Jeff Dean leaves Google for Discovery Loop — roadmaps are people

Four of the people most responsible for the infrastructure under your AI stack walked out of Google this morning. Jeff Dean left after 27 years to co-found Discovery Loop, and he took Sanjay Ghemawat, Quoc Le, and Oriol Vinyals with him. Alphabet is one of the investors. If you buy AI from a vendor — and you do — this is the clearest reminder available that a product roadmap is a staffing chart with a logo on it.

What actually happened

Per TechCrunch, Discovery Loop is structured as a public benefit corporation and intends to automate scientific research — running thousands of experiments in parallel rather than answering questions. Dean's framing: more experiments at higher quality produce breakthroughs faster.

The initial round is co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed, Doerr Capital, and Alphabet participating. Wilson Sonsini confirms it advised on the launch and initial funding.

Here is what nobody has published: a round size, a valuation, a product, a launch date, or a customer. Reports indicate the round is not closed and the company declined to give a valuation. Treat every number you see attached to this in the next week as a rumor until someone names a filing.

The departures themselves are not ambiguous. Dean was Google's chief scientist. Ghemawat is a senior fellow and co-author of the systems work — MapReduce, Bigtable — that the modern data stack was copied from. Le was a founding member of Google Brain. Vinyals was a senior research scientist at DeepMind. Google has not named replacements.

Why vendor brain drain matters for your business

Your contract covers uptime, not continuity of judgment. Read your AI vendor agreement and you will find latency, availability, data handling, and a price. You will not find a clause about who decides what the model does next, or what happens when that person leaves. The roadmap you bought into is an unenforceable promise made by people who can resign.

This is not a prediction that Gemini gets worse. Big labs are deep, and Alphabet funding the spinout looks like a deliberate move to keep the relationship rather than lose it outright. It is a prediction that roadmaps drift, models get retired, defaults change, and none of it will be announced in a way that fits your release schedule.

So build for drift. Pin the model version in config, not in a prompt file. Keep a small eval set — twenty real inputs and the outputs you accept — and run it every time a version moves. Route model calls through one internal module so swapping providers is an afternoon, not a quarter. And buy on what a vendor ships today, never on what a keynote says arrives next year.

Sources: TechCrunch, Wilson Sonsini.