How Demis Hassabis Went From Idealist to Realist

Below top-k · 77

• mechanismHassabis's two-sided aversion to dominationc 0.90

  Hassabis genuinely refused to control others, but he equally refused to be controlled. Suleyman's repeated challenges to his authority over DeepMind's direction crossed a red line because DeepMind was Hassabis's identity.

• claimIf safety can't be negotiated inside one company, cross-lab governance is hopelessc 0.90

  Three years of failed governance talks inside an unusually open company suggested that negotiating common safeguards across multiple labs in multiple countries is essentially impossible.

• claimThe idealist-to-realist arc framed as growing upc 0.90

  Hassabis frames the shift not as defeat but as maturation: a transformative technology with unknown consequences demands adaptability and rejects black-and-white thinking, even while keeping one's values.

• evidenceHoffman pledged $1 billion to the visionc 0.85

  Fresh off selling LinkedIn to Microsoft and worth $3.8 billion, Hoffman committed more than a quarter of his net worth — an astonishing $1 billion — to DeepMind's plan.

• claimOpenAI faced the mirror-image governance problemc 0.85

  DeepMind was a for-profit seeking nonprofit-style guardrails around powerful AI; OpenAI was a nonprofit needing capitalist machinery to raise capital. Both groups were searching for the same capitalist/post-capitalist hybrid.

• claimThe Google-DeepMind impasse was structural, not personalc 0.85

  Google saw too much commercial potential in AI to release control. DeepMind saw too much existential risk to let commerce drive deployment. Neither side could let go, yet neither could agree.

• implicationTrustless negotiation actively destroys trustc 0.85

  Hassabis came to see that demanding trustless structures signals distrust of one's partners, burning the goodwill needed to influence decisions when real safety questions arise.

• implicationA well-meaning insider is flimsy comfort but perhaps the best availablec 0.85

  Once multiple labs in multiple countries are racing, no individual or charter can hold back civilian and military deployment — leaving the seat-at-the-table strategy as a thin but perhaps best-available reassurance.

• evidenceThe SpaceX safety board collapsed and seeded OpenAIc 0.80

  DeepMind's first AGI safety board meeting at SpaceX in August 2015 produced no agreements, derailed by personal tensions between Musk and Larry Page. Worse, Musk used what he learned there to found OpenAI as a direct rival, killing external oversight before it began.

• mechanismThe 3-3-3 board and Project Marioc 0.80

  Google's Alphabet restructuring offered a path: DeepMind could become a semi-independent 'bet' governed by three DeepMind directors, three Alphabet directors, and three independents. The DeepMind leaders code-named the talks Project Mario.

• evidenceGoogle reversed course in November 2016c 0.80

  After a cordial meeting with Sundar Pichai, chief legal officer David Drummond arrived in London with 'concerns' about a spin-out. Days later Pichai told Hassabis and Suleyman that AI was too strategically important to Google's core products to be hived off.

• mechanismPlan B: $5 billion and a walk-out threatc 0.80

  Hassabis and Suleyman planned to raise $5 billion from safety-minded investors — five times what OpenAI had claimed at launch — as leverage to walk away from Google if governance demands were not met.

• mechanismCapitalist intensity in service of post-capitalist impactc 0.80

  The structure was meant to combine the operational drive of a for-profit with an outcome that returned no equity gains to owners — run hard like a company, but with its benefits flowing to humanity rather than shareholders.

• claimPichai's counter was a split: spin out research, fold in Appliedc 0.80

  Over a vegetarian curry at his home, Pichai proposed splitting DeepMind in two—Hassabis would chase AGI in a semi-independent entity while Suleyman's Applied team would be absorbed into Google, with Suleyman running its applied AI from California.

• implicationThe end of DeepMind's ambition to ship its own productsc 0.80

  With Suleyman gone and Jim Gao's climate team also quitting to launch a startup, DeepMind retreated from marketing its own products. It would still feed applications into Google, but the Applied experiment was over.

• mechanismIndependent overseers have their own distorting incentivesc 0.80

  Google's leaders foresaw that independent directors would face skewed incentives just as the SpaceX safety meeting participants had — a structural problem with oversight, not a failure of personnel selection.

• mechanismEarning trust by leaning in beats demanding it from outsidec 0.80

  Hassabis redirected the energy from adversarial negotiation into building real relationships with Google — getting a seat at the table when safety issues arise and accumulating shared successes.

• claimSuleyman wanted a post-capitalist reconciling mechanism for AIc 0.75

  With external oversight stillborn, Suleyman set out to design a novel governance form that could balance profit, existential risk, and social justice. It was an explicitly post-capitalist ambition for an unprecedented technology.

• implicationBillionaires answer to nobodyc 0.75

  The episode dramatizes how the governance of transformative AI was being shaped by private individuals operating with no accountability — the same vacuum the global interest company was supposed to fill was filled in practice by unilateral billionaire judgment.

• implicationThe negotiation drained attention as the transformer arrivedc 0.75

  Constant 11-hour London–San Francisco trips left Suleyman distracted and Hassabis less alert to the transformer architecture when it appeared in summer 2017—a costly opportunity cost of the governance fight.

• claimMusk's bid exposed his power hunger and backfiredc 0.75

  The Tesla absorption proposal echoed Musk's 2014 pitch to swallow DeepMind, and it handed Altman the leverage to consolidate Brockman and Sutskever against Musk. OpenAI refused to attach itself to Tesla.

• exampleHassabis nearly secured $15 billion and independencec 0.75

  A converged plan would have given Suleyman command of Applied inside Google and Hassabis an independent global interest company running Research, with $15 billion in Google funding over a decade. Hassabis welcomed it partly because he was sick of legal negotiations.

• exampleDeepMind's failure to deny the Bloomberg leakc 0.75

  When Bloomberg called with "placed on leave," DeepMind's leadership debated whether to fight and never landed on a plan. They didn't plant the story but failed to deny it convincingly, which Suleyman read as Hassabis's tacit approval.

• claimYears of governance negotiation were worse than uselessc 0.75

  Hassabis and Suleyman, despite being rivals, agree the talks achieved nothing, were bound to achieve nothing, and were positively harmful in consuming energy and goodwill.

• contextSolving governance inside Google would have largely solved it everywherec 0.70

  At the time, DeepMind was the world's top AI lab and its strongest rival, Google Brain, sat within the same parent company. If Suleyman could fix governance internally, he would go most of the way to fixing it for the field.

• claimAI was too strategic to be treated like a moonshotc 0.70

  Pichai argued the 'bet' structure was meant for projects unrelated to Google's core business, like autonomous cars. AI, destined to be central to search and cloud, did not qualify — a position that effectively foreclosed the spin-out.

• contextThe DeepMind commitment dwarfed Hoffman's OpenAI pledgec 0.70

  The $1 billion figure was 100 times what Hoffman had pledged to OpenAI roughly a year earlier, signaling how seriously he took DeepMind's specific governance model.

• claimStartup pundits sold a failed science of entrepreneurshipc 0.70

  The body of advice marketed as a science of entrepreneurship by startup pundits has failed to deliver on its promise.

• claimPage and Brin had effectively checked out of the fightc 0.70

  When the next meeting came, Page arrived two hours late and Brin even later, leaving Pichai as the only Alphabet figure DeepMind could actually negotiate with—and quietly puncturing the Silicon Valley myth of permanent founder control.

• exampleThe Aviemore reveal landed as a spin-in plus spin-outc 0.70

  At DeepMind's Scottish off-site, Suleyman unveiled a 'Global Interest Company' org chart showing Applied AI absorbed into Google's Mountain View core and a separate independent research entity—a structure that left staff confused about whether they were getting independence or being absorbed.

