Jeff Dean and Google legends launch Discovery Loop

If you had to pick the ultimate coding duo in tech history, Jeff Dean and Sanjay Ghemawat would top almost everyone's list. Now, they are finally leaving Google to build something new.

Dean, Google's 30th employee and longtime chief scientist, announced on August 5 that he is stepping down to launch Discovery Loop. He is taking an all-star crew with him: legendary systems engineer Sanjay Ghemawat, Google Brain co-founder Quoc Le, and DeepMind senior research scientist Oriol Vinyals. The four co-founders share between 14 and 30 years of history together, having built the foundation of modern Google and its Gemini models. Dean will serve as CEO.

Discovery Loop is registered as a Public Benefit Corporation, and it is going straight after the bottleneck of human intelligence.

Automating the scientific loop

Science and engineering have always relied on a slow, human-led loop: propose a hypothesis, run a test, read the log, repeat. Discovery Loop wants to wipe that bottleneck out.

The startup plans to deploy high-scale algorithms capable of launching and iterating thousands of experiments at once. The goal is to automate complete experimental loops, starting with large-scale machine learning research before tackling problems from the NAE Grand Challenges.

Crucially, the team also wants to explore recursive self-improvement: letting AI systems iterate and improve future AI models with minimal human interference.

The industry is backing them heavily. Discovery Loop's seed round is co-led by Radical Ventures and Khosla Ventures, alongside Lightspeed, Kleiner Perkins, and Doerr Capital. In a telling move, Google's parent company Alphabet also participated as an investor.

Why it matters

We think this is one of the clearest signals yet that frontier AI is shifting away from static chatbots toward automated discovery engines.

For researchers and engineers, the promise here is monumental. If machine learning experimentation can be automated, researchers will spend less time babying training runs and more time setting high-level hypotheses. The bottleneck moves from human bandwidth to pure computational capacity.

For the tech giants, it is a wake-up call. When Google's most revered architect leaves after 27 years to build this elsewhere, it suggests big-tech bureaucracy might be too sluggish for this level of ambitious research. Alphabet investing in the round looks like a clever way to keep a foot in the door, but it also shows they could not retain the vision internally.

There is, however, a catch. Running thousands of parallel experiments works beautifully when everything lives in software. When Discovery Loop eventually tries to tackle broader physical science and engineering problems, real-world lab throughput and verification will remain stubborn, analog roadblocks.