## Background
Published on YouTube by Sequoia Capital on August 4, 2026, this interview hosted by Pat Grady and Sonali Singh features Chai Discovery co-founders Josh Meier and Matt McPartlon. The page description outlines how Chai Discovery approaches drug design as an engineering and scaling problem rather than a bespoke science, highlighting their Chai-2 model's leap in de novo antibody hit rates and their vision for a molecular CAD suite that dramatically compresses discovery timelines.
## TL;DR
Josh Meier and Matt McPartlon explain why scaling data, compute, and model capacity transforms biological drug discovery into a predictable engineering discipline. By shifting from multi-module complexity to radical architectural simplicity and diffusion models, Chai Discovery elevated de novo antibody design hit rates from sub-0.1% to 15–16%. They detail their choice to build a CAD software platform for biopharma rather than operating as a traditional pipeline drug company, demonstrating how faster design loops and biological verifiability unlock previously undruggable targets.
## Atmosphere & style
The conversation carries a highly intellectually rigorous, technical, and forward-looking energy, balancing the visionary ambition of frontier AI with the pragmatic realities of enterprise biopharma. The dynamic between the Sequoia hosts and the co-founders is fluid and collaborative, featuring lighthearted moments about GPU clusters overheating alongside deep technical dives into structural biology. Listening to the discussion feels like sitting in on an insider masterclass where deep learning expertise, mathematical rigor, and engineering conviction converge to reshape medicine.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=wv53mDmY-k0&t=0s) Biology should be treated as an engineering discipline rather than a messy, serendipitous science, shifting drug discovery to a zero-shot drug design model.
- [00:00](https://www.youtube.com/watch?v=wv53mDmY-k0&t=0s) Higher AI model accuracy increases the ROI of wet lab testing, raising the demand for real-world lab verification rather than replacing it.
- [03:25](https://www.youtube.com/watch?v=wv53mDmY-k0&t=205s) Deep learning in structural biology evolved rapidly from low-resolution inter-residue distance estimation in 2018 to atomic-resolution prediction with AlphaFold 2 in 2020, culmining in joint sequence-structure diffusion models.
- [06:04](https://www.youtube.com/watch?v=wv53mDmY-k0&t=364s) Chai overcome the perceived data bottleneck for antibody modeling by treating antibodies and mini-proteins under a unified sequence-structure architectural foundation.
- [10:13](https://www.youtube.com/watch?v=wv53mDmY-k0&t=613s) Diffusion models outperform VAEs in molecular generation because their iterative denoising process gives the model 'time to think' and learn local structural rules.
- [15:22](https://www.youtube.com/watch?v=wv53mDmY-k0&t=922s) Chai-2 increased de novo antibody design hit rates from an industry baseline of ~0.1% to roughly 15%–16%, drastically accelerating hypothesis testing.
- [18:18](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1098s) Chai adheres to the 'Bitter Lesson' by building specialized models from scratch and pursuing radical architectural simplicity over hand-crafted domain heuristics.
- [20:05](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1205s) Biological target binding provides a physical, objective binary readout (binds or does not bind), making it more verifiable than software engineering benchmarks despite longer lab feedback cycles.
- [22:40](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1360s) Scale enables neural networks to naturally learn complex biological features (like glycosylation and functional motifs) without adding bespoke architectural submodules.
- [25:01](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1501s) Chai aims to build a Computer-Aided Design (CAD) suite for molecular engineering to arm the entire pharma industry rather than operating as a traditional biotech pipeline company.
- [25:48](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1548s) Collapsing the hypothesis-to-test loop from nine months to nine days enables high-specificity treatments for previously intractable or uneconomical diseases.
- [28:37](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1717s) Chai deliberately chose a platform infrastructure model over asset development to avoid overfitting to a few pipeline targets and continually re-invest revenue into compute and model scaling.
- [31:14](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1874s) Major biopharma partners adopt technological tools rapidly when proven, creating a powerful data flywheel that combines structure prediction, language modeling, and synthetic data generation.
- [38:20](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2300s) Chai focuses on benchmarking against nature and veteran wet-lab standards, demanding production-grade engineering over short-term blue-sky research.
- [45:28](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2728s) The near-future goal of molecular CAD is controllable parameter design, moving from probabilistic binding to exact binding affinity specifications (e.g., requesting a 10 nM binder on demand).
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=wv53mDmY-k0&t=0s) From Discovery to Design: Industrializing Biology
Josh Meier and Matt McPartlon introduce Chai Discovery with a simple thesis: biology should be treated as an engineering discipline rather than a messy, serendipitous science. In the early days of biotech, drug discovery relied heavily on trial and error, screening millions or billions of molecules like looking for a needle in a haystack. Chai's goal is to turn drug discovery into *drug design* by using generative AI models to materialise target molecules directly from high-level specifications, similar to how large language models generate code from plain text.
The co-founders address the misconception that better AI models mean eliminating wet lab testing. Rather than shrinking the lab's role, higher model accuracy dramatically increases the return on investment (ROI) for real-world verification. Just as AI coding tools expanded the overall demand for software engineers by making them more productive, AI molecular design will likely increase total lab testing as researchers scale up experiments to verify promising hits.
- **Foundational Architecture**: Chai-1 was built on 23 distinct submodules; the team emphasizes radical system simplicity to make models easier to iterate on and scale.
- **Verification Paradigm**: Lab experiments act as the objective verifier in an iterative hill-climbing loop to refine AI-generated molecular designs.
---
[03:25](https://www.youtube.com/watch?v=wv53mDmY-k0&t=205s) The Evolution of AI in Protein Science
Matt traces the rapid evolution of deep learning in structural biology over the last decade, pointing to the CASP (Critical Assessment of Protein Structure Prediction) bi-annual competition as a major benchmark. Before 2018, predicting 3D protein structures was largely unsolved. Deep learning first stepped in by estimating distances between amino acids to output blurry, low-resolution 3D renders—resembling early low-quality image generation models.
By 2020, AlphaFold 2 delivered a massive step-change in structure prediction resolution. Between 2021 and 2022, research shifted toward the inverse problem: given a target protein structure, could a model design a matching amino acid sequence? The true synthesis arrived with diffusion models, allowing systems to simultaneously generate both protein structure and sequence while accepting complex real-world prompts and binding constraints.
- **2018**: Introduction of deep learning to predict inter-residue distances in proteins.
- **2020**: AlphaFold 2 achieves atomic-resolution protein structure prediction.
- **2021–2022**: Shift toward inverse folding (sequence design given a target backbone shape).
- **2023– Present**: Diffusion models enable joint sequence-structure generation under complex prompt constraints.
---
[06:04](https://www.youtube.com/watch?v=wv53mDmY-k0&t=364s) Overcoming the Antibody Bottleneck in 2024
When Chai was founded in 2024, conventional wisdom held that predicting antibody-antigen structures was too difficult due to the sparsity of high-quality antibody data in public repositories like the Protein Data Bank (PDB). As a result, early machine learning efforts concentrated on "mini-proteins"—a class of engineered structures that were academically interesting but lacked broad therapeutic applicability in the pharmaceutical industry.
