Discovery Loop will point its first experiments at its own machine-learning algorithms, and Google — a founding investor — supplies the computing power for the opening year.

Jeff Dean, who helped start Google Brain and is chief scientist of both Google DeepMind and Google Research, is leaving after just under 27 years to build machinery that runs the scientific process without a person in it: propose an experiment, implement it, run it, evaluate the result. Three senior colleagues are going with him, and losing all four at once is a devastating blow to Google as it tries to keep pace in the competition over AI models.

They are Sanjay Ghemawat, with whom Dean wrote large portions of the systems that underpin computing and search at Google; Oriol Vinyals, a research VP at DeepMind and one of Gemini's technical leads; and Quoc Le, another Google Brain cofounder and the principal scientist behind AutoML-Zero, a project in which machine learning algorithms build AI products on their own.

Run thousands of those loops, Dean has argued, and they could unlock advances in science, biology, chip design and AI models themselves. Discovery Loop means to be its own first customer, aiming the loops at its own advanced algorithms before putting the upgraded software at the centre of a system pointed at materials, drugs and chips. Le said the prospect of automating machine learning excites him, and that they might turn up a different transformer architecture. Vinod Khosla, who funded them, put the distinction plainly: people have been using AI to carry out research rather than to be a researcher, and what is fundamental here is that the AI is the researcher.

The founders are equally plain about what is missing. What they will have to concentrate on, Vinyals said, is how these models arrive at new ideas worth trying — not something they are currently very strong at. At the start, ideas are to be developed jointly with people.

Khosla met the four at his Sand Hill Road office on a Saturday, timed so nobody would notice the loss Google was about to take. A team of that quality, he said, was fundable without being told what they planned to work on. They brought a few slides covering their backgrounds and a rough outline of why the approach would work; the deck was not made with AI. Khosla Ventures and Radical Ventures put money in, along with a small number of other firms, and neither the amount nor the valuation has been disclosed. Some companies have paid tens or hundreds of millions of dollars to hire a single star researcher.

Google sent them off with its approval and its money. Sundar Pichai said the company will stay on as a founding investor and as the venture's Cloud partner, and that the two will jointly develop a research framework covering machine-learning systems and the infrastructure behind them. Negotiating the exit proved more painful than raising the funding, Dean said; Pichai spent multiple meetings trying to keep them.

So far the startup is founders and funding. It has not begun hiring or rented office space. Asked who runs it, Dean hesitated, then said he supposed he was the CEO, since everyone had pointed at him.

Cover image: “Googleplex (4842867784)” by John Marino from Pittsburgh, The fine US of A, Wikimedia Commons, CC BY-SA 2.0, resized, re-encoded.