Chai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem
In a Nutshell
Chai Discovery's core thesis is that drug design is fundamentally a scaling problem: by simplifying their AI architecture to a unified sequence-based model and scaling compute, data, and parameters, they've achieved a 150x improvement in zero-shot antibody design success rates (from 0.1% to 15% hit rates). The company is building infrastructure rather than drugs, treating biology as an engineering discipline where the goal is to specify desired molecular properties and have the model generate them directly, reducing the idea-to-testable-molecule timeline from 9 months to 9 days. This "bitter lesson" approach—scaling laws over bespoke solutions—has enabled them to reach commercial adoption with major pharma partners like Eli Lilly and Pfizer.
These notes were generated by AI and may contain inaccuracies.
One of the big guiding principles is simplicity. When looking at a model like Chai 1, there are 23 distinct submodules. Trying to iterate on something with that complexity becomes difficult because each submodule needs to be understood independently, including their behaviors and dynamics. This approach does not scale well. The focus shifts toward simplifying the system by identifying what is really important, which makes the research process and identifying scaling directions much simpler.
Chai is engineering molecules with AI and functions as a foundation model lab for biology. The big idea is to make the drug discovery process look more like engineering. Code generation with language models works well because code is a simple abstraction. Biology today is characterized by trial and error. Early modern biotech medicines were discovered randomly through serendipity. The goal is to industrialize the process by developing tools that create abstraction layers similar to those in code and modern engineering, enabling faster iteration in biology and drug discovery.
There is a boundary between what can be engineered and what needs to be tested in the real world. This boundary has been moving over time, with an increasing proportion of the drug development process becoming engineerable rather than reliant on trial and error. The lab serves as an important verification component. Verification is a major theme in AI as well. When a model can be evaluated and verified, progress can be made through hill climbing. The objective is to transform drug discovery into drug design. The term drug discovery reflects the current paradigm of screening millions or billions of molecules to find one that works. The vision is to input the desired molecule properties and have the model materialize it. This approach may not reduce lab testing and could potentially increase it, as higher ROI leads to greater demand, similar to how software engineers became more in demand after becoming more productive.
The field has evolved significantly over the last decade. The biannual protein folding competition involves holding out protein targets that have never been seen before and are not deposited publicly, with academic groups competing to predict protein structures. A significant step change in performance occurred in 2018, followed by another breakthrough in 2020 with AlphaFold 2. The advent of deep learning initiated this progress. Early systems predicted distances between amino acids in proteins and used a renderer to produce protein structures from noisy, incomplete distance information. With relatively minimal information, machine learning systems could produce proteins that were largely correct, though resolution was limited compared to current capabilities. The evolution paralleled early image generation models that were grainy and blurry, progressing to high-definition outputs.
Once protein folding became realizable, other sub-problems emerged. Given a protein structure, the task became designing a sequence that would fold to that structure, moving closer to the drug design problem. The focus shifted to engineering proteins that match specific shapes or perform specific functions. This work occurred around 2021-2022. The advent of diffusion models enabled simultaneous generation of protein structure and sequence. Prompts became increasingly realistic, allowing models to be prompted on targets with specific shapes and binding locations while incorporating real-world constraints.
Early discussions about starting the company included work on antibody-antigen structure prediction. For a long time, protein folding for antibodies was considered too difficult due to insufficient data in the Protein Data Bank. This led to the belief that antibody design would remain out of reach. Early deep learning protein design work focused on mini proteins rather than antibodies. The question was whether antibody proteins with therapeutic function could be designed computationally. The thinking was that if antibody structures cannot be predicted, designing them would not be possible.
Biology appears complex with many distinct concepts, but these are all sequences of amino acids that function as different types of prompts for the model. Just as ChatGPT can handle both math problems and English homework, the same model can handle different protein types through a unified representation of amino acid sequences for both design and prediction.
Josh has been interested in AI and biology since childhood. He attended a high school with a stem cell lab and grew up as a programmer. He started his career at OpenAI on the early team when it was a nonprofit, contributing to GPT-1, GPT-2, and scaling laws research. The research question was whether models that can learn to speak English, German, and French could also learn to speak DNA and protein. Matt's background was in theoretical computer science before transitioning to deep learning and protein structure prediction.
Discussions about starting the company occurred around the time when the first signs emerged that antibody design was becoming feasible. This progress was fueled by new architectures, particularly diffusion models. The first time anyone was able to generate reasonable-looking proteins was with the advent of diffusion models.
Diffusion is not a one-step process. The previous generative design paradigm used variational autoencoders that attempted to compress the input distribution into fuzzy Gaussian vectors, which proved difficult. Diffusion models worked by giving the model more time to think through iterative refinement. The approach involves showing the model slightly broken proteins and teaching it to make incremental improvements. The protein can be made progressively more noisy and broken, with the model learning shortcuts to make it slightly better repeatedly.
