They Aimed for 1%. The Lab Hit 20%. Chai Discovery Moved Antibody Design Onto the Computer.

Joshua Meier, CEO of Chai Discovery, former chief AI officer at Absci, and previously a researcher at Meta, sits down with Redpoint's Jacob Effron on *Unsupervised Learning*. Effron frames biology as the application everyone from Dario Amodei downward claims will matter most for human health. Meier's job in the hour is to separate that ambient excitement from what actually works. He dates his own start to 2013, a year after AlexNet, when the tooling was clumsy and nothing moved at today's pace.

By 2015 the first deep-learning models were being aimed at protein structure and design. Meier's co-founder built some of those early folding networks, and they could already do more than the compute budget suggested. The joke, Meier says, is that biologists had been running a language model for two decades. Pre-deep-learning folding took a sequence, found similar sequences, and looked for co-evolutionary patterns. Mask a residue, predict what belongs there: the same math NLP later used to mask a word. Healthcare is usually 20 years behind. Here it was 20 years ahead.

On the commercial side he counts three waves. Wave one treated data as the bottleneck and went to produce it. Recursion's phenotypic screens threw random molecules at cells, photographed the result, and let a network filter what changed, boil-the-ocean biology with a neural net on top. Wave two tried to model the molecules themselves, especially small-molecule property prediction, which Meier still thinks was the right problem. Many of those companies were founded before GPT-3, some before GPT-1, so they never had today's machine-learning stack.

Wave three, the cohort Chai belongs to, started after the generative-AI craze and could steal lessons from NLP, computer vision, and video. The result, he says, is reliability that would have sounded like science fiction: success rates of 10 to 20% and up. Pick ten molecules and a couple work on the first try. That is the number the rest of the conversation has to explain, because two years ago nobody would have put it in a plan.

Biotech, Meier reminds Effron, is three problems stacked: find a target, find a drug against it, then run trials. AI is noisy in all three. On the trial end, companies use ChatGPT-style models to draft the mountain of regulatory paperwork, and the FDA has said it will use AI to read some of those documents. Meier's dry joke is that we will have AI talking to AI. He is glad for the speed, and he does not pretend the bureaucracy is optional. You are putting a molecule into a person.

Chai plays the middle: drug discovery. Someone has spent a decade in a lab on a therapeutic hypothesis, make a drug against GLP-1R for diabetes, metabolic disease, or weight loss, and the question is how to get the best molecule to the patient. A drug, he says, is atoms organized a certain way, either by chemical synthesis or by a DNA sequence that folds into that shape. Learn the interactions at atomic scale and you can invent life-saving drugs. That is why this is, in his view, a perfect AI problem.

AlphaFold 1 arrived around 2018, AlphaFold 2 in 2021: sequence in, 3D structure out. Solving a structure in the lab can still cost on the order of $30,000, down from hundreds of thousands. Predict 100 million proteins and you have done billions of dollars of research in a few months of GPU time. Useful for basic biology. Not yet design. Design means inventing a new molecule that perturbs an existing one, which requires knowing how proteins, small-molecule drugs, DNA, and RNA interact as complexes. The last two years, Meier says, graduated models from single proteins to those macromolecular assemblies. Chai-1 is the atomic-level microscope they needed before they could move atoms around under it.

The two dominant therapeutic classes are antibodies, proteins made by biology, and small molecules, chemistry. Gene editors and CRISPR therapies exist, but those two still own the markets, which is why most AI-bio companies live there. Pharma companies usually have a team that has spent a career in one modality. Foundation models, Meier argues, do not care. Chai-1 is an all-atom model. It takes an input and returns an output whether the atoms belong to a small molecule, a protein, DNA, or RNA. Data can even transfer across those domains. Commercial focus still exists because manufacturing and partner teams differ. Chai has concentrated on antibodies so far. Long term he sees no AI reason to stay in one silo.

Biology does not have "the whole internet," Effron says. Meier pushes back on volume. Count the DNA base pairs you can already download and you are in double-digit petabytes, comparable to language-model corpora. Quality and diversity are the real constraints. He imports the pre-training versus post-training split. Early GPT-3 and GPT-3.5 hallucinated in part because they were trained on Reddit, where "I don't know" is not an upvote. Biology's analog is academic sequence dumps, including a New York University study that swabbed the subway and sequenced whatever bugs turned up. That data teaches what makes a protein a protein. It does not teach what makes a protein a drug. Drug-likeness is post-training data you have to build.