• evidenceOpenAI's founders splintered over who would control AGIc 0.70

  By 2017 Sutskever and Brockman had fallen out with Musk and Altman; Altman was maneuvering to push Musk out, Musk poached an OpenAI scientist for Tesla, and Sutskever feared both men wanted absolute control of AGI.

• evidenceMusk crushed the rebellion by threatening to cut fundingc 0.70

  Musk responded 'Guys, I've had enough' and threatened to stop funding OpenAI. Within two days Sutskever and Brockman capitulated, Altman quickly cozied up to Musk, and the for-profit restructuring was shelved.

• exampleMusk's 3am counteroffer to fold OpenAI into Teslac 0.70

  Musk proposed that OpenAI spin into Tesla, using OpenAI talent to accelerate self-driving cars and self-driving profits to fund AGI. He framed Tesla as the only entity with a non-zero shot at counterbalancing Google.

• evidenceDeepMind invoked Putin and existential risk to lobby Alphabetc 0.70

  In 2018, DeepMind pitched Alphabet's board with rival AI hype, Putin's "ruler of the world" line, and warnings of public backlash. The dual ask: empower DeepMind to sprint for AGI and build a governance structure that could survive scrutiny.

• mechanismHassabis preferred internal comrades to managerial outsidersc 0.70

  DeepMind's scientific culture demanded humility from non-scientists, so Hassabis avoided hiring managerial stars from outside. He relied on long-standing personal ties stretching back to Cambridge and Elixir Studios, which is why Suleyman was indispensable.

• contextBullying complaints against Suleyman emerge in 2019c 0.70

  DeepMind employees alleged Suleyman used harsh language and intimidating messages to frighten subordinates. There were no claims of violence or harassment, but the pattern was serious enough to force a reckoning.

• exampleSuleyman's coerced acceptance of the chargesc 0.70

  Given a few hours to decide, Suleyman was told that accepting the complaints meant a sabbatical, a coach, and a possible return as ambassador, while disputing them meant likely firing and forfeited compensation. He accepted, believing his reputation would survive.

• implicationThe Bloomberg headline shredded the implicit contractc 0.70

  Suleyman believed accepting the charges had bought him his reputation. The leak destroyed any path back to DeepMind, and with the health group gone there was little left of Applied to return to.

• exampleDeepMind Health's review panel attacked the company to prove its independencec 0.70

  Even after DeepMind had fixed its data practices and was producing lifesaving algorithms, its hand-picked Independent Review Panel publicly impugned Google's motives — burnishing their own credibility rather than informing the public.

• exampleOpenAI's board collapse exposed nonprofit oversight as fictionc 0.70

  When OpenAI's nonprofit board tried to assert authority by firing Altman in 2023, he rallied financiers and ejected three directors — showing that hybrid governance gives no real power to the safety-minded.

• mechanismBright-line principles drawn in advance get drawn wrongc 0.70

  A safety charter negotiated years ahead of the technology will fix red lines in the wrong places; flexible judgment from people at the table works better than pre-committed rules.

• caveatThe no-weapons promise quietly brokec 0.70

  The DeepMind sale extracted a pledge that the technology would never be used for weapons or surveillance, but by 2025 Google was eagerly supplying AI to the national security complex.

• mechanismA company limited by guarantee, bound by charterc 0.65

  The reconstituted DeepMind would issue no shares and pay no dividends, taking the nonprofit-style legal form of a company limited by guarantee. Its obligations would run to the principles in its charter, not to investors.

• exampleHoffman acted without telling Altman or his own VC firmc 0.65

  He made the billion-dollar commitment without informing Sam Altman or his Greylock partners — a vivid illustration of how unconstrained a billionaire's solo decision-making can be.

• exampleThe good-cop/bad-cop pattern with Pichai and Drummondc 0.65

  Pichai received Hassabis warmly and seemed open to everything, only for David Drummond to arrive the next day and declare Pichai entirely opposed to Alphabetization. The pattern repeated, leaving DeepMind unable to find a real counterpart.

