Aug 11, 2026 · 1h 35m · latent-space
🔬They Thought the Model Was Broken — Matt McPartlon & Neil Patil, Chai Discovery
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Latent Space podcast episode, Chai Discovery co-founder Matt McPartlon and Head of Platform Neil Patil discuss how their startup applies generative all-atom diffusion models and CAD-style software tooling to transform drug discovery into an agile engineering discipline. They detail their technical architecture, enterprise pharmaceutical partnerships, and empirical validation breakthroughs that enable de novo precision molecular design.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.9 Guest teaching 4.4 Guest disagreement 1.4 The hosts pushing back 3.0
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:20:37 Neil rejects the 'no data moat' premise
Neil firmly pushes back on Brandon's assertion that Chai lacks a proprietary data moat, countering that compute can generate data and comparing their strategy to Anthropic's enterprise defensibility.
Hardest push from the hosts ▶ 56:50 RJ challenges the commercial economics of structural AIRJ refuses the conventional pitch that computational hit discovery is transformative, pointing out that saving a couple million dollars is trivial within a half-billion-dollar drug development campaign.
Biggest teaching moment ▶ 53:50 Matt highlights the 90% failure rate of docking modelsMatt corrects the common misconception that AlphaFold solved structure prediction, pointing out that multimer versions fail on nearly 90% of antibody-antigen prediction tasks.
The host holds their own ▶ 1:09:47 Brandon cites triangle layers and inductive bias trade-offsBrandon demonstrates deep domain knowledge of structural ML architectures, drilling into the necessity of triangle layers and inductive biases versus pure transformer tokenization.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Chai Discovery's Pure Software Model and 50-Target Challenge | 4 | 5 | 1 | 4 | RJ presses the guests on why pharma partners are specifically compelled to buy from Chai instead of competing structural biology companies. Matt explains their pure software platform thesis, the timing of multimer prediction breakthroughs, and their 50-target validation challenge. | |
| Antibody Biology and Precision Molecular Engineering | 6 | 4 | 1 | 2 | RJ and the guests trade explanations of antibody structure, contrasting natural immune recognition with synthetic modalities like ADCs and induced proximity. Both sides demonstrate deep familiarity with CDR loops and GPCR targets. | |
| Cross-Reactivity, Selectivity, and Counter-Screening | 6 | 4 | 2 | 3 | Brandon asks about traditional hit discovery methods and how Chai handles selectivity and counter-screening against off-target proteins. Neil and Matt clarify cross-reactivity requirements across monkey/human homologs and how their model directly designs for conserved regions. | |
| Chai-1 Development: MSA Infrastructure and Startup Sprint | 5 | 5 | 0 | 1 | RJ asks Matt to explain Multiple Sequence Alignments (MSAs) in two sentences, prompting Matt to describe co-evolutionary signals. The guests recount the early sprint at OpenAI's offices to build Chai-1 and open-source it alongside a web server. | |
| Technical Architecture: Tokenization, Transformers, and Diffusion | 6 | 4 | 1 | 2 | RJ and Brandon explore the architectural transition from Chai-1 folding to Chai-2 co-design. Matt details how atom tokenization feeds into transformers and diffusion to concurrently co-generate 3D coordinates and amino acid sequences. | |
| Confidence Calibration, Diversity Metrics, and Cryo-EM Validation | 6 | 5 | 1 | 3 | Brandon questions how Chai avoids self-reinforcing generative loops without ground truth, prompting Matt to explain confidence calibration and diversity metrics. Neil shares an anecdote about a 0.33 angstrom Cryo-EM validation result that initially looked like an error. | |
| Chai-3 Scaling and Multi-Property Developability | 5 | 4 | 1 | 2 | RJ inquires about what distinguishes Chai-3, focusing on developability attributes like stability, aggregation, and manufacturability alongside affinity. Matt describes their decision to scale model capacity rather than hand-crafting target-specific heuristics. | |
| Designing the Platform: Single-Tenancy, CAD UI, and Pharma Pragmatism | 5 | 5 | 2 | 3 | Brandon asks how Chai overcomes medicinal chemists' deep skepticism toward AI tools. Neil and Matt describe building a CAD/Photoshop-style UI with single-tenant data isolation and winning over pharma partners by empirically hitting previously unsolved targets. | |
| Agile Biology: Transitioning from Waterfall Drug Stages to Iterative Loops | 6 | 4 | 1 | 4 | Brandon probes how a pure platform company battle-tests lead optimization without advancing internal therapeutic assets. Neil and Matt explain that one-shot generative models collapse traditional waterfall discovery gates into rapid agile loops. | |