Matt, coming from a pure mathematics and theoretical computer science background, realized that antibodies and mini-proteins are fundamentally just sequences of amino acids. Much like ChatGPT can seamlessly switch between solving math problems and writing English essays, a single unified model can learn to handle diverse biological prompt types once given the right architectural foundation.
> "People were like, there was not enough data for antibodies in the Protein Data Bank to solve this problem... The holy grail was whether we could design antibody proteins on the computer, especially ones that had therapeutic function."
---
[10:13](https://www.youtube.com/watch?v=wv53mDmY-k0&t=613s) Why Diffusion Models Work for Biological Generative AI
Before diffusion models took over, variational autoencoders (VAEs) were the primary paradigm for generative molecular design. VAEs tried to compress complex structural distributions into a simplified latent space in a single step—a task that proved extraordinarily difficult for dense biological structures. Diffusion models succeeded because they break generation down into an iterative process, giving the model "more time to think."
By training a network to reverse noise step-by-step—taking a corrupted, broken protein structure and making it slightly better at each step—the model learns simple local rules that aggregate into high-precision, photo-realistic molecular designs. The host jokes that this process sounds like prompting ChatGPT to continually make a picture of a unicorn stronger, to which the team lightheartedly agrees with the conceptual similarity.
- **Variational Autoencoders (VAEs)**: Single-step compression models that struggled to capture complex 3D molecular distributions.
- **Diffusion Process**: Iteratively removes noise to refine noisy, incomplete structures into valid amino acid geometries.
---
[11:41](https://www.youtube.com/watch?v=wv53mDmY-k0&t=701s) Building the Multidisciplinary "Avengers" Team
Josh outlines Chai's hiring strategy, emphasizing that building a category-defining biology lab requires a cross-disciplinary team of AI researchers, structural biologists, antibody engineers, and software product designers. Prior to Chai, Josh was on the early team at OpenAI working on GPT-1, GPT-2, and scaling laws. His core thesis was straightforward: if language models can learn to speak English, French, and German, they can learn the structural languages of DNA, RNA, and proteins.
To validate their AI outputs, Chai hired prominent domain experts, including antibody scientist Andy Young and protein design prodigy Nathan Rollins. The company also recruited experienced product builders from leading tech companies like Stripe to create seamless user interfaces, ensuring that complex foundation models can be deployed intuitively by practicing drug discoverers.
- **Nathan Rollins**: Joined David Baker's Nobel Prize-winning protein design lab at age 14 and started his PhD at 18.
- **Product Strategy**: Bridges raw foundation model performance with intuitive user software to make biological CAD tools accessible to pharma partners.
[13:23](https://www.youtube.com/watch?v=wv53mDmY-k0&t=803s) Building the Avengers Team and Engineering Rigor
Building out the specialized team at Chai Discovery required bringing together diverse disciplines, ranging from deep infrastructure experts to top-tier AI researchers. Co-founder Josh Meier shares that one of their team members previously ran a cybersecurity company—an expertise that became invaluable as data security and deployment standards grew increasingly paramount when onboarding large pharmaceutical partners. To handle massive computational workloads, Chai brought in engineers who specialize in low-level GPU optimizations to prevent systems from breaking at scale. Illustrating the sheer intensity of their compute workloads, Josh recounts a lighthearted exchange with one of their hyperscaler cloud providers, who messaged them after a cluster issue saying, *"Oh, like the GPUs got too hot. I think you guys are running too many."* To which the Chai team humorously responded, *"Isn't that the point?"*
Beyond infrastructure, Chai maintains a hyper-lean team where every member is slightly over capacity, forcing strict prioritization on high-impact tasks. On the research front, co-founder Matt McPartlon highlights founding engineer Kevin Wu, creator of one of the very first protein diffusion models, citing his phenomenal execution speed. The co-founders emphasize that despite working on complex biology and frontier AI, Chai operates fundamentally as a high-quality software company where clean codebases and engineering excellence are non-negotiable standards.
- **GPU Stress Testing:** Hyperscalers reported physical hardware overheating due to the extreme compute density and load demands of Chai's workloads.
- **Strategic Hiring Sequence:** Chai intentionally delayed hiring dedicated antibody engineers until after their core generative models reached sufficient maturity to execute complex multi-specific antibody case studies.
- **Founding Engineering Feats:** Founding engineer Kevin Wu authored one of the earliest protein diffusion models in the field.
[15:22](https://www.youtube.com/watch?v=wv53mDmY-k0&t=922s) Hit Rates and Scaling Laws in Drug Design
When Chai Discovery set out, the state of the art in machine learning for antibody design suffered from an abysmal 0.1% hit rate—meaning only about 1 in 1,000 generated molecules actually bound to its intended target in wet-lab testing. Such a low hit rate meant labs had to screen thousands of candidates just to find a single viable hit, providing a weak learning signal to improve the underlying algorithms. Chai set out to transform drug discovery from incremental lab-in-the-loop tweaks into a true zero-shot *de novo* design paradigm on the computer.
With the release of their **Chai-2** model, the team achieved a dramatic step-function improvement, elevating success rates to roughly 15% to 16%. Screening 1,000 molecules now yields around 150 successful hits, providing rich statistical feedback to directly bake drug-like properties into the model. Matt admits that both he and Josh were surprised by the speed of this breakthrough; during initial budgeting, they believed reaching even a 1% success rate (1 in 100) would be a groundbreaking achievement, and estimated that hitting 20% would take 3 to 4 years.
- **Hit Rate Progression:** State-of-the-art antibody design moved from ~0.1% (1 in 1,000) to ~15%–16% (150+ in 1,000) with Chai-2.
- **Initial Expectations vs. Reality:** The team originally projected a 1% hit rate as a major victory and expected higher hit rates to take 3 to 4 years of development.
- **Zero-Shot Design Philosophy:** Shifting away from iterative, incremental lab-in-the-loop modifications toward generating complete, highly optimized candidate molecules directly on a computer.
[18:18](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1098s) The Bitter Lesson and Model Simplicity
Chai Discovery embraces a core philosophy heavily inspired by Rich Sutton's "Bitter Lesson"—the firm belief that long-term breakthroughs stem from scaling data, compute, and model capacity rather than hand-crafting domain-specific heuristics. However, scaling biology models requires finding true scaling laws, which in turn demands radical architectural simplicity.
Matt points out that their earlier **Chai-1** model contained 23 distinct submodules. Managing, iterating on, and scaling a system with that many moving parts made it exceptionally difficult to decipher internal dynamics or establish clear scaling vectors. By aggressively stripping away complex submodules, the team streamlined their architecture, making scaling directions vastly clearer. Furthermore, Josh clarifies that Chai builds its models completely from scratch from the ground up, explicitly avoiding shortcuts like fine-tuning general LLMs (e.g., Llama) on protein sequence data.
- **Chai-1 Complexity:** Contained 23 distinct submodules, which created friction when analyzing individual behaviors and trying to establish scaling laws.
- **Architectural Philosophy:** Emphasizes extreme simplicity over adding bespoke submodules, allowing scaling compute and data to drive feature discovery.
- **Ground-Up Modeling:** Rejects fine-tuning off-the-shelf LLMs on biological text, opting to build specialized molecular architectures from scratch.
