Skip to content

AI

Google's Jeff Dean Quits to Build a Lab That Runs Itself

Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals are leaving Google for Discovery Loop, a startup that wants thousands of experiments running at once. Humans in the loop are optional, apparently.

Cover photo: Gregory Varnum / CC BY-SA 4.0

TL;DR

  1. Jeff Dean leaves Google to reportedly serve as CEO of new startup Discovery Loop.
  2. Co-founders: Sanjay Ghemawat, Quoc Le, Oriol Vinyals. That is a lot of Google résumé.
  3. The pitch: automate science, run thousands of experiments in parallel, maybe let AI improve AI.

What happened

Jeff Dean, described as one of Google's longest-serving and most influential executives, is stepping down to start his own AI company. He reportedly plans to be CEO.

His co-founders include Sanjay Ghemawat, the Google senior fellow and long-time engineering partner of Dean. Quoc Le, one of the researchers who helped start Google Brain. And Oriol Vinyals, who was a senior scientist at DeepMind.

The company is called Discovery Loop. It is a public benefit corporation, and its mission is to use AI to speed up scientific research.

The plan, per the startup: use powerful algorithms to kick off and refine thousands of experiments at the same time. The goal is to partially automate research and push experimentation to a scale human labs cannot reach by hand.

There is a second ambition, and it is the spicy one. The startup is also interested in using AI to help build more powerful AI, a process known as recursive self-improvement. That would remove humans from the iteration loop entirely.

Yes, the company is literally named Discovery Loop. Subtle, it is not.

Why it matters

Talent moves markets. When several founding-generation names leave the same employer at once, the rest of the industry reads it as a signal about where the interesting work, and the interesting equity, now lives.

The target is also telling. Research itself is the bottleneck: experiments are slow, sequential and human-limited. A company built to attack that constraint is betting that the next big gain comes from faster discovery, not just bigger chatbots.

Then there is the recursive self-improvement line. It is the oldest sci-fi trope in the AI book, and now it sits in a startup's stated interests, with a public benefit corporation badge on top. We are not panicking. We are noting it, sapiens.

One caution: the details so far come from a single report and the startup's own description of its plans. No product, no funding figure, no timeline. Ambition is not output.

What to do with it

  • Watch the hiring.Who follows these four out of Google will tell you more than any launch post.
  • Track the claims."Thousands of parallel experiments" is a pitch until Discovery Loop shows results, so wait for the numbers.
  • Audit your own loop.If your work is slow, sequential and manual, that is exactly the shape of task this startup wants to automate.
  • Read the structure.A public benefit corporation promises a mission beyond profit, so check what that mission commits them to once details appear.

Ciao, sapiens. The loop is running, and your invitation to it is still pending.