Google AI Legends Walk Out to Build Autoresearch

William Palmer

August 5, 2026. Alphabet stock drops four percent in a single morning. And, uh, look, Wall Street reacts to a lot of things, but this, this was different. This was the news that Jeff Dean is leaving Google after twenty seven years. And he's not going alone. He is walking out the door with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le.

William Palmer

I mean, as a computer scientist, I, I, I genuinely had to sit back in my chair when I read that. You have to understand who these people are. Jeff Dean and Sanjay Ghemawat are not just senior executives or famous researchers. They are the literal builders of the modern internet. Back in the early days, working out of a small office in Palo Alto, they built MapReduce. They built BigTable. They built TensorFlow. They constructed the compute architecture that allowed Google to index the entire web and train the largest neural networks on earth.

William Palmer

So why leave now? Why leave when Google DeepMind has the largest trove of training data and more compute infrastructure than almost anyone in human history? Well, you start to see it when you look at how DeepMind operates today. Every flagship paper for models like Gemini now carries over one thousand coauthors. One thousand! When you get to that scale of corporate consensus, even the brightest minds end up spending more time coordinating mega structures than running lean experiments. It is like trying to turn a supertanker when what you really need is a nimbler machine.

William Palmer

And look at the internal reshuffle that happened right alongside this exodus. Demis Hassabis is moving from Chief Executive Officer to Chair of Google DeepMind and Chief Scientist of Alphabet. He is stepping back from day to day operations to focus on long term AGI strategy and his work at Isomorphic Labs. Meanwhile, Koray Kavukcuoglu steps up to Senior Vice President to take direct operational control of DeepMind and keep Gemini execution on track. It is a massive governance reset inside Alphabet, but the subtext is undeniable. The foundational architects of modern AI decided that if they wanted to build the next breakthrough, they had to walk away from the mega lab and go back to a clean sheet of paper in Palo Alto.

William Palmer

So what are Jeff, Sanjay, Oriol, and Quoc actually building at their new Palo Alto startup, Discovery Loop? They are not building another generic chat assistant. They are not trying to train a ten trillion parameter general purpose model to answer trivia or write polite emails. They are targeting what many of us in the field consider the true frontier: automated scientific discovery.

William Palmer

The core mission of Discovery Loop is automating machine learning, science, and engineering workflows through closed loop autoresearch. Instead of a human scientist sitting down to design a test, code it up, analyze the logs, and tweak the hyper-parameters, you build an agentic system capable of proposing experiments, running them, evaluating results, and iterating, thousands of times over. That exact loop, proposing experiments, running them, evaluating results, and iterating, thousands of times over, is how you compress ten years of scientific trial and error into a single week.

William Palmer

Now, here is the fascinat, the fascinating paradox in all of this. Who is one of the key investors in Discovery Loop's seed round alongside Radical Ventures and Khosla Ventures? Alphabet! Google is actually putting money into the startup formed by the very team that just left them. Plus, they are providing access to compute. Which raises a huge, unavoidable question: why couldn't Google just build Discovery Loop inside DeepMind?

William Palmer

I, I think the answer reveals a fundamental truth about where artificial intelligence is heading. The frontier of capability is shifting. It is moving away from brute force parameter scaling inside centralized corporate studios and toward self improving algorithmic loops. But big tech bureaucracies are optimized for risk mitigation, product integration, and massive thousand author papers. They are not built for four superhumans working in an agile garage setup to blow up existing paradigms.

William Palmer

When John Jumper left DeepMind for Anthropic, or when Noam Shazeer left for OpenAI before his brief return, people thought those were isolated moves. But coupling those departures with David Silver, Denny Zhou, and now Jeff Dean and his core team... it feels like a turning point. If the world's most capable infrastructure builders have to leave the biggest lab on Earth just to push automated science forward, then maybe the era of the centralized tech megastudio holding a monopoly on innovation has officially hit its ceiling. It is going to be incredible to watch what they build next.