The Builders of Google AI Left to Automate What They Built It For

On August 5, Jeff Dean announced he was leaving Google after 27 years to co-found a company called Discovery Loop. He took Sanjay Ghemawat with him, along with Quoc Le and Oriol Vinyals. Those four people built the infrastructure that made current AI possible: the distributed file systems, the training clusters, the model architectures, the software stack. MapReduce. BigTable. TensorFlow. TPUs. AlphaFold. Gemini. Now they are leaving to build something those systems cannot build on their own.

The funding coalition confirms this is serious: Radical Ventures and Khosla Ventures are co-leading the seed round, with Kleiner Perkins, Lightspeed, and Doerr Capital participating. Alphabet is both a founding investor and cloud partner. No valuation or dollar amount has been disclosed, which means the round is either oversubscribed and they see no reason to anchor expectations, or the architecture of the deal is unconventional enough that a headline number would mislead. Neither interpretation is bearish.

The right response is not to cover this as a talent story. It is to ask what these four specific people — who understand better than almost anyone alive what large-scale AI can and cannot do — believe the constraint is now. Their answer, as stated directly and embedded in every technical decision in Jeff Dean’s own words: “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.”

The Experiment Loop

The experimental loop Dean is describing is the backbone of every research program that has ever produced a real result. A researcher forms a hypothesis. They design an experiment to test it. They run the experiment, collect data, analyze the results, revise the hypothesis based on what they learned, and repeat. Every step in that cycle — except sometimes the running of the experiment — is currently done by a human. Not because humans are uniquely qualified to do it, but because the software to automate the judgment at each handoff point hasn’t existed at sufficient quality to be trusted.

Discovery Loop’s plan is to automate the entire cycle, running thousands of iterations in parallel. Their Radical Ventures investors describe the system as one where “the AI proposes the experiment, conducts the run, learns from the result, and iterates recursively.” The recursion is the important word. It’s not a system that proposes experiments for humans to run — it’s a system that closes the loop computationally, end to end, without approval at each step.

Their initial domain is ML research itself. This is the decision that reveals the most strategic clarity. ML experiments are code. Running an ML experiment means writing a training script, executing it on accelerators, reading evaluation metrics, and deciding what to change. Every step in that loop is software. There’s no wet lab, no physical synthesis, no biological system whose behavior can’t be captured in a number. Discovery Loop is starting in the one scientific domain where the full experimental loop — from hypothesis to result — is a computation problem, and computation problems are what the founding team has spent three decades making faster.

The plan after that: chip design, drug discovery, biology, materials science. Those domains still require physical experiments. The strategy is to prove the loop works in ML first — to demonstrate that automated iteration at scale produces better results than human-directed iteration at human speed — and then expand into domains where the feedback signal comes from physical reality rather than benchmark metrics. That sequencing is careful. It doesn’t promise to solve wet-lab validation on day one. It promises to demonstrate the method works where the method is cleanest, then carry the credibility into harder domains.

What This Team Knows

The credentials here matter in a specific way. Jeff Dean and Sanjay Ghemawat published the MapReduce paper in 2004 and the BigTable paper in 2006. Those weren’t research contributions in the normal sense — they were existence proofs that problems previously considered intractable at scale were tractable if you re-architected the computation. Training runs that would have taken years on single machines became feasible on distributed clusters in weeks. The co-authored papers on Google’s infrastructure stack are, in retrospect, the engineering decisions that made modern AI training possible. Dean went on to help launch Google Brain, oversee TensorFlow’s development, and serve as Chief Scientist before this week.

Quoc Le built large-scale language model pretraining systems at Google Brain and co-developed sequence-to-sequence learning architectures — the foundations on which neural machine translation and eventually the large language model pretraining paradigm were constructed. He knows what scaling those systems looked like from inside the lab that was doing it. Oriol Vinyals led AlphaCode, DeepMind’s competitive programming model, and AlphaStar, which reached grandmaster-level play in StarCraft II. AlphaStar is particularly relevant context: it reached grandmaster by running approximately 200 years of game experience across 16 weeks, playing against itself millions of times, evaluating results, and updating policy without a human in the loop. That is an automated research loop applied to the problem of game strategy. Discovery Loop is the same pattern applied to the problem of ML research.

The simultaneous restructuring at Google DeepMind compounds the signal. Sundar Pichai’s announcement on August 5 moved Demis Hassabis from CEO of Google DeepMind to Chair of Google DeepMind and Chief Scientist of Alphabet, with a mandate to “actively shape the future of AGI.” Koray Kavukcuoglu — formerly CTO, 13 years at DeepMind — becomes SVP of Google DeepMind and takes operational control of Gemini model development, the Gemini app, and developer teams. Google is reorganizing its AI leadership around two distinct axes: near-term product delivery and long-run AGI strategy. The infrastructure architects are now outside, building what they believe is the next layer.

Google retaining a stake as founding investor and cloud partner is not a contradiction. It’s consistent with Google’s history of funding and hosting ventures that extend its infrastructure reach even when they technically compete. What it clarifies is that this is not an acrimonious departure — it’s a structured exit that preserves Google’s access to whatever Discovery Loop produces while freeing the founders to pursue a bet that doesn’t fit inside a product organization built around Gemini quarterly releases.

What the Critics Have Already Found

The objection to automated research loops is not theoretical. There is real evidence, published this year, that current AI scientist systems fall well short of their marketing. A May 2026 analysis of Google DeepMind’s Co-Scientist and Future House’s Robin found that both systems “rely heavily on human input to define the key scientific question, sense-check predictions, and prioritize predictions for further investigation.” Neither one validates hypotheses through physical experiments — they stop at the proposal stage. Of five cancer treatment candidates Co-Scientist identified in one drug repurposing exercise, only three showed positive results in lab testing, and the results weren’t compared against specialized computational methods, meaning even the positive results are not contextualized.

The NCBI’s survey of AI for scientific discovery identified the primary constraint with precision: “validating AI-generated designs through laboratory experimentation — a time-consuming, costly human-dependent process.” In physical sciences, the problem isn’t generating hypotheses — AI systems can generate hypotheses faster than researchers can test them. The problem is that testing requires physical equipment that generates data at physical-world speed, operated by humans who are in limited supply. Automation that accelerates hypothesis generation without accelerating validation just creates a larger backlog.

There is also a quality-of-ideas problem distinct from the quantity problem. Research in the field shows that LLM-generated ideas are often perceived as novel but tend to underperform human-generated ideas after implementation. A system producing ten thousand experimental proposals is not ten thousand times more useful than one that produces ten — it’s only more useful if the proposals are more likely to be correct, testable, and interpretable. And interpretability is its own problem: can a fully automated system determine whether an experimental result is informative or noise? Can it distinguish between an architecture that outperforms a benchmark because it solves the underlying problem and one that outperforms because it overfits to the test distribution? That is a judgment call. Judgment doesn’t disappear because you run the loop faster.

Discovery Loop’s starting domain — ML research — partially insulates the company from the physical-validation objection. An ML experiment produces a training curve, a benchmark number, a loss function. Those signals are clean, measurable, and fast. Running a thousand ML experiments in parallel is an infrastructure problem with infrastructure solutions. Whether automated iteration over those signals produces better research than human-directed iteration is still an open empirical question — but it’s at least a well-posed one, with a feedback signal you can trust. The same cannot be said for drug discovery at this stage, which is why starting there would have been the wrong call.

What remains is the quality problem. A system optimizing against training loss or benchmark metrics will find ways to improve those numbers that may not correspond to genuine research progress. That’s not a hypothetical — it’s the core of Goodhart’s Law applied to automated research. Discovery Loop’s founders know this. Oriol Vinyals built AlphaStar, which had exactly this problem at several points in its development: self-play optimization producing strategies that exploited game mechanics rather than developing genuinely stronger play. The fix required re-designing the reward signal. Building a good reward signal for automated ML research is probably the hardest part of what Discovery Loop is attempting.

That’s the question the company is actually trying to answer. Not whether AI can run experiments faster — it obviously can. Whether AI can be trusted to evaluate which experimental results are worth building on. That’s the loop Discovery Loop is really trying to close, and it’s harder than the founding deck probably makes it look.

The Vote

Jeff Dean knows what another order of magnitude of training scale buys. He has spent the last decade watching the returns on scaling. Sanjay Ghemawat has seen what the infrastructure ceiling looks like from inside the people who built it. Quoc Le has watched which research investments at Google Brain produced real architectural breakthroughs and which produced benchmark gains that didn’t transfer. Oriol Vinyals ran the AlphaStar research loop for years and knows what happened when automated self-play worked, and when it didn’t. When those four people leave, together, to build a system that automates the research loop they used to run by hand, the correct interpretation is not that they believe this is easy. It’s that they believe the bottleneck has moved — from compute and scale to research iteration speed — and that the tools to address the new bottleneck do not currently exist inside the companies that built the scale machinery. The experiment is now running.

AI-generated editorial illustration · TemperatureZero · August 6, 2026