. Google Bought DeepMind by Winning Over Scientists With Vision and Closing the Deal on Talent Pricing

Musk bragged about his DeepMind investment to Larry Page while the two were on a private jet. Page quietly wrote down the name on his Android phone, and Hassabis soon received an email from Google’s investment team. The program traces an ironic chain: Musk, a potential competitor, ended up putting DeepMind on Google’s radar.
Zuckerberg also offered generous terms and was even prepared to give several founders “an amount of money they couldn’t refuse” personally. But when he listed AI, VR, AR, and 3D printing together as hot sectors, Hassabis was disappointed. He wanted someone who treated AGI as the central mission, not a buyer cataloging every trend.
Larry Page reached him directly: “Why don’t you come to Google? Isn’t your goal to pursue AGI? I’ve already prepared everything for you here.” Google could provide capital, compute, and a scientific culture, while Hassabis could finally escape the exhausting cycle of fundraising.
Pricing was still a finely calibrated negotiation. Hinton’s 3-person company sold for $45M, creating a benchmark of $15M per top researcher. Google initially priced DeepMind at $10M per person for its 30 to 40 researchers; according to the program’s correction, the final transaction was roughly $500M-$650M, not a definitive $650M.

8. AlphaGo Chose Go Because It Was the Hardest and Clearest Public Proof

After Google acquired DeepMind, Hassabis proposed to Page that it beat a Go world champion and casually gave himself “2 years.” Chess had already been conquered by IBM’s Deep Blue, while Go’s branching factor was widely considered impossible to brute-force. Winning there would prove that DeepMind was a world-class AI lab.
Hassabis rated himself only an amateur-level Go player. He understood the game, but was nowhere near professional strength. In 2016, AlphaGo defeated Lee Sedol in Seoul, giving the world its first look at DeepMind’s algorithms, Google’s compute, and concentrated research under one roof.

9. AlphaFold’s Breakthrough Came From Repeatedly Admitting the Previous Model Was Wrong

Hassabis was drawn to the Anfinsen problem during his PhD: nature uses roughly 20 common amino acids, and if their folding rules could be decoded from sequence, large numbers of protein structures might become predictable. During the 2016 Go match in Seoul, he was already discussing with David Silver whether proteins should be the next target now that Go had been conquered.
Foldit turned protein folding into a scored game, while CASP converted prediction into a global competition—an ideal setup for Hassabis, who had to win whatever game he played. Traditional X-ray crystallography often required multiple PhDs working for a long time. The structures accumulated across human history were later described as numbering roughly 100,000.
The team first treated the problem as sequence modeling and used recurrent neural networks. When it realized that amino acids also had folding and contact relationships, it switched to convolutional neural networks and image recognition. John Jumper brought experience using molecular dynamics at Schrödinger, but ultimately chose machine learning: if the model itself is wrong, greater simulation precision can simply take you farther down the wrong road.
On the data side, the team combined roughly 100,000 existing structures with about 20,000 human proteins and millions of plant, animal, and bacterial proteins from UniProt. It also converted “contact” into “distance” and generated its own predictions of the spacing between amino acids. It beat 97 teams in its first competition in 2018; in 2020, AlphaFold 2 scored 92.4 and upended the field. The program identifies the Transformer architecture as the decisive element.

10. The Real Post-Acquisition Conflict Was That Independent Research and Google Control Could Not Both Be Maximized

Hassabis once said DeepMind had led Google to establish an “AI First” strategy. But when asked how DeepMind and Google Brain divided responsibilities or measured performance, he could offer only broad references to prediction and recommendation. The two elite teams initially operated independently: Google had the Transformer invented by Google Brain and DeepMind had reinforcement learning, but the company did not combine the two assets early enough.
The AI-safety and ethics committee envisioned in the acquisition talks used a “three-three-three” structure—3 external members, 3 from Google, and 3 from DeepMind. Its first meeting was even convened by Musk. But Google’s legal and other departments objected, and the mechanism never truly operated. Writing safety principles into a transaction did not translate into effective governance rights.
An Alphabet restructuring briefly created an opening for DeepMind to go independent. Google had promised $15B in research funding over the following 10 years, while Suleyman was to lead applications and seek a vice-president role. After the arrangement changed, they launched a secret “Mario Plan,” seeking funding from Reid Hoffman, 蔡崇信, and others and designing a global public-benefit structure modeled on OpenAI. The spinout ultimately failed because it could not raise enough money.

