Jeff Dean Just Left Google to Automate Science — and Alphabet Is Helping Fund It

On the same Wednesday in August, Google announced two things at once. Its most famous engineer is leaving the company he helped build. And its most famous AI executive is stepping back from running the day-to-day machine. Jeff Dean is going off to start a company. Demis Hassabis is handing over the CEO seat at DeepMind. And here’s the part that made me do a double take: Google’s parent company is funding the first move with real money.

>This isn’t a story about a company losing a war. It’s a story about the people who built modern AI quietly rearranging the map — and the giant they’re leaving deciding to buy a ticket instead of standing in the way.

Jeff Dean and Demis Hassabis: Two Announcements, Back to Back

On August 5, 2026, Jeff Dean — Google’s chief scientist, employee number 30, a man who has worked there since 1999 — announced he was leaving after 27 years to launch Discovery Loop, an independent public benefit corporation that wants to use AI to automate scientific research. He’s taking three heavyweights with him: Sanjay Ghemawat, a Google Senior Fellow and his collaborator for more than two decades; Quoc Le, a founding member of Google Brain; and Oriol Vinyals, a VP of research at Google DeepMind. Dean plans to serve as CEO.

Hours earlier or later — depending on your timezone — Alphabet CEO Sundar Pichai announced a shakeup of DeepMind’s top leadership. Demis Hassabis, who built DeepMind from a London startup into the world’s most decorated AI lab, is stepping down as CEO to become the chair of Google DeepMind and chief scientist of Alphabet. He’ll keep running Isomorphic Labs, DeepMind’s AI drug-discovery spinoff, and work with Pichai on AGI strategy. His successor: Koray Kavukcuoglu, formerly DeepMind’s CTO, now SVP reporting directly to Pichai.

Two stories, same day, one message: the founding generation of Google AI is moving into a new phase — and Google wants to stay in the room for all of it.

The Counterintuitive Part: Google Is Funding the Exodus

Here’s where my brain snagged. Alphabet is a founding investor in Discovery Loop. It’s also the startup’s Cloud partner, and it plans to collaborate on research into machine learning systems and infrastructure. The initial funding round is being co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed, and Doerr Capital also in.

Think about what that means. The most valuable researcher Google ever had walked out the door — and Google wrote a check to help him leave. That seems backwards until you look at it the way a chess player looks at a trade: Google is giving up a piece, but it’s keeping the initiative.

This is the same pattern I traced when I looked at Nvidia’s $50 billion lease and $250 billion guarantee — the AI industry increasingly runs on circular money. Labs fund the infrastructure they depend on, and investors back the people they once employed. If Discovery Loop becomes a major consumer of compute, guess which cloud gets the bill? If it cracks automated chip design, Google wants the first look.

You can’t handcuff talent. You can only decide whether the person leaving becomes a competitor or a partner. Google chose partner.

What Discovery Loop Actually Wants to Build

Strip away the press release and the pitch is remarkably simple: automate the scientific method itself. Instead of a human researcher proposing an experiment, running it, analyzing the result, and starting over — slowly, one iteration at a time — Discovery Loop wants AI systems to propose, run, evaluate, and iterate on experiments thousands at a time, in parallel, using massive compute.

Per the company’s own description, it starts by automating machine learning research and engineering: using frontier models to generate hypotheses, run evaluations, and learn from the outcomes. The stated expansion path runs from ML research into chip design, biology, drug discovery, and materials science.

Dean’s quote to the New York Times captures the ambition: “We think there is opportunity for AI to more fully automate what has traditionally been a very human-intensive experimental loop. You will get both a higher quantity and a higher quality of experiments, and that will lead to scientific breakthroughs and advances.”

And then there’s the part that should give you pause. The company is also interested in recursive self-improvement — using AI to help build more capable AI, cutting human iteration out of the loop entirely. That’s the phrase the industry uses when it’s talking about the machines improving themselves faster than we can improve them.

Khosla Ventures’ Vinod Khosla put the shift bluntly in an interview with Wired: humans have spent the last few years using AI to do research; Discovery Loop’s premise is that AI is the researcher. That’s a quiet but enormous change in who we think is doing science.

AI Has Already Started Doing Some of This

This isn’t science fiction, and it’s not even that new. I wrote earlier this summer about an AI system that cracked a 35-year-old math problem and found something no human predicted — the kind of result that used to be a once-a-generation event for a research team. What Discovery Loop is proposing is industrializing that: making discovery a factory process instead of an artisan craft.

The founding team’s collective résumé is the reason people are taking this seriously. Their work spans Google Search, Google Translate, the Google File System, MapReduce, BigTable, Spanner, and TensorFlow. Wired’s Steven Levy — who interviewed the four founders — reported the whole idea came together only a few weeks ago. Four of the most accomplished systems builders in the industry decided, in the span of weeks, that the single most impactful thing they could do was leave the biggest company in tech and automate the lab.

