Original title: (Hassabis steps down, Jeff Dean leaves to start a business, and Google’s morale has waned)
Original author: Insightful Beating

On August 5, Google DeepMind CEO Demis Hassabis stepped down from his daily management duties, moving to become Chairman of DeepMind, and also taking on the role of Alphabet Chief Scientist. He will still be responsible for long-term AI research and continue to lead the new drug development company Isomorphic Labs. His title is higher than before, but in reality it’s a promotion in name with a demotion in substance—he has less real management authority in hand.

The R&D and delivery of Gemini will be handled by Koray Kavukcuoglu, the former DeepMind Chief Technology Officer, who will report directly to Pichai. Hassabis remains a scientific icon of Google AI and continues to participate in judgment calls on the direction of AGI, but he will no longer control DeepMind’s day-to-day operations. After this adjustment, the management chain between DeepMind and Alphabet’s headquarters will be further shortened, and Pichai will also take more direct control of the Gemini R&D progress.

On the same day, Jeff Dean announced he was leaving Google.

He took Sanjay Ghemawat, Oriol Vinyals, and Quoc Le with him to found a company called Discovery Loop. They hope to use AI to automate the entire research workflow: posing questions, designing experiments, executing experiments, and evaluating results. Google participated in the seed investment; it will become the company’s cloud service provider and provide computing power for the first year.

Google did not strongly retain this team; instead, it chose to maintain relationships in terms of capital, cloud services, and research collaboration. For a group of top researchers who have already decided to start their own ventures, this counts as a rather decent way to handle it.

The market reacted quickly. Alphabet’s share price fell more than 5% at one point during the trading day, closing down nearly 4%. Based on the company’s size, a 3.8% drop corresponded to a market value of about $175 billion. At the peak of intraday decline, the amount wiped out was close to $260 billion.

In the past six weeks, Alphabet has already been punished by the market for AI for the fourth time.

Jeff Dean and Hassabis: two traditions of Google AI

Jeff Dean is hard to categorize as a typical technical executive.

When he joined Google in 1999, he was the company’s 30th employee. Over the next nearly 30 years, during several of Google’s most important technology transitions, he was almost always there.

Early search and advertising systems had to process rapidly expanding data. He and Ghemawat wrote MapReduce and Bigtable. These systems later became the foundation for large-scale distributed computing, with an impact far beyond Google. His involvement was also behind the building of Google Brain, TensorFlow, and TPU. In the Gemini era, he became the technical co-lead of the models again.

Behind the fact that Google can now train large models, run them on its own chips, and plug them into search, cloud services, and various consumer products, you can find the infrastructure he helped build.

Hassabis represents a different kind of capability.

In 2010, he founded DeepMind in London. In 2014, Google completed the acquisition. Two years later, AlphaGo defeated Lee Sedol, bringing AI into the global public eye in such a direct way for the first time. In 2024, he won the Nobel Prize in Chemistry with John Jumper for AlphaFold.

In 2023, Google Brain merged with DeepMind, and Hassabis took over the integrated Google DeepMind. He is both the leader of the Gemini project and the most important person in Google’s public storytelling about the future of AI.

Dean and Hassabis, respectively, have supported Google’s engineering tradition and scientific tradition in AI. Now that Dean has left the company, Hassabis has moved into the background.

Four sudden plunges in six weeks

This reshuffle has triggered such a strong market reaction partly because it happened after a sustained run of setbacks for Google’s AI.

In June, one of the authors of the Transformer paper and Gemini technical co-lead Noam Shazeer left Google to join OpenAI. In 2024, Google had just brought him back from Character.AI through a special deal of about $2.7 billion and arranged for him to enter Gemini’s core team. Two years later, Shazeer left again.

Soon after, John Jumper—the head of AlphaFold and the 2024 Nobel Prize winner in Chemistry—joined Anthropic.

Within two days, Google lost two highly influential researchers. On June 22, Alphabet’s stock price fell by more than 7% at one point during the day, and its market capitalization evaporated by over $200 billion.

On July 16, news broke that Gemini 3.5 Pro would be delayed. The flagship model had originally been planned for release in June, but because performance benchmarks such as coding capability did not meet expectations, the timeline had already been pushed back by several months. After the news was announced, Alphabet’s stock price fell again by more than 3% at one point during the day.

In late July, the pressure shifted to the financial statements.

Alphabet’s capital expenditures in the second quarter reached $44.9 billion, doubling year over year. The full-year capex guidance was raised to between $195 billion and $205 billion. In the same period, the company’s free cash flow turned negative by $5.9 billion—its first quarter of negative free cash flow in history. On the first trading day after the earnings release, the stock price fell more than 7% at one point during the day.

Google is putting in more and more resources, but the models have not formed a leading advantage proportional to that investment. And the flagship products haven’t been delivered on time; core figures responsible for research and engineering have continued to leave one after another.

After Jumper left, media outlets quoted DeepMind employees saying that in fields such as text, images, video, speech, and vision, it was becoming difficult for Google to find models clearly at the very forefront of the industry.

Google is never short on resources—indeed, it could be said that it has the most complete set of resources across the entire AI industry.

It has its own TPU, data centers on a global scale, and accumulated data from search and YouTube. It also has distribution channels provided by Android, Chrome, Workspace, and cloud services. In an internal email, Pichai also listed a series of achievements, saying that Gemini’s monthly active users for its app reached 950 million, Gemma downloads exceeded 900 million, and Gemini Robotics is still moving forward.

