Josh Birkley

1. Key Themes

Biology Is an Engineering Problem, Not a Discovery Problem

Chai's central thesis is that drug discovery — historically a serendipitous, trial-and-error process — can be transformed into a deterministic engineering discipline. The analogy is to software: just as code has abstraction layers that allow rapid iteration, biology should too.

"One of the exciting things we're trying to do at Chai is to make the drug discovery process look a little bit more like engineering. And we've seen all the work happening with LLMs for code generation, for instance, right? And that works really well because code is a very simple abstraction... Biology doesn't look like that today." — Josh Birkley 00:01:04

The Bitter Lesson Applied to Biology: Scale Wins

Chai is explicitly "bitter lesson-pilled" — meaning they believe scaling compute, data, and model size will outperform hand-crafted biological heuristics, just as it did in NLP and computer vision. This is their core research philosophy and the organizing principle behind every modeling decision.

"We're a very bitter lesson-pilled company. So like, we really believe in like scaling data, scaling models, scaling compute. In order to do that, obviously you need to identify scaling laws. Otherwise you're just kind of like wasting time and resources." — Matt McPartlon 00:18:28

Simplicity as a Scaling Prerequisite

Rather than adding complexity to solve hard problems (e.g., adding a 24th module to Chai 1's 23), the team relentlessly simplifies their architecture. They argue that complexity obscures scaling behavior and prevents compounding improvement.

"When you look at a model like let's say Chai 1, I think there are 23 distinct sub-modules in Chai 1. And like when you're trying to iterate on something like that, it gets really hard... How do I simplify this? How do I identify what's really important? And once you have that, the whole research process and identifying these types of scaling directions becomes a lot simpler." — Matt McPartlon 00:00:00

Hit Rate as the Unlock: From 0.1% to 15%

The practical breakthrough that made Chai's business model viable was a 150x improvement in antibody binding success rates. This wasn't just a performance metric — it changed the statistical feedback loop, enabling real drug-like property analysis.

"When we started the company, the state of the art for antibody design was about like a 0.1% binding rate. So one in a thousand of the molecules you design would actually bind in the lab... We got to, with our Chai-2 model, about like a 15% success rate. So now if you screen a thousand molecules, you're getting 150 back. Now you can start to get some interesting statistics on the properties of the molecules." — Josh Birkley 00:16:46

The Infrastructure Play: Partner, Don't Pipeline

Chai deliberately chose not to build their own drug pipeline (unlike Isomorphic), instead positioning as infrastructure for the pharma industry. This keeps capital focused on model development and creates a rigorous feedback loop through real-world partner deployments.

"Our bet when we started... was that this is how most future drugs are going to be discovered. And if that's the case, somebody needs to go and like build that infrastructure and make it happen... As the models get better, actually wins us the right to continue investing more in them. Right. And you have partnerships with these pharma companies that are paying off today and allow us to invest in it. So it's a much more scalable business in that way." — Josh Birkley 00:29:00

Verification as the Core Scientific Discipline

Chai treats biological lab readouts as the ground truth evaluator — analogous to unit tests in software — and uses that verifiability to hill-climb model quality. The rigor required by demanding pharma partners enforces this discipline externally.

"The lab is a very important part to verify that what you're doing is correct. And actually verification is a very big theme in AI as well. If you can evaluate that your model works, you can verify it, then you can start to hill climb that and you can make progress on it." — Josh Birkley 00:02:22

Data Compounds as Models Improve

Chai's data flywheel mirrors what happened with LLMs: better models generate higher-quality structural predictions from sequence databases, which become training data for the next generation of models. Lab experimental exhaust also feeds back into model training.

"What you can do once you have really good models is just run them on the sequence databases to get new structures out. So again, like you have this compounding effect as your models get better, they get more and more accurate at predicting these structures. And then you have more and more training data for the next series of models." — Matt McPartlon 00:35:47

"There's a similar analogy that's starting to happen in our world now as well, where the models have reached a point where there's actually like a renewed interest in data. And like how do we actually bring the models into the loop on like making that happen?" — Josh Birkley 00:37:00

Pharma Is More Tech-Forward Than the Market Believes

Contrary to conventional wisdom that pharma is slow to adopt technology, Chai found pharma companies to be rigorous, fast-moving adopters once evidence was compelling — driven by a structural competitive imperative to continuously produce blockbuster drugs.

"A lot of people told us that pharma doesn't know how to use AI... And to be honest, that hasn't really been our experience... Eli Lilly is a trillion dollar pharma company right now. If they don't get more blockbuster drugs, they will not be a trillion dollar pharma company forever. And I think that forces these companies to really be on their game of adopting new technologies." — Josh Birkley 00:31:35

Zero-Shot Drug Design Is Now Real

Chai crossed a threshold where a molecule can be computationally designed from scratch — zero-shotted — with sufficient confidence to advance directly to a clinical program, compressing what was a multi-year discovery cycle.

"A year ago, you could not zero shot a molecule and like, you know, have a good sense that your program was going to work. Like now that's changed. Someone might zero shot a molecule and be like, I think we're going to bring this program to the clinic now. And then a year later, you know, they might even have some of those first molecules going into patients." — Josh Birkley 00:45:28

2. Contrarian Perspectives

More AI Will Increase Lab Testing, Not Decrease It

The conventional assumption is that AI drug design reduces the need for wet lab work. Chai argues the opposite: as the ROI on each experiment increases, demand for lab testing will go up, just as more productive software engineers increased total demand for engineering.

"It's not even a question of like reducing lab testing. I mean, maybe that happens as a result. I actually might even take the opposite side of the coin... maybe we'd actually do even more lab testing because the ROI will increase the same way. There's more demand for software engineers now that they become more productive. There might be more demand for the lab." — Josh Birkley 00:02:45

Biology Is One of the More Verifiable Domains, Not Less

The popular narrative is that biology is too messy and complex for rigorous AI benchmarking. Chai inverts this: molecular properties like binding, manufacturability, and stability are objectively measurable. Code generation, by contrast, struggles to evaluate taste and maintainability.

"I would say, first of all, that I actually think this is one of the more verifiable domains... When we think about designing a molecule in the lab, we can actually be quite specific about many of these properties... Maybe it takes a little bit longer to validate it... But at least you can be honest with yourself." — Josh Birkley 00:21:21

The Infrastructure Model Is Harder Than the Drug Pipeline Model, but More Defensible

The market assumes that owning a drug pipeline creates more value and is safer (you can fix things in the clinic). Chai argues the opposite: building infrastructure forces your models to be rigorously good because there is no cleanup step. This makes it harder to execute but ultimately more scalable.

"When you have this partnering based model, you have to be really rigorous about your models. Your models have to work really well because otherwise those partners are not going to come easily. So it's made our life harder, I think, in many ways. But I think it's also the more rewarding path if we can get it to work." — Josh Birkley 00:30:55

There May Be More Biological Sequence Tokens on the Internet Than English Language Tokens

This is a non-obvious data abundance claim that reframes the data scarcity narrative around biology AI.

"You'd be surprised, but there might even be more biological sequence tokens on the internet than English language tokens. And now a lot of that data is not that useful. It might not be redundant, it might be very noisy, but there's a lot of data out there." — Josh Birkley 00:36:02

Working Against Nature, Not Competitors

Rather than framing competition as model-vs-model, Chai's competitive frame is AI-vs-nature. The real baseline to beat is decades of optimized wet lab protocols, not other AI companies.

"It's not like we are head to head like with other model providers... Where all of us, I think, are working against nature... People have added like module 240 to the existing wet lab protocols and they have been tuned quite considerably." — Josh Birkley 00:38:44

3. Companies Identified

Chai Discovery

AI-native foundation model lab for biology; builds computer-aided design tools for molecule engineering. Core company of the episode. Partners include Eli Lilly, Novartis, Argenix, and Pfizer. Chai 2 achieved a 15% antibody binding success rate vs. the prior state of the art of 0.1%.

"We look at our partners, Eli Lilly, Novartis, Argenix, Pfizer. Like these are not companies that are, you know, they take this stuff for granted. Like you have to really deliver on these partnerships for them to take you seriously." — Josh Birkley 00:30:26

OpenAI

Early-stage AI research lab where Josh Birkley worked on the founding team, contributing to GPT-1, GPT-2, and scaling laws research. Cited as the origin of key intellectual frameworks that Chai now applies to biology.

"I really started my career at OpenAI. So it was on the early team there, was a nonprofit back then... We did GPT-1, GPT-2, scaling laws." — Josh Birkley 00:08:51

Isomorphic Labs

DeepMind spinout pursuing full-stack AI-driven drug development. Mentioned as the contrasting business model to Chai — building a drug pipeline rather than enabling infrastructure.

"A lot of times folks think about Chai and Isomorphic in the same neighborhood. Isomorphic is developing drugs. You guys are enabling the existing industry to develop drugs more efficiently." — Pat Grady/Sonali Singh 00:28:37

Eli Lilly

Major pharma company and Chai partner. Cited as an example of why pharma is structurally incentivized to adopt AI rapidly — they must continuously produce blockbuster drugs to maintain their trillion-dollar valuation.

"Eli Lilly is a trillion dollar pharma company right now. If they don't get more blockbuster drugs, they will not be a trillion dollar pharma company forever." — Josh Birkley 00:32:04

Novartis

Global pharma company and Chai partner. Named as one of the rigorous enterprise customers validating Chai's models in production.

"We look at our partners, Eli Lilly, Novartis, Argenix, Pfizer." — Josh Birkley 00:30:26

Argenix

Antibody-focused biotech and Chai partner. Particularly relevant given Chai's specialty in antibody design.

"We look at our partners, Eli Lilly, Novartis, Argenix, Pfizer." — Josh Birkley 00:30:26

Pfizer

Global pharma company and Chai partner. Andy Young, Chai's lead antibody scientist, spent significant time at Pfizer before joining Chai.

"Andy Young... has got 20 years experience in like Pfizer and Genentech, like really honing these methods, has a drug approval to his name in antibodies." — Josh Birkley 00:39:09

Genentech

Biotech pioneer. Mentioned as part of Andy Young's career background, lending scientific credibility to Chai's science team.

"Andy Young... has got 20 years experience in like Pfizer and Genentech, like really honing these methods, has a drug approval to his name in antibodies." — Josh Birkley 00:39:09

Stripe

Payments infrastructure company. Two of Chai's product team members — co-founder Jack and Munaz — came from Stripe, bringing elite product engineering discipline to the biology software stack.

"Our co-founder Jack worked at Stripe. Munaz, who was one of the top 10 code contributors at Stripe." — Josh Birkley 00:13:22

4. People Identified

Co-founder, Chai Discovery. Background spans OpenAI (GPT-1, GPT-2, scaling laws) and early protein language model work (ESM). Brings the "bitter lesson" scaling philosophy from LLM research into biology.

"I really started my career at OpenAI... We did GPT-1, GPT-2, scaling laws... the question was like, if the models can learn to speak English, German, French, why can't they learn to speak DNA and protein?" — Josh Birkley 00:08:51

Matt McPartlon

Co-founder, Chai Discovery. PhD background in protein structure prediction and diffusion models. Built early protein diffusion architectures and brings deep structural biology modeling expertise.

"One of the founding engineers, Kevin Wu, he had the first, I think it was the first protein diffusion model like ever. And that speaks to Kevin's speed of execution." — Matt McPartlon 00:14:36

Andy Young

Lead antibody scientist at Chai. One of the original developers of yeast display technology at MIT; 20+ years at Pfizer and Genentech; has a drug approval to his name.

"One of the scientists on our team, Andy Young, was one of the first people working on yeast display at MIT actually like two decades ago. And he's got 20 years experience in like Pfizer and Genentech, like really honing these methods, has a drug approval to his name in antibodies." — Josh Birkley 00:39:09

Nathan Rollins

Protein design researcher at Chai. Child prodigy who was homeschooled, joined David Baker's Nobel Prize-winning protein design lab at age 14, started his PhD at 18.

"Nathan was actually homeschooled and then started college very early on. So he joined like David Baker's lab who won the Nobel Prize for protein design. Like when he was 14, started his PhD when he was 18 and has so many creative ideas." — Josh Birkley 00:12:53

David Baker

Nobel Prize winner in protein design; ran the lab at which Nathan Rollins trained. Foundational figure in the scientific lineage behind Chai's work.

"He joined like David Baker's lab who won the Nobel Prize for protein design." — Josh Birkley 00:12:53

Kevin Wu

Founding engineer at Chai. Built what is believed to be the first protein diffusion model ever. Cited specifically for speed of execution and engineering quality.

"One of the founding engineers, Kevin Wu, he had the first, I think it was the first protein diffusion model like ever. And that speaks to Kevin's speed of execution. Like he is a heck of an engineer." — Matt McPartlon 00:14:36

Jack (co-founder, Chai Discovery)

Co-founder of Chai, came from Stripe. Leads product development and champions the philosophy of slowing down short-term to build durable, compounding systems.

"Our co-founder Jack worked at Stripe... Something that one of our other co-founders, Jack, likes to say is that if you want to move fast in the long term, you sometimes have to just move a little bit slower in the short term." — Josh Birkley 00:13:22 / 00:43:28

Munaz

Senior engineer at Chai. Was one of the top 10 code contributors at Stripe before joining Chai, bringing elite software engineering discipline.

"Munaz, who was one of the top 10 code contributors at Stripe." — Josh Birkley 00:13:22

Frank Slootman

Former CEO of ServiceNow and Snowflake. Cited for his framing of company pain: "You either have the pain of failure or the pain of growth."

"I remember one of the CEOs that we've worked with a couple of times, Frank Slootman, had this line about you either have the pain of failure or the pain of growth. You'd much rather have the pain of growth." — Pat Grady/Sonali Singh 00:41:55

5. Operating Insights

Hire Ahead of the Model, Not Ahead of the Problem

Chai deliberately did not hire antibody engineers until they had an antibody design model. Premature hiring of domain experts before the enabling technology is ready creates confusion and wasted cycles — and the technology must be ready before specialized talent can contribute meaningfully.

"We didn't hire antibody engineers before we had an antibody design model. Like what are those folks going to do?... It actually took a couple of weeks when some of those people showed up before the models could work at a point that they could work on some of these interesting case studies." — Josh Birkley 00:13:51

Keep the Team Slightly Over Capacity to Force Prioritization

Chai runs a deliberately small team where everyone operates at slightly above capacity. This is a feature, not a bug — it forces the organization to work only on what truly matters and prevents diffusion of effort.

"We've got the team very small as a result too. So this way, you know, everyone is a little bit like slightly over capacity, I think, which means we have to prioritize. It forces us to work on the things that really matter most." — Josh Birkley 00:14:19

Never Add a Module to Solve a Problem You Should Scale Through

The temptation in model development (and company building) is to add bespoke solutions for each hard problem. Chai's discipline is to resist this — if a general scaling approach should eventually solve a problem, adding a special-case module creates technical debt that obscures true progress and impedes future scaling.

"There have actually been plenty of times where, man, if we had a 24th module, like we can actually unlock that new target. And we're like, is that really something that we want to maintain long term? Is this incremental or is this like actually a compounding improvement?" — Matt McPartlon 00:22:39

Slow Down Short-Term to Compound Long-Term

As Chai transitioned from research lab to production partner for major pharma companies, they adopted a discipline of investing in code quality and maintainability even when it slowed immediate progress — because compounding returns on clean systems outweigh short-term speed.

"If you want to move fast in the long term, you sometimes have to just move a little bit slower in the short term. Right. And make sure that you are building something again that goes back to that compounding idea." — Josh Birkley (quoting co-founder Jack) 00:43:28

6. Overlooked Insights

The Real Competitive Moat Is Foolproofing Against Self-Deception in Biology

Chai made an offhand but profound observation: the error bars in biology wet lab experiments are so large (±5%) that gains which look like model progress may be statistical noise. This creates a massive structural advantage for teams with genuine scientific rigor — and a massive vulnerability for competitors who are optimizing against noisy signals without knowing it. Most AI drug companies are likely fooling themselves, and won't know until late-stage trials.

"In biology, if you're like plus or minus 5 percent in your lab, like that's all that might all be the same. So it actually just means the bar is really high in terms of the step changes that you want to see with the models. But you also need to be really honest with yourself about whether you're making progress or not. You could come up with some fancy model that looks like it works well in like one or two new tasks. But it's very important to show that that works more generally." — Josh Birkley 00:20:33

The Protein Language Model Silently Learns 3D Structure Without Being Trained On It

Josh Birkley's early work on ESM (the protein language model) revealed something deeply non-obvious: a model trained purely on protein sequences — with no 3D structural data — internally develops representations of 3D structure as an emergent property of next-token prediction. This is a profound signal that biological sequence data alone may be sufficient to unlock structural and functional understanding at scale, reducing dependence on the scarce and expensive protein structure database entirely.

"If you can train a language model to understand protein sequences, what ends up happening is it ends up kind of like representing the 3D structure internally... In order to predict like missing amino acids... you really need to understand, okay, what does that amino acid's immediate micro environment look like? And in order to do that well, you need to understand the protein's 3D shape... Josh was even more bitter lesson-pilled than me. He's like, we're just going to look at the sequences and this is just going to emerge." — Matt McPartlon 00:34:55