Eight months after I wrote "the loop is closing," Jeff Dean left Google after 27 years to close it. On Wednesday he co-founded Discovery Loop, taking with him Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The new public benefit corporation's mission: automate machine learning research, then use it to automate the rest of science.

The plan is three stages. First, automate ML research itself. Second, use that to optimize Discovery Loop's own technology stack. Third, generalize to any measurable learning loop in science and engineering, with chip design, biology, drug discovery, and materials science as named targets. Every one of those stages assumes the one before it works, which is a lot of assuming, but the names attached to it are heavy enough that the assumption is worth taking seriously.

Jeff Dean was Google's chief scientist and employee number 30, there since 1999. Sanjay Ghemawat was his longtime collaborator. Oriol Vinyals built DeepMind's Flamingo model and co-leads Gemini, their largest language model. Quoc Le co-created Google Brain. The funding is equally heavy-weight: Radical Ventures and Khosla Ventures co-led the round, with Kleiner Perkins, Lightspeed, and Doerr Capital participating. Alphabet is a founding investor and Cloud partner, providing the compute for year one. They didn't disclose the amount raised or the valuation, but with that roster it isn't a seed round.

The timing isn't a coincidence. In November I covered the White House Genesis Mission, which directed the Department of Energy to build robotic laboratories and AI-directed experimentation across all 17 national labs. In February I wrote about Anthropic using Opus 4.6 to debug the infrastructure that measures Opus 4.6, and Anthropic's own acknowledgment that "confidently ruling out these thresholds is becoming increasingly difficult." The more capable the model, the harder it is to honestly evaluate, because above a certain level it begins to participate in its own development.

Anthropic is already using models to debug the systems that measure models. DeepMind just saw Oriol Vinyals, who built Flamingo and co-leads Gemini, leave to build AIs that generate novel research ideas. This is Daniel Kokotajlo's AI 2027 scenario happening in front of us: a research automaton capable of inventing and testing hypotheses beyond what it was trained to consider. It doesn't matter if the year on the scenario was wrong; the sequence is playing out in order.

We are about to see scientific progress at a rate we have no reference for. Dean said it directly: "You will get both a higher quantity and a higher quality of experiments, and that will lead to scientific breakthroughs and advances." The quantity is obvious; thousands of parallel experimental cycles running continuously will produce more raw attempts than any human lab. Quality is the part I can't get my head around. These aren't random tries, they're a search that learns from its own previous rounds. The "loop" in Discovery Loop isn't poetry, it's a wiring diagram: models propose experiments, laboratories run them, previous results inform the next cycle, and around we go.

Vinyals admitted that getting models to generate truly novel ideas is the core problem, but he didn't sound worried about it. I think he's right not to be. Language models are already shockingly effective at open-ended ideation; they only need a bridge back to physical-world experimental validation. Building that bridge is a hardware and robotics problem, not a computer science one. The underlying models are, as far as I can tell, done. I don't mean finished. I mean the remaining work is wiring them up rather than making them smarter.

DeepMind restructuring its leadership around this effort is the other sign of where we are. Demis Hassabis, the co-founder and CEO, stepped back from daily operations to become chair and Alphabet's chief scientist. Koray Kavukcuoglu took over as the Senior VP of Google DeepMind, overseeing Gemini. Moving a founding CEO out of daily ops usually means they need to focus on something more consequential than running the existing business. My read is that Hassabis wanted back in the lab, and I'd like to know what he thinks is worth leaving the corner office for.

I've believed for a while that the critical path to transformative AI has two parts: a language model capable enough to meaningfully participate in AI research, and a way for that model to iterate on itself. Discovery Loop is explicitly founded to do exactly that, and they're building on Gemini, Flamingo, and more than a decade of collaboration between Dean, Ghemawat, and Le. This isn't a side project.

When the White House aimed the national laboratories at this same goal, it was important. Now it's imminent, and the people who built the underlying technology are sprinting at it themselves. Every other contender in this field is going to have to follow, because no research lab can afford not to have a loop if Alphabet has one. Six months ago this was a scenario; as of Wednesday, it's an investment thesis with Kleiner, Khosla, and Doerr attached. I give it a year before it's just how ML systems get built, because the day after a Discovery Loop model invents something important is the day this becomes an all-out race.

I said capability is outrunning evaluation, evaluation is outrunning oversight, and the 2027 timeline is getting tighter. We just lost a year on all three of those sentences. The only way to win an iterated game is to not play it, and opting out of this one isn't on the menu. August 5 won't feel like a milestone in the moment. These dates never do. But it will be the reference point we look back to and say: there, that is when it started.