11. Suleyman’s Exit Turned a Governance Rift Into Competition Between Microsoft and Google

泓君 notes that an NHS project led by Suleyman sought to use 1.6M patient records to identify kidney failure, triggering privacy concerns and a backlash over British medical data falling into the hands of a Silicon Valley giant. 周健工 relays Mallaby’s view that the media amplified the episode and that Suleyman also did substantial data-governance work. Whatever the allocation of responsibility, DeepMind’s and Hassabis’s reputations were damaged.
Internal complaints also focused on Suleyman’s management style. Hassabis launched an investigation rather than burying the matter. Suleyman was subsequently marginalized, spent 2 years at Google holding a largely nominal vice-president role, then left to found Inflection AI and later became Microsoft’s AI chief. He and Hassabis, now running Google’s AI effort, ended up on opposite sides of the competition.
周健工 stresses that these organizational relationships shaped the industry. He believes Musk helped create OpenAI because he could not control DeepMind after Google acquired it; OpenAI’s nonprofit structure then encouraged DeepMind to seek independence. What looks today like model competition between companies often began with unresolved fights over control and safety governance a decade earlier.

12. ChatGPT Exposed Hassabis’s Biggest Strategic Blind Spot

DeepMind’s research priorities were long explicit: reinforcement learning first, neuroscience modeled on the human brain second, and induction from data third. Large language models were only a branch of that third layer. Even after repeated hires from OpenAI issued warnings, the team did not start paying attention until GPT-2 and did not become truly alarmed until GPT-3.
DeepMind then developed Sparrow, while Google Brain developed LaMDA, later reworked into Bard. Both aimed to surpass GPT-3’s 175B parameters. The program says that when Google was contemplating models with roughly 400B-600B parameters, OpenAI was secretly training a larger system reportedly approaching 2T parameters. Beyond research cadence, the larger difference was willingness to release.
Influenced by an incident in which an engineer claimed that an AI was conscious and was subsequently fired, Google decided to spend several more months on safety, human-feedback learning, and alignment. ChatGPT launched first. 泓君’s conclusion is that science without engineering, models without products, and R&D without commercial understanding are all insufficient to advance AGI: “If you don’t let more people use it, what does ‘general’ even mean?”
On the question of grounding, Hassabis insisted that AGI had to be “grounded” and that language itself was not grounded. Ilya Sutskever argued that language indirectly contains large amounts of human experience and logic. ChatGPT forced Hassabis to acknowledge the possibility of another route and triggered his competitive instinct: “They’ve driven a tank onto our lawn.”

13. Google’s Counterattack Began With Organizational Restructuring, and Only Then With Bigger Models

After GPT-4 launched and Bard stumbled, Google merged DeepMind and Google Brain in April 2023 and named Hassabis—not Jeff Dean—CEO of Google DeepMind. Sergey Brin also returned to the front line to inspect code, data, and models. 周健工 believes this showed that Pichai had always viewed DeepMind as a strategic asset to be activated sooner or later.
The merger unlocked an extreme concentration of resources. The book says DeepMind at one point consumed more compute than Gmail’s billions of users. The new organization cut some blue-sky research, stopped immediately publishing frontier results, and moved more people toward engineering and products, with research openness giving way to competitive speed.
The working style also reverted to its early skunkworks model: the company’s best people concentrated on one objective, and any breakthrough in any module was immediately shared internally and adopted. 周健工 attributes the gradual impact of this high-intensity collaboration to first-generation Gemini, Flash Thinking, later Gemini 2 and Gemini 3, native multimodality, and long context.
泓君 believes the balance began to reverse after Google I/O 2025, with Google’s models moving to the top of multiple benchmarks. But she also stresses how quickly leadership can change—every “3 months, 6 months”—and says Claude could take the lead next. The conclusion is not that Google will win forever, but that it has finally reorganized its core advantages in talent, compute, and distribution.

