Hassabis Steps Down: The Final Act of the Pre-LLM Era

Original Title: "Hassabis Steps Down: The Final Act of the Pre-LLM Era"
Original Author: Dongcha Beating

Hassabis was also set to leave.

According to industry insiders cited by Pathfounders, Demis Hassabis, co-founder of Google DeepMind and winner of the 2024 Nobel Prize in Chemistry, originally planned to leave Google at the same time as Jeff Dean. Management was concerned that the simultaneous departure of the two AI leaders would severely impact stock prices, so they sought to keep Hassabis on board. Insiders indicate that he may still leave once the situation stabilizes.

On August 5, Google presented a much more dignified story to the public.

Hassabis voluntarily handed over the day-to-day management of Google DeepMind, taking on the roles of chairman and chief scientist of Alphabet, while continuing to oversee AGI, scientific research, and the drug development company Isomorphic Labs. Koray Kavukcuoglu, who has worked at DeepMind for thirteen years, took over the daily operations; he is Google’s chief AI architect and one of the creators of DeepMind’s early deep learning team, and will now report directly to Sundar Pichai.

Big Companies Know How to Write Farewells.

Farewells from big figures often start with longer titles. Chairman, chief scientist—there are more words on the business card, but fewer people directly managed. Google retained Hassabis’s reputation and Nobel Prize, leaving a name that could temporarily reassure investors.

This most powerful tech company in the world has found it increasingly difficult to accommodate the scientist who brought it AlphaGo, AlphaFold, and a Nobel medal in the same way it used to. Now, what it can do is to keep him on board first, and then find a less awkward day for a potential breakup.

Google and Hassabis maintained the last complete lifestyle of the pre-LLM era together. Scientists made decisions, the company paid the bills, research was measured in years, products could be launched later, and profits were not the primary concern.

The unfinished breakup between Google and Hassabis has also become the final act of the pre-LLM era.

An Hour

Hassabis was originally supposed to leave on the same day as Jeff Dean. Jeff Dean is one of Google’s most legendary engineers, having joined Google in 1999 and worked on the underlying systems for search, advertising, and distributed computing, later co-founding Google Brain. Google employees have made countless jokes about him, as if this person could snap his fingers and half the internet would need to reboot.

After he announced his departure that day, someone joked that Google had finally "open-sourced Jeff Dean to humanity."

Fourteen years ago, he did the exact opposite. He brought in one of the world’s most scarce groups of AI scientists for Google.

In December 2012, at Lake Tahoe, in room 731 of Harrah's hotel, one of the most famous talent auctions in modern AI history took place at a table at the end of two double beds.

The auctioneer, Geoffrey Hinton, is a British-Canadian scientist who has studied neural networks for decades and later became known as the godfather of AI. He suffered from severe back pain and had been unable to sit normally since 2005; once seated, his disc could slip again, leaving him bedridden for weeks.

Thus, he transformed his life to standing and lying down; in the car, he lay across the back seat, kneeling at the table while eating, and when he arrived at the hotel, he would turn a trash can upside down on the table and set his laptop on top, standing by to wait for new auction bids to arrive in his email.

Hinton and two students had just founded a company called DNNresearch, with no products and no revenue; the company’s entire assets 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, suddenly reviving the neural networks that had been neglected for decades.

Another of Hinton’s students, Ilya Sutskever, would later become a co-founder and chief scientist of OpenAI. But at this moment, he was still standing in room 731 with his teacher, watching the bids come in via Gmail.

The auction rules were simple: whenever a new price appeared, other companies had an hour to decide whether to follow, with each increase being at least one million dollars.

Baidu came, Microsoft came, Google came, and there was also a London company that had only been established for two years, DeepMind. It couldn’t compete with the cash of those giants and could only bid with its own shares, quickly getting pushed off the table.

The remaining three companies continued to raise the price, shouting all the way to 44 million dollars. Hinton called off the auction, took a nap, and the next day decided to sell the company to Google.

He didn’t continue to raise the price. At this number, the money was already enough; where he and his two students would work next was the more important matter.

Jeff Dean and Google engineering head Alan Eustace spent Hinton’s 65th birthday with him before the auction. They mainly wanted to see how Hinton’s two students were doing. Later, Google spent 44 million dollars to buy this company with no products, essentially to acquire these three people.

The Best Decades

DeepMind did not buy Hinton at Lake Tahoe. More than a year later, it grew into what Google wanted to buy.

The story unfolded at a chaotic birthday party.

In June 2013, Elon Musk’s wife, actress Talulah Riley, rented a castle in Tarrytown, New York, to celebrate her husband’s birthday. Guests dressed in samurai costumes, and a sumo champion weighing about 350 pounds also attended. Musk personally stepped in to compete with him and injured his neck.

Google co-founder Larry Page also attended the party. During those years, he was troubled by a vocal cord disease, and could only speak in a very low voice. Page called Hassabis outside, and the two walked along the outside of the castle.

Once outside, he asked:

"Why don’t you leverage everything I’ve already created?"

Hassabis had an epiphany. The meaning of this sentence was "You don’t have to build another Google" because that would take the best decades of your career. Google already had data centers, engineers, cash, and products spread across the globe. What you need to solve is intelligence; why bother with the hassle of starting a company?

Hassabis was a chess prodigy as a child, designed 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 very beginning.

Just a few years after the company was founded, he was already tired of fundraising.

Researching intelligence is like throwing stones into a bottomless pit, and you have to keep coming out every now and then to explain to investors why this bottomless pit hasn’t echoed back yet. Moreover, after completing one round of financing, you have to rush into the next round of financing.

Page’s answer to him was to save decades of building another Google, with Google taking on the time that may not yield immediate returns.

In 2014, Google acquired DeepMind for about 650 million dollars. The lab remained in London, retaining its name and research culture. Google also promised to establish an AI ethics and safety committee to oversee how these technologies are used.

Jaan Tallinn, co-founder of Skype and an early investor in DeepMind, later recalled that Google’s purely financial offer was not the best; the commitment to the safety committee was the key reason DeepMind ultimately chose Google.

This is the most charming marriage contract of the pre-LLM era.

Google promised money and patience, while DeepMind delivered the possibility of AGI. As for when to make a profit, let’s push that back a bit.

At that time, Google was large and wealthy enough to afford losses and wait.

Two 10 Billion

Before Christmas 2024, British journalist Sebastian Mallaby met Hassabis in a bar in London. They had already talked many times before.

That time, Hassabis pulled out a leather box from his backpack. He opened it to reveal a heavy golden disc inside.

He had just returned from Stockholm with the Nobel Prize in Chemistry.

Hassabis and DeepMind scientist John Jumper jointly won the Nobel Prize in Chemistry that year for AlphaFold. Dr. Jumper took over the AlphaFold project just six months after graduating, and later, this team solved the protein structure prediction problem that had plagued the biological community for decades.

The Nobel Prize is already the highest medal that the secular world can award to scientists. Hassabis told Mallaby:

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

In June 2026, Alphabet initiated a maximum 80 billion dollar equity financing to expand its AI infrastructure. Buffett's Berkshire bought 10 billion dollars of Alphabet stock through a targeted issuance.

By July, the investment firm led by Buffett for sixty years had a position in Alphabet exceeding $31 billion.

Buffett did not discuss AGI; he remains the person who calculates in Omaha, adhering to an old principle: a good company must be able to earn significantly higher capital returns than risk-free assets over a long period.

When talking about Google and other AI companies, he is concerned about the hundreds of billions they are pouring in and whether it will ultimately generate more money.

"This is real money," he said.

Real money must at least make a sound when it is spent.

Two investments of $10 billion, just eighteen months apart, collided like two different value languages within Google.

Hassabis used $10 billion to illustrate the pricelessness of a medal, while Buffett used $10 billion to question how much this company would be worth in the future.

When Google acquired DeepMind, the money followed the scientific questions, buying a period of research time that could not be priced. By 2026, Alphabet's annual capital expenditure is expected to approach $200 billion, most of which will be directed towards AI computing power and data centers. Money is starting to precede research, first transforming into chips, land, electricity, servers, and depreciation.

Data centers have no sentiment; every server room that lights up adds a debt to the future on the balance sheet. LLM has pushed intelligence from scientific questions in laboratories into a massive civil engineering project.

On both sides of the table

The marriage between Google and DeepMind has not been as peaceful as it appears from the first year.

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

Suleyman has long been concerned about the social impact of AI and wants to establish a governance structure that is not entirely controlled by Google shareholders.

No agreements were reached at the meeting.

While discussions on human destiny took place on the table, conflicts of interest between Google and Musk, personal grievances between Musk and Page, and each person's vastly different understanding of AI risks were brewing underneath.

After the meeting, Hassabis and Suleyman engaged in a lengthy negotiation with Google, codenamed "Mario Project." They hoped to transform DeepMind into an institution with independent directors serving the global public interest.

In the following years, the independent plan faced repeated setbacks and was ultimately abandoned in 2021. Google was reluctant to let go of AI, and DeepMind could not survive without Google's computing power and funds.

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

It took him a while to realize that he was the main character of the dinner.

The issues they discussed were straightforward; they were worried that Google and DeepMind were too far ahead and wanted to build a new lab to serve as a counterbalance.

A few months later, OpenAI was born.

