AIWatch.

Google is not falling apart, and I do not think this means it is losing the AI race. It still has arguably the strongest full stack in the industry: its own TPUs and data centers, (still) world-class research teams, Google Cloud, and distribution through Search, YouTube, Android, Chrome and Workspace.

Very few companies can turn a research breakthrough into a product used by billions of people as quickly as Google can.

But (!) the departure of Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le exposes a deeper problem: Google is struggling to retain some of the people who created that advantage in the first place, some of the most influental key reserachers in ML.

Jeff Dean did not merely help build Google’s AI division. Together with Ghemawat, he helped build the infrastructure that made Google possible, from MapReduce and Bigtable to Spanner. He later helped establish Google Brain and contributed to DistBelief, TensorFlow, TPUs, Pathways, PaLM and Gemini. Vinyals became a key figure in sequence-to-sequence learning, AlphaStar and Gemini, while Quoc Le was a founding member of Google Brain and spent years working on automated machine learning.

Their new company, Discovery Loop, wants to automate machine learning, science and engineering. The idea is to build systems that can formulate hypotheses, design and run experiments, evaluate the results and use them to generate the next experiment. It is a bet that the next major AI product will not simply answer questions, but continuously produce new knowledge.

That ambition is not unique. OpenAI has publicly targeted an AI research intern by September 2026 and a “legitimate AI researcher” by March 2028 (image down below). And Anthropic has also repeatedly made similar ambitions public. Google itself is pursuing automated discovery through projects such as AlphaEvolve and Isomorphic Labs, while Anthropic is recruiting heavily from scientific and frontier-model teams.

So the important question is not why these researchers believe in automated science. Almost every frontier lab now believes in it.

The question is (and comes down to) why they concluded that the best place to pursue this vision was outside Google!

According to Wired, Sundar Pichai tried repeatedly to persuade them to stay. Oriol Vinyals described the inertia that comes with large organizations and the difficulty of making radical changes inside them. That does not prove that compensation or internal disagreements played no role, but it tells us something important about the tension inside Google: extraordinary resources do not automatically create an environment in which a small group can pursue one radical idea with complete focus.

Google must allocate talent and compute across Gemini, Search, Cloud, advertising, consumer products and long-term research. Discovery Loop can build its entire organization around a single objective. Frontier AI is not simply a headcount contest. A relatively small number of researchers decide which architectures deserve enormous amounts of compute, which failures contain useful signals and which ideas are worth pursuing for years (remember: just 2025 Meta paid up to 100m for key reseracher, individually). Google retains the technology these people built, but it loses some of their tacit knowledge and judgment about what should be built next.

Google’s response shows that it understands both the risk and the opportunity at the same time. It will become a founding investor and cloud partner in Discovery Loop and provide compute during the first year.

This looks less like a hostile breakaway than a negotiated spinout!

Google loses direct control over the team and its research agenda, but preserves a financial and technical relationship with whatever emerges.

The same ambiguity applies to Demis Hassabis stepping away from day-to-day management. He is not leaving Google. As Chair of Google DeepMind and Chief Scientist of Alphabet, he will focus more closely on AGI, long-term strategy, its societal consequences and Isomorphic Labs. Koray Kavukcuoglu will take over operational leadership as Senior Vice President, with a clearer focus on execution and turning research into products.

The positive interpretation is that Google is separating two jobs that have become too large for one person: long-term scientific leadership and the operational delivery of frontier models across a global product ecosystem. The more critical interpretation is that DeepMind is becoming more tightly integrated into Google’s product machine, potentially reducing the independence that once made it such an unusual research institution. Both interpretations could be true at the same time.

The timing nevertheless raises legitimate questions: Gemini 3.5 Pro has not arrived within the expected timeframe, while attention is already shifting toward Gemini 4. That does not prove that Hassabis was moved because of execution problems, but combined with the recent departures of David Silver, Noam Shazeer, John Jumper, Jonas Adler and Alexander Pritzel, it suggests that Google is trying to solve a difficult structural problem. But its not a good sign about their new frontier model tho.

How does a company preserve the freedom, concentration and willingness to take risks that produce foundational breakthroughs, while also turning those breakthroughs into reliable products for billions of users? Thats the big question.

Google is not collapsing. It may still be better positioned than almost anyone to win the next phase of AI. But its greatest challenge is no longer simply building better models. It is creating an organization in which its best researchers still believe their most ambitious work can happen inside Google.

The silence surrounding Gemini 3.5 Pro and the repeated references to the new Gemini 4 (including new pre-training, as far as I know) speak volumes and are both a good and a bad sign.

Both at the same time.