Google Names Gemini 4 Lead as Both Co-Heads Quit to Launch Discovery Loop

Google’s AI Brain Drain: Gemini Co-Leads and Key Researchers Exit as Hassabis Steps Back

On Wednesday, Google CEO Sundar Pichai announced a sweeping overhaul of the company’s AI leadership. Koray Kavukcuoglu, a 13-year DeepMind veteran who built the lab’s deep learning team, was named senior vice president of Google DeepMind and tasked with building Gemini 4, the next generation of Google’s flagship AI model. Simultaneously, Demis Hassabis stepped back from daily operations to become chair of DeepMind and chief scientist at Alphabet, focusing on long-term AGI strategy and drug discovery at Isomorphic Labs.

The same day, two of Gemini’s co-technical leads—Oriol Vinyals and Quoc Le—left the company, along with Google legends Jeff Dean and Sanjay Ghemawat. The four are founding Discovery Loop, a startup that aims to automate the scientific method by running thousands of experiments in parallel using AI. The departures, first reported by CNBC and Wired, were confirmed by Pichai’s internal memo, which praised Dean and Ghemawat’s contributions but notably omitted any mention of Le or Vinyals.

Alphabet shares fell about 5% following the news as investors digested the loss of some of the company’s most celebrated AI researchers. The exits come at a moment when Google faces intensifying competition from OpenAI and Anthropic, as well as reported delays in its next-generation models—including a Gemini Pro setback over coding performance—and the company’s first cash-flow-negative quarter on record.

Discovery Loop, formed just weeks ago, has already raised a seed round co-led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet itself. Google will serve as the startup’s cloud partner for its first year and has also entered a joint research framework on machine learning systems. With only four founders and no office yet, Discovery Loop’s ambition—to build AI that can propose, run, and iterate on its own research—directly challenges the model of the massive research teams they just left behind.

Why This Reshuffle Signals Trouble for Google’s AI Race

A Leadership Void at the Heart of Google’s AI Strategy

The departure of Vinyals and Le, who co-led the Gemini program, creates a critical knowledge gap just as Google tries to accelerate Gemini 4. While Kavukcuoglu is highly respected and has deep roots in DeepMind’s technical work (WaveNet, DQN), he inherits a team that has lost its original architects. The omission of Le and Vinyals from Pichai’s public statements suggests an attempt to downplay the scale of the brain drain, but the market’s 5% sell-off shows investors are not buying it.

Discovery Loop’s Bet on Automated Discovery

Discovery Loop’s thesis is that scientific progress is bottlenecked not by the scientific method, but by its manual execution. The founders, whose prior-work slides include TensorFlow and TPUs, plan to build an AI system that acts as the researcher—proposing experiments, running them, interpreting results, and looping on its own. Starting with machine learning research and using the system to improve its own stack, the startup aims to tackle grand challenges like medicine and clean water. The irony is sharp: a handful of people now claim the same “full-stack depth” that Pichai boasted about Google’s tens of thousands of employees.

Google’s Investment: Hedge or Conflict?

Alphabet’s decision to invest in Discovery Loop and provide its first-year compute through Google Cloud blurs the lines between partner and potential competitor. The joint research framework on ML systems suggests Google wants to keep the founders close, perhaps to benefit from any breakthroughs. But the arrangement also raises questions about conflicts, especially if Discovery Loop develops technology that competes directly with Google’s own AI research division.

The Bigger Picture: Incumbent Resistance vs. Agility

The exodus feeds a narrative that Google, despite its vast resources, struggles to move quickly in the AI race. Reports of delayed Gemini models, combined with the departure of core technical leaders, reinforce the perception that internal friction and bureaucracy may be stifling execution. Discovery Loop’s four founders are betting that a small, focused team can out-innovate the giant where they came from—a test of whether the incumbent’s full stack is truly an advantage or a burden.

The Stakes for Google, Discovery Loop, and the AI Industry

For Google and Alphabet

• Kavukcuoglu must move quickly to stabilize the DeepMind team and demonstrate tangible progress on Gemini 4. Any further delays could accelerate the loss of enterprise and developer trust to OpenAI and Anthropic, especially given the recent 5% share drop.
• Google’s next quarterly earnings call will be scrutinized for updates on AI product timelines and any signal that the reshuffle has impacted developer morale or hiring. The company’s credibility on AI execution is now on the clock.
• Alphabet’s equity stake in Discovery Loop provides a potential upside, but the board should watch for conflicts of interest as the startup scales. Clear firewalls between the joint research framework and Google DeepMind’s own work will be essential.

For Discovery Loop

• The immediate priority is to hire a core team and secure office space, as Jeff Dean noted. The Google Cloud compute deal offers a running start, but the startup must quickly translate its reputation into a working prototype that demonstrates an automated research loop—ideally within its first year.
• The founders should transparently address how they will manage the relationship with Alphabet. Having a cloud partner that is also a competitor and investor creates risk; governance provisions should be clear.

For the AI Industry

• The exodus reinforces a trend of top researchers leaving big tech to build lean, mission-driven startups. Other AI labs may need to rethink retention incentives and internal agility to prevent a similar brain drain.
• If Discovery Loop succeeds in automating the scientific research loop, it could reshape how R&D is done—shifting value from massive headcounts to highly autonomous systems. Competitors should monitor early signals of progress and consider how such a shift would affect their own research models.