| The Difficulty of Epitope Prediction Without Evolutionary Templates | 7 | 5 | 1 | 4 | Brandon pushes on epitope prediction as the true bottleneck, noting that antibodies lack conserved evolutionary MSA templates. Matt agrees it is a harder problem and notes that AlphaFold-Multimer historically failed 90% of antibody-antigen docking benchmarks. | |
| The Economics of Bio AI: Unlocking Unreachable Drug Modalities | 7 | 5 | 3 | 6 | RJ challenges the core economic thesis of structural AI, arguing that saving a few million in hit discovery is negligible in a half-billion-dollar clinical campaign. Neil explicitly challenges the premise, explaining that AI unlocks entirely unreachable modalities like agonists and bispecifics. | |
| Infrastructure Challenges: Data Microheterogeneity, GPUs, and Temporal | 5 | 5 | 2 | 3 | Matt explains data microheterogeneity and parser complexity in biological formats, while Neil describes GPU cluster procurement bottlenecks and using Temporal for distributed durable execution. | |
| ML Philosophy: Core ML Frameworks and Architectural Simplicity | 8 | 4 | 2 | 4 | Brandon demonstrates strong ML expertise by discussing inductive biases, triangle inequality layers, and data distillation papers. Matt argues for radical simplicity and trimming architectural sub-modules to benefit from scaling. | |
| Competitive Dynamics: Closed Frontier Models and Data Moats | 7 | 4 | 4 | 4 | Brandon asks how Chai maintains defensibility without an internal proprietary pipeline and proprietary data moat. Neil pushes back directly against the 'no data moat' framing, drawing parallels to Anthropic's enterprise model. | |
| Biopharma Economics: Eroom's Law and Capital Allocation | 7 | 3 | 1 | 2 | The conversation covers biopharma economics, Eroom's law, and treating internal ML researchers as portfolio capital allocators directing compute toward high-conviction ideas. | |
| Removing Bottlenecks and Final Takeaways for AI in Biology | 4 | 4 | 0 | 1 | In closing, Matt and Neil identify wet-lab validation cycle times and talent obscurity as the primary remaining bottlenecks, summarizing why AI in biology is transitioning into a deterministic precision engineering discipline. |
Statements from this episode (32)
Patil: Chai operates as a neutral software factory for medicines
“There's a lot of bio companies, AI for bio companies that are like making their own drugs. We really don't see ourselves that way, right? We see ourselves as almost a neutral software factory. For making medicines.”
McPartlon: Designing an antibody binder is often easier than predicting binding
“And in some cases it might actually be even easier to design a protein binder that is an antibody than to actually predict how it might bind that target in general.”
Patil: Generative models enable atom-level precise antibody epitope design
“And I think, like, one of the things that's really exciting about where we're getting to with some of these models is we can start to get that precise, right? Epitope, right? Meaning like binding spot, right? A very specific set of atoms to have the antibody g…”
Patil: Chai is raising capital to train more general generative models
“Like a lot of the money that we're raising now will let us train bigger models that can maybe be even more general and start to account for even more things at the same time, right?”
McPartlon: OpenAI co-led Chai Discovery's seed funding round
“Actually OpenAI co-led our seed round.”
McPartlon: Chai-2 generated binders for 25 targets with 20% hit rate
“So we designed antibodies to 50 targets for that paper, got binders to about half of them with being on average around a 20% hit rate for binding.”
McPartlon: AI biology problems are solved like standard machine learning problems
“People think you can't work on like AI bio unless you're a biologist, but it's kind of like you can't work on like video models unless you're like a director or something. Like there are all these like super domain specific things like, oh yeah, to understand …”
Patil: Chai-2 crossed the performance threshold for antibody design last year
“The ultimate goal here is to design medicines, right, and design new molecules, and I think CHI-2 really crossed the threshold of performance for doing that with antibodies a year ago.”
Patil: Modern structure prediction models achieve sub-angstrom experimental accuracy
“We're getting the point now where these structure prediction models are within, you know, a few angstroms or less of the actual atomic positions that you validate.”
McPartlon: Chai-2 achieved 0.33 angstrom error in cryo-EM testing
“And in this case, it was a 0.33 angstrom error, which is one third the width of an atom.”
McPartlon: Chai-2 was evaluated on targets with no known antibody binders
“We actually chose these targets specifically like to have no known antibody binder. So like if we did get a hit, like it was definitely the first antibody hit to this target.”
McPartlon: Chai bet on model scaling over analyzing individual target failures
“Just be bitter, less impaled in that sense, and just really bet on the models getting better. And we definitely took the latter approach. Like we bet on the models getting better and we just pushed as hard as we could on that front.”