[20:05](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1205s) Biological Verifiability and Ground-Truth Evals
Addressing the common perception that biology is an inherently messy and un-verifiable domain, Josh makes the counterintuitive argument that biological target binding is actually more objective and verifiable than software engineering benchmarks like SWE-bench or code generation. While evaluating generated code often involves subjective assessments of style, readability, or technical debt, a molecular assay provides a hard, physical, and unambiguous readout: either the molecule binds with high affinity and can be manufactured, or it cannot.
The primary difference lies in feedback latency. Code compilers return unit test results in seconds, whereas wet-lab validation requires days or weeks. Furthermore, because wet-lab measurements carry inherent noise and error bars (where a ±5% variation might simply be lab noise), model iterations cannot rely on minor incremental gains. To prove real progress, model improvements must demonstrate massive, robust step-changes that generalize across diverse target classes rather than over-fitting to narrow benchmark tasks.
- **Verifiability Comparison:** Unlike code generation (where code quality and design taste are subjective), biology yields binary, physical metrics like binding affinity and manufacturability.
- **Lab Error Bars:** Experimental wet-lab error margins mean small model fluctuations (e.g., ±5%) are statistically meaningless, requiring massive step-function improvements to prove efficacy.
- **Feedback Loops:** Evaluation cycles take days or weeks in the wet lab rather than seconds on a server, necessitating rigorous internal evaluation standards.
[22:40](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1360s) Unlocking Hard Targets via Model Scale
When evaluating difficult biological targets—such as undruggable membrane proteins or complex post-translational modifications like glycosylation sites—the team often faces the temptation to add special-purpose modules to handle specific edge cases. However, Chai's core philosophy resists adding short-term "24th modules." Instead, they rely on model scaling to force the neural network to discover intrinsic biological features on its own.
Matt explains that even if glycosylation or functional motifs are not explicitly hard-coded into the architecture, a scaling model aiming to reduce loss is naturally forced to learn the underlying sequence motifs and structural rules that signal where those modifications occur. Josh notes that "hard targets" mean completely different things across Chai's internal departments, and treating these multi-dimensional challenges in parallel creates a smooth, continuous rate of progress without single point-of-failure bottlenecks.
- **Departmental Definitions of "Hard Targets":**
- **Biology / Science Team:** Modulating traditionally "undruggable" targets, such as functional membrane-bound GPCRs.
- **Machine Learning Team:** Targets the model struggles to fold or accurately predict 3D structures for.
- **Product Team:** Properly representing complex physical modifications (e.g., glycosylations, membrane interactions) within a user software interface.
- **Feature Emergence:** By pushing loss down through compute and data scaling, models automatically learn complex biological phenomena without requiring hand-engineered architectural rules.
[25:01](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1501s) The Computer-Aided Design (CAD) Vision for Biology
Ultimately, Chai Discovery is not aiming to operate as a traditional biotechnology company focused on bringing a small pipeline of 5 proprietary drug candidates to market. Instead, their mission is to build a comprehensive Computer-Aided Design (CAD) suite for molecular engineering that permanently changes how medicines are discovered across the entire pharmaceutical industry.
Rather than relying on brute-force physical library screening or endless trial-and-error, a biological CAD suite allows researchers to specify desired molecular properties, constraints, and binding characteristics up front in software. The AI engine then directly generates engineered molecules ready for rapid lab validation, turning biology from a slow, bespoke discovery science into a predictable engineering discipline.
<ElicitationsGroup message="Explore more from this interview:">
<Elicitation label="Summarize Chai Discovery's business model and pharma partnership strategy" query="Summarize Chai Discovery's business model and pharma partnership strategy from the rest of the interview."/>
<Elicitation label="Explain how Chai Discovery's data flywheel works" query="Explain how Chai Discovery's data flywheel works based on the Sequoia podcast transcript."/>
</ElicitationsGroup>
Here is the detailed, faithful walkthrough covering the provided transcript section:
[25:48](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1548s) Faster Design Loops and the Molecular CAD Vision
The overarching goal of Chai Discovery's technology is to create a computer-aided design (CAD) suite that dramatically speeds up the cycle from initial idea to testable hypothesis. In traditional drug discovery, this iterative loop can take around nine months to discover a viable molecule. By collapsing that timeline down to nine weeks or even nine days, each order-of-magnitude reduction vastly scales the sheer volume of hypotheses and ideas a research team can test. Ultimately, this acceleration is how the field will converge on significantly better human medicines.
This speed directly intersects with the challenge of hard drug targets. The debate over whether to prioritize speed or target difficulty is a false dichotomy: if researchers can iterate through hypotheses vastly faster, previously intractable and "hard" targets become far easier to crack.
- **Design Loop Timeline:** Compressing the hypothesis-to-testing loop from 9 months down to 9 weeks or 9 days.
- **Scaling Hypothesis Testing:** Order-of-magnitude time savings directly multiply the number of molecular ideas that can be explored.
[26:32](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1592s) Future Outlook: The Drug Discovery Landscape in 2035 and Beyond
Looking forward ten to fifteen years into a future where molecular CAD suites and advanced wet-lab automation are standard practice, the pharmaceutical industry will undergo a fundamental transformation. Rather than settling for candidates that are merely "80% of the way there" just to meet clinical timelines, companies will design medicines with immense precision and intentionality. Today, suboptimal candidates are often pushed into clinical trials only to be surpassed a year later by a competitor—an inefficient use of capital. In contrast, the future promises high-specificity drugs tailored to target diseases without the harmful off-target side effects that plague current treatments.
Beyond refining existing targets, this paradigm shift brings previously inaccessible areas within reach:
- **Intractable Diseases:** Complex conditions like Alzheimer's, which have stymied researchers for decades, become solvable through rapid computational hypothesis testing.
- **Uneconomical Therapies:** Personalized medicines and rare disease treatments with smaller patient populations become economically viable when discovery costs drop and speed increases.
- **"Last-in-Class" Strategy:** The historical mindset of choosing between being "first-in-class" or "best-in-class" shifts toward being "last-in-class"—designing the definitive, final molecular solution for a disease.
[28:37](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1717s) Platform Infrastructure vs. Asset Development
Chai Discovery deliberately chose a platform infrastructure model over the full-stack drug development approach pursued by peers like Isomorphic Labs. While both models hold immense value, Chai's founders—drawing from prior experience building full-stack drug pipelines—embraced a "bitter lesson-pilled" perspective: as AI models improve, resources should be funneled back into making better models rather than being absorbed by expensive clinical trials.
This infrastructure-first business model creates a self-reinforcing financial and operational engine. High-performing models secure lucrative partnerships with established pharmaceutical companies, generating revenue that directly funds further model investment and compute scale.
- **Core Bet:** Infrastructure enablement allows Chai to impact a far broader range of medicines across the entire biotech ecosystem.
- **Avoiding Self-Deception:** Developing a small internal pipeline of 5 or 10 targets risks overfitting; a model might produce exquisite results on those few targets while failing to generalize to the broader 500 targets the industry tackles annually.
- **Tier-1 Pharma Partners:** Collaborating with industry giants like Eli Lilly, Novartis, Pfizer, and Regeneron forces absolute rigor. Because these partners rigorously test every claim, Chai cannot rely on post-hoc chemical cleanup—their shipped models must work reliably out of the box.