The founding team consisted mostly of AI researchers who had worked on scaling models or applying them to this domain. With each generation of model, the team composition has expanded. For Chai 2, antibody engineers and scientists were brought in, including Andy Young. Nathan Roland joined the team after being homeschooled, starting college early, joining David Baker's lab at age 14, and beginning his PhD at age 18. As models improved, a product team was built with people who had built major products, including co-founder Jack from Stripe and Munaz, one of the top 10 code contributors at Stripe. Neil, who ran his own cybersecurity company, was brought on as security became important for deployment to partners. GPU specialists were added to scale systems without breaking at large scale.
Hiring is pragmatic and milestone-driven. Antibody engineers were not hired before an antibody design model existed. For some next-generation formats like multispecifics, early models did not work initially, but progress was fast enough to bring capabilities online. The team has been kept small, with everyone slightly over capacity, which forces prioritization on what matters most.
Founding engineer Kevin Woo created the first protein diffusion model. Engineering has been important since day one. Researchers are also excellent engineers. The focus is on building a high-quality codebase. The deliverable is software, making the company fundamentally a software company despite being AI researchers and protein designers.
The thesis when starting the company was to focus on de novo generation of molecules. Previous drug discovery AI work focused on making existing molecules slightly better through lab-in-the-loop approaches. The question was whether molecules with desired properties could be generated entirely on the computer through zero-shot generation. The first requirement was designing molecules with high success rates. When the company started, state-of-the-art antibody design had approximately 0.1% binding rate. With Chai 2, success rates reached approximately 15%. This improvement means screening 1,000 molecules yields 150 binders, enabling meaningful statistics on molecular properties and iteration on drug-like characteristics.
Progress exceeded expectations. Budgeting originally projected 20% hit rate in three to four years, with 1% considered ambitious. The approach follows principles that have worked in the rest of machine learning. Biology is treated as following the same principles as other domains rather than as a bespoke problem.
Chai is a bitter lesson compelled company that believes in scaling data, scaling models, and scaling compute. Scaling laws must be identified to avoid wasting time and resources. The guiding principle is simplicity. Identifying what is really important simplifies the research process and makes scaling directions clearer.
The models at Chai are built completely from scratch rather than fine-tuning existing models like GPT on protein data. The company building process involves taking a philosophy and iterating on it with guiding principles. Chai operates as a blue sky research company with a big AI problem that opens opportunities for customers as progress is made.
Simplicity is a core principle, reflecting the bitter lesson that scaling compute, data, and models drives progress. Rigor is essential because error bars in wet lab biology are quite large. In biology, a plus or minus 5% result in the lab may represent no meaningful difference, unlike software engineering benchmarks where incremental improvements like moving from 16% to 17% to 18% on SWE bench are trackable. The bar for step changes is therefore very high, and teams must be honest about whether genuine progress is occurring. Fancy models that appear to work on one or two tasks must demonstrate generalizability to broader applications before they can be considered production-ready.
Biology is actually one of the more verifiable domains because it provides objective readouts. When designing molecules in the lab, properties such as binding affinity, manufacturability, and other developability characteristics can be measured with specificity. While validation takes longer than running unit tests, these measurable properties make the domain more verifiable than tasks like code generation where subjective qualities like code maintainability are harder to quantify.
There are targets addressable with traditional screening methods, which are slow and rudimentary, and targets that are not druggable with existing methods. Model progress must be evaluated across both categories: generating molecules faster for existing targets, and unlocking novel biology and previously undruggable targets.
The team frequently considers whether adding modules or features represents incremental improvement or genuine compounding progress that will generalize to broader target classes. Their philosophy centers on identifying scaling laws and continuing to push loss down, believing that models will learn intrinsic features of targets. For example, even without explicitly representing glycosylation sites, sequence motifs should enable models to learn these patterns as loss decreases. Hidden features of targets should unlock with better models, though rigorous validation remains essential.
What constitutes a hard target varies by team. The science team might define it as an undruggable GPCR target with no functional modulation achieved. The ML research team might define it as a protein the model cannot fold correctly. The product team might define it as challenges representing modifications like glycosylations on membrane proteins to users. This diversity is considered a feature rather than a bug because it enables parallel progress across different dimensions, creating smooth advancement without single bottlenecks.
The company's goal is building a computerated design suite for molecules rather than developing one or two molecules or a pipeline of therapies. This means shifting biology toward an engineering discipline where principles for molecule design are specified upfront, and an engine realizes those specifications into lab-testable molecules. The aim is reducing the idea-to-testable-hypothesis loop from approximately 9 months to 9 weeks or 9 days, enabling orders of magnitude more ideas to be explored.
In the future, medicine quality should improve because teams will no longer advance molecules that are only 80% optimal due to timeline pressures. Hard-to-drug diseases like Alzheimer's may become tractable. Personalized medicines and rare diseases with smaller patient populations may become economically viable through faster iteration, faster launches, and lower costs. Drug design will become more intentional, with molecules designed to be highly specific to disease targets while avoiding negative interactions that current drugs often have.
