This article is written based on publicly available information and is intended solely for informational exchange, not as any investment advice.

On August 5th, Google announced two things. Demis Hassabis stepped down as CEO of Google DeepMind and became the chair of the department and Chief Scientist of Alphabet. Jeff Dean, Google's 30th employee, officially left after 27 years at the company. Joining him were Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The four co-founded a new company called Discovery Loop.

On the day of the announcement, Google's stock price fell by more than 4%. On the same day, Google stated that the Gemini app had surpassed 950 million monthly active users. However, the media added a sentence in the same report: the progress of Gemini 3.5 Pro was several months behind schedule.

What the market clearly cares about is not those 950 million users. It cares about who left and why.

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

But the story he told when leaving is not the same one Google is telling now.

On July 25th, Dean described an automated version of the scientific method to 6,000 aspiring founders at Y Combinator's startup school: a model proposes experimental hypotheses, builds the necessary tools, conducts experiments, evaluates results, and then designs the next round of experiments itself. After thousands of cycles, the system might find breakthroughs in biology or chip design on its own. The audience that day didn't know he was describing his newly secret startup.

Discovery Loop's mission is clear: to enable AI to navigate the entire scientific research process on its own. This is not a new idea. Dario Amodei has said it, Sam Altman has said it, and Jensen Huang has said it—AI will become the engine of scientific discovery, solving problems that humans alone cannot tackle.

But Dean and his team are the first to truly leave a major company for this purpose. Their approach is also unique. The new company is registered as a Public Benefit Corporation, a structure that allows management to make decisions "not necessarily aligned with pure financial interests" while still being profit-driven. Dean said something candid in an interview: "We might make decisions that aren't purely about making money."

The four have not yet rented an office or hired their first employee. When asked who the CEO is, Dean paused, a bit embarrassed, and said, "I guess I'm the CEO. Everyone's pointing at me."

Investors don't care about these details. After meeting with them in his Dune Road office, Vinod Khosla said, "With this team, I don't need to know what they're going to do to invest." Khosla Ventures and Radical Ventures invested, Alphabet itself invested, and Google committed to providing computing power for the new company's first year. The parent company investing in a competitor founded by former employees is, in itself, a microcosm of the current AI talent war.

Hassabis's exit is another clue.

He no longer oversees the day-to-day operations of Google DeepMind but retains his position as CEO of Isomorphic Labs. This company focuses on AI-driven pharmaceuticals, the most direct commercial application of DeepMind's technology. His new title, Chief Scientist of Alphabet, sounds more powerful than it is. Koray Kavukcuoglu succeeded him as head of DeepMind, reporting to Pichai. Kavukcuoglu has been at DeepMind for over 13 years, founded the deep learning team, and led research on WaveNet and DQN. His resume is impressive, but the situation he inherited is not smooth.

Axios reported that DeepMind is experiencing a wave of departures, with employees anxious about progress. The delay of Gemini 3.5 Pro is partly attributed to low morale. These anonymous sources cannot be independently verified, but combined with the talent exodus over the past few months, they paint a picture of instability.

Earlier to leave was Tim Rocktäschel. He left Google earlier this year to found Recursive Superintelligence, also focused on AI self-improvement. His reason for leaving was straightforward: the startup environment is completely different—faster, more focused, with smaller teams and less bureaucracy and politics. This sounds less like criticism and more like an irreversible trend: when the frontier of AI has evolved from "can we build large models" to "can we make models improve themselves," those who truly want to answer this question would rather do it in their own garage.

These departures cannot be explained by salary alone. Dean's compensation at Google was sufficient to secure a prosperous life for him and several generations of his family. Hassabis's reassignment was not a marginalization; he remains in Alphabet's core decision-making circle. What truly motivated them was the same thing: they saw the next step, and their organizations were not yet prepared for it.

That step is called recursive self-improvement. Letting AI systems optimize themselves, propose experimental hypotheses, build tools, evaluate results, and enter the next cycle on their own. This logic appears repeatedly in Discovery Loop's mission statement, in the name of Recursive Superintelligence, and in Dean's description to YC students. It is not the fantasy of a single individual but a shared belief among the most elite group at the frontier of AI today.

And Google, precisely, is the company that first encountered this threshold but finds it hardest to cross. Its search advertising empire requires stability, its cloud business needs to serve major clients, and its Gemini needs to catch up to competitors. All these require one prerequisite: no loss of control. Recursive self-improvement inherently carries the risk of losing control. No one can predict where a model that has modified itself thousands of times will end up. For a publicly traded company responsible for hundreds of millions of users and billions in advertising revenue, this is not an easily acceptable proposition. For a startup, it is the entire value proposition.

Pichai stated in the announcement that Google is the only company with full-stack capabilities. He is not wrong. Google has TPUs, data centers, the largest user base globally, and DeepMind's research capabilities. But the cost of full-stack is that no layer can fail. Recursive self-improvement, however, requires tolerance for failure. Its entire logic is to find the optimal solution through continuous trial and error. In a large company, this trial-and-error rhythm is slowed down by layers of safety reviews, ethical reviews, and legal compliance. Dean and his team do not lack computing power or funding; what they lack is the freedom to make all their mistakes within a sufficiently short time window.

This is the true meaning behind Vinod Khosla's statement, "I don't need to know what they're going to do to invest." He is not investing in a business plan but in a judgment: recursive self-improvement is the next frontier of AI, and Jeff Dean is one of the few people on Earth who could pull it off. As for what specifically to do or how to do it, these can remain unknown at the outset. What matters is the journey itself and the absence of baggage at the starting point.

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

Full-stack is Google's strength. It is also its weight. Discovery Loop has only four people, no office, and no employees. Their only assets are the names of the four individuals and a shared belief: AI can improve itself. Whether this belief can be transformed into a product remains unknown. But at least one person believes it is worth pursuing—Jeff Dean, who for the first time in 27 years, is no longer a Google employee.

Full-stack is too heavy. To run fast, you must travel light.