The rule he wants in the latest wave of companies is model-driven, not dataset-vain. If the subway bacteria already answer the question, do not reinvent the wheel. Do not build data for the sake of having the biggest set. Build it where the data does not exist or the model is not performing. He will not bet against base models getting better, and he also will not wait five years on a beach if a post-trained model can put a meaningful drug in front of patients now. That, he says, is close to a social obligation. In-house wet labs are not a status symbol. Post-COVID antibody-engineering capacity sits underused; Chai rents it. The Chai-2 report used multiple external labs both for elastic capacity and so outsiders could reproduce the result. A few years ago the flex was headcount and square footage. Now the flex is a lean team.

Chai has talked about two models. Chai-1, the atomic microscope, was open-sourced late last year. Meier says they expected models at that capability to go open anyway, so they did it themselves, and they needed the microscope to build the design tools. When the company started, state-of-the-art methods predicted high-accuracy antibody–antigen complexes only about 10% of the time, even though antibodies are roughly half the drugs in development. Chai-2, announced a couple of weeks before the taping, is the design layer: prompt with a target protein, a site, and desired properties, and the system emits molecules to synthesize and test. Meier's metaphor is Photoshop for molecules.

The breakthrough was the wet-lab rate. One in five tested molecules bound the target as expected. The team's goal for the year had been 1%, because 1% is the point at which you can brute-force ideas on a computer instead of standing up a high-throughput screen you cannot ship to hundreds of industry programs. They beat the OKR by 20 times. They did not believe it for weeks. Most AI-biology papers test one to five projects; Chai ran 50 targets, ten times the usual, plus follow-up characterization. The emotional residue, Meier says, is suspended disbelief. Crazy ideas now default to "it will probably work in the lab," which is a wild way to do drug discovery, and the company's energy has gone up with it.

The machine-learning turn was diffusion. Earlier folders, including the AlphaFold generation, behaved like old image models: sequence in, one structure out. If the model was unsure, it collapsed toward a mean. Diffusion can sample different guesses and let another model pick. Effron restates the implication: previous systems compressed many ideas into a bland average; diffusion brainstorms, including designs that make no biological sense and designs that are novel and good. Meier agrees that diffusion and biology are a natural pair. He also hears an older echo. David Baker's Rosetta was an energy model, a pile of terms describing a protein. Diffusion, in his telling, rediscovered that physics-based view after the field had spent years on statistical sequence methods. The machine-learning community looped back to a 20-year-old protein tradition.

Has biology had its AlphaGo move 37? Meier thinks so. You prompt in ordinary language for a function on a protein; the model is not reasoning through the biology so much as translating a hypothesis into sequences to make. Chemists look at the map and say it is a good idea, but will it work. Two weeks later the lab says yes. He immediately undercuts the romance. Creativity is not what matters. What matters is solving problems that were not solvable, or making the process orders of magnitude more efficient. Validation in a dish or an animal is easier than generation. Clinical validation is still slow and expensive. Pre-AI "drug discovery" meant screening billions of random molecules in a space larger than the atoms in the universe. Most were duds. Generation is now the model's job, so the bottleneck is ideas, and ideas can be brute-forced too.

Will AI mean fewer lab experiments? Meier says more. Move lab capacity onto brute-forcing biological hypotheses and you can attempt undruggable targets that old methods could not touch. One example he likes: make molecules to several sites on a target and stitch them into a biparatopic antibody, a class that is miserable to discover at the bench. People at Chai get asked why they tried this idea rather than that one. The honest answer is often that they tried a lot of them, because it is cheap enough. His advice to scientists with a hypothesis is not to polish it. Try ten.

Who should call about Chai-2. Anyone on an antibody project, he says, which is about half of current drug work. There are two no-brainers. First, the easy programs you already believe you can design: speed them up and explore better versions. Second, and this is the one that excites him more, the programs that have been intractable for years. Companies spend years failing to find a molecule for a hypothesis that Chai-2 might hit at the click of a button. Legendary drug hunters do not disappear in that world. If hypotheses can be brute-forced, the people who are creative about which problem to pose become more valuable, even as general-purpose LLMs get better at scientific brainstorming.

Effron quotes Meier's co-founder Jack from another podcast: be a product company, not just a model shop. Meier's definition is deliberately un-glamorous. Product means solving a real problem, working backward from it. If the answer turns out to be ordinary software or a physics method and not AI, that is fine. You do not need a sticky web app. In pharma, "product" usually means the marketed drug a patient swallows. At Chai it might mean the tool a drug hunter uses to make that drug. The whole team, he says, is supposed to stay open-minded about whichever solution actually works.

The current landscape, Isomorphic Labs, Profluent, EvolutionaryScale and the rest of wave three, can be sliced two ways: technical approach, and how you get paid. The industry's real revenue is an approved, marketed medicine. A single product can do billions a year; the largest do double-digit billions. From that fact you can fully own a drug, partner on development, or sell tools that help other people find drugs. Meier thinks it is too early to freeze those choices. Technology is moving faster than the business map. Most biotechs never take a therapy end to end anyway; they exit earlier. The asset is still the thing. You have to show you raised the odds of approval, or opened a class of drugs that could not be made. One path is to take a molecule to phase one and sell a de-risked data package and IP, sometimes for billions.

Why this wave is different is the GPT-2 analogy. You can build a business around a model that does not work as well as the customer's existing tools. Some people will try it. They may not come back. Previous AI-bio partnerships, in his telling, often looked like that. Now the models actually compete. The milestone he wanted at the start of the year was getting antibody design above 1% so discovery could move from the lab onto a computer you can ship. Chai-2's 20% means that move is happening, even if the molecules still need work before anyone would put them in a patient. Watch how much of discovery runs on the computer, he says, because that loop compounds: models generate data and enumerate hypotheses that go back into models. The first "AI-designed drug in the clinic" is, to him, an overrated headline. Computation has already touched approved therapies. The question is whether you can make drugs that were not possible before.

Feedback speed is why the field is suddenly moving. Design 100,000 antibodies and you wait weeks to months. If the model is good enough that you test 20, you know in two weeks. That is also his answer to the skeptic who says binding is cute, call me when a full battery of studies beats what we already make. Recreating a known drug with AI does not change the game. The aim should be to uplevel the molecules, biparatopics and other designs that were too hard at the bench. He has a concrete sales slide for "this time is different": a company spent $5 million to $10 million and years on a program; Chai prompted the model with the same therapeutic hypothesis and had an answer in about two weeks. If someone has worked five years and the model does it in two weeks, he says, that is pretty clear.

Look out a couple of years, Meier says, and you can imagine zero-shot drug candidates: not a 100% hit rate, but a first batch that already contains the molecule you would take into the clinic. Patients win through better molecules. Pharma wins as a capital allocator. He cites a split of about two-thirds in-licensed drugs and acquisitions versus one-third internal pipeline. Biotech's edge is the open question. Discovery getting easier could pull more of it in-house at Pfizer-scale firms, or push more of it out to AI companies and small biotechs that sell molecules after a lab test. He expects a mix. He does not expect "throw the model over the wall" to work. Discovery process inside biopharma has not changed much in 20 or 30 years, unlike tech, which already absorbed mobile and then AI. Adoption will start slow. Then, if Pfizer gets working AI through a partner, everyone else has to follow or fall behind.

In the quickfire, underhyped is molecular discovery itself: it works now, and people should pay attention. Overhyped is "AI-designed molecule in the clinic." It does not matter whether the molecule came from AI. It matters whether better drugs reach patients. The research he watches outside biology is interdisciplinary on purpose: LLMs, video models for dynamic proteins, and the physics and chemistry of how atoms interact. Effron asks whether Eroom's law, Moore's law spelled backward, the decades-long decline in R&D efficiency, reverses in five years. Meier hopes so. Then he demotes the metric. A high-probability copycat that does not change the standard of care is not the win. The industry needs more risk, not less. Work backward from what patients actually get.

For people who want to do something tomorrow, the pointers are specific. Chai-1 is on GitHub for structure and complex prediction. Antibody teams, especially on programs that have failed for years, should talk to the company about Chai-2. Judge AI-bio outfits by whether they solve previously impossible problems, not by whether they rediscover an old drug with a new logo. And if you have a biological hypothesis, stop sanding it. Generation is cheap enough to test ten of them. The site is chaidiscovery.com, with a technical report and video on Chai-2.