When Google legend Jeff Dean left the company in August 2026 to co-found Discovery Loop with Oriol Vinyals, Sanjay Ghemawat, and Quoc Le, he took aim at the biggest prize in technology: automating the scientific research process itself. Months earlier, AlphaFold co-inventor and Nobel laureate John Jumper had joined Anthropic. But Cao Yuan, former senior research scientist at Google DeepMind and now co-founder of Unreasonable Labs AI, offers a more sober view of this so-called AI for Science (AI4S) breakthrough - the real bottleneck isn't compute or data, it's verification.
Speaking on the Silicon Valley 101 podcast, Cao Yuan argued that AI-driven science is at a pivotal moment. Large models' reasoning abilities in mathematics and coding have been validated, and agent "harness" infrastructure - the systems that string workflows together - has matured, so AI can now propose hypotheses, analyze data, and self-correct. But physical experiments remain expensive and slow, acting as a brake on the loop. "The hardest part, for AI for Science, is definitely verification. It is a bottleneck. Once this bottleneck is broken, you can iterate very quickly," Cao Yuan said.
The talent exodus is structural, not symbolic
Cao Yuan reads Jeff Dean's Google exit on three levels. Organizationally, the Gemini project consolidated previously independent DeepMind teams into a "concentrate resources on big goals" structure, which created internal friction. On product, Gemini 3.5 Flash and 3.6 Flash underperformed competitors between April and July 2026, and repeated delays hit the 3.x Pro model. But the most important signal is the third: Dean's new venture - teaching AI to have a "Bayesian moment," where yesterday's experiments genuinely update today's priors - is the natural next extension once math and coding capabilities mature.
Cao Yuan rejects the narrative that an era ended with Dean's departure. Google still has full-stack advantages: TPUs, infrastructure, data centers, toolchains, models, talent, and distribution. "It's not that Google can't do it well; it's that priorities and resource allocation need adjustment," he said. The current priority at DeepMind is getting Gemini to state-of-the-art level first. Meta is seeing similar dynamics - going "from 80 to 100" is far harder than "from 0 to 80."
What is AI4S, exactly?
Three terms often get conflated: AI4S, AI4AI, and RSI. Cao Yuan defines them crisply. Both AI4S and AI4AI treat "AI as a researcher" - the difference is the subject, scientific problems versus AI itself. RSI (recursive self-improvement) isn't a task but a method: a model iteratively improving based on feedback from its own outputs.
The explosion in 2026 has two prerequisites: reasoning capabilities in math and coding, plus mature agent harness infrastructure. "If you can let AI do it itself, it can probably try many new ideas and many new experiments much faster than humans," Cao Yuan said. But AI currently plays two distinct roles: the "co-scientist" that accelerates human analysis, and the "autonomous discoverer" (like AlphaFold). The problem definition still must come from humans. As host Chen Xi summarized: "Now we have the right time, the right place, and the right people."
Biopharma leads, but the fundamentals aren't AI
The most capital-attractive AI4S domain is biopharma - but not because of AI. "Traditionally, biotech, life sciences, and drug discovery represent a huge market - and that has nothing to do with AI," Cao Yuan noted. Drug development routinely takes years and costs hundreds of millions of dollars, so any acceleration in intermediate steps creates enormous value. Materials science, by contrast, has a fragmented value chain stretching across metals, leather, biomedical materials, semiconductors, and rare earths, making it harder to capture AI-created value through a standardized process.
For selecting problems suitable for AI research, Cao Yuan offers three criteria: computable, cleanly modelable, and rapidly verifiable. AlphaFold is the canonical example - input an amino acid sequence, output a three-dimensional structure, verify easily. "What experiment to do next" is a search problem, with models evaluating potential returns and balancing exploring the unknown against exploiting what's known.
Finding a path forward
Two paths are running in parallel to break the verification bottleneck. One is automated labs, where standardized processes go to robots. Google DeepMind's collaboration with UC Berkeley and Lawrence Berkeley National Laboratory - called A-Lab - showings 353 experiments in 17 days, synthesizing 36 of 57 target materials. The other path is improving model intelligence so that fewer physical experiments are needed. OpenAI's collaboration with robotics lab Ginkgo Bioworks, using GPT-5 to propose thousands of protein implementations and execute experiments, is what Cao Yuan calls one of the "most convincing closed-loop experiments" discovered to date. But robotic manipulation precision now limits speed, and in silico simulation works advancing faster than physical lab work.
A deeper ceiling is conceptual. Nobel laureate Jennifer Doudna found that no AI-generated proposal was something her team didn't already know. Cao Yuan's explanation: "A language model's entire universe is determined by its training data." Current AI "discoveries" are mostly recombinations within the existing knowledge space.
What the Big Three aren't doing
The three major AI labs have starkly different AI4S strategies. Google/DeepMind has invested the earliest and deepest, with Alpha series, Co-Scientist, Isomorphic Labs, and A-Lab - but Gemini must reach SOTA first, tilting resources toward near-term products. OpenAI is catching up with GPT-5, GPT-Rosalind, and the new Astra math model, but Ben Weil's original AI4S lead departed around April-May 2026, and the team was merged into Codex - betting that general GPT plus agents will naturally solve scientific problems. Anthropic is the late entrant with Claude Science, launched less than a month ago - and the notable absence: the only one of the three without a vertical scientific model yet.
But in Cao Yuan's view, none of the three institutions treat AI4S as a top priority. Commercializing coding agents and preparing public offerings take precedence, and Chinese open-source models are pressuring Anthropic and OpenAI to move faster. That's the startup window. Jeff Dean's Discovery Loop and Cao Yuan's own Unreasonable Labs AI (addressing the hybrid approach with "large language models + symbolic reasoning") are both positioning into AI for Science & Research.
Why mathematics is the nearest-term frontier
The model that Nano scale works: language models propose proof strategies, which get automatically formalized into Lean (a formal proof language) and verified by a compiler. Once it compiler, the result is correct. This approach has three bottlenecks: the language model's math abilities, natural-language-to-Lean translation capability, and the coverage of mathlib, the Lean library. For professional-mathematician-level proofs, Lean isn't optional - it's the only write way to trust correctness. Cao Yuan cites 2018 Fields Medalist Peter Scholze, who had to manually formalize one of his own proofs into Lean over perhaps a two-year period until its compilation confirmed the proof was still correct. "Even the most elite scientists, when dealing with extremely complex mathematical problems, still find it very hard to determine whether something is actually correct," said said. "You can only trust logic and computers."
The real gap is "concept invention" - definitions like number, eigenvalue, or Riemannian metric, without which derivation can't even begin. If no new mathematical objects are required, a problem is just a search; with a strong enough model, a proof should theoretically exist. It's "concept abstraction" - what Matteo considers the "last mile of AGI." This creates a controversial phenomenon known as "proof surplus": numerous online problem statement, problems that humans failed to solve for years now have dozens of AI-generated solutions, but no human expert wants the verification burden. "A correct result that doesn't equal a valuable one," Cao clarified. "Correctness of the result doesn't mean it's automatically right-you need to understand it, and then judge it."
Instant ripple effects and the long timeline
Physics won't be fully automatable. "The most fundamental difference between AI for Physics and AI for Math is that you must step outside the logical world and into the material world-you must verify," Cao Yuan said. Even exceptions like AlphaGo's "Move 37" are actually still recombinations inside existing representation space; they involve Novo new concept - "recombination," not leap. Purely random events don't produce knowledge; "abductive" reasoning does-inferring causes from phenomena and generalizing-which demands abstraction, curiosity, and a world model.
For your own company, Cao Yuan applies a 2% symbolic layer-a rule skeleton-outside a language model to help it step outside of probability. He predicts final integration will be roughly 98% connectionist, 2% symbolist. AI for math already uses this exact pattern: neural nets in the "idea" stage, Lean in the "verification" stage.
Where AI4S actually matters
For researchers in any life-science domain, the important takeaway is first to preserve the loop. The fastest genuine win is the near-term: AI compiling proposals, DNA sequences, and ICI hypotheses at scale - but they will be plausible, not novel. You need a human gatekeeper that can recognize that a proposal falls within the sphere of your existing knowledge, even if it appears in an unfamiliar shape.
Second, the right thesis is standard, verifiable workflows. If "what you do" (an assay, a drug target, a material property) has a clear "in/out" API, and model can make a difference today. If your domain's experiments are expensive, messy interaction, robotic labs are still neck-and-neck with model improvements. The most short-term value alternates between those who standardize their own exchange-rate- five minutes, not five years - and researchers that borrow the disciplines of software engineering (checkpointed, Lean-based, peer-verified).
It's worth tracking: Discovery Loop's first product, whether John Papa Jumper at Anthropic can produce a vertical biology model, and whether the "OpenAI model-codex"-merger check - survives the next six months.
For researchers: Where you stage your career
Cao Yuan's timeline is explicit: causal reasoning, long-term memory, and continual learning "may take five or six years." The best case that AI autonomously discovers Nobel-caliber science is "two to three decades" out. Meanwhile, "two to three decades" makes a timeline that puts most readers mid-career. So if you're in it for the long game, the concrete skills that make you irreplaceable are defining problems, verifying results, and making value judgments. The ecosystems that will reward "AI4S" as something more than a boss named a bringing also include a "digital assistant" that won't be able to see the whole picture view alone-you'll be that controlled environment, that stopping point. Pick a domain with a quickly verifiable loop, stay on math-ish problems if you want legitimately connected, and remain AI-literate - but don't programatically surrender your judgment to the model. Protect your accountability.
Your membership also unlocks:
