This essay is adapted from my special lecture at AIAS⁺ 2026, the Chen Institute Symposium for AI Advancing Science and Society, held November 5–7 in San Francisco. Sources for the figures appear at the end of the essay.
Today, every scientist knows that generating an answer or a hypothesis has become easy. Knowing whether it is true has become harder.
Last year at AIAS⁺, I spoke about Discoverative AI¹: AI that helps us go beyond what is already known. This year, I want to take that question one step further. If AI proposes something no one has known before, how do we know whether it is true?

Six Things We Learned About Verification
In the Lab
To answer this question, we ran many experiments over the past year. To keep them simple, we reduced verification to its smallest form. One model, which we call the solver, produces an answer. Another model, which we call the critic, judges it. Then we changed one thing at a time and watched what happened. From these experiments, we arrived at the first four of our six conclusions.
First, producing an answer and recognizing a good one are different abilities.
In one public study, a model tried to fix real software bugs. With one try, it fixed about 16 percent. With 250 tries, more than half the bugs got a correct fix.² The right answer was there: the model just couldn’t tell which one it was.

Share of real software bugs fixed by at least one of N attempts. Only the two endpoints are measured; the line between them is drawn on a log scale.
Source: Brown et al., “Large Language Monkeys”, 2024. Model: DeepSeek-Coder-V2-Instruct.
In our own pilot on 21 real repair tasks, a weak model working alone solved 8 percent. Then we added a strong critic model that only read the work and pointed out problems. It never wrote the fix itself. With the critic, the weak model solved 56 percent.¹⁰

Pass rate on 21 real code-repository repair tasks, one submission each, with the same strong read-only critic supervising each solver. Strong models were already near the top and gained only 2–4 points, so they are not drawn.
Source: Apodex internal experiment, October 2026 (not yet published).
Second, verification must be independent.
In 1785, Condorcet showed that many judges beat one, but only if they judge independently.³ Copies of one judge add nothing.
We saw this directly. With one solver fixed, a critic from the same model family made results about 12 percent worse. It shared the solver's blind spots. A critic from another family made them about 19 percent better.¹⁰ But different names do not guarantee different blind spots. Real independence comes from evidence: a fresh calculation, a test the solver did not see, an experiment neither model controls.

Solver held fixed; change in its average score with a critic from the same model family versus another family. The same-family drop is concentrated in one problem and turns slightly positive without it; the sample is still small.
Source: Apodex internal experiment, October 2026 (not yet published).
Third, verification must be able to argue and to act.
When we asked models to judge sixty recent papers, adding a defender, whose only job was to argue for the paper, raised accuracy from 45 to 70 percent.¹⁰ When critics could choose tests and inspect their results, success on our hardest tasks rose from 68 to 84 percent.¹⁰ For discovery, the most useful output of verification may be the next experiment: the one that separates two explanations.

Left: prediction accuracy when judging 60 recent papers, without and with a defender whose only job is to argue for the paper. Right: success rate on our hardest tasks when critics can choose tests and inspect the results. Preliminary; samples are small.
Source: Apodex internal experiments, 2026 (not yet published).
Fourth, verification works best in layers.
No single check catches everything. When a cell copies its DNA, the copying enzyme makes about one error in a hundred thousand letters. Separate layers of proofreading and repair bring that to about one in a billion.⁴ Human science works the same way: a calculation, then a review by peers, then an independent replication, then trials in the real world. Each layer catches what the one before it missed.

Each independent layer catches errors the one before it missed: from about one error in 100,000 bases to about one in a billion. Orders of magnitude as commonly cited in textbooks; they vary by organism and site.
Source: molecular biology textbooks (e.g. Alberts, Molecular Biology of the Cell).
In History
These experiments reduced verification to two models. But what does verification look like in the real world? Experiments alone cannot tell us. So let us look at the order in which AI has surpassed humans: chess, then Go, then protein structure, then competition code, and now math. That order follows where good verification already existed. A game has a winner within minutes, and a machine can play millions of games. Code can be run against its tests. A proof can be checked line by line.
Protein structure is the most interesting case. AlphaFold looked like a success of a model. But behind it stood the Protein Data Bank, started in 1971⁵: fifty years of structures that scientists had solved and checked by experiment. And behind it stood CASP, a blind competition where predictions were judged against structures no one had seen.⁶
In every field where AI has surpassed humans, the means to check the answer was in place long before the model arrived. So history gives us a fifth conclusion: Build verification first. Then build the model.

Each field where AI has surpassed or matched the best humans already had a way to check the answer. In protein structure, the check came decades first: the Protein Data Bank in 1971 and CASP in 1994, before AlphaFold in 2020.
Sources: Berman 2008 (PDB); CASP; IBM (Deep Blue, 1997); DeepMind (AlphaGo, 2016; AlphaFold, CASP14 2020; IMO 2025); OpenAI (ICPC World Finals 2025).
Build verification first. Then build the model.
Now AI is moving into drugs, materials and medicine. Here, verification needs real-world conditions that are far more expensive and complex to build. Whether a drug works. Whether a material holds up. Whether a new training method really makes a better model.
You might say drug research already has a very thorough verification system: animal studies, an IND, three phases of clinical trials. That is true. But consider this. In about twenty years, sequencing a genome became about a hundred thousand times cheaper.⁷ Yet the number of new drugs per billion dollars of R&D has halved roughly every nine years.⁸ We became much better at generating candidates, not at verifying them. The verification exists. It is too slow and too expensive to keep up. And a signal that comes back in a day can teach a thousand times more than one that comes back in three years.

New drugs per billion dollars of R&D follow the trend Scannell et al. fitted for 1950–2010: halving about every nine years. Genome sequencing cost falls from about $95 million (2001) to about $500 (2022). Only these endpoints are drawn.
Sources: Scannell et al., Nature Reviews Drug Discovery, 2012; NHGRI DNA Sequencing Costs data.
And that brings us to the sixth conclusion: verification must not only be possible. It must be cheap, fast and able to scale.

A loop that returns a signal in a day can run about 1,095 times in three years; a loop that takes three years runs once. That is the thousandfold difference in how much it can teach.
Arithmetic on the essay’s own figures (3 × 365 days).
What a Verification System Needs
Put these six together. Verification must recognize a good answer, stay independent, argue and act, work in layers, be built before the model, and be cheap and fast. No single judge can be all of these. So I believe the key scientific instrument of this era is not a bigger solver, and not a better judge. It is a verification infrastructure.
Concretely, it has four parts: a critic that judges, a record that keeps every decision and its reasons, the means to test, such as compute, instruments and laboratories, and people and benchmarks that set the standard before the results arrive. Connected, they form a loop, and every result flows back to make the next judgment better.

Verification infrastructure · four parts, one loop
In math and code, this infrastructure largely exists: a proof checker or a test suite settles the question in seconds. For almost every other field of science, no one has built it yet. That is what I mean by a world that can verify. Waiting for the real world to verify our conclusions is slow and expensive. So we have to build a world that verifies them for us: cheaply, quickly and independently.
Putting verification to work at Apodex
This year, I founded Apodex, a company built for discovery. We are based in Silicon Valley, with offices in Singapore, Seattle and Tokyo. From the very first day, we asked ourselves one question: how can an AI company actually succeed at discoveries that matter? In the end, we chose two fields. Each one matters, and in each one we are building the verification infrastructure first, so that the loop can close. The first is AI training AI. The second is drug research.
Building Verification Into AI Research
We start with AI research because we are an AI company. Training AI is what we do every day, so our own company can become the training ground for an AI that can train AI.
Not every AI company can do this. When we started Apodex, the first thing we did was not to train a model. It was to build our verification infrastructure. We built the tools ourselves, end to end:
Our own messaging, documents and meetings, in place of Slack, Notion and Zoom, so every discussion and decision stays in one record.
Our own experiment platform and Artifact system, where every GPU job is tied to a work order and every experiment must end with a written conclusion and its evidence.
A data constitution, so every dataset has a known source, quality and purpose.
Compute: at our peak, close to ten thousand GPUs.
People: about 150 scientists, so ideas can be tested quickly and judged well.
A benchmark, our own standard, where we write down what good research looks like.
~10,000GPUs at our peak | ~150scientists |
None of this was easy. The hard part was not the software or the GPUs. It was people. Our researchers came from big labs and companies with years of habits. When we asked them to leave Slack and Notion, almost no one accepted it. They said our tools lacked features and would slow them down. We did not win that argument with words. We set deadlines, we closed the old doors, and we held to one rule: if a decision or a dataset is not in the system, it does not exist. A verification loop with holes in it is not a verification loop.
Why does this matter? To learn how to do research, an AI must see how research is actually done: what people knew at the time, which alternatives they weighed, why they chose one test over another, and what happened next, all in one context. Scattered across tools, those links break. What is left is like a finished paper: it keeps the result and erases the decisions.
Because the AI works where people work, every decision it makes gets a quick human answer: right or wrong. Researchers can appeal any AI review, and every disagreement becomes a lesson. That human feedback is verification too, and it is how our AI learns to train models better.
With this infrastructure in place, we trained a model whose job is to train models. It will not match the frontier models on many things like Chinese history, dozens of languages, or making videos. That is not its job. Its job is to train models.
Many systems described as recursive self-improvement (RSI) still rely on a frontier model to do the training. We think a truly recursive loop needs a model that can train its successor on its own. Our model is going to train its successor, without relying on any frontier model. We believe it may be one of the first. We will share the evidence as it comes in.
Then we think if AI can train AI, it can help train a discovery model for any field. The work of training a model repeats from field to field. What changes is the domain knowledge and the evidence. Put that knowledge together with a strong verification infrastructure in that field, the same system can build models for other fields.
Testing the Same Idea in Drug Research
So we decided to go a step further, we chose AI-driven drug discovery (AIDD). Drug research is where verification gets hard. Biology brings new measurements, new sources of error, new costs, and processes we cannot hurry. We cannot build all of it ourselves. But I believe the earlier we start building its verification infrastructure, the more it will matter later. So we are building three pieces of it.
First, we are building automated laboratories. We are building a network of automated equipment and partner laboratories, which we call a Contract Verification Organization, or CVO. Whatever can be automated, we will build, buy or partner for, and connect into one network. Only automated labs can keep pace with what AI can generate. Without automation, verification will always be the bottleneck. In short, wherever lab automation reaches, we will invest and build.
Required animal studies and human trials remain essential, with scientific, ethical and regulatory standards intact. Better methods should make the process more effective, not lower the standard of evidence.
Second, we are making drug computation free. We set up the Drug Research Computing Center to give biotech and pharmaceutical companies free compute and free algorithms. It will offer sixteen labs, including clinical trial prediction and drug repurposing, which are already live. If you run on our platform, the compute is free. In return, our system is tested against what matters most: real problems and validation signals from the front lines of the industry.
Third, we keep people in the loop. Full automation is not possible, and people remain an essential layer. So we are forming an Apodex Verification Board of working scientists. Before an experiment begins, the Board writes down what the result would support, what would count against the claim, and which uncertainties would remain. Its disagreements with the AI become cases we study. Two Nobel laureates, two former university presidents and three pharmaceutical CEOs have already joined. We also hope to bring in a hundred scientists from the front lines of medicine and research to work with the Board on verification.
When these three pieces connect into one verification infrastructure loop, our AI can quickly train a drug research model, the same way it learned to train AI. That loop becomes a flywheel: a steady flow of targets and lead candidates, and a model that gets stronger with every experiment.

The three pieces of drug-research verification infrastructure feed one stream of verified data, from which our AI trains the drug research model. Link widths show structure only; they are not measured quantities.
Source: Apodex.
Our First Wet-Lab Project
I’m glad to share that our first wet-lab project runs through all of this infrastructure. Apodex Lab designed a library of more than 200,000 AAV capsid sequences de novo, aiming to deliver gene therapy to the heart and muscle rather than the liver. The library entered synthesis at AAVnerGene⁹ on October 4, with cell and mouse tests to follow.
We have no results yet, but AAV experts we consulted believe AI-driven rational design is one of the most promising approaches. The size of the library tells you how much we proposed. The experiments will tell us what those proposals were worth.
16drug research labs, free to run | 200,000+AAV capsid sequences designed de novo |
Why We Started With Verification
People ask: Why would a younger company, with a smaller model, reach what the biggest models reach?
Because we believe strong verification plus a specialized model can win in a focused field.
Discovery is limited by judgment, not by size. In a focused field, a model's knowledge can be narrow and deep, as long as its judgment is not missing. Our model is an expert in machine learning, code, mathematics, statistics and judgment. It does not need to speak fifty languages, make videos or tell jokes. It needs to be able to train the next model.
But a model alone, however much it knows, is not enough. Discovery needs a loop that runs all the way to reality and back, and every piece of it must connect: a model trained from scratch to do research and to judge, a complete record of how decisions are made, automated labs that test claims in the physical world, and human scientists who set the bar.
Because we built every piece, our loop runs fast: a decision, a test, a result, a better model, and back again. Because our AI can help train AI, each turn can produce a better model for the next one.
Big companies can build this too. But once the work reaches reality, big and small companies are equal. Cells do not grow faster for a bigger company. An experiment takes the same weeks, and a failed one hurts the same, no matter who runs it. Money can buy more GPUs, but it cannot buy back time. It takes time, it is painful, and it has to be started before you need it. That is why it was the first thing we did.
Build the World That Can Verify
But what gives me the most confidence is not that no one can catch us. That is a comforting story. It is that we chose discovery. Discovery is not a zero-sum game. The unknown has no end. It is not about taking jobs from people, or beating another company. If you find a drug for cancer, I can find one for obesity. Even for the same disease, we will not find the same molecule.
So the real question is not whether others will do this one day. It is how well we do it ourselves, now.
For three years, AI got better at sounding right. The next three years must be about being proven right. That will not happen by itself. Someone has to build it.
Don’t just build the model. Build the world that can verify it.
Biotech and pharmaceutical teams can bring research questions to the Drug Research Computing Center. Approved projects run free when their results are shared publicly.
If you’re a scientist working in medicine or drug research and want to help set standards for evidence, I’d like to hear from you about the Verification Board.
Subscribe for updates on what we learn from our self-training model and our first wet-lab study.
Notes and sources
Discoverative AI: my closing lecture at AIAS 2025, Chen Institute news release.
Bug-fixing with repeated tries: Brown et al., “Large Language Monkeys: Scaling Inference Compute with Repeated Sampling”, 2024. On SWE-bench Lite, DeepSeek-Coder-V2-Instruct solved 15.9% of issues with one sample and 56% with 250.
Independent judges: Condorcet, Essai sur l’application de l’analyse à la probabilité des décisions rendues à la pluralité des voix, Paris, 1785; see the Stanford Encyclopedia of Philosophy, “Jury Theorems”.
DNA copying errors: about 1 wrong base per 100,000 before proofreading; repair layers bring the overall rate to roughly one in a billion or better.
Protein Data Bank: H. M. Berman, “The Protein Data Bank: a historical perspective”, Acta Crystallographica, 2008: “began as a grassroots effort in 1971.”
CASP: a double-blind structure prediction experiment held every two years since 1994.
Sequencing costs: NHGRI DNA Sequencing Costs data: about $95 million per genome in 2001 and about $500 by 2022.
Drugs per R&D dollar: Scannell et al., “Diagnosing the decline in pharmaceutical R&D efficiency”, Nature Reviews Drug Discovery 11, 191–200, 2012.
AAVnerGene: aavnergene.com, where the AAV library entered synthesis.
Apodex figures: the 21-task repair pilot, the critic-family comparison, the paper-judging and test-choosing results are from Apodex’s internal experiments and have not yet been published.


