Hashmasks Exit, Final Scene of the Pre-LLM Era

Original Title: "Hassabis Exits, the Final Scene of the Pre-LLM Era"
Original Author: Watcher Beating

Hassabis was originally supposed to leave as well.

According to industry sources cited by Pathfounders, Demis Hassabis, co-founder of Google DeepMind and the 2024 Nobel Prize in Chemistry laureate, had initially planned to depart Google at the same time as Jeff Dean. The management was concerned that the simultaneous departure of the two AI leaders would significantly impact the stock price, so they managed to convince Hassabis to stay. Insiders suggest that he may still leave once the situation stabilizes.

On August 5th, the story presented by Google to the public was much more dignified.

Hassabis voluntarily relinquished the day-to-day management of Google DeepMind to assume the roles of Chairman and Alphabet's Chief Scientist, continuing to oversee AGI, scientific research, and the drug research company Isomorphic Labs. Koray Kavukcuoglu, who has worked at DeepMind for thirteen years and is Google's Chief AI Architect and one of the early creators of DeepMind's deep learning team, took over the day-to-day operations. He will now report directly to Sundar Pichai.

Big companies are good at saying goodbye.

The farewell of a prominent figure often begins with an elongated title. Chairman, Chief Scientist—more words on the business card, but fewer direct reports. Google retained Hassabis' prestige and Nobel Prize, providing investors with a name that could temporarily reassure them.

The world's most powerful tech company found it challenging to continue accommodating the scientist who brought them AlphaGo, AlphaFold, and a Nobel Prize in the same old way. Now, all they can do is retain the person and find a less ugly day for the possible departure.

Google and Hassabis together maintained the last intact set of lifestyles of the pre-LLM era. Scientists decided on the problems, and the company footed the bill. Research was conducted on an annual basis, products could be delayed, and profits were not an initial concern.

The ongoing separation between Google and Hassabis has become the final scene of the pre-LLM era.

One Hour

Hassabis was originally set to leave on the same day as Jeff Dean. Jeff Dean is one of Google's most legendary engineers, having joined Google in 1999. He worked on the underlying systems of search, advertising, and distributed computing before co-founding Google Brain. Google employees have created numerous jokes about him, as if at the snap of his finger, half the Internet would need to restart.

After he announced his resignation that day, someone joked that Google had finally "open-sourced Jeff Dean to the entire human race."

Fourteen years ago, he did the exact opposite. He brought in the world's scarcest group of AI scientists to Google.

In December 2012, at Lake Tahoe, USA, in Room 731 of the Harrah's Hotel. One of the most famous talent auctions in modern AI history took place at a table at the end of two twin beds.

The auctioneer, Geoffrey Hinton, a British-born Canadian scientist who had been researching neural networks for decades and later became known as the Godfather of AI, sat there. He had a severe back injury and had not been able to sit properly since 2005. Once he sat down, his intervertebral disc might slip again, leaving him bedridden for weeks.

He transformed his entire life to stand and lie down, lying across the back seat when driving, kneeling at the table to eat, and in the hotel room, he would flip a trash can onto the table, then place his laptop on top of it, standing by to receive new auction bids in his inbox.

Hinton and two students had just founded a company called DNNresearch, with no product, no revenue, the entire assets of the company were the three people in the room, and the AlexNet they had just created.

AlexNet was named after Hinton's student Alex Krizhevsky. That year, it significantly outperformed traditional methods in the ImageNet image recognition competition, causing the long-neglected neural networks to suddenly rise to prominence.

Another student of Hinton, Ilya Sutskever, would later become a co-founder and chief scientist of OpenAI. But at this moment, he stood with his teacher in Room 731, watching the bid emails flood into Gmail.

The auction rules were simple: with each new price, other companies had an hour to decide whether to follow or drop out, with an increment of at least $1 million.

Baidu arrived, Microsoft arrived, Google arrived, and another London-based company that had only been around for two years, DeepMind. It didn't have as much cash on hand as those giants, so it could only bid with its own shares, and soon was squeezed out of the table.

