Jeff Dean left Google to start a company called Discovery Loop.
Not Discovery Lab. Not Discovery AI. Loop.
The day after that announcement, a Chinese tech podcast recorded a 109-minute interview with Cao Yuan, a former senior research scientist at Google DeepMind who now co-founds Unreasonable Labs AI. It is nominally about AI for Science. What it actually answers is a question I had been stuck on for three articles: same loop shape, why do some spin fast and others refuse to move at all?
The bottleneck is not intelligence
Asked which step is technically hardest, Cao Yuan does not hesitate: verification.
His argument runs backwards, from coding. Why is the coding agent currently the most successful commercial loop in all of AI, the one producing exponential growth? Because it closes. A programmer writes a program and it can be verified immediately — is this right or not. No waiting, no outside party. Because it can be verified, users are willing to hand over the entire task.
Science does not get that. Its verification lives in the physical world.
Then the line I have not been able to put down:
If we could ideally get one experimental verification per minute, the problem would be solved.
You could generate infinite data and train on it directly, the same way we train on code and math.
So what is blocking science is not that AI is not smart enough. It is that feedback is too slow and too expensive. Cao Yuan pushes this further: the universe's basic elements are information, matter and energy. AI for math and AI for coding both live in the information layer. Even AI for Science — proposing hypotheses, running simulations — is still information. But to do science you must reach the matter layer. You must run the wet experiment.
Which is a statement about your company, not just about science
A line of business does not move at the speed of your model. It moves at the speed and price of your verification.
Ours, concretely. The Flatkey gateway code is written almost entirely by Claude: change a line, run the tests, glance at the dashboard, and you know within minutes whether it was right. So we let agents make large changes there. Our paid-acquisition line needs three days of data before a creative can be judged good or bad. Same model, same agents, iteration speed two orders of magnitude apart.
The ads agent is not dumber. The ads verification is more expensive.
I have written before that quality gating is the fourth layer of a working loop. This interview corrected me on what that layer is for. It is not only the filter that keeps garbage out. It is also the gear that sets the loop's speed.
Two ways to drag verification back into the information layer
If that is the bottleneck, there are exactly two moves, and Cao Yuan named both.
Automate the physical experiment. Where the protocol is standard, replace the hands with robots — AutoLab, Cloud Lab, where you call an API and a remote machine pipettes for you. He cites the A-Lab work with Lawrence Berkeley: 353 experiments in 17 days, 36 of 57 targets achieved. The more recent example is OpenAI pairing GPT-5 with Ginkgo Bioworks, a robotic lab in Cambridge — GPT proposes formulations for generating new proteins, the robots run them, results flow back. He calls this the most convincing closed-loop experiment so far.
Run fewer experiments. If the bottleneck is physical, then every experiment has to count. Where it used to take ten runs to hit a protein structure with the required properties, the intelligence has to improve until it takes two. And that requirement lands on a specific capability: analyzing the history of failures well enough to pick a better next move.
Which is why he insists the field should not be called AI for Science but AI and Science — not science as an application handed to an already-strong AI, but AI's own capability and science capability growing together. He goes further: science is not an application of AI, it is a catalyst for AI. A model that can do something this hard will be better at everything else too.
Surplus proofs
The interviewer brings up Terence Tao's recent framing: mathematics has moved from an age of scarce proofs to an age of surplus proofs.
The picture is concrete. Websites that collect unsolved problems now have dozens of AI-generated solutions piled under each one — and no human expert is willing to pick them up and verify them. Many of those problems may already have answers. Humans lack the time, the bandwidth, the energy, or the ability to check.
Cao Yuan's addition is the harder half: a correct result is not automatically a correct thing to have. Even if automated verification says it holds, a human still has to understand it and judge it. Humans carry value judgment.
Then the example everyone shipping software will recognize. AI does write a great deal of code now, and most of it is correct. And the cost and time of maintaining code went up. Engineers cannot read what was written. The result is probably right, but when a bug appears, you first have to understand it before you can locate anything.
Let the AI debug it, then? Sure — and that iteration scrambles the code further. A language model is a stochastic machine; run the same task twice and you get different code. That uncertainty is a trust problem.
His conclusion is blunt: if AI is poor at reliability and interpretability, it makes the whole economy more expensive to run. If AI one day handles 80% of economically valuable activity but every piece needs a human to verify it, you would be better off not using AI for it.
This is the same phenomenon as the unreviewed PRs in your repo and the folder of AI-written drafts nobody opens. Different industry, one mechanism: when review is the only bottleneck, adding output is pure waste.
