Stalled Exit Signals End of Google's Pre-LLM Scientific Era

Woofun AI reports that the stalled departure of Demis Hasabis from Google signals the definitive end of the pre-LLM era, a period defined by scientific curiosity rather than immediate commercial pressure. Originally intended to coincide with the exit of Jeff Dean, Hasabis' planned departure was halted by management fears that the simultaneous loss of two top AI leaders would severely impact stock prices.

On August 5, Google announced a restructuring where Hasabis voluntarily handed over day-to-day management of Google DeepMind to become chairman and chief scientist at Alphabet, while Koray Kavukcuoglu, who had worked at DeepMind for thirteen years, assumed daily operations and now reports directly to Sundar Pichai. This arrangement allows Google to retain Hasabis' reputation, his 2024 Nobel Prize in Chemistry, and the stability of Isomorphic Labs, effectively preserving the last intact model of the pre-LLM era where scientists made decisions, the company covered costs, and profits were not a priority.

The strategic shift in titles reflects a broader change in management philosophy, where longer titles on business cards mask fewer people under direct management. By keeping Hasabis in a high-profile role overseeing AGI development and scientific research, Google retains a name that can temporarily reassure investors, even as the world's most powerful tech company finds it increasingly difficult to accommodate the scientist who brought it AlphaGo, AlphaFold, and Nobel medals in the old ways. The decision to delay his separation is not merely about reputation but about managing the transition from a research-driven culture to one focused on industrial scaling. This unfinished separation marks the final act of an era where the company could afford to wait for returns, a luxury that is rapidly disappearing as the focus shifts to capital expenditure and hash rate.

Jeff Dean's legacy is deeply intertwined with this transition, having joined Google in 1999 to work on the underlying systems for search, advertising, and distributed computing. He later helped found Google Brain, becoming one of the company's most legendary engineers, with employees joking that a snap of his fingers could restart half the Internet. When Dean announced his resignation, some joked that Google had finally 'open-sourced' him to the world, a stark contrast to fourteen years earlier when he brought some of the world's rarest AI scientists to the company. This reversal highlights the changing dynamics of talent acquisition and retention, where the value of individual scientists is now measured against the backdrop of massive industrial infrastructure rather than pure research potential.

The pivotal moment in this history occurred in December 2012, in Room 731 of the Harrah's hotel by Lake Tahoe, where one of the most famous talent auctions in modern AI history took place. The auctioneer was Geoffrey Hinton, a British-Canadian scientist known as the 'godfather of AI,' who had studied neural networks for decades. Suffering from severe back pain since 2005, Hinton had adapted his life to standing or lying down, using an upside-down trash can as a table surface to place his laptop on while waiting for bid emails via Gmail.

He and two students had just founded DNNresearch, a product-less company with no revenue, just the three of them and their newly created AlexNet. Named after student Alex Krizhevsky, AlexNet had far outperformed traditional methods in the ImageNet image recognition competition, suddenly boosting neural networks after decades of neglect. Hinton's other student, Ilya Sutskever, would later become co-founder and chief scientist of OpenAI, but at that moment, they were in Room 731, watching bids arrive.

The mechanics of the DNNresearch acquisition were simple yet intense: each new price required other companies to match within an hour, with bids increasing by at least $1 million. Baidu, Microsoft, Google, and DeepMind, a London-based company founded just two years earlier, all participated. DeepMind, with less cash, bid with shares and was quickly pushed off the table. The remaining three companies kept raising the bid until it reached $44 million.

Hinton stopped the auction, slept, and decided the next day to sell to Google, prioritizing where he and his students would work over pushing the price higher. Jeff Dean and Google's engineering lead, Alan Eustace, celebrated Hinton's 65th birthday before the auction, aiming to assess the capabilities of the two students around Hinton. Ultimately, Google paid $44 million for this product-less company, essentially acquiring those three people, a move that DeepMind, which didn't buy Hinton at Lake Tahoe, would later regret as it grew into exactly what Google wanted.

Woofun AI data shows that just over a year later, in June 2013, the story shifted to a poorly organized birthday party in Tarrytown, New York, where Elon Musk's wife, Talulah Riley, rented a castle to celebrate his birthday. Guests wore samurai costumes, and a sumo champion weighing around 350 pounds was brought in, whom Musk personally challenged, spraining his neck. Google co-founder Larry Page, who suffered from vocal cord issues and could only speak in a low voice, took Demis Hasabis aside outside the castle. Page asked Hasabis if he needed to create another Google, arguing that it would cost him the best decades of his career.

Google already had data centers, engineers, cash, and products worldwide; what needed addressing was intelligence. Hassabis, a chess prodigy who designed a video game at 17 and studied cognitive neuroscience, had founded DeepMind in 2010 with Mustafa Suleiman and Shane Legg, aiming for AGI. Tired of fundraising, which felt like throwing stones into an endless pit, Hassabis was persuaded by Page's offer to save the decades needed to build another company. In 2014, Google acquired DeepMind for around $650 million, keeping the lab in London and promising an AI ethics and safety committee, a commitment that convinced Jaan Tallinn, Skype co-founder and early investor, to choose Google over purely financial offers.

The clash of values between science and capital expenditure became evident in the two $100 million deals that occurred eighteen months apart. Hassabis used $100 million to illustrate the priceless value of a medal, while Buffett used $100 million to question the company's future worth. When Google acquired DeepMind, money followed scientific questions, purchasing untouchable research time. By 2026, Alphabet's annual capital expenditure was expected to approach $200 billion, with most going into AI hash rate and data centers. Money began to take precedence over research, turning into chips, land, electricity, servers, and depreciation. Data centers have no sentiment; every time a server room lights up, it adds debt to the books owed to the future. LLMs transformed intelligence from a laboratory science problem into a massive civil engineering project, marking a fundamental shift in how AI development is funded and executed.

Early conflicts emerged in August 2015, when the AGI safety committee agreed upon during the DeepMind acquisition held its first meeting at SpaceX. Hassabis, Suleiman, Musk, Google executives, and Reid Hoffman, co-founder of LinkedIn, were present. Suleiman wanted a governance structure not entirely controlled by Google's shareholders, but no agreements were reached. Beneath discussions of human destiny lay intense conflicts of interest between Google and Musk, personal disputes between Musk and Page, and differing views on AI risks.

After the meeting, Hassabis and Suleiman launched 'Project Mario,' aiming to transform DeepMind into an organization with independent directors serving global public interests. This independence plan struggled repeatedly before being completely abandoned in 2021, as Google was reluctant to let go of AI, and DeepMind couldn't do without Google's hash rate and funding. Also in 2015, at the Roswood Hotel on Sand Hill Road, Sam Altman, then head of Y Combinator, invited Ilya Sutskever to dinner.

Musk, Greg Brockman, previously in charge of technology at Stripe, and others were already there. They worried that Google and DeepMind were gaining too much advantage and wanted to rebuild a lab as a balancing force. OpenAI was born a few months later, with Sutskever leaving Google to become co-founder and research leader. The early version was a non-profit, unconstrained by financial returns, allowing researchers to publish papers, blogs, and code, with patents meant to be shared globally.

There was no ChatGPT, no subscription fees, no enterprise APIs, and no endless data center bills. Scientists could focus on the future, leaving money-making to later. OpenAI emerged out of fear of Google's dominance but inherited DeepMind's original ideals. Later, Scaling Law dominated AI development, with models growing larger and bills longer. In 2019, OpenAI established a for-profit subsidiary; in 2022, ChatGPT appeared on screens.

The golden age of research was epitomized by AlphaGo, AlphaFold, and other breakthroughs. In March 2016, AlphaGo played a five-game match against world Go champion Lee Sedol in Seoul. By the 37th move of the second game, AlphaGo placed a black stone in a position rarely chosen by humans, causing Lee Sedol to leave for the smoking room for over ten minutes. He lost the game anyway. The $1 million prize money reflected DeepMind's ethos, with Hassabis promising it without Google's financial approval. After winning, the prize was donated to charity organizations and Go associations.

Scientific questions came before financial procedures. Then came AlphaZero, MuZero, and AlphaFold. AlphaFold turned amino acid sequences into 3D structure predictions, partnering with the European Bioinformatics Institute to create a free database with over 200 million predicted structures, covering most known proteins. More than 3 million researchers from over 190 countries have used it. AlphaFold didn't charge monthly fees, nor did it have a business growth curve for earnings reports. It saved scientists months and years, brought Google two Nobel medals, but generated almost no commercial revenue.

Under pressure from ChatGPT, Sergey Brin, a co-founder who had faded from daily work, returned. Brin recalled meeting an OpenAI employee who asked him a question, leading him to become increasingly involved in Gemini. In 2023, Pichai merged Google Brain and DeepMind into Google DeepMind, with Hassabis becoming CEO and Jeff Dean becoming chief scientist. This seemed to be the peak of Hassabis' power but also the end of the old DeepMind as an independent lab, integrated into Google's models, applications, and business systems.

The exodus of key scientists accelerated this transition. John Jumper, leader of the AlphaFold project, went to Anthropic; Noam Shazeer, author of the Transformer paper and co-leader of Gemini, went to OpenAI. Later, Jeff Dean, along with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, founded Discovery Loop. Dean and Ghemawat built one of Google's earliest distributed systems, while Vinyals and Le were at the heart of modern AI research. Discovery Loop was registered as a nonprofit organization.

In his farewell letter, Dean explained that an independent company could make decisions that 'might not align perfectly with the company's pure financial interests.' Google became the founding investor and cloud computing partner of Discovery Loop, funding them to continue working outside since it couldn't keep them within its organizational structure. This marks how the pre-LLM era ended: no one publicly tore down ideals, no one stormed out in anger.

The old lab was integrated into the product line, and old scientists used parent company funds to research outside, retaining dignity. Hasabis hasn't left yet, but the 2013 agreement has lost its underlying condition—time. What Page referred to as 'everything' is much more today than thirteen years ago. Google has more money, machines, engineers, and data centers, so much so that Buffett is willing to invest $100 million, betting on returns. But the core assets Page handed to Hassabis are dwindling.

From the upside-down trash can in Room 731 by Lake Tahoe to Alphabet's nearly $200 billion in annual capital expenditure, it only took AI fourteen years. The room turned into a data center, the small company of three people into an industrial system supported by hundreds of thousands of chips, and hourly talent bidding into quarterly model competitions. Those people turned AI from a small circle into the largest industrial system of our time. Once everything grew too big to imagine, they began to leave one after another.

Hasabis was also planning to leave, but Google turned the clock back on his departure.