[31:14](https://www.youtube.com/watch?v=wv53mDmY-k0&t=1874s) Partnering Reality Check and Pharma Adoption Dynamics
Contrary to the common narrative that legacy pharmaceutical companies are slow to adopt technology or lack AI sophistication, Chai's real-world experience reveals highly rigorous, tech-forward partners. Because pharmaceutical companies face strict patent exclusivity windows, a trillion-dollar enterprise like Eli Lilly must continuously innovate and discover blockbuster drugs to maintain its valuation.
This existential commercial pressure makes pharma an aggressively competitive, innovation-driven industry. While partners perform exhaustive internal validation before deploying Chai's platform, once the performance data is verified, adoption scales rapidly. Partners frequently find creative, unanticipated ways to apply the models across their pipelines.
- **Internal "Dogfooding":** Alongside commercial partnerships, Chai maintains an internal team of scientists who actively use the models on challenging real-world tasks to identify failure modes and push capabilities.
- **The Data Flywheel:** Unlocking new model capabilities reveals fresh data needs, which in turn fuels the training of next-generation models.
[33:44](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2024s) Data Sources, Language Models, and the Sequence-Structure Synergy
In AI for biology, data compounding relies on two primary paradigms that were historically viewed as opposing approaches:
1. **Structural Data (The Protein Data Bank):** The gold standard source since the 1970s, containing experimentally determined 3D crystal structures deposited by structural biologists.
2. **Sequence Data (Language Models):** Massive databases containing hundreds of billions to trillions of biological sequence tokens—potentially outnumbering English language tokens on the internet.
Matt McPartlon initially favored a structure-based approach focused on predicting 3D spatial coordinates. Conversely, Josh Meier brought a "sequence-first" perspective from his foundational work on the Evolutionary Scale Modeling (ESM) team at Meta. Sequence language models operate by predicting masked amino acids; to successfully predict a missing amino acid, the model must implicitly learn the local 3D microenvironment and spatial geometry of the protein.
Chai intentionally avoids dogmatic adherence to either paradigm, pragmatically combining sequence and structure methodologies. High-accuracy prediction models run over massive sequence databases to generate synthetic structural data, creating a self-reinforcing data flywheel. Furthermore, experimental "exhaust" generated from Chai's internal lab experiments feeds directly back into model training.
> "I think you'd be surprised, but there might even be more biological sequence tokens on the internet than English language tokens... once the base model got good enough, then you could start to do those kinds of experiments."
[37:16](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2236s) Staying at the Frontier in a Competitive Field
As interest in AI-driven protein and small molecule design has exploded, maintaining a lead at the frontier requires rigorous evaluation rather than marketing hype. Chai focuses heavily on hill-climbing model performance metrics to ensure genuine step-function improvements over existing baselines.
Key breakthroughs demonstrated by their platform include:
- **Chai-2 Performance:** Pushed model hit rates high enough that researchers no longer need to perform massive physical library screenings to find binding molecules.
- **Direct Optimization for Developability:** Moving beyond raw binding affinity to directly model downstream drug properties like manufacturability and solubility right inside the design phase.
<ElicitationsGroup message="Explore related topics from this transcript:">
<Elicitation label="Summarize Chai-2 hit rates & scaling laws from Part 2" query="Summarize the previous transcript section covering Chai-2's hit rates, scaling laws, and de novo antibody design performance."/>
<Elicitation label="Detail Chai Discovery's culture and founding story" query="Provide details on Chai Discovery's culture, team composition, and original founding story as discussed in the Sequoia podcast."/>
</ElicitationsGroup>
[38:20](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2300s) Benchmarking Against Nature and Fine-Tuned Wet Labs
Josh Meier explains that when new technical capabilities emerge within their models, the team quickly locks them in, making them robust and production-ready to ship to pharma partners. He emphasizes a core philosophy: Chai Discovery isn't primarily competing head-to-head against rival AI model providers, but against nature itself. Nature has set a very high baseline, and traditional drug discovery methods have been honed over decades. While the team avoids adding bloat—such as a hypothetical "module 24"—to their clean AI architecture, traditional wet lab protocols have evolved by adding "module 240," resulting in highly optimized, bespoke pipelines.
To illustrate the bar they must clear, Meier highlights Andy Young, a veteran scientist on the Chai team who was among the pioneers working on yeast display at MIT two decades ago. With 20 years of experience across pharma giants like Pfizer and Genentech, as well as an approved antibody drug to his name, Young knows how to engineer high-quality antibodies using traditional methods. Meier stresses that clearing this traditional benchmark is mandatory for real-world adoption, an inflection point Chai crossed recently, sparking their major pharma partnerships. Ultimately, the ceiling for AI must be exponentially higher—they didn't found Chai just to build a "10 times faster mouse," but to design breakthrough medicines that were previously impossible.
- Traditional wet lab protocols have been tuned over decades with dozens of stacked procedural modifications.
- Andy Young bring 20+ years of biopharma experience from MIT, Pfizer, and Genentech, including an approved antibody therapeutic.
- Clearing the high benchmark set by expert human scientists and traditional protocols was the prerequisite for Chai's recent major pharma partnerships.
---
[39:44](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2384s) Culture, Scale, and the Pain of Growth
When asked about the best and worst aspects of working at Chai Discovery, co-founder Matt McPartlon points to the team's extreme mission-driven focus as the clear highlight. The dedication across the company resembles the intense, shared motivation found in top PhD laboratories. Everyone is aligned around a clear philosophy and roadmap, creating a highly motivated environment.
When reflecting on the toughest challenges, McPartlon points to the operational realities of engineering at scale. He candidly shares that when Chai launched with 128 GPUs, coming from a PhD group where four researchers shared eight GPUs, he felt "GPU rich." However, operating massive cluster infrastructure introduces complex systems problems: maintaining cluster health, keeping multi-month training runs stable, auto-resuming after node crashes, minimizing inter-node communication latency, and maximizing compute utilization. Host Pat Grady references a favorite quote from executive Frits van Slooten: leaders must choose between the "pain of failure or the pain of growth," and growth pain is always preferable.
- **PhD resource baseline:** 4 researchers sharing 8 GPUs.
- **Initial Chai setup:** 128 GPUs (which initially felt like unlimited compute).
- **Current compute scale:** Managing large-scale distributed GPU clusters that run continuous, multi-month training workflows.
---
[42:06](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2526s) Production-Grade Engineering over Blue-Sky Research
For Meier, the most fulfilling part of Chai is seeing tangible results. After working in AI and biology for a decade, he often questioned whether publishing cool academic papers was genuinely improving the world or impacting patients. Today, breakthrough model outputs, smoother software user experiences, and fresh wet lab validation data provide exhilarating proof that their platform is making a real-world difference.
The flip side of this impact is the heavy responsibility of serving major global biopharma partners who rely on Chai's platform for active drug pipelines. The team can no longer treat their codebase like a casual research sandbox where they get a paper result and move on. They must maintain strict production standards, prevent regressions, write clean code that scales across a growing team, and eliminate bugs. Co-founder Zach contextually reminds the team 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" to build sustainable, compounding architectures.