These assets are enough to prove that Google is still powerful, yet they can’t replace the next-generation flagship model itself. The market is willing to pay for long-term investment, as long as the money ultimately turns into a lead. After model delays and the departure of key personnel started happening one after another, investors began to wonder whether Google is missing time or—rather—the ability to turn its massive resources into a single force.

DeepMind says goodbye to its founder-led management

Koray Kavukcuoglu, who took over the day-to-day operations from Hassabis, has worked at DeepMind for 13 years, where he was mentored by Yann LeCun. He has worked on DQN, WaveNet, and multiple generations of the Gemini projects, and in the past served as CTO of Google DeepMind.

Koray knows DeepMind well and has long been responsible for connecting research, infrastructure, and products. With him taking over Gemini, it indicates that Google will place even more emphasis on model delivery, product coordination, and commercial execution.

This reshuffle also changed DeepMind’s position within Google.

In the DeepMind era of Hassabis, there was always a distinct founder-like character. It started in London, and even after Google’s acquisition, it maintained a relatively independent research tradition. After Google Brain merged with DeepMind, Hassabis became CEO of the integrated team, and also ensured DeepMind’s culture dominated within the new organization.

Now that Hassabis has stepped back from day-to-day management, Koray reports directly to Pichai, and the relationship between DeepMind and Alphabet’s headquarters has become tighter. For Google, this is a move to strengthen control. Pichai needs to have a clearer grip on Gemini’s R&D cadence, and he also needs someone to take direct responsibility for things like model delays, resource allocation, and bringing products to market.

For DeepMind, this also means the end of a chapter.

Hassabis is still at Google, still has a lofty title, and still participates in long-term research. But that DeepMind—directly led by its founders and shaped into an organization by scientific ideals—is gradually becoming a more standardized AI R&D unit within Alphabet.

Google is learning how to send talent away

Jeff Dean’s departure shows another way Google handles talent exodus.

Google participated in Discovery Loop’s founding investment, continuing to provide cloud services and computing power, and maintaining research collaboration with the new company. Although Dean and his team left the organization, they still remain within Google’s capital and technology ecosystem.

This arrangement benefits both sides.

Discovery Loop got expensive computing power, infrastructure, and early funding, so it didn’t have to build a research platform from scratch. Google, in turn, retained the possibility of investment returns, cloud orders, and future collaboration, and also avoided sending a group of core talents who had spent more than two decades working there directly to competitors.

From the perspective of corporate governance, this is almost an ideal exit plan.

But it still can’t answer one question: for a team that is familiar with Google, helped build Google, and can draw on Google’s resources, why did it ultimately conclude that after leaving, it would be easier to carry out its own research?

They did not abandon their past research directions, and they were not forced to leave because of failed internal competition. They simply moved work that was originally carried out inside Google to a new company that they could fully control themselves.

When they discussed the reasons for leaving, they said that within large organizations, there is always too much inertia to overcome in order to drive radical change. They wanted to build something different.

In the past, Google attracted top researchers by offering the resources they could get by staying here—enabling them to do work that other places couldn’t.

Now, a group of people who understand Google’s resources best are starting to make the opposite judgment. They still believe those research efforts are worth investing in, and they still believe AI will change science. What they no longer believe is that a large company’s organizational approach is best suited to accomplish it.

Outside the lab, new companies begin to grow

Google DeepMind brings Bell Labs to mind.

Bell Labs relied on long-term stable profits from AT&T, bringing together some of the best scientists and engineers of the 20th century, which led to the creation of the transistor, the laser, information theory, and Unix in succession. Its greatness wasn’t preserved just by the organization itself; it spread through the people who left.

After co-inventor of the transistor, William Shockley, left Bell Labs and founded Shockley Semiconductor in California, in 1957 eight engineers left as a group to establish Fairchild Semiconductor. The founding teams behind Intel and AMD later also spun out of the Fairchild ecosystem. Much of what later became known as the industrial network of Silicon Valley was formed along the path of these talent flows.

Google DeepMind hasn’t reached the fate of Bell Labs. Gemini has a large user base, and Google Cloud is still growing. The similarity between the two is that a top lab nurtured by a mature commercial system has started to take its people outside to build new labs and companies.

When a laboratory is at its height, people look at how many geniuses it attracts. In the next stage, people will look at what those geniuses build after they leave.

This is both an achievement and a loss. For the entire tech industry, both can be true at the same time, but for Google, it’s hard to just write it off so easily.

Google’s investment in Discovery Loop allows it to share in the future returns of this company and provides cloud services—turning the company’s computing expenditure into revenue again. But while a company can invest in external outcomes, it cannot use that to replace the creative capacity it lost internally.

Dean worked at Google for 27 years. He said he only began seriously considering starting his own venture about five weeks ago.

A person who had spent nearly 30 years at the company decided to leave in five weeks. At Discovery Loop, it’s still the same familiar teammates: Ghemawat handles systems, Vinyals leads research, and Quoc Le continues to work on automated machine learning. There hasn’t been much change in how people are divided, nor has the research direction suddenly shifted—only that they moved out of Google’s buildings.

In his farewell letter, Hassabis said that AGI is near at hand. The new company Dean founded also bets on AI changing scientific research.

They probably still believe in the same future, but they’ve started setting out on their own paths.

Original link