• evidenceGoogle walked the Aviemore plan back within ten daysc 0.65

  Days after the all-hands reveal, Google returned a heavily red-lined document showing Pichai was nowhere near approving the announced plan, forcing DeepMind's leaders to consider retracting the vision they had just shared with the company.

• exampleSutskever's ultimatum email called out dictatorship risksc 0.65

  On September 20, 2017, Sutskever wrote to Musk that the current structure would give him 'unilateral absolute control over the AGI,' and questioned Altman's motives given his simultaneous flirtation with a California gubernatorial run.

• contextApple's poaching of Giannandrea collapsed the dealc 0.65

  When Apple hired John Giannandrea in April 2018, Jeff Dean's promotion eliminated the slot Suleyman was meant to fill at Google. Suleyman had to retract promises already made to his deputies about moving to California.

• implicationAI governance needed nonprofit oversight to fill a political vacuumc 0.60

  Hoffman, having just watched Trump's election, believed AI required nonprofit oversight with democratic buy-in because politicians were too slow on technology. That conviction made him open to backing a DeepMind walk-away.

• contextHoffman framed AI as the defining technology of his lifetimec 0.60

  Hoffman justified the commitment by calling AI the most impactful technology of his lifetime, one that 'shouldn't be used to entrench a monopoly' — the same reasoning behind his OpenAI support.

• implicationThe Red Queen framing offers a better alternativec 0.60

  A Red Queen model of competition is proposed as a more useful lens for understanding entrepreneurship than the prevailing pundit playbook.

• exampleMusk quits OpenAI and calls an intern a jackassc 0.60

  At an all-hands, Musk announced his departure and tried to poach researchers for Tesla. When an intern asked whether building AI at for-profit Tesla contradicted his own principles, Musk insulted him and stormed out.

• evidenceThe 25-page report ruled the conduct misconductc 0.60

  After three months, an outside lawyer concluded Suleyman's management style amounted to misconduct, citing public humiliations and instructions to communicate off Google channels. Colleagues later argued it was standard founder behavior, but the charge sheet held.

• exampleGoogle's AI ethics council collapsed within daysc 0.60

  Google's 2019 Advanced Technology External Advisory Council disintegrated almost immediately after social media activists attacked one politically conservative appointee, forcing the rest to withdraw.

• mechanismAlphabet's logic made the spin-out plausiblec 0.55

  Google wanted to separate a money-gusher ad business from cash-burning moonshots, both to manage them better and to boost the stock price. The fact that the plan served Google's commercial interests made Hassabis and Suleyman believe it might actually happen.

• caveatThe walk-away threat had real legal weaknessc 0.55

  DeepMind staff were Google employees bound by noncompetes and IP clauses, and Alphabet could simply have told them to 'get back in your boxes.' Advisers knew that lifting a hundred people out to start a new company would be legally messy.

• mechanismNever threaten, just keep the option visiblec 0.55

  Advisers explicitly avoided ultimatums, never telling Google 'do this or we leave.' The art was to keep the walk-away option credible enough that Google took the negotiation seriously without forcing a showdown.

• claimHassabis preferred Alphabetization over a messy legal breakc 0.55

  Wary of a drawn-out court battle that would consume his energy, Hassabis judged that spinning out as an Alphabet company was the cleaner path to independence.

• contextPage still backed Alphabetization but deferred to Pichaic 0.55

  Larry Page told Hassabis he continued to favor spinning DeepMind out, but said the plan now required Sundar Pichai's buy-in—signaling that the founders had ceded operational authority.

• contextAltman explored the same restructuring menu as DeepMindc 0.55

  Needing capital to chase AGI, Altman called Reid Hoffman for money, considered a public-interest corporation, and even contemplated a cryptocurrency—two of the three paths mirroring DeepMind's parallel deliberations.

• exampleHassabis doubled down on Suleyman after the deal collapsedc 0.55

  They jointly courted Joe Tsai for a $1 billion walk-away offer, revived Pichai's spin-out plan, and forged a pact to avoid public recriminations. Hassabis even maintained their Sunday mint-tea ritual at the pub.