- **10-year perspective:** Transitioning from publishing academic papers to deploying models that generate real clinical candidates.
- **Production standards:** Shifting focus toward robust code bases, regression testing, and reliable software deliveries for enterprise pharma workflows.
---
[44:03](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2643s) What's in a Name? The Origins of "Chai Discovery"
When asked about the origins of the company's name, the co-founders reveal that "Chai" stands directly for **Ch**emistry and **AI**. Beyond the acronym, they admit to loving chai tea, which remains a constant fixture around their office.
Meier adds that traditional biotechnology and pharmaceutical companies often choose complex, dense, or obscure names. In contrast, Chai Discovery wanted a simple, user-friendly name that mirrored their underlying mission: simplifying and streamlining the entire drug design process.
---
[45:28](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2728s) Looking Ahead: Zero-Shot Candidates and Controllable Affinity
Looking toward the next 6 to 12 months, Meier expresses immense excitement for the real-world deployment of Chai's models across partner programs. Rather than sitting on a shelf, the models are actively designing molecules for devastating human diseases. Just a year ago, the industry could not zero-shot a candidate molecule with high confidence; today, partners are taking zero-shot computer-generated designs and preparing to advance them toward clinical trials in human patients. Reflecting on patents McPartlon generated at a previous company, Meier notes how mind-blowing it is to contemplate a future where dozens or hundreds of computer-designed Chai molecules are tested in patients.
McPartlon shares that the rapid pace of model improvement makes the long hours effortless and highly energizing. Technical goals for the immediate future focus on achieving precise, fine-grained control over molecular design parameters. Rather than simply hoping a model generates a high-affinity binder against a target, the team is unlocking the ability to explicitly specify constraints—such as requesting an exact 10-nanomolar ($10\text{ nM}$) binder—turning biological discovery into a precise engineering discipline.
- **Zero-shot transition:** Moving from theoretical predictions to generating zero-shot candidate molecules directly destined for clinical development within 12 to 24 months.
- **Engineering target:** Transitioning from probabilistic binding affinity to exact target specifications, such as designing precise $10\text{ nM}$ binders on demand.
## Notable quotes
> [06:04](https://www.youtube.com/watch?v=wv53mDmY-k0&t=364s) People were like, there was not enough data for antibodies in the Protein Data Bank to solve this problem... The holy grail was whether we could design antibody proteins on the computer, especially ones that had therapeutic function. — Matt McPartlon
> [13:23](https://www.youtube.com/watch?v=wv53mDmY-k0&t=803s) Oh, like the GPUs got too hot. I think you guys are running too many. — Hyperscaler Cloud Provider
> [13:23](https://www.youtube.com/watch?v=wv53mDmY-k0&t=803s) Isn't that the point? — Chai Discovery Team
> [33:44](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2024s) I think you'd be surprised, but there might even be more biological sequence tokens on the internet than English language tokens... once the base model got good enough, then you could start to do those kinds of experiments. — Josh Meier
> [42:06](https://www.youtube.com/watch?v=wv53mDmY-k0&t=2526s) if you want to move fast in the long term, you sometimes have to just move a little bit slower in the short term — Zach
## Mentioned
Josh Meier, Matt McPartlon, Chai Discovery, Chai-1, Chai-2, Sequoia Capital, Pat Grady, Sonali Singh, CASP, AlphaFold 2, Protein Data Bank, ChatGPT, OpenAI, GPT-1, GPT-2, Andy Young, Nathan Rollins, David Baker, Stripe, Kevin Wu, Rich Sutton, Llama, SWE-bench, Isomorphic Labs, Eli Lilly, Novartis, Pfizer, Regeneron, Meta, Evolutionary Scale Modeling, MIT, Genentech, Frits van Slooten
Chai Discovery co-founders Josh Meier and Matt McPartlon argue that therapeutic drug discovery is evolving into a predictable engineering problem driven by scaling laws, diffusion models, and radically simplified architectures. By abandoning hand-crafted heuristics and prioritizing platform software over internal drug pipelines, the startup aims to provide a molecular computer-aided design suite for the pharmaceutical industry.
Refactoring Biological Models Around the Bitter Lesson (18:18)
Chai Discovery's foundational engineering follows Rich Sutton's "Bitter Lesson," asserting that model capacity, data scale, and compute yield greater long-term progress than bespoke, domain-specific rules. The company built its earlier Chai-1 architecture around 23 distinct submodules, but the co-founders found that managing so many moving parts clouded scaling dynamics and made iterating difficult. Subsequent work stripped away these hand-crafted submodules in favor of building architectures from scratch, avoiding shortcuts such as fine-tuning generic large language models on biological tokens.
Rather than adding specialized architectural modules to capture challenging biology—such as membrane-bound G-protein coupled receptors or post-translational glycosylation—the team relies on scaling compute and objective loss reduction. According to McPartlon, a scaling neural network naturally learns the underlying sequence motifs and structural rules required to handle these edge cases without explicit hard-coding. The co-founders view biological assays as uniquely verifiable benchmarks compared to software coding tasks, because physical target binding yields unambiguous, physical readouts, even though wet-lab noise requires models to demonstrate robust step-function gains rather than minor margin improvements.
Hit-Rate Expansion and De Novo Zero-Shot Design (15:22)
When Chai was founded in 2024, the state of the art in machine learning for antibody generation suffered from hit rates near 0.1%, forcing wet labs to screen roughly 1,000 candidates to identify a single viable binder. The team's Chai-2 model pushed these success rates to approximately 15% to 16%, yielding around 150 hits per 1,000 generated molecules. This acceleration surpassed the founders' original expectations, which had anticipated reaching a 1% hit rate initially and taking three to four years to approach 20%.
This performance lift enables true zero-shot de novo candidate design, moving beyond slow, iterative lab-in-the-loop screening. The technical shift was unlocked by adopting diffusion models over variational autoencoders. While variational autoencoders struggled to compress dense 3D molecular distributions in a single pass, diffusion models iteratively de-noise structural representations, allowing the model to learn localized assembly rules that scale into high-precision, photo-realistic molecular designs.
Platform Infrastructure Versus Asset Pipelines (28:37)
Chai Discovery deliberately operates as a software infrastructure provider rather than developing an internal pipeline of proprietary drug assets. Meier and McPartlon argue that running a small internal pipeline of five to ten drug targets introduces the risk of model overfitting, giving the illusion of progress while failing across the broader set of hundreds of targets evaluated across the biopharmaceutical sector. Instead, Chai partners with enterprise pharmaceutical developers like Eli Lilly, Novartis, Pfizer, and Regeneron.
This commercial strategy is designed to create a self-funding flywheel: high-accuracy foundation models generate partner licensing revenue, which is directly funneled back into compute infrastructure and model scale. The founders note that large pharmaceutical firms operate under urgent patent-cliff pressures, making them sophisticated evaluators who rigorously validate platform capabilities before deployment. This external pressure forces the team to avoid post-hoc chemical cleanup and maintain production-grade engineering standards.
Operational Realities and Near-Term Engineering Milestones (39:44)
Transitioning foundational research into an enterprise-grade platform has surfaced significant infrastructure challenges. McPartlon notes the operational friction of moving from an academic setup where researchers shared eight GPUs to operating continuous, multi-month training runs across dense GPU clusters prone to node crashes and thermal stress. The engineering team has prioritized codebase stability, regression testing, and low-level GPU optimization to maintain high compute utilization.
Over the next 6 to 12 months, Chai plans to deploy models directly into clinical development pipelines with commercial partners, testing zero-shot computational designs in human settings. The primary technical milestone is shifting from broad affinity optimization to explicit parametric control, allowing researchers to specify deterministic design constraints—such as engineering an exact 10-nanomolar binder on demand—and compressing the drug design cycle from nine months down to weeks or days.
Chai Discovery's ambition rests on proving that computational scaling can reliably bypass decades of bespoke wet-lab trial and error. While early zero-shot results with Chai-2 demonstrate significant hit-rate improvements, the ultimate viability of this molecular CAD paradigm depends on whether computationally generated candidates achieve real-world clinical safety and therapeutic efficacy in upcoming human trials.
[6:04] Matt McPartlon: People were like, there was not enough data for antibodies in the Protein Data Bank to solve this problem... The holy grail was whether we could design antibody proteins on the computer, especially ones that had therapeutic function.
[13:23] Hyperscaler Cloud Provider: Oh, like the GPUs got too hot. I think you guys are running too many.
[13:23] Chai Discovery Team: Isn't that the point?
[33:44] Josh Meier: I think you'd be surprised, but there might even be more biological sequence tokens on the internet than English language tokens... once the base model got good enough, then you could start to do those kinds of experiments.
[42:06] Zach: if you want to move fast in the long term, you sometimes have to just move a little bit slower in the short term
Original transcript
[0:00] One of our big guiding principles is
[0:01] just simplicity. Um so like when you
[0:03] look at a model like let's say a Chai 1,
[0:05] I think there are 23 distinct submodules
[0:08] in Chai 1. Um and like when you're
[0:10] trying to iterate on something like
[0:11] that, it gets really hard cuz you're
[0:12] like I kind of I need to understand each
[0:14] of these submodules independently. I
[0:15] need to understand all of their
[0:16] behaviors, their dynamics, and like that
[0:18] doesn't actually scale that well. Um so
[0:20] like you can start to think, okay, how
[0:22] do I simplify this? How do I identify
[0:23] like what's what's really important? And
[0:25] once you have that like kind of the
[0:26] whole research process and like
[0:28] identifying these types of scaling
[0:29] directions uh becomes a lot simpler.
[0:33] >> [music]
[0:48] >> We're here with Josh and Matt, two of
[0:51] the co-founders of Chai Discovery. Chai
[0:53] is engineering molecules with AI. So
[0:57] it's something of a foundation model lab
[0:58] for biology. Sounds like a big idea, but
[1:01] let me just start there. What is the big
[1:03] idea?
[1:04] >> Thanks for having us on on the show. Um
[1:07] one of the exciting things we're trying
[1:08] to do at Chai is to make the drug
[1:11] discovery process like a little bit more
[1:12] like engineering. And we've seen all the
[1:14] work happening with LLMs for code
[1:16] generation for instance, right? And that
[1:18] works really well because code's a very
[1:19] simple abstraction, right? Of like to to
[1:21] kind of get across what you want to do.
[1:23] Uh biology doesn't look like that today.
[1:24] It's a lot of like trial and error.
[1:26] Um I think if we go back to uh the early
[1:29] days of uh of just like modern biotech,
[1:31] right? A lot of the medicines that we
[1:33] have today were kind of discovered quite
[1:34] randomly. It was a bit serendipitous.
[1:36] And what we're trying to do is allow us
[1:38] to industrialize that process a little
[1:39] bit more
[1:41] uh and try to come up with the tools
[1:42] that we'd need in order to have
[1:45] abstraction layers the same way we have
[1:46] that in code and in modern engineering
[1:47] allows us to iterate really quickly and
[1:49] try to bring that uh into into biology
[1:51] and into drug discovery.
[1:52] >> So So let me ask you a question on that
[1:54] then. So there's And tell me if this is
[1:56] a reasonable way to frame it. There's
[1:58] almost this boundary between that which
[1:59] can be engineered
[2:01] and that which needs to be tested in the
[2:03] real world.
[2:04] And it feels like that boundary has been
[2:06] moving over time. Like the the the
[2:08] proportion of the drug development
[2:10] process that can be engineered as
[2:12] opposed to trial and error'd seems to be
[2:14] increasing. Is that a reasonable way to
[2:16] think about it? If so, can you talk
[2:18] about like what specific developments
[2:20] have pushed that boundary over time?
[2:22] >> I think that's one way to think about
[2:23] it. Um the way that we often uh look at
[2:26] this is like the lab is is an important
[2:28] part to to verify that what you're doing
[2:30] is correct. And actually verification is
[2:31] a very big theme in AI as well. If you
[2:33] can evaluate that your model works, you
[2:35] can verify it, then you can start to
[2:37] hill climb that and you can make
[2:38] progress on it. So I think the lab is a
[2:39] very important part of this. And then
[2:41] the question is how do we take drug
[2:42] discovery and make it look a lot more
[2:44] like drug design. So one of the reasons
[2:46] why we call this field drug discovery is
[2:47] we're often looking for a needle in a
[2:49] haystack. We'll screen millions,
[2:50] billions of molecules try to find one
[2:52] that works. And if we can instead put in
[2:54] what is like the dream state of the
[2:56] molecule you want and then the model can
[2:58] materialize that, that's going to be
[2:59] really powerful. So it's not even a
[3:01] question of like reducing lab testing. I
[3:03] mean, maybe that happens as a result. I
[3:05] actually might even take the opposite
[3:06] side of the coin and we can talk about
[3:07] that where maybe we'd actually do even
[3:09] more lab testing cuz the ROI will
[3:10] increase. The same way there's more
[3:12] software engineers
[3:13] or there's more demand for software
[3:14] engineers now that they become more
[3:15] productive, there might be more demand
[3:17] for the lab.
[3:18] But I think the key part is how do we
[3:20] just change the paradigm here and how do
[3:22] we we make it more design oriented?
[3:24] >> Mhm.
[3:25] Can we talk about what is state of the
[3:27] art today?
[3:29] And then maybe let's take a little trip
[3:30] down memory lane. 5 years ago, 3 years
[3:32] ago, 1 year ago. Like what have been
[3:35] some of the major breakthroughs and like
[3:37] how has state of the art changed over
[3:39] the recent years?
[3:40] >> Yeah, I think the the field has like
[3:42] it's really evolved especially like over
[3:44] the last decade. Um so it really wasn't
[3:46] until like uh you can actually look
[3:48] back. There's like this bi-annual
[3:50] protein folding competition. So every 2
[3:52] years uh
[3:53] usually a bunch of academic groups would
[3:55] compete in this protein folding
[3:56] competition. What would happen is like
[3:58] you kind of hold out these protein
[3:59] targets that no one's ever seen before.
[4:01] They never get deposited publicly. And
[4:03] then all these groups compete so you can
[4:04] predict the proteins the best. And like
[4:07] really it wasn't until 2018 where you
[4:10] started to see this like big step change
[4:11] in in performance and then finally again
[4:13] in like 2020 with AlphaFold 2. But
[4:16] really like it was kind of like the
[4:17] advent of deep learning that really like
[4:19] kicked off this whole field. Um so at
[4:21] first what you do
[4:23] these are kind of like bring me back to
[4:24] my original days in my advisor's lab. So
[4:27] we worked on like one of the first
[4:28] systems to do this with deep learning.
[4:30] What we would try to do is predict like
[4:31] the distances between amino acids in a
[4:33] protein. And then we'd have some render
[4:35] which then took in like these kind of
[4:37] noisy incomplete looking distances and
[4:39] then emitted some protein from that. Um
[4:41] and what's pretty remarkable is that
[4:43] with
[4:44] relatively minimal information like you
[4:46] could build a machine learning system
[4:48] that could actually produce something
[4:49] that really looks like a protein and for
[4:51] the most part was was correct. Like
[4:54] there are still pretty big gaps it
[4:55] didn't quite have the resolution of like
[4:57] what what we have today. Almost like
[4:59] with the early image generation models
[5:01] where it was like a little bit grainy a
[5:02] little bit blurry and now it's like oh
[5:04] my gosh like this is this is crazy high
[5:06] definition images like in seconds. Uh so
[5:08] we kind of like saw that same evolution
[5:10] happen over time. So it started with
[5:12] protein folding and then I guess once
[5:14] that became more realizable
[5:17] there were kind of these other sub
[5:18] problems that people wanted to solve. So
[5:20] now given a protein structure can I
[5:21] design say like a sequence which might
[5:23] fold to that? And that's getting closer
[5:25] and closer to like the drug design
[5:27] problem. So you're you're kind of like
[5:29] now thinking okay
[5:31] how do I start engineering proteins that
[5:32] like match a certain shape or might
[5:34] perform a specific function?
[5:36] And so people kind of study that problem
[5:38] independently and that was kind of like
[5:40] 2021 2022 time. And then kind of these
[5:43] ideas began to merge together. It was
[5:45] really like the advent of diffusion
[5:46] models where we started to be able to
[5:47] like, "Okay, I can now generate a
[5:49] protein structure and a sequence kind of
[5:51] simultaneously.
[5:52] Um and I can start making these like
[5:54] what we would call prompts more and more
[5:56] realistic." So, I can now prompt these
[5:57] models on a target that might have a
[5:59] certain shape and I can say, "Hey, I
[6:00] want to bind over here." Kind of like
[6:02] add add more real-world constraints on
[6:04] the problem.
[6:04] >> What did you guys see in 2024 that made
[6:06] you think that was the right time to
[6:08] start the company?
[6:09] >> Yeah, there were a lot of discussions
[6:10] that went into it. I remember one of
[6:12] these uh early discussions actually back
[6:14] in Matt's house when we were uh you
[6:16] know, Matt was was showing me some of
[6:18] these results on like antibody antigen
[6:20] like structure prediction. And we Matt
[6:21] was just talking about protein folding.
[6:23] Um but for for a long time people
[6:24] thought that protein folding for
[6:26] antibodies was just like too hard of a
[6:27] problem. People were like there was not
[6:29] enough data for for any in antibodies
[6:31] like in the protein data bank for
[6:32] instance to like solve this problem. And
[6:34] people as a result thought that like
[6:35] antibody design was going to be out of
[6:37] reach. A lot of the work that people
[6:39] even did on protein design in the early
[6:40] days with deep learning, it wasn't
[6:42] antibodies. It was uh these different
[6:43] class of proteins called mini proteins,
[6:46] um which are uh actually really
[6:47] interesting in their own right, but
[6:48] they're not what most of the uh drug
[6:50] industry is is looking at. So, the holy
[6:52] grail was whether we could design
[6:54] antibody proteins on the computer,
[6:55] especially ones that had all that
[6:56] therapeutic function. And our thinking
[6:58] was that if you couldn't predict what an
[7:00] antibody looks like, how are you ever
[7:02] going to design one?
[7:03] >> Traditionally, people have thought like
[7:05] like biology is scary. Um like there's
[7:07] so much to know.
[7:08] >> of biology.
[7:09] >> I I am as well. Honestly, [laughter]
[7:11] like
[7:12] uh like my my background was was never
[7:14] biology. I I studied like pure math and
[7:16] started my PhD in uh theoretical
[7:18] computer science. And it was only like
[7:20] after my third year that I ended up
[7:21] switching into like deep learning,
[7:23] protein structure prediction, all of
[7:24] this stuff. So, it was like totally new
[7:26] field to me. Seemed insane, but like at
[7:28] the end of the day, it's much simpler
[7:29] and like the the problems are much more
[7:31] like interconnected uh than one might
[7:33] think. So, like
[7:34] I think people like when they start
[7:36] getting into the field, they're like,
[7:37] "Oh man, what's an antibody? What's a
[7:38] mini protein? What are These are all
[7:40] just like sequences of amino acids." At
[7:42] the end of the day, like these are just
[7:43] like different types of prompts for the
[7:45] model. Um but like in the same way where
[7:47] you might like have a math problem that
[7:48] goes into chat GBT, chat GBT can both
[7:51] answer your math math problem and like
[7:53] help you with your English homework. Uh
[7:55] so like really we we have the same type
[7:56] of thing going on with our models. Like
[7:58] we just have some way of representing
[8:00] these sequences of amino acids, then we
[8:01] have a way of like designing, predicting
[8:03] those as well. Um and I think like in
[8:05] that lens, like things things become a
[8:07] lot more clear.
[8:09] >> And Matt, you mentioned your background
[8:10] a little bit. Um Josh, talk a bit about
[8:12] your background and then more broadly,
[8:14] in order to pull this off, you have a
[8:15] bunch of different disciplines that kind
[8:17] of come together. So can you just talk a
[8:19] bit about like your background and that
[8:21] of some of the other core members of the
[8:23] team and how these things all fit
[8:24] together?
[8:25] >> Yeah, I've been excited about AI and
[8:27] biology uh since I was a kid. So I guess
[8:29] I'm like Matt Matt, I didn't start with
[8:31] like theoretical CS and get into that
[8:33] way, but um I went to a high school with
[8:35] a stem cell lab. So I was just always
[8:36] excited about biology as a kid and I
[8:38] grew up as a programmer. Um I uh I
[8:41] really started my career at Open AI, so
[8:43] it was on on the early team there. It
[8:44] was a non-profit back then. So it was a
[8:46] a pretty uh good time to be there. We
[8:48] did GPT-1, GPT-2, scaling laws. Um and
[8:53] it was uh the question was like if the
[8:54] models can learn to speak English,
[8:56] German, French, why can't they learn to
[8:57] speak DNA and protein? Um and that was
[9:00] kind of my research agenda since then. I
[9:02] think that sort of intersects with
[9:03] around the time like Matt got into the
[9:04] field as well. And it was a uh I think
[9:07] it was a pretty important time as well,
[9:09] right? Because if you look at the kind
[9:10] of methods that we were bringing in,
[9:11] like there's been a lot of these, you
[9:13] know, changes on the edges if you will,
[9:14] right? And you know, Matt talked a bit
[9:16] about the history of what's happened in
[9:18] the fields here. Um but you know, it's
[9:20] all about like how do we find like the
[9:21] right deep learning architectures with
[9:23] the right compute configuration and the
[9:25] right model architectures this happen.
[9:27] What are the right tasks to apply it to?
[9:29] We're talking about how we even knew
[9:30] that like, you know, 2024 is the right
[9:32] time to start the company. As Matt was
[9:34] saying, you know, these are all
[9:34] different like kinds of amino acid
[9:36] sequences and people thought that, you
[9:38] know, antibody class of of problem was
[9:40] going to be too hard and we started to
[9:42] see the first signs of life that
[9:43] actually this was starting to work. I
[9:44] think a lot of it fueled by some of the
[9:47] new architectures we're bringing to the
[9:49] problems. I think it was diffusion
[9:50] models back then.
[9:51] >> The first time like anyone was able to
[9:53] generate reasonable looking proteins
[9:56] was the advent like of diffusion models.
[9:59] So it was pretty crazy like there were a
[10:00] bunch of generative modeling approaches
[10:02] that like would kind of work if you if
[10:03] you had a bunch of data. So like people
[10:05] got these working for images. There were
[10:07] like a bunch of tricks to make this
[10:08] better and better along the way. Like
[10:10] we've definitely borrowed a lot of those
[10:11] ideas in our domain as well.
[10:13] But it was really like once diffusion
[10:15] models came around
[10:16] >> And is there is there an intuition for
[10:17] why diffusion models work?
[10:19] >> Yeah, so diffusion is is not a one-step
[10:22] process. So like I think kind of up
[10:24] until this point like the the main
[10:26] generative design paradigm was like
[10:29] called variational autoencoders. And in
[10:31] that case you're like saying I want to
[10:33] just like compress my input
[10:34] distribution. So like you you have some
[10:36] like proteins you want to make these
[10:38] look like kind of like fuzzy Gaussian
[10:40] vectors. That task is just like really
[10:42] hard. And maybe today if we tried like
[10:45] super hard I think we'd crack it. But at
[10:47] the time like I I think like it wasn't
[10:49] really as developed enough to to work
[10:52] for our problems. What turned out
[10:53] working really well was just kind of
[10:54] giving the model more time to think and
[10:57] showing it like more examples like
[10:58] here's like a slightly broken looking
[11:00] protein. How do you make it better? And
[11:02] you can kind of break that protein more
[11:03] and more and more and you can make it
[11:05] look more and more noisy, more and more
[11:06] broken and teach the model just quick
[11:08] little shortcuts. All right, here's how
[11:10] I make it slightly better. You can just
[11:11] keep asking over and over again. Make it
[11:13] slightly better. Make it slightly
[11:14] better. And kind of like breaking the
[11:16] problem down to that scale worked really
[11:18] really well for for biology.
[11:20] >> Yeah, the make it slightly better
[11:21] reminds me of a game that I like to play
[11:22] with my daughters where we have chat GPT
[11:25] give us a unicorn and then we make it
[11:27] stronger. And we just keep telling it to
[11:29] make it stronger. And by the time we're
[11:30] done we have the strongest unicorn in
[11:32] the world. So, about the same, right?
[11:35] >> Yeah. That's uh
[11:36] >> [laughter]
[11:37] >> It's that easy.
[11:38] >> Yeah.
[11:39] >> Okay, maybe not the same.
[11:40] >> Yeah. So, we feel like in this domain,
[11:43] you need to assemble a kind of like a
[11:44] quadrilingual group of people, like an
[11:47] Avengers squad of chemistry people,
[11:50] biologists, AI folks. And so, that's a
[11:52] challenge. Um how have you guys gone
[11:54] about finding people, convincing people
[11:56] to join the team, and and who are your
[11:57] superstars?
[11:58] >> So, we've been really pragmatic about
[12:00] this at Character. If you look at the
[12:01] founding team, it was mostly AI
[12:02] researchers. Um so, people who had uh
[12:05] worked on either scaling models or
[12:07] getting them to work in this domain.
[12:09] But, really with each generation of
[12:10] model, uh the kind of people we've
[12:12] needed for the next milestone has has
[12:13] changed. Uh or or I'd say probably has
[12:16] expanded, right? Uh so, if you look at
[12:18] Character 2, right? That's the point
[12:19] when we started to bring in um some of
[12:21] the most incredible like, you know,
[12:23] antibody engineers and and scientists in
[12:25] the world. Um
[12:26] uh one of the scientists from Andy Andy
[12:28] Young Actually, when we hired him, uh
[12:30] people asked us if we had pivoted into
[12:32] building a full-stack drug pipeline cuz
[12:33] they're like, "You'd be crazy not
[12:34] [laughter] to do that if like Andy
[12:35] joined your team." Um but, uh I think
[12:38] Andy has has enjoyed running more
[12:40] antibody campaigns in the past couple of
[12:41] months than he's probably run in his
[12:42] whole whole career, uh which is very
[12:44] cool to see. You have folks like Nathan
[12:46] Rollins on the team. Like, Nathan was
[12:47] actually homeschooled and then started
[12:49] college very early on. So, he joined
[12:51] like David Baker's lab who won the Nobel
[12:52] Prize for protein design like when he
[12:53] was 14, uh started his PhD when he was
[12:56] 18, and has so many creative ideas. On
[12:58] our uh as the model started to get
[13:00] better, we needed to build up a product
[13:01] team uh because while the researchers
[13:04] might get the models to point that
[13:05] they're uh they're very powerful, you
[13:07] need to build the right product
[13:08] interfaces so that the models are
[13:09] actually useful. Um and uh that's where
[13:12] we started to bring on uh people who
[13:14] have, you know, built some of the most
[13:16] exciting products we know about today.
[13:17] Uh like our co-founder Jack worked at
[13:19] Stripe, Monaz who was one of the top 10
[13:20] code contributors at Stripe, uh Neil who
[13:23] ran his own cybersecurity company before
[13:25] security started to become very
[13:26] important as we deploy this to our big
[13:27] partners. As as we started to scale up,
[13:30] we brought in people uh who've
[13:32] uh really done a lot of like the GPU
[13:34] hacking, if you will, in order to like
[13:35] scale up our systems that they they they
[13:37] don't break when we're running them at
[13:38] scale. We had an email from uh or a
[13:40] Slack message from one of our
[13:42] hyperscalers the other day uh where you
[13:44] know we had a cluster, I think that had
[13:45] some issue to it and they're like, "Oh,
[13:46] like the GPUs got too hot. I think you
[13:47] guys are running too many." And we're
[13:48] like, "Isn't that the point?" Right?
[13:50] That's probably I was like, "Good, we're
[13:51] doing our job at least." Yeah, so like
[13:53] how do you convince those people to
[13:54] join? I think again a lot of it comes
[13:55] back to uh the results and like a clear
[13:57] need. We didn't hire antibody engineers
[13:59] before in an antibody design model. Like
[14:01] what are those folks going to do? We
[14:02] were even I think worried when we
[14:03] started that trend because for some of
[14:05] the um next generation formats, the
[14:07] complicated antibodies, like
[14:09] multispecