Jeff Dean's Discovery Loop Should Automate Chip Design First

Discovery Loop is one day old and, as someone who worked at Google (2014-2022), in my view the most interesting "neolab". Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le left Google this morning to found it, to automate machine learning, science, and engineering. Jeff's announcement says they will automate the experimental loop itself, proposing parallel experiments, running them, evaluating results, and iterating. Their first target is, of course, ML research and engineering, with a stated ambition, "important subproblems in nearly every one of the fourteen NAE Grand Challenge problems."

So what type of science should they automate? I forecasted, conditional on them focusing on one of the 14 domains they mention, their odds of a breakthrough by 2028. My theory is that our model of AI 2027 timelines forecast, e.g. how AI speeds up R&D feedback loops, will be useful here.

My conclusion is that Discovery Loop should focus on chip design, with a 38% chance of a major breakthrough by 2028. Second is secure cyberspace and third are tools of scientific discovery.

This is broadly consistent with my and other forecaster's views on AI 2027. FutureSearch predicted that superhuman coding and other types of AGI would not come until 2031-2032, because of how hard automating these feedback loops can be. Discovery Loop will be a direct test of some of this theory.

First, though, since these are all former Google Brain / DeepMind people, we should think about their priors on this based on how well DeepMind did when it tried to apply AI to speed up scientific discovery.

DeepMind using AI for science

Before that actually, some forecast-relevant facts about Discovery Loop. Their seed round was co-led by Radical Ventures and Khosla Ventures, with Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet participating, terms undisclosed. Unsurprisingly Google Cloud supplies compute. No employees, no office, just a plan to hire a lean in-person founding team. I forecast the seed valuation in my companion piece about the post-Jeff Dean Google on the reorg at a median of $5.8 billion in its first non-Google-led round.

Back to DeepMind. The best reference class for a lab automating science is the applied-science record of the lab these four just left. DeepMind's fast wins do in fact seem to have come in domains where the experimental loop closes, as per our AI 2027 modeling. AlphaGo of course. But even AlphaFold had a public archive of over a hundred thousand solved protein structures to learn from, and CASP, a biennial blind contest, to judge it. GraphCast and GenCast had decades of reanalysis weather data and clean skill scores. AlphaTensor, AlphaDev, and the IMO systems had formal verification. Data-center cooling had a thermostat.

The slow projects and the failures ran through the physical world or through institutions. DeepMind Health's Streams app was shut down after the NHS data controversy. GNoME predicted millions of crystals, and independent labs disputed how many were real. Isomorphic Labs was founded on AlphaFold in 2021 and reached its first clinical trials only in 2026, with no approved drug.

I seeded FutureSearch with this research and my view on this. Here are the resulting forecasts, unconditional on whether they focus on chips/cyber as I suggest:

FutureSearch thinks the cold start is hard. Even OpenAI took years to get going. Thinking Machines, which launched with a sizable team already assembled, took about seven months to publish its first technical blog post, and SSI took close to a year to reach twenty employees. The August 2027 median on a first loop-authored result, with a 90th percentile stretching to late 2028, also factors in attribution of the result. Claiming your system found the result, rather than your researchers, is a tall order, see Sakana with its AI CUDA Engineer. And the median headcount a year from now is 28 people beyond the founders.

The January 2029 commercialization median is a statement about strategy, not capability. Discovery Loop calls itself its own first customer, the careers page lists no sales or partnerships roles, and free Google compute removes the revenue pressure that forces early deals. If a named paid deal shows up anyway, I would update. By the way, FutureSearch found there is only a 21% chance their first published artifact carries Google's name. The forecasts expect the break from Google to be real.

So why Chip Making?

Dean says ML research first, then outward toward the 14 Grand Challenges. I picked only 6, then I forecast the same outcome, an independently recognized breakthrough in that field by the end of 2028, conditioned on Discovery Loop announcing that field as its primary applied focus by mid-2027.

The ranking lands where the DeepMind record points. Chip design tops the board at 38%, mostly because of AlphaChip (Jeff Dean was a co-author on the Nature paper), the reinforcement-learned TPU floorplans that shipped in production TPU generations. Secure cyberspace is at 28% and materials and scientific tooling is at 28%, roughly double the fields whose loop runs through a wet lab or a clinic, medicines and health informatics at 15% each. Fusion, which needs machine time on hardware that mostly does not exist yet, sits last at 12%. Focus is worth a lot everywhere: each field's conditional probability runs four to six times its unconditional one, which is a good sanity check. Even fully committed, a team of this caliber gets barely one-in-three odds of an AlphaFold-class recognized breakthrough on a 29-month clock. The AlphaFold reference class, four years inside a staffed lab, says that is about right.

One caveat before treating those bars as advice. These are conditional forecasts, not causal ones. The condition is Discovery Loop's own choice, and they would only pick a field their early experiments favored, so every bar is somewhat flattered by selection. The flattery is uneven, and it favors the chips conclusion. A pivot to cybersecurity would be a surprise, so its 28% leans hardest on what the choice itself would reveal. Chips is the default extension of the founders' systems work, so choosing it reveals almost nothing, and its 38% is the closest thing on this chart to a pure read of where the loop works. Meanwhile medicines and fusion cannot clear 15% even with the selection boost, which is the damning version of the same point.

Isn't automating ML research extremely dangerous?

Of course it is. Automating ML research and engineering is what recursive self-improvement is all about. This is probably a Very Bad Idea for AI Safety, and it's notable that the original DeepMind people, who have cared a lot about AI safety historically, are not involved in this. The forecasted median date for a formal safety framework governing that loop is March 2029, and the curve gives just 11% to one appearing within the first year. A lead investor described the mission as building "recursively self-improving superintelligence", they have zero safety or policy roles among the openings. I mean, the founders could use the Frontier Safety Framework they just left at GDM right away if they wanted.

An interesting data point is that Thinking Machines took eighteen months to publish even a high-level safety document, and SSI, with safe in the company name, has published no operational framework at all (!). A softer leading indicator is more hopeful: there is a 32% chance at least one person with an explicit safety, alignment, or policy role works there within a year. For contrast, the parent they left has been shedding its own governance structures all year, which is the subject of my companion post on the Google AI reorg.

Grading these forecasts

Like AI 2027, these are some multi-year complex forecasts that could be disputed based on certain milestones. Still, I think like AI 2027, we'll see whether this is directionally correct before too long. I'm curious if they commit to chip design early, or maybe cybersecurity (if they get hacked early during the AI takeoff?). It just makes a lot of sense, now that I've run all the forecasts.

Forecast these yourself in the FutureSearch app, whenever information about Discovery Loop comes out that updates this view.