14. Hassabis’s Advantage Is Turning Cross-Disciplinary Intelligence Into Executable Goals

周健工 interviewed Hassabis for about 1 hour in May 2017. His first impression was of a short, balding man who looked like an ordinary research-institute scientist. Once he began speaking, however, his pace was rapid and his thinking fluid, clear, and sharp; he moved freely across fields of knowledge. “Any intellectual challenge might be no obstacle at all for this person.”
He chose the word “wisdom” rather than “smart,” because Hassabis did not deliver a PR presentation played countless times; he gave “the freshest idea” for the specific problem at hand. 周健工 thinks Jensen Huang’s communication has a somewhat polished, PR-like quality, while Mallaby’s more than 30 hours of private conversations come closer to Hassabis’s subtle and authentic inner voice.
What Hassabis wants most is not wealth or corporate power, but a return to university research in physics and life sciences and a machine capable of making discoveries greater than those of Einstein and Newton—“decoding the entire universe and reverse-engineering God.” 周健工 therefore places him in the tradition of Europe’s Scientific Revolution and humanism, rather than among typical Silicon Valley entrepreneurs.

15. Optimism Has Not Removed the Risk; It Has Made the Race Harder to Stop

Compared with the AI-doom warnings from Geoffrey Hinton and Yoshua Bengio, Hassabis has long believed AI could be the most beneficial technology in human history because major challenges “have no solution other than AI.” He imagines AI taking humanity into a post-scarcity era: mining asteroids, solving nuclear fusion, extracting fuel from seawater, and pushing civilization to a new level.
周健工 rejects pure apocalypticism and says the “Oppenheimer complex” may exaggerate both inventors and the power of their inventions. He favors a human-centered approach in which people and machines collaborate and AI is kept in a sandbox: “Let them explore, let them try,” but never let them harm human society.
The scientist’s dilemma is captured by a von Neumann-style formulation quoted in the book: technology’s power is enormous, but refusing to fulfill the duty of discovery also violates scientific ethics. Once that power is released, the inventor may become “the most hated person in the world.” 泓君 adds that the pleasure of exploration is so strong that top scientists often cannot order themselves to stop.
A postscript at the end of the program further shifts the balance: Hassabis has begun publicly warning that AI systems are becoming increasingly autonomous and that the consequences could be severe if they depart from their intended path. Existing international cooperation “may simply be insufficient.” The optimism remains, but the risk has moved from abstract philosophy onto the urgent governance agenda.

16. Being a “Good Person” Is Not a Governance Mechanism, and the Biography Does Not Resolve the Paradox of Control

泓君’s reservation about Mallaby’s book is that its portrait of Hassabis is “a little too positive.” It spends substantial space presenting his ideas while revealing fewer contradictions at decisive moments; people may also construct more flattering explanations for themselves after success. 周健工 acknowledges that the author treats him as a “national treasure of Britain” and is visibly sympathetic to Hassabis and his British outlook.
More than 30 hours of conversation gave the book an unusual sense of candor, but may also have amplified in full the self-narrative Hassabis wanted to project. 周健工 is inclined to believe most of it, while agreeing that if readers do not find the statements sincere, they can look like a form of PR subtly engineered by Hassabis himself.
The sharpest counterevidence is his appetite for victory: outwardly easygoing and claiming not to want control over others, he nonetheless says that in chess, poker, table football, shogi, and backgammon, “once I play, I have to beat the other person,” while possessing the ability to persuade everyone. The program ends not with a verdict on his character, but with an institutional question: even if he truly is “a good person,” can goodwill withstand an AI machine accelerated by capital and competition?

Ep. 003 - Deep Dive on NVIDIA Vera Rubin VR NVL72 (AI Supply Chain) | Jordan Nanos, Myron Xie, Copper Wei (Wega), Howie Universal Medical Intelligence: OpenAI's Plan to Elevate Human Health, with Karan Singhal