Sutskever left Google to become a co-founder and research lead at OpenAI. The early OpenAI was a non-profit research organization, stating in its founding announcement that it was not bound by financial returns, allowing researchers to publish papers, blogs, and code, and preparing to share patents with the world.

At that time, there was no ChatGPT, no subscription fees, no enterprise APIs, and no endless data center bills.

That group believed that scientists could first research the future and worry about making money later.

OpenAI was born out of fear of Google's excessive power and inherited the initial ideals of DeepMind. Later, Scaling Law began to dominate the development path of AI, with models growing larger and bills becoming longer. In 2019, OpenAI established a for-profit subsidiary; in 2022, ChatGPT burst onto the public screen.

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

After the thirty-seventh move

In March 2016, AlphaGo played five games against Go world champion Lee Sedol in Seoul. In the second game, after the thirty-seventh move, AlphaGo placed a black stone in a position rarely chosen by human players. The commentators on-site initially suspected the machine had made a mistake.

Lee Sedol left his seat and stayed in the smoking room for over ten minutes. When he returned to the board, he ultimately lost the game.

Later, people repeatedly discussed this move, viewing it as the moment when machines first publicly stepped beyond the boundaries of human experience.

The $1 million prize for that match also reflected the spirit of DeepMind at the time. Hassabis even promised the prize without completing Google's financial approval in advance.

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

Scientific questions took precedence over financial processes; that was the golden age of laboratories.

Then came AlphaZero, MuZero, and AlphaFold.

In the past, scientists often took months or even years to analyze a protein structure through experiments. AlphaFold transformed amino acid sequences into three-dimensional structure predictions, and DeepMind established a free database with the European Bioinformatics Institute, housing over 200 million predicted structures, covering the vast majority of known proteins in the scientific community.

Over 3 million researchers from more than 190 countries have used it. AlphaFold did not charge monthly fees and had no commercial growth curve to showcase in financial reports. It saved countless scientists years in the lab, brought Google two Nobel Prizes, but generated almost no corresponding revenue.

Under the pressure brought by ChatGPT, co-founder Sergey Brin, who had faded from Google's daily work, returned.

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

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

Brin felt the person was right, and thus became increasingly involved in Gemini.

In 2023, Pichai merged Google Brain with DeepMind to form Google DeepMind, with Hassabis becoming CEO and Jeff Dean becoming chief scientist. This seemed to be the pinnacle of Hassabis's power and precisely marked the end of the old DeepMind as an independent laboratory.

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

Pichai repeatedly emphasized "acceleration," "rapid action," and the upcoming new models in the announcement.

Thirteen years ago, Google gave DeepMind time.

Thirteen years later, what Google lacks the most is precisely time.

Turning back the clock

In the summer of 2026, almost all those who left were from the old stories of the pre-LLM era.

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

Subsequently, Jeff Dean, along with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, founded Discovery Loop. Jeff Dean and Ghemawat had built some of Google's earliest and most important distributed systems, while Vinyals and Quoc Le had long been at the center of Google's modern AI research.

Discovery Loop was registered as a public benefit corporation. In his farewell letter, Jeff Dean explained that one of the meanings of an independent company is to make some decisions that "do not necessarily align with the company's pure financial interests."

Google became the founding investor and cloud computing partner of Discovery Loop. It could not keep these people within its organizational structure, so it decided to spend money to let them continue their work outside.

This is how the pre-LLM era came to an end.

No one publicly tore apart ideals, nor did anyone storm out. The old laboratory was integrated into the product line, and the old scientists took the parent company's funding to continue their research outside, leaving everyone with dignity.

Hassabis has not left yet, but the agreement that began outside the castle in 2013 has lost the conditions on which it was founded: time.

"Why don’t you utilize everything I have already created?"

What Page referred to as "everything" is far more than it was thirteen years ago.

Google has more money, machines, engineers, and data centers than it knows what to do with, so much so that Buffett is willing to bet $10 billion that these massive investments will eventually pay off.

But the core assets that Page handed to Hassabis are becoming fewer and fewer.

That was the courage to dedicate a decade to a single problem, a luxury that allowed a group of scientists to temporarily ignore the calendar.

From the trash can overturned in room 731 at Lake Tahoe to Alphabet's nearly $200 billion in annual capital expenditures, AI took just fourteen years. The room transformed into a data center, a small company of three grew into an industrial system supported by hundreds of thousands of chips, and talent bidding that happened once an hour turned into model competitions calculated quarterly. That group of people turned AI from a small circle into the largest industrial system of this era, and when everything finally became unimaginably large, they began to leave one by one.

Hassabis was also set to leave, but Google turned back the clock on his departure.

But it cannot turn back to 2013.

Original link