Google's Farewell: Large Models Are Entering the "Garage Era"

On August 5, Google announced two major pieces of news. Demis Hassabis stepped down as CEO of Google DeepMind and transitioned to the role of Chairman of the division and Chief Scientist of Alphabet. Jeff Dean, the 30th employee of Google who has worked at the company for 27 years, officially left the company. Along with him, Sanjay Ghemawat, Oriol Vinyals and Quoc Le also departed. The four of them have jointly founded a new company named Discovery Loop.

On the day the news was released, Google's stock price fell by more than 4%. On the same day, Google announced publicly that the monthly active users of the Gemini app had exceeded 950 million. However, media outlets added a note in the same report: the development of Gemini 3.5 Pro is several months behind schedule.

What the market cares about is obviously not the 950 million users. It cares about who left, and why they left.

Jeff Dean's 27 years at Google almost cover the entire technical history of the company. He designed MapReduce, BigTable and Spanner — the three sets of systems that enabled Google's search engine to handle global traffic and later became the infrastructure of the entire industry. He and Ghemawat are the only two people at Google who have received the highest technical honor of "Senior Fellow". It is no exaggeration to say that he is part of the technical soul of Google.

But the story he told when he left is no longer the same story that Google is telling.

On July 25, when Dean spoke to 6,000 aspiring founders at Y Combinator's Startup School, he described an automated version of the scientific method: the model puts forward experimental hypotheses, builds the required tools, runs the experiments, evaluates the results, and then designs the next round of experiments on its own. After tens of thousands of cycles, the system may independently find breakthroughs in biology or chip design. The audience at the time did not know that what he described was the new company he had just founded in secret.

The mission of Discovery Loop is clearly stated: to let AI run through the entire scientific research process on its own. This is not a new idea. Dario Amodei has mentioned it, Sam Altman has mentioned it, and Jensen Huang has also mentioned it — AI will become the engine of scientific discovery, solving difficult problems that humans cannot overcome alone.

But Dean and his team are the first group of people who actually left big tech companies for this goal. The way they chose is also very special. The new company is registered as a public benefit corporation, a for-profit structure that allows management to make decisions that "do not necessarily align with pure financial interests". Dean said a very honest sentence in the interview: We may make some decisions that are not purely for making money.

The four of them have not rented an office or recruited their first employee yet. When asked who the CEO is, Dean paused for a moment and said a little embarrassedly: "I guess I am the CEO. Everyone pointed at me."

Investors don't care about these details. After meeting the four of them in his office on Sand Hill Road, Vinod Khosla said one sentence: With this team, I don't need to know what they are going to do to invest. Khosla Ventures and Radical Ventures have invested, and Alphabet itself has also invested. Google also promised to provide computing power for the new company in its first year. The structure of the parent company investing in a competitor founded by former employees is a microcosm of the ongoing AI talent war right now.

Hassabis's exit is another clue.

He no longer manages the daily operations of Google DeepMind, but retains the position of CEO of Isomorphic Labs. This company focuses on AI drug discovery, which is the most direct commercial application scenario for DeepMind's technology. His new title is Chief Scientist of Alphabet, which sounds more powerful than the actual authority he holds. Koray Kavukcuoglu took over as the head of DeepMind, reporting directly to Pichai. Kavukcuoglu has worked at DeepMind for more than 13 years, founded the deep learning team, and led the research of WaveNet and DQN. His resume is impressive enough, but the situation he took over is far from stable.

An Axios report says that DeepMind is currently experiencing a wave of resignations, and employees are anxious about the progress of the business. The delay of Gemini 3.5 Pro is partially attributed to low morale. The statements from these anonymous sources cannot be independently verified, but together with the talent drain in the past few months, they form a picture of instability.

Tim Rocktäschel left even earlier. He left Google earlier this year and founded Recursive Superintelligence, which also focuses on AI self-improvement. His reason for leaving is very straightforward: the environment of a startup is completely different, faster, more focused, with a smaller team, less bureaucracy and less internal politics. This statement is not so much a criticism as a description of an irreversible trend: when the AI frontier has evolved from "can we build large models" to "can we make the model improve itself", the people who really want to answer this question would rather go to the garage to find the answer themselves.

The departure of these people cannot be explained by compensation. Dean's salary at Google is enough for him and several generations of his family to live a wealthy life. Hassabis's position adjustment is not a marginalization, and he is still in Alphabet's core decision-making circle. The real thing that drives them to make this move is the same thing: they have seen the next step, but the organization they are in is not ready for that step yet.

That step is called recursive self-improvement. Let the AI system optimize itself, put forward experimental hypotheses on its own, build tools on its own, evaluate results on its own, and then enter the next cycle on its own. This logic appears repeatedly in Discovery Loop's mission statement, in the company name of Recursive Superintelligence, and in Dean's description to YC participants. It is not the fantasy of a single person, but the common belief of a small group of top people at the current AI frontier.

And Google happens to be the company that first reached this threshold but finds it the hardest to step over. Its search ad empire requires stability, its cloud business needs to serve large customers, and its Gemini product needs to catch up with competitors. All of these require a premise: no loss of control. Recursive self-improvement naturally carries the risk of getting out of hand, and no one can predict where a model that has modified itself thousands of times will end up. For a public company that is responsible for hundreds of millions of users and billions of dollars in advertising revenue, this is not a proposition that can be easily accepted. For a startup, it is the source of all its value.

Pichai said in the announcement that Google is the only company with full-stack capabilities. He is not wrong. Google has TPUs, data centers, the world's largest user base, and DeepMind's research capabilities. But the price of having a full stack is that no layer of the stack can make mistakes. Recursive self-improvement, on the contrary, requires tolerance for errors. Its entire logic is to find the optimal solution through continuous trial and error. In a large company, the pace of such trial and error will be slowed down layer by layer by security reviews, ethical reviews, and legal compliance. Dean and his team do not lack computing power or capital, what they lack is the freedom to make all the mistakes within a sufficiently short time window.

This is the true meaning of Vinod Khosla's sentence "I don't need to know what they are going to do to invest". He is not investing in a business plan, but a judgment: recursive self-improvement is the next stop for AI, and Jeff Dean is one of the few people on Earth who has the potential to make it happen. As for what to do specifically and how to do it, these do not need to be known when you set off. What matters is the act of setting off itself, and setting off without any baggage.

Hassabis said in his farewell memo that Google is the only company with a full stack. This statement is a fact. But it also explains why this group of people chose to leave.

The full stack is Google's advantage. It is also its burden. Discovery Loop has only four people, no office, no other employees. Their only assets are the names of the four people and a shared belief: AI can improve itself. Whether this belief can be turned into a product is still unknown. But at least one person believes it is worth it — Jeff Dean, for the first time in 27 years, is no longer a Google employee.

The full stack is too heavy, and you can only run fast if you travel light.

This article is from the WeChat Official Account "Silicon Starlight", written by Xing Mang, and published with authorization from 36Kr.