The remaining three companies continued to raise the price, shouting all the way to $44 million. Hinton called off the auction, went to sleep, and the next day decided to sell the company to Google.

He did not continue to increase the price. At this point, the money was already enough; what was more important was where he and his two students would work next.

Jeff Dean and Google's Chief Engineer, Alan Eustace, accompanied Sutton to his auction after celebrating his 65th birthday. They were mainly interested in seeing how the two students next to Sutton were doing. Later, Google spent $44 million to acquire this product-less company, essentially to acquire these three individuals.

The Best Decades

DeepMind did not acquire Sutton at Lake Tahoe. Over a year later, it grew into what Google wanted to acquire.

The story took place at a chaotic birthday party.

In June 2013, Musk's wife, actress Talulah Riley, rented a castle in Tarrytown, New York, to celebrate her husband's birthday. Guests wore samurai costumes, and a 350-pound sumo wrestling champion even attended. Musk personally took on the champion and ended up injuring his neck.

Google's co-founder Larry Page also attended the party. During those years, he was struggling with a vocal cord issue, speaking with a very low voice. Page called Hasabis outside, and the two took a walk around the outside of the castle.

Outside, he asked:

“Why don't you use what I've already built?”

Hasabis had a sudden realization. This statement meant, “You don't have to build another Google,” because that would take away the best decades of his career. Google already had data centers, engineers, cash, and products worldwide. What needs to be addressed is intelligence, so why rush to busy oneself with starting a company.

Hasabis was a chess prodigy in his youth, participated in designing a video game at seventeen, and later studied cognitive neuroscience. In 2010, he co-founded DeepMind with Mustafa Suleyman and Shane Legg, with the goal of AGI from the start.

With the company only a few years old, he was already fed up with fundraising.

Researching intelligence is like throwing stones into a bottomless pit and then having to come out every so often to explain to investors why there is still no echo from the pit to this day. Moreover, after each round of funding, he had to relentlessly start the next round of fundraising.

Page's answer to him was to skip decades of building another Google and let Google take on those times of invisible returns for now.

In 2014, Google acquired DeepMind for approximately $650 million. The lab remained in London, keeping its name and research culture. Google also pledged to establish an AI Ethics and Safety Board to oversee how these technologies are used.

Skype co-founder and early DeepMind investor Jaan Tallinn later recalled that Google's purely financial offer was not superior; it was the commitment to the safety board that was a key reason DeepMind ultimately chose Google.

This was the most charming pre-LLM era engagement.

Google promised money and patience, while DeepMind delivered the potential for AGI. As for when to make money, that could wait.

At that time, Google was a behemoth, able to afford losses and wait.

Two 100 billion

Before Christmas 2024, British journalist Sebastian Mallaby met Hassabis at a London pub. They had spoken many times before.

On that occasion, Hassabis pulled out a leather case from his backpack. Inside was a heavy golden disc.

He had just received the Nobel Prize in Chemistry from Stockholm.

Hassabis and DeepMind scientist John Jumper jointly received the Nobel Prize in Chemistry that year for AlphaFold. Six months after Dr. Jumper graduated, Hassabis handed him the AlphaFold project to lead. Later, this team solved the protein folding problem that had plagued the biology community for decades.

The Nobel Prize was already the secular world's highest award for scientists, Hassabis told Mallaby:

"Even if you offered me $10 billion for this Nobel Prize, I would refuse."

In June 2026, Alphabet initiated a $80 billion equity financing round to expand its AI infrastructure. Buffett's Berkshire Hathaway acquired $10 billion worth of Alphabet stock through a private placement.

By July, the investment firm led by Buffett for sixty years held over $31 billion in Alphabet positions.

Buffett didn't talk about AGI. He is still the abacus guy in Omaha, sticking to the old adage that a good company should be able to earn a significantly higher return on capital than risk-free assets over a long period of time.

When discussing Google and other AI companies, he is concerned about the hundreds of billions of dollars they are pouring in and whether they will ultimately be able to make more money.

“This is real money,” he said.

When real money is poured in, at the very least, you should hear a sound.

Two investments of $100 billion, only eighteen months apart, seemed like two different value languages colliding within Google.

Haasabis used $100 billion to illustrate the priceless nature of a medal, while Buffett used $100 billion to question how much this company would actually be worth in the future.

When Google acquired DeepMind, money followed behind scientific questions, buying a piece of unquantifiable research time. By 2026, Alphabet's annual capital expenditure is expected to be close to $200 billion, with most of it going into AI compute and data centers. Money started leading research, first transforming into chips, land, power, servers, and depreciation.

Data centers have no sentiment; every lit-up server room adds another entry of indebtedness on the balance sheet. LLM has turned intelligence from a scientific problem in a lab into a giant civil engineering project.

On both sides of the table

The marriage of Google and DeepMind was not as smooth as it seemed from the first year.

In August 2015, the AGI Safety Committee established during the acquisition of DeepMind held its first meeting at SpaceX. Haasabis, DeepMind co-founder Mustafa Suleyman, Musk, Google executives, and LinkedIn co-founder Reid Hoffman were all present.

Suleyman had long been concerned about the social impact of AI and wanted to establish a governance structure not entirely controlled by Google shareholders.

The meeting did not reach any agreements.

While discussions on the table revolved around the fate of humanity, underneath the table, there were tense conflicts of interest between Google and Musk, personal animosities between Musk and Page, and everyone's starkly different understanding of AI risks.

After the meeting adjourned, Haasabis and Suleyman engaged in a protracted negotiation with Google, codenamed the "Mario Plan." They hoped to transform DeepMind into an entity with independent directors, serving the global public interest.

In the following years, the Independent Plan was repeatedly pulled back and was eventually abandoned in 2021. Google was reluctant to let go of AI, and DeepMind could not leave behind Google's computing power and money.

Also in 2015, at the Rosewood Hotel on Sand Hill Road in Silicon Valley, Sam Altman, then head of Y Combinator, sent an email to Ilya Sutskever inviting him to dinner. When Sutskever arrived, Musk, Greg Brockman, former CTO of Stripe, and others were already there.

It took him a while to realize that he was the focus of this dinner.

They had a very straightforward discussion. They were concerned that Google and DeepMind were too far ahead and wanted to establish a new lab to be a balancing force.

Months later, OpenAI was born.

Sutskever left Google to become a co-founder and research director of OpenAI. The early OpenAI was a nonprofit research organization, and their founding announcement stated that they were not constrained by financial return, researchers could publish papers, blogs, and code, and patents were also intended to be shared with the world.

At that time, there was no ChatGPT, no subscription fees, no enterprise API, and certainly no endless data center bills in sight.

That group of people also believed that scientists could focus on researching the future first, and worry about making money later.

OpenAI was born out of fear of Google's excessive power and also inherited DeepMind's original ideals. Later, the Scaling Law began to dominate the development path of AI, with models getting bigger and the bills getting longer. In 2019, OpenAI established a for-profit subsidiary; in 2022, ChatGPT entered the public's screens.

Since then, everyone had to start running, including Google.

After the thirty-seventh move

In March 2016, AlphaGo played against Go world champion Lee Sedol in Seoul in a five-game match. In the second game, at move thirty-seven, AlphaGo placed a black stone in a position very rarely chosen by human players. The live commentators briefly doubted the machine's move.

Lee Sedol left his seat and spent over ten minutes in the smoking room. Upon returning to the board, he eventually lost that game.

Later, people would often discuss this move, considering it the moment when the machine publicly stepped beyond human experience for the first time.

The $1 million prize for that match was very much in DeepMind's style at the time. Hassabis even committed to this prize before obtaining Google's financial approval.

When AlphaGo won, the prize money was eventually donated to charity and Go organizations.

Scientific questions took precedence over financial processes; it was the golden age of the lab.

This was followed by AlphaZero, MuZero, and AlphaFold.

In the past, deciphering the structure of a protein through experiments often took several months or even years. AlphaFold translates amino acid sequences into three-dimensional structure predictions. DeepMind then established a free database with the European Bioinformatics Institute, incorporating over two billion predicted structures, covering the vast majority of known protein structures in the scientific community.

Over 3 million researchers from more than 190 countries have used it. AlphaFold does not have a monthly fee nor a business growth curve that can be shown on financial reports. It saved countless scientists months or years in the lab, earned Google two Nobel Prizes, and yet had almost no corresponding revenue on the commercial side.

Under the pressure brought by ChatGPT, Sergey Brin, who had already faded out of Google's daily operations, returned.

Brin later recalled that at a small gathering, he met an OpenAI employee who asked him:

"What are you really doing? This may be the most transformative moment in computer science history."

Brin felt the other person was right, so he became more deeply involved in Gemini thereafter.

In 2023, Pichai merged Google Brain with DeepMind to form Google DeepMind, with Hassabis as CEO and Jeff Dean as Chief Scientist. This seemed to be the peak of Hassabis's power but also precisely the end of the old DeepMind as an independent lab.

It was incorporated into Google's models, applications, and business systems, becoming the entire company's AI engine. Koray, who has taken over the day-to-day management of DeepMind, is now responsible for Gemini models, cutting-edge research, Gemini applications, and the developer team.

In the announcement, Pichai repeatedly emphasized "acceleration," "quick action," and the upcoming release of a new model.

Thirteen years ago, Google gave DeepMind something it lacked: time.

Thirteen years later, what Google lacked the most happened to be time.

Turn the clock back one notch.

In the summer of 2026, those who left were mostly people from the pre-LLM era.

The AlphaFold project leader, John Jumper, went to Anthropic; Noam Shazeer, the author of the Transformer paper and co-lead of Gemini, joined OpenAI.

Subsequently, Jeff Dean, along with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, founded Discovery Loop. Jeff Dean and Ghemawat built some of Google's early critical distributed systems, while Vinyals and Quoc Le have long been at the heart of Google's modern AI research.

Discovery Loop was established as a nonprofit corporation. Jeff Dean explained in his farewell letter that one of the meanings of an independent company is the ability to make decisions that may not necessarily align with the company's purely financial interests.

Google became Discovery Loop's founding investor and cloud computing partner. Unable to keep this group of people within its organizational structure, Google decided to spend money to let them continue their work outside.

That's probably how the pre-LLM era concluded.

No one publicly tore down ideals, nor did anyone storm out. The old labs transitioned into product lines, the old scientists took the parent company's funds outside to continue their research, and everyone maintained their dignity.

Hasabis hasn't left yet, but the agreement that started outside the castle in 2013 has lost the one condition it needed to stand on: time.

"Why aren't you leveraging everything I've already created?"

The "everything" in Page's words is much more today than it was thirteen years ago.

Google has more money, machines, engineers, and data centers, so much so that Buffett is willing to put up $10 billion, betting that these massive investments will eventually pay off.

The core asset that Keypay handed to Hasabis is becoming increasingly scarce.

That is the courage to hand over a decade to a question, the luxury of allowing a group of scientists to temporarily ignore the calendar.

From the trash can overturned in Room 731 of Lake Tahoe to Alphabet's nearly $200 billion in annual capital expenditures, AI has taken only fourteen years. The room has transformed into a data center, a small company of three has evolved into an industrial system supported by hundreds of thousands of chips, hourly talent auctions have turned into quarterly model competitions. That group of people has transformed AI from a small circle into the largest industrial system of this era. When all of this finally grew beyond imagination, they began to leave one by one.

Hasabis was also supposed to leave, but Google turned back the resignation clock by one notch.

But it cannot turn back to 2013.

Original Article Link