The Bigger Pattern: The Talent Exodus Isn’t New

What makes this week feel different is that Dean’s exit lands on top of a longer trend. In June, Noam Shazeer — co-author of the Transformer paper, the architecture underneath basically every modern LLM — left Google’s Gemini team to join OpenAI. The same week, John Jumper, the AlphaFold lead who shared the 2024 Nobel Prize in Chemistry with Hassabis, announced he was joining Anthropic after nearly nine years at DeepMind. Shazeer had been through this dance before: Google spent roughly $2.7 billion in 2024 to bring him back from Character.AI. Two years later, he left again.

Put it together and the picture is stark. Google is losing the people who invented the Transformer, built the search infrastructure, founded Google Brain, and led AlphaFold — all within about seven weeks. Some to rivals, some to their own startups, and in every case, the money to leave came from somewhere.

The frontier labs are fighting over the same finite pool of people who know how to build this stuff. I’ve argued before that open-weight models are catching up to the frontier, which means the models themselves become less of a differentiator over time. What’s left as the real moat? The people. And those people are mobile.

Why the AGI Language Matters

Hassabis framed his own move in unmistakable terms. “I’ve been working towards AGI my whole life, and as we enter this pivotal moment, I’m stepping into a new role as Chair of Google DeepMind and Chief Scientist of Alphabet,” he wrote. Pichai’s announcement used the same vocabulary — this move, he said, lets Hassabis put his full attention on actively shaping the future of AGI. “We have arrived at a pivotal moment in human history,” Hassabis added. “I feel it is close at hand.”

Read those quotes again. The two most senior AI figures at Google are not reorganizing because of a quarterly earnings miss. They are telling you, in public, that they believe AGI is close enough to warrant restructuring the company around it. That’s the same debate I covered when Sam Altman declared we’re in the singularity while Satya Nadella warned against betting everything on it — except now the “pivotal moment” language is coming from the leadership transition itself.

What This Looks Like From Where I Sit

I run an ICT division in a Philippine government institution. We don’t have billion-dollar talent wars, but I know exactly what it feels like when your best senior developer gives notice. The first thing you feel is the knowledge walking out the door with them — the undocumented system, the unwritten decision, the context that lives in their head and nowhere else.

In the Philippines we have a name for this: brain drain. Our best engineers leave for Singapore, for the Middle East, for remote roles at companies that pay in dollars. I’ve watched agencies try to fight it with retention bonuses and job titles, and I’ve watched those fail. You know what actually works? Two things. First, documentation and knowledge transfer — making sure the institution keeps the memory even when the person moves on. Second, treating the departing person like a future ally, not a traitor. Keep the relationship warm. You’d be surprised how often they come back — or send work your way.

Google, in its clumsy, multi-billion-dollar way, just did both. It funded the knowledge transfer (literally — Alphabet is a founding investor) and it kept the relationship warm (Cloud partner, research collaboration). It’s the chess move of a player looking at the whole board: you can’t stop the talent from leaving, so make sure you profit from where they land.

The boxing equivalent is rolling with a punch instead of blocking it. Block hard enough and you break your own hand. Roll, and the force carries you somewhere useful.

The Uncomfortable Question Nobody’s Asking

Everyone is focused on whether Discovery Loop will work. I want to ask something different: if it does work, who checks the work?

An AI that runs ten thousand experiments a day will find things no human predicted — that’s the whole point. But it will also hallucinate discoveries, chase artifacts, and optimize for metrics that don’t mean what they seem. The “human gate” doesn’t scale to ten thousand results a day. We are about to find out what happens when the pace of claimed discovery outruns the pace of verification.

I keep a human gate on my own AI agents for the same reason — the machines are fast, but they are not accountable. That tension gets a hundred times sharper when the experiments are running in a biology lab or a chip foundry instead of a Python notebook.

And let’s be honest about what recursive self-improvement means in that world. The system that designs the next system is a system no human fully reviewed. The people building it are the ones who know best how much that should scare us — and they’re the ones racing toward it.

The Bottom Line

There’s a through-line from Meta’s launch of Muse Code, its AI agent for huge codebases, to Discovery Loop’s plan to automate the lab: the industry’s center of gravity is shifting from models to builders. The companies that win the next phase won’t be the ones with the biggest training run. They’ll be the ones with the best people — and the ones that figured out what to do when those people leave.

Google just showed its hand. It’s not trying to keep everyone forever. It’s trying to make sure that wherever the founders go, the giant has a seat at the table.

So here’s my question for you: if your most talented person gave notice tomorrow, would your organization turn them into a competitor — or a partner?