The robot at the piano
Following "is a correct result enough," Cao Yuan offers the best analogy in the interview.
Suppose you are going to a piano recital. I tell you the piano will be played by a robot, and I guarantee the pressure and duration of every key is perfect, exactly as the composer specified.
Do you still want to go?
The interviewer says no.
Even with perfect tone, he says, you might not. But if a pianist plays it, you go — because what you came for is not the result, not the timbre. It is the performer's own emotional state, the thing that resonates.
Mapped onto science: the result of a proof matters, of course. But if you inspect the argument and find nothing interesting inside, nothing that can be interpreted — then however correct it is, to a mathematician it is a robot at the piano. No insight that moves anyone.
Asked whether a discovery humans cannot understand is meaningless, his answer is careful: it has partial meaning — instrumental meaning. He cites the four color theorem, machine-proved in the 1970s. The conclusion is certainly true and you can go color maps with it. But the proof is essentially exhaustive enumeration, and it produced no deep new insight. Whether it has value depends entirely on which axis you evaluate: the instrumental one, or the epistemic one.
The last mile is abstracting a new concept
He splits human knowledge production into three steps.
- Perception. You see a person push a cart and the cart moves. That is a surface phenomenon.
- Concept. You extract "force" from it, and symbolize it as F.
- Formalization. F = ma.
Today's AI operates inside an existing representation space, recombining and searching over concepts it already has. Cao Yuan grants this is already far more efficient than humans and solves a large class of problems. It is not creating new concepts.
Was AlphaGo's move 37 a discovery? Yes, he says, definitely a new discovery — but the mode of that discovery was rearranging existing internal representations and sampling an uncommon path. A discovery, not a new concept.
His counterexample is sharper. Imagine a language model never trained on mathematics. A caveman who sees three geese, four birds, five trees can abstract the concept of number — and that concept then applies to any object whatsoever. Feed the same sensory input to AI and it will not extract "number."
So: concept abstraction may be the last mile to AGI, and that mile may never be crossed. He suspects the process is not even Turing-computable, which is also why he doubts the brain is equivalent to a Turing machine at its base.
There is a blunt reality check attached. Asked whether her lab uses AI for research, Nobel laureate Jennifer Doudna said yes — and that not one of the proposals it produced was something they did not already know. Some were things they routinely overlook, but all of them were already in the literature.
Anyone shipping agent products should pin that above their desk.
Three things I took away
One: treat verification cost as a variable you design, not weather you accept. If loop speed is set by verification, then when a line of business will not move, the correct question is not "let me try a stronger model," it is "can I get this from three days to three hours." Automated labs are exactly that move — not making AI smarter, but dragging physical-world feedback back into the information layer. Translated to a normal company: fix monitoring, fix A/B, fix staged rollout, before you swap models. That is a statement you can prioritize against, not a nice-sounding truism.
Two: surplus proofs have already arrived at every company using AI; nobody calls it that. The unverified solutions under math problems and the unreviewed AI-written PRs in your repo are the same event. Our own answer is adversarial gating — one agent hunts for flaws in another agent's output, and only what survives reaches a human. But that only pushes the bottleneck one notch downstream. How many notches it can be pushed, I do not know yet.
Three: "a correct result is not automatically correct" is a hiring standard. The piano analogy sounds like aesthetics, but the real claim is that beyond the result, whether the process contains understandable, reusable insight is itself valuable. Once AI produces the result, the human's remaining job is not to re-check whether it is right — the machine computes better than you. The human's remaining job is to judge whether the result is worth having. As output volume explodes, that is the only thing still getting more expensive.
Which brings it back to the company name. Four people walk out of Google to make AI do science on its own, and they call it Discovery Loop.
Not Lab. Not AI. Loop.
I do not think that is a coincidence.
Source: a video podcast published 2026-08-15 (about 1h49m), interviewing Cao Yuan, former senior research scientist at Google DeepMind and co-founder of Unreasonable Labs AI. All claims attributed to Cao Yuan above are his views as stated in that episode. The Terence Tao framing and the A-Lab figures (353 experiments in 17 days, 36 of 57 targets) were given by the host. Figures on OpenAI's Astra and an unreleased research Claude, mentioned in the episode's opening narration, are not independently verified here; Cao Yuan himself notes the Astra results are disputed. Doudna's remark is Cao Yuan's paraphrase of her interview. Details about the author's own gateway and ads workflows are first-hand.
Top comments (0)