Chai chose to build infrastructure enabling the existing industry rather than developing drugs internally like Isomorphic. The founding team had previously built full-stack drug pipelines and observed that as models improve, resources should flow toward making better models rather than clinical trials. Partnerships with pharma companies provide revenue that funds continued model investment, creating a scalable business. Working with partners like Eli Lilly, Novartis, Organon, and Pfizer prevents self-deception because these rigorous customers demand consistent performance across many targets, not just success on a small number of molecules.
Contrary to expectations that pharma lacks AI sophistication, these companies proved to be rigorous customers who thoroughly test claims but commit fully when data supports them. Pharma's competitive dynamics and limited drug exclusivity periods force continuous innovation adoption. Once models demonstrate value, adoption scales quickly with creative applications emerging across partner organizations.
The primary gold standard data source is the protein database, containing crystal structures deposited by scientists since 1970. A second major source is massive sequence databases containing trillions of tokens. Models trained on sequences learn to represent 3D structure internally because predicting the next amino acid requires understanding the protein's three-dimensional microenvironment. As models improve, they generate more accurate structures from sequence databases, creating additional training data for subsequent models.
The team takes a pragmatic stance rather than committing to sequence-only or structure-only paradigms. There may be more biological sequence tokens available than English language tokens. Once models reached the capability to design molecules for new targets, they began generating data from experimental exhaust, creating a flywheel effect similar to reinforcement learning in language models where base model capability enables new data generation approaches.
Staying at the frontier requires rigorous evaluation focused on hill-climbing model performance rather than head-to-head competition. Capabilities like achieving success rates eliminating large library screening requirements, and incorporating developability properties such as manufacturability, represent first-in-field achievements. The relevant baseline is nature and existing wet lab protocols that have accumulated hundreds of refinements over time, not other model providers.
Andy Young joined the team with two decades of experience in yeast display methods, having started this work at MIT. His background includes time at Pfizer and Genentech, where he developed expertise in antibody engineering. He holds a drug approval for antibodies developed using traditional methods. The team views this level of expertise as the baseline standard that AI systems must match or exceed to achieve adoption in pharmaceutical development.
The company aims to surpass existing capabilities rather than simply accelerate current processes. The goal extends beyond creating a faster version of traditional methods to developing breakthrough medicines that were previously impossible to create.
The team reached an inflection point several months prior to this discussion, marked by multiple major pharmaceutical partnership announcements. This milestone indicates that AI-generated antibodies have achieved parity with traditional methods, clearing the adoption threshold required for industry acceptance.
The most valued aspect of working at Chai is the exceptional mission-driven dedication across the entire team. This level of commitment exceeds what has been observed at other companies, with the closest comparison being highly motivated PhD laboratory environments. The team shares an obvious common goal and works with intense focus, supported by a clear philosophy and defined plan for advancement that has remained consistent since the company's founding.
The most difficult aspect involves managing systems at production scale. Initial operations began with 128 GPUs, which initially seemed like maximum scale compared to previous academic environments where four researchers shared eight GPUs. Current operations require maintaining substantially larger infrastructure and compute resources.
Key operational challenges include maintaining cluster health over extended periods, executing large training runs that continue for months, implementing automatic recovery when systems fail, minimizing communication overhead, and optimizing models for maximum resource utilization.
The most rewarding aspect is seeing tangible results that will impact patients. Team members with a decade of AI research experience now observe their work transitioning from academic publications to actual clinical applications. Recent breakthrough results and new product features that improve customer workflows provide continuous motivation, confirming that the technology will create meaningful change.
The responsibility of serving real pharmaceutical partners creates new operational demands. When releasing new model generations, teams must ensure absence of bugs and prevent performance regressions. This differs from pure research environments where experimental results could be published without production-level validation requirements.
The team has implemented production-grade codebases as the organization scales, ensuring that code remains accessible and maintainable as more contributors join. One co-founder emphasizes that moving fast long-term sometimes requires moving slower initially to build sustainable systems.
The company name "Chai" derives from Chemistry and AI, while also referencing the popular tea. The name was selected for its user-friendly simplicity compared to typical complex biotechnology company names, aligning with the company's goal of making drug discovery more accessible.
Primary excitement centers on ongoing partnership deployments and anticipated results from pharmaceutical collaborators. Partners are applying models to real drug development programs targeting serious diseases, with some programs potentially advancing molecules to clinical stages within accelerated timelines.
A year ago, zero-shot molecule generation with confidence in program success was not feasible. Current capabilities allow teams to generate molecules and advance programs toward clinical development, with some molecules potentially entering patients within a year of initial design.
Future technical focus includes improving controllability of generated molecules. Rather than hoping for good binding affinity, the goal is to specify exact parameters such as precise 10 nanomolar binding affinity. Several technical capabilities appear close to resolution, with team members motivated by the prospect of achieving these specific performance targets.
Keep Sequoia Capital in your library
Save the videos and channels worth coming back to, and find them again in one place.
