Summary

Chai Discovery is deliberately not building its own drugs; it sells a modeling and product layer to pharma. Matt McPartlon says that thesis was controversial when Chai started. The four partnerships named in the transcript are Eli Lilly, Pfizer, Novartis, and Genentech. Neil Patil calls Chai “almost a neutral software factory for making medicines.”
The commercial unlock was Chai-2: antibodies designed against 50 targets, with binders for about half and an average binding hit rate of around 20%. The team first tried interesting targets, then shifted toward targets validated in CRO catalogs after many early targets did not work. The transcript’s “no known antibody binders” point applies to the targets used in the cited cryo-EM/data-leakage check, not necessarily all 50.
A cryo-EM validation in the Chai-1 paper produced a 0.33-angstrom error, roughly one-third the width of an atom. The team initially thought the result had to be wrong because the prediction overlaid the electron-density point cloud with almost no visible difference.
The value proposition is not only faster discovery but access to modalities and mechanisms that traditional immunization cannot readily produce. Matt points to precise GPCR agonists, multispecific formats, and bispecifics, where finding two independent binders by traditional screening creates a multiplicative challenge.
Epitope prediction—deciding where a therapeutic should bind—is described as harder and still largely unsolved. Matt cited, with uncertainty, roughly 11% accuracy for AlphaFold 2/the multimer version on antibody–antigen prediction cases, meaning most such cases were wrong. He said Virtual Cell might be the closest state-of-the-art direction, but is still a ways out.
Compute is a structural headwind for bio-AI. Neil says startups may be competing for scraps while hyperscalers and major AI labs buy more than 95% of roughly 10,000 B300 units in his example. The models’ pair representations and L-cubed batching create different compute and memory needs from LLMs. Chai had also raised another $40 million, not $400 million.
Chai’s product is intentionally CAD-like rather than chatbot-like: Autodesk, SolidWorks, or Figma for molecules, with an epitope paint tool and a content-aware-fill-like binder-generation workflow. Single-tenant deployments helped address pharma’s IP concerns, and the company works with partners on specialized or fine-tuned model versions.
The field’s desired transition is from a waterfall of target discovery, hit discovery, and optimization to a model-assisted loop. Matt’s research north star is drug-like molecules directly from models; the product north star is iterative campaigns that can eventually operate at higher levels of abstraction, from epitopes to pathways. Matt’s fiat bottleneck is validation latency, while Neil’s is talent scarcity. The discussion of bending Eroom’s law is a host’s framing, not a stated Chai result.

Deep dive

1. Chai sells the modeling layer, not the drug — and treats that as the thesis

Matt says Chai’s thesis has always been to be the software and modeling layer, which was “very controversial at the time.” The team was betting that the models would become capable enough to design useful molecules. Multimer structure prediction only became available around 2021, and Matt says it was needed to unlock design; inverse folding then began working in real-world experiments, with credit to the Baker lab for its validation work.
Neil’s positioning is explicitly non-competitive with customers: Chai does not see itself as an AI-bio company making its own drugs. He calls it “almost a neutral software factory for making medicines,” which lets it support pharma companies in their drug-discovery journeys. Matt says Chai depends on its partners succeeding.
The four partnerships named in the transcript are Eli Lilly, Pfizer, Novartis, and Genentech. The case for Chai is generality at scale: after Chai-2, the team did not show success on only one or two targets and then stop; it decided to “go all in.”

2. Why antibodies — and why the incumbent competitor is a mouse

The discussion uses a lock-and-key analogy: a disease-related protein is the lock, and the antibody is a key designed to stick to it. The speakers describe antibodies as flexible, general proteins with relatively uniform framework regions and variable binding tips.
Antibodies are Y-shaped proteins. The binding occurs at the tips, while the framework region is relatively constant and can, for the most part, be selected from frameworks the immune system already recognizes. That makes antibodies attractive design targets: the model can focus on the fingertips while retaining a familiar scaffold.
Matt says Chai CEO Josh likes to call the company’s biggest competitor “the mouse—or nature, in certain ways.” Traditional discovery can involve immunization campaigns or very large yeast-display screens. The latter may search at least billions of candidate molecules against one target, eventually producing perhaps one, two, or a dozen hits.
The limitation is that a traditional hit may only be known to stick to the target. You may not know where it binds, whether it is drug-like, or whether it has the therapeutic effect you want.
Matt also notes an asymmetry: predicting how an existing antibody binds a target is notoriously difficult, but design can be more selective about which structures to attempt. If the model has freedom to choose, it may be able to select easier cases.

3. Precision, not only speed — epitopes, selectivity, and cross-reactivity

A host describes one Chai differentiator as intentionality: a user can specify the area where the binder should engage and later inspect whether the validated antibody binds in the expected pose. That can support both therapeutic reasoning and selectivity design.
Matt describes the target as a specific epitope—a binding spot—or even a specific set of atoms. Historically, brute-force screening might find a binder somewhere on a molecule without giving the user control over where it pokes the target.
GPCRs are the showcase example. Matt describes them as “doorbell proteins” in the cell membrane. An antibody engineered to poke a GPCR in a particular way could trigger a downstream chain reaction, producing agonist behavior rather than merely blocking a target.
Neil defines cross-reactivity with a development example: a drug may need to bind both the human and monkey variants so it can be tested in monkeys. The product and models can identify a conserved region and target it. Selectivity is the inverse problem—binding the intended protein while avoiding a similar human protein whose accidental blockage could cause toxicity or side effects.
When a host frames this as a large counter-screen, Matt narrows the claim: the framing is that users can be specific about what they want to bind and what they want to avoid. He says larger models may eventually account for more such constraints simultaneously.

4. Chai-1 was open-sourced as an infrastructure forcing function

Neil says that, early on, the team was working on protein design while realizing it needed an MSA pipeline and substantial infrastructure. When AlphaFold 3 came out, the team decided to open-source a model and build the infrastructure required to support it. Neil agrees with the host’s framing that the project was also a way to learn how to build production infrastructure.
Chai had five people at the time. The open-source Chai-1 project gave them a clear goal and forced them to build infrastructure at company scale rather than only research scale.
An MSA, or multiple sequence alignment, gives a structure-prediction model many related protein sequences. Conserved amino acids and correlated mutations can indicate which positions are close in three-dimensional space. A host summarizes this as learning from evolution which changes preserve or break a protein, and Neil agrees that it is remarkable that the method works.
The early launch story captures the constraints: OpenAI co-led Chai’s seed round, and the five-person team worked for a period out of a mostly vacant OpenAI office in the Mission. They stayed up 48 hours to finish the paper, web server, and technical report. Josh then interviewed with Bloomberg TV or a similar outlet around 7:00 in the morning while the rest of the team hid, leading the interviewer to remark that the company did not appear to have any employees.

5. Inside the models: a tokenizer, an LM-like trunk, and a diffusion component

Neil’s sketch of Chai-1 is roughly a tokenizer, a transformer or language-model-like trunk, and an image-diffusion-like component stitched together. The tokenizer is not a conventional word tokenizer: atoms in a molecule are grouped into tokens, while properties such as charge and element type are also represented.
The model ultimately has to return to three-dimensional atomic coordinates. The diffusion component converts the internal representation into the predicted structure.
Matt describes Chai-2 as an all-atom diffusion model. Rather than only predicting positions for a fixed sequence, it can design atoms, place them, and decide which atoms are present, then map those atom choices back to amino acids.
Chai-1 is a folding model: given a sequence, it predicts the structure. Chai-2 is a design model: given a target structure, it generates candidate molecules intended to bind it. Matt says this crossed the threshold of usefulness for antibody design about a year before the recording.
Matt argues that a biology degree is not a prerequisite for computational work in the field. He compares AI-bio to video models: one need not be a film director to work on the underlying machine-learning problems. The transcript identifies Matt’s background as theoretical computer science and later says that, apart from Matt and Kevin, the research team largely lacks a formal biology background.

6. Chai-2 crossed the usefulness threshold: 50 targets, half with binders, and a roughly 20% hit rate

A host’s analogy describes Chai-1 as recognizing that an image contains a cat and Chai-2 as generating a cat in a specified setting. The deeper point is that Chai-2 must generate a sequence and a compatible structure together.
Matt describes the process as iterative: the model can adjust the structure, ask what sequence would support it, adjust again, and eventually converge on a self-consistent sequence–structure pair. Neil compares the process to an EM-like algorithm.
The 50-target campaign was selected partly pragmatically. The team first chose interesting targets, but many did not work while the wet-lab process was still being developed. Matt says about half of the targets simply did not work, so the team shifted toward targets that CROs had already validated and selected an interesting set from those catalogs.
The reported result was binders for about half of 50 targets, with an average binding hit rate of around 20%. The transcript says that this was when pharma began to see signs that the approach might work in its programs.
The “no known antibody binders” point belongs to a later validation example. For the cited Chai-1 cryo-EM work, Matt says the targets were chosen specifically to have no known antibody binders, so a hit would be the first known antibody hit to that target. The transcript does not establish that all 50 Chai-2 targets were selected on that basis.
The commercial consequence was inbound interest from pharma and biotech companies asking to use the model, which led Chai to build a product and secure the compute needed to serve it.

7. Validating designs when there is no direct ground truth — and a 0.33-angstrom surprise

Protein design lacks an obvious ground-truth metric. A common workaround is to run the designed sequence through an independent structure-prediction model and check whether it predicts a structure similar to the design.
Matt is candid about the failure mode: researchers can game self-consistency, especially if a model produces nearly identical structures. Chai therefore also considers model confidence and generation diversity. A model that is consistent and confident but produces the same sequence and structure every time has not demonstrated broad problem-solving ability.
Physical validation is slow. Cryo-EM can provide structural evidence, but a single validation may take months. Chai also works on in-silico metrics that might predict laboratory success. The wet-lab feedback loop has improved from months to weeks, which is slower than scaling an LLM evaluation but fast enough to begin recursive self-improvement.
In the Chai-1 paper, the team overlaid a predicted structure on an electron-density point cloud and initially saw no difference. Neil gives the error as 0.33 angstrom, roughly one-third the width of an atom. The team thought the result could not be right and that the lab must have sent back the wrong design.
When asked about data leakage, Matt says the targets in that example were chosen to have no known antibody binders. The result therefore could not simply be explained by reproducing a known antibody–target complex.

8. Chai-2.5 to Chai-3: they bet on the models rather than autopsying the misses

After Chai-2, the team debated what to do with the 25 targets that had not produced hits. One option was to study their shared properties in detail; the other was to bet that scaling and improving the models would eventually solve them. Chai chose the latter.
Matt describes the release history as incremental: Chai-2, Chai-2.5, Chai-2.7, and eventually Chai-3, with progressively better performance.
Binding affinity is one major axis. A weak binder is not enough; Chai wants molecules at or near therapeutic grade. Developability matters as well: whether a molecule is safe, stable, manufacturable, and resistant to self-aggregation.
Chai-2.5 included a developability study, and Neil says the team has been pleasantly surprised by how much it has improved those properties.
A host argues that such ancillary properties may matter more to product usefulness than raw structure-prediction accuracy. The guests agree that these properties are important, while Matt says structure prediction remains a useful benchmark because it has a comparatively clear ground truth.

9. The product is CAD for molecules, and single tenancy addressed the IP objection

Neil says the obvious interface would have been a chatbot, but the team instead built something visual. He compares Chai’s product to Autodesk, SolidWorks, or Figma rather than ChatGPT.
The design suite resembles a Photoshop-like environment: users can load a molecule, paint an epitope, and use a content-aware-fill-like workflow to generate binders. It also includes scientific analysis and plotting so users can inspect results and avoid conditioning the models incorrectly.
Pharma is highly sensitive to IP. Neil says people initially told him that customers would never put proprietary data into a shared platform and have new medicines generated there. His security background helped him design aggressive data segmentation and single-tenant deployments, in which each customer has a separate version or account.
Eli Lilly was one of the early partners to work closely with Chai on the first version of the design suite. The transcript says that, in many deals, Chai works with a partner to train or fine-tune a model version for that partner; it does not establish that these details are public for every deal.
On scientist adoption, Matt says results create the activation energy. Pharma scientists are pragmatic and may begin by giving Chai a target they have struggled with before. Once the results are convincing, they are willing to try the product.
Matt contrasts this with his experience selling security products, where he found customers less technically sophisticated. At Chai, he says, partners often include scientists who have spent 5, 10, or 20 years studying one target.
Matt tells the story of a pharma scientist who began crying during a review because Chai had helped produce an initial binder to a target she had worked on for 10 years.
Internal rigor comes from hires such as Nathan Rollins. RJ describes Rollins as having started in the Baker lab at 14, graduated from Harvard at 18, and earned a PhD by about 21 in the Marks lab. He was initially skeptical, then said after seeing Chai-2 results that he needed to “bulletproof this” and that nobody should celebrate yet.

10. From waterfall to loop: two north stars and a product designed to be replaced

RJ describes target discovery, hit discovery, and optimization as a waterfall in which each stage has a gate and can take months to years. The cost of trying things early is high. If models produce promising candidates, the process could become a loop more like agile software development.
Matt says Chai does not want to preserve the stage distinctions merely because initial molecules are usually poor. Its north star is to produce drug-like molecules directly from the models. He acknowledges that this will involve many roadblocks, including better conditioning and reinforcement-learning systems, but calls it achievable.
RJ describes two apparently conflicting goals. Research aims to produce increasingly good de novo medicinal candidates, potentially starting with easier classes such as antagonists. Product aims to support iterative workflows in which laboratory results condition the next model run.
The two goals can coexist because harder modalities will require higher-level product workflows. RJ gives agonists, bispecifics, and ADCs as examples. Neil asks how a model can reliably one-shot a switch on a cell; these are the next abstraction levels the product may need to support.
Neil says he became “very existential” about the possibility that Chai-4 could obsolete the current molecule-visualization product. He accepts that the product may last only about a year while serving as a bridge to deliver value and enable better research.
Today’s product is more like Cursor for molecules: users inspect bonds and molecular properties. Later versions could orchestrate campaigns across epitope choices, then campaigns across targets in a pathway, eventually reaching the “outer loop of science.”
Chai does not run its own drug-development pipeline. Internal science teams maintain benchmark sets containing targets with known therapeutics and targets selected to push the models. They run experiments to validate and improve the models, not to develop those drugs.

11. Epitope prediction is the harder problem, and both guests say it remains difficult

A host argues that epitope prediction may be harder than finding binders. Matt agrees, calling it “a ridiculously hard problem” because it requires global biological context to determine what is interacting and how.
Matt distinguishes several levels: identifying which proteins are responsible for a disease, understanding how those proteins interact, and determining which interaction should be disrupted. The specific site to block is the epitope.
Asked about a hypothetical SARS-CoV-3, a new flu virus, or a similar event, Matt offers a hedged workflow: run a structure-prediction model, inspect where it predicts binding, and use that site if the model is highly confident. He immediately says that structure prediction is not solved in general.
Matt cites, with “I think” and “like,” roughly 11% accuracy for AlphaFold 2/the multimer version on antibody–antigen prediction cases, implying that most such cases are incorrect. The exact scope of the 11% figure is presented uncertainly in the transcript.
The hosts explain that antibodies are designed to recognize unfamiliar pathogens by recombining components, so they lack the conserved evolutionary template that helps with other proteins. Matt agrees with that framing.
Matt says Virtual Cell might be the closest thing to the state of the art for the broader context problem, but it is still a ways out.

12. The economics pushback: this is about new modalities, not only saving $2 million

RJ lays out the economic puzzle: antibody discovery may cost a couple of million dollars inside a successful drug-development campaign that could cost roughly half a billion dollars, while the commonly cited $2.6 billion figure includes failures. If Chai only saves the discovery cost, why is the product so valuable?
Matt is the first to challenge that premise. Neil emphasizes that precise antibody design is very hard. Matt’s broader answer is that Chai can unlock targets, mechanisms, and modalities that are not readily available through traditional immunization.
Matt cites the Chai-2 GPCR agonist example. An antibody can be engineered to hit a GPCR switch precisely rather than merely bind somewhere on the protein.
He also cites multispecific and four-headed formats that must be designed from first principles. In a bispecific, both arms need to bind different targets. If each arm has a one-in-a-billion chance of producing a binder, the traditional approach faces a multiplicative problem.
RJ adds the portfolio argument: Chai is not helping with only one drug but potentially with a partner’s portfolio of targets. Neil says the platform lets the partner concentrate learning across a subdomain and benefit from that learning across programs.
The company’s mission is to turn drug discovery “from a scientific experiment into an engineering discipline”: users should be able to define the desired therapeutic properties declaratively and have the model fill in the gaps.

13. Compute is LLM-oriented, and engineering primitives are a bottleneck

Matt’s first answer to what blocks science-to-engineering is data infrastructure. Biological file formats contain multiple copies of structures, unresolved regions, alternative placements, information about how structures were solved, and other edge cases. The engineering question is whether to encode many special cases or use a simpler, more principled strategy.
When a host suggests LLMs might handle the parsing, the response is that LLMs do not necessarily know all the relevant edge cases. Chai favors code that humans can understand and audit.
Neil’s answer is compute. After capacity crunches in spot and on-demand markets, Chai decided to buy its own hardware. He describes the market with an example of roughly 10,000 B300 units shipping broadly, with hyperscalers and the largest AI labs buying more than 95% and startups competing for the remainder.
Neil says newer systems such as B300 and Vera Rubin have been designed in an LLM-oriented way, with large KV caches and configurations in which 72 GPUs need to communicate with one another. These systems may provide useful performance, but the compute and software markets are not yet optimized for biology models.
Matt explains that an AlphaFold-2/3-style model attends over a pair representation that is roughly length L-squared rather than length L. Batching can therefore reach L-cubed complexity. The result is high compute per token and substantial memory-bandwidth overhead; even layer normalization can become a meaningful bottleneck.
Neil separately notes that triangle layers are costly and inefficient on modern GPUs because they combine small hidden dimensions with large sequence dimensions, the opposite of the shape GPUs handle most naturally. He describes them as trading parameters for compute.
Neil’s infrastructure solution is durable execution through Temporal. Distributed model calls can fail because of flaky buckets, databases, or GPUs. Without durable execution, teams can spend most of their time on queues, retries, and orchestration.
Chai uses Temporal for database calls, side effects, model calls, and long-running data pipelines so that failures can be retried without bespoke queue logic.
Near the end, Neil says Chai has raised another $40 million and needs to buy another compute cluster for larger inference and training runs. The transcript does not say $400 million. His conclusion is that reliable engineering primitives will enable more ambitious biology work.

14. Simplicity versus the Bitter Lesson — and the data-efficiency pushback

Matt says complexity and being “Bitter Lesson-pilled” are fundamentally in tension. He gives AlphaFold 3 as an example, hedging that it may have had 23 submodules. At that level, it becomes difficult to understand how changing one module affects the whole system.
Chai keeps a picture of SpaceX’s Raptor 1 and Raptor 2 engines on its office wall as a reminder to delete unnecessary components and simplify systems.
Neil pushes back that AlphaFold 2 and AlphaFold 3 succeeded partly because they were relatively small, data-efficient, compute-intensive models built around many inductive biases. Removing those biases would require new data sources or a fundamentally different treatment of data.
Neil cites an Apple paper that distilled on a very large ESMFold-derived dataset. He says it obtained useful signal but failed to generalize because it was pattern matching rather than reasoning.
Matt agrees that protein structure data is difficult: there is abundant sequence data but much less experimental structure data. Sequence data helped ESM-style models, while applying the same approach directly to experimental structural data is much harder.
His resolution is to abstract the useful ideas from AlphaFold rather than copy its entire architecture. Triangle layers provide an example of a valuable inductive bias that can be modified and built upon, even if they are expensive.
Matt connects this to his theoretical-computer-science background. His adviser told him never to underestimate polynomial time, and Matt’s first paper used an n-to-the-20th-time algorithm. The broader lesson is that one can first allow a broad solution space and simplify later.

15. Commoditization, the data-moat question, and pharma as a capital allocator

The hosts describe protein design as crowded, with perhaps 10 or 15 startups. They also recall an example from shortly after RFdiffusion in which someone produced picomolar binders and validated them with cryo-EM. The transcript does not establish that those were specifically mini-binders.
Neil expects some modalities to become commoditized, while more ambitious tasks remain served by frontier models. He compares the situation to LLMs: open models may cover some use cases, but closed frontier models capture substantial value because they solve harder tasks and are paired with better products.
He says the product layer matters as much as the model layer. Users may not choose an open model if it lacks the surrounding workflow and tools. Even if AGI eventually one-shots everything, he expects a substantial period in which the product remains important.
Matt says biology’s slow measurement loop is a defense against instant commoditization. Public sequence data can create a base model, but useful measurements take time and must be iterated on. He says AGI may not solve those technical blockers immediately.
Neil rejects the premise that Chai has no possible data moat. He compares the situation to Anthropic working with enterprises whose data cannot necessarily be used to train a general model. Chai is investing in ways to turn compute into data and uses partner requests to guide research.
In many deals, Chai works with partners to train or fine-tune specialized model versions. Partners may have experimental data or preferred properties that help the model perform better on particular target classes. Sometimes the lift is as concrete as ensuring that designs have a partner’s preferred property.
Neil says Chai has no plans to start its own drug-development pipeline. He values the incentive alignment of improving the models, helping partners succeed, and using those results to improve the product.
Neil frames the downstream value of a pharmaceutical token as unusually high because successful drugs can become multibillion-dollar assets. He says the two GLP-1 drugs combined may represent a trillion-dollar asset, with the appropriate hedge.
A host, not Neil, says that until roughly three months earlier, GLP-1 revenue exceeded the combined revenue of the AI labs. Another speaker notes that Genentech was one of the major early venture outcomes in Silicon Valley.

16. Eroom’s law, the fiat wishes, and the takeaways

The discussion defines Eroom’s law as Moore’s law backward: the cost of developing a drug is rising rather than falling. A host says this would eventually make the marginal return on new drug development negative. The statement is presented as the host’s framing, not as a measured Chai result.
Another speaker speculates that companies such as Chai might help flip or bend that curve. The transcript does not establish that Chai has already done so.
The group repeatedly uses capital allocation as an operating metaphor. Chai describes its roughly 10-person research team as allocating ideas to compute, while the broader company—about 30 people in the transcript—is described as allocating attention and compute with the help of AI tools.
Matt’s bottleneck to remove by fiat is the validation loop: being able to know instantly whether a protein-design hypothesis works. He says the field still involves some walking around in the dark.
Neil’s answer is talent scarcity. Many technically strong people move into LLMs, software, or SaaS, while comparatively few move into computational biology. He attributes part of this to the field’s obscurity and lack of visual accessibility, which Chai tries to address through its website and product.
Neil’s closing takeaway is that biology is beginning to move toward declarative precision engineering. With folding models within an angstrom and design models producing hit rates above 50%, he says a 96-well plate can contain 48 interesting binders. He compares this transition with software, circuit design in Cadence, and mechanical CAD.
Matt’s closing takeaway is that the field is working: the signs of life have already been shown, commercial traction exists, and there are still many research problems and low-hanging opportunities. He says the combination of three-dimensional geometry, diffusion models, language-model-like trunks, and core machine learning makes the field unusually broad and potentially impactful.

Vol.229 Macro Talk 109 | We Are at the Tail End of an Upward Dollar-Dominated Capital Cycle (recorded Aug. 6)Sergey Brin Retakes Gemini, 4 Labs Lose Containment, Compute Trades at NYSE w/ Kush Bavaria | EP 278