The People Who Built Google's AI Are Leaving to Build AI That Does the Research Itself. Meet Discovery Loop.

On 5 August, Google confirmed a departure that would have been unthinkable a year ago. Jeff Dean, its chief scientist and 30th-ever employee, is leaving after joining in 1999, and he is not going alone. Sanjay Ghemawat, a senior fellow, Quoc Le, a founding member of Google Brain, and Oriol Vinyals, a senior research scientist at DeepMind, are going with him. Between them these four have fingerprints on much of the modern stack: the infrastructure beneath Google Search, the research that became Google Brain, and the work behind Gemini. They are leaving to build a startup called Discovery Loop.

What Discovery Loop is actually trying to do

Discovery Loop, structured as a public benefit corporation, is a bet on the biggest idea in AI right now. Its goal is to automate the scientific research process itself: use AI to initiate and iterate thousands of experiments at once, and explore recursive self-improvement to cut humans out of the slowest loops. "You will get both a higher quantity and a higher quality of experiments, and that will lead to scientific breakthroughs and advances," Dean said. The plan is to start by automating machine-learning research and engineering, the work these four know best, then push into hardware design, drug discovery, and clean-energy problems.

Why this is bigger than a talent story

It is easy to file this under famous-people-leave. The more important read is what they chose to build. The frontier of AI is quietly shifting from training a better model to automating the loop that trains better models. If the research cycle itself, form a hypothesis, run the experiment, judge the result, can be run by machines thousands of times over, the rate of progress stops being bounded by how many experiments humans can supervise. That is the wager. It is also why choosing which model to run and how much to spend on it, as we covered with GPT-5.6's three tiers, keeps shifting under everyone's feet: the people who make the models are now trying to make the model-making faster.

The part that says the most about the industry

Look at the cap table. The seed round is co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed and Doerr Capital in, and Alphabet itself participating as an investor and cloud partner. Google is helping fund the company its own stars left to start, keeping a stake and a compute contract in the process. That is not sentiment. It is how the talent-and-compute economy now works: the scarce inputs are a handful of researchers and a mountain of GPUs, and the incumbents would rather own a slice of the breakaway than lose it outright. Bloomberg called Dean the most Google person there is. Google still made sure it kept a seat.

What it means if you are not a frontier lab

For everyone building on top of AI rather than inside a lab, the signal is direction, not disruption. Expect the base models to keep getting better and cheaper on a faster clock, because some of the sharpest people alive are now working full-time on speeding that clock up. The durable advantage does not sit in the model. It sits in the specific, grounded problem you point it at, the proprietary data and the decision only you can define. General research automation races ahead. Applied optimisation in your own operation, deciding what to actually do with all that capability, is a different and quieter craft, and it is the layer Cognilium works in.