• contextSafety was baked into DeepMind from the startc 0.50

  Hassabis first bonded with co-founder Shane Legg over a 2009 lecture warning that superintelligent computers could subjugate or annihilate humans. That preoccupation with existential risk shaped the company's culture long before the Google negotiations.

• mechanismPublic-interest law as the lever against Googlec 0.50

  The team planned to argue that the British public had an interest in DeepMind breaking free because a spin-out would improve AI safety. The bet was that Google cared too much about its reputation to contest such a claim in court.

• contextAsilomar and the search for outside backersc 0.50

  At the January 2017 Asilomar AI conference — modeled on the famous genetics meeting of the 1970s — Hassabis and Suleyman pitched Reid Hoffman on financing a public-interest AI company if they broke free from Google.

• mechanismSuleyman lobbied via Kent Walker with credibility signalsc 0.50

  At Asilomar, Suleyman courted Alphabet's top lawyer Kent Walker by introducing him to UN weapons official Angela Kane and name-dropping Obama and Gore, trying to show that a governed spin-out would attract serious legitimacy.

• evidenceMusk's chart: OpenAI on a path of certain failurec 0.50

  In January 2018 Musk sent his co-founders a chart showing DeepMind and Google Brain produced most AI research, declaring OpenAI on 'a path of certain failure relative to Google' unless it changed course.

• caveatBrockman insisted moral high ground was OpenAI's real assetc 0.50

  Pushing back on Musk, Brockman argued conference papers were a poor metric and that 'our biggest tool is the moral high ground,' with fiduciary duty owed to humanity rather than to corporate competitors.

• evidenceBrin's prediction that outsiders use platforms for themselvesc 0.50

  At an Alphabet board meeting, Sergey Brin argued the panel's self-interested behavior was entirely predictable: give outsiders a platform and they will use it to advance their own reputations, not the company's projects.

• contextGovernment regulation was never coming to fill the gapc 0.50

  Hoffman's 2017 bet that governments would not act has held up, meaning the failure of company-level oversight leaves no backstop.

• exampleHassabis's secret hedge fund to fund independencec 0.45

  To bankroll a future independent DeepMind, Hassabis assembled around 20 researchers to train high-frequency trading algorithms, hoping to beat Jim Simons's Renaissance Technologies. Google did not approve, and the project was quietly disbanded.

• contextOpenAI's own governance ambitions echoed DeepMind'sc 0.40

  Sam Altman, inspired by Madison's notes on the Constitutional Convention, mused about letting the world elect representatives to OpenAI's board. The episode shows how the same governance problem was being grappled with across labs at the same moment.

• contextHoffman had already earmarked most of his wealth for philanthropyc 0.40

  He had previously decided that around 90 percent of his wealth would flow to philanthropic causes, which made a billion-dollar commitment feel within his existing plan rather than a radical departure.

• contextDeepMind advisers urged seizing Hoffman's offer to force a spin-outc 0.40

  Some advisers wanted to accept Reid Hoffman's anchor investment and proceed with independence, arguing the operational agility and equity upside would justify any legal fight with Google.

• contextThe public square punishes broad debatec 0.40

  Activists dominating online discourse were out to crush opponents rather than encourage debate, making any pluralistic advisory body politically untenable.

• contextHoffman as the American counterpart to Suleyman's idealismc 0.35

  Hoffman positioned himself as an unabashed idealist focused on helping humanity flourish — described as a grander, American version of Suleyman's aspirations, which is why the pitch landed.

• contextSuleyman returned as a powerless Google VPc 0.30

  After Google's review judged Suleyman's conduct as gray-zone, he was made a vice president in California but barred from managing anyone — a grand title with no court.

• contextBoth protagonists now run AI at rival US giantsc 0.30

  By 2024-25, Hassabis ran the expanded Google DeepMind while Suleyman, after a startup detour, headed Microsoft AI — two North Londoners atop AI at two American giants, despite their bitter split.