McPartlon: Structure prediction is a speed-run benchmark compared to de novo design
“The nice thing about structure prediction is there is a ground truth that you can compare against. For design, you, don't really have that. You're like, here's some new, like disease molecule. Give me a binder for that. And like, if you want to know if this th…”
Patil: Eli Lilly was an early design partner for Chai's V1
“We'd been working with you know or talking to Eli Lilly and, you know, they were, you know one of the first partners to really work with us closely on that kind of you know, it made that V one of that that design suite, right. That you can use to engineer some…”
Patil: Enterprise cybersecurity buyers are surprisingly non-technical and unsophisticated
“I come from a cybersecurity background or, you know, have worked on security products before and those were dark, dark years because you spend a lot of your time actually selling to people who are surprisingly not that technical. You think cybersecurity people…”
McPartlon: Chai prodigy hire Nathan Rollins entered Baker lab at 14
“One of our first hires on that realm was Nathan Rollins, who I think he started working in the Baker lab at 14. Graduated from Harvard at like 18 and got his PhD by like 21 or something like this in the Marx lab.”
Patil: Chai runs wet-lab experiments only to validate its models
“We don't care about going and developing those drugs. Like we just do that in service of validating and making our models better.”
McPartlon: AI models are nearing direct output of viable drug molecules
“We're kind of at the inflection point now. We're really seeing this internally at CHI, where the models are getting pretty close to, like, producing Molecules that could eventually, or like are very close to drugs.”
Patil: Enterprise software lifespans have shrunk from 20 years to one year
“Like maybe you'd have built software that was supposed to last like 20 years. Now it's supposed to last maybe one year, but it is the bridge to deliver value and kind of enable the research that then gets you to the next thing.”
McPartlon: AlphaFold-Multimer gets antibody-antigen prediction right only 11% of the time
“Not really like outfold to got like, I think, 11%, the multiple version of this got like 11% of antibody antigen prediction cases. Correct. That means 90% of the time it's wrong.”
Patil: Chai-2 demonstrated precise antibody-based GPCR agonist activity
“For example, in CHI-II, we showed like GPCR agonist activity, right? Where you can really hit the switch on a, you know, on a cell doorbell protein, so to speak, right? In a very precise way. Very, very, very hard to do that with antibodies if you can't be tha…”
McPartlon: Multi-specific antibody modalities must be designed computationally from first principles
“There are drug modalities that you just can't discover with immunization. Like you're not gonna design your like crazy multi-specific Warheaded, super intense formats. These are really things where you kind of have to design these from first principles. Even j…”
Patil: Hyperscalers and top AI labs buy 95% of top-tier GPUs
“The hyperscalers and the, you know, the biggest the biggest AI labs are buying 95 plus percent of it, right? And then you kind of have the startups, like, fighting over the scraps.”
Patil: Biomolecular AI models will be as massive and impactful as LLMs
“And, you know, I think this class of models is going to be like just as big, just as impactful as LLMs, but it's almost like the compute market, like kind of doesn't realize that yet, both in the capacity sense, but also in like the software stack sense.”
McPartlon: AlphaFold 3's high architectural complexity contradicts the Bitter Lesson
“One thing that I like to say is kind of like complexity and being bitter lesson pill, they're like fundamentally at odds. For example, I think like outfold three, I might get this number wrong, but I think it was like 23 sub modules. And at that point, that's …”
McPartlon: Chai's research team largely lacks formal biology backgrounds
“Like the whole research team at CHI except for me and Kevin, really like we're the only people with quote bio background. Even still, like we're pretty far removed. So I think like we try to like look at every problem as a core ML problem.”
McPartlon: Triangle layers run inefficiently on modern GPU architectures
“These layers are pretty costly and like that kind of limits what you can do with the architectures. They're not like, Not only are they like costly in terms of compute, they're just like not efficient on modern GPUs either. You have small hidden dimensions, la…”
Patil: Closed frontier models capture the majority of market value
“And actually, if you look at the amount of value captured, it's actually the closed source frontier models. You know, the whole pie is growing, but it's growing so fast that even as the open source models like share expands, the frontier models are still able …”
Patil: Chai fine-tunes specialized foundation models for biopharma enterprise partners
“A lot of these deals you know, and this is all public, right? We are working with them to, you know train or fine tune a version of our model for them.”
Patil: GLP-1 drug revenue recently exceeded all AI labs combined
“Yeah, I mean, I think up until I think three months ago, right, GLP one's like total revenue was more than all of the AI labs put together.”
Patil: Chai Discovery maintains a total company headcount of 30 people
“We're only 30 people. And that's because everyone we hire onto the research team or the engineering team, you know, especially now that they're in some ways like very empowered with AI, A lot of it is just, like, allocating, you know, their attention into the …”
Patil: Biology is rapidly transitioning into a precision CAD engineering discipline
“You start to get to the point where now you can declaratively precision engineer what you want, rather than betting on, you know, nature or trial and error to get you there. And I think that Look, we had the same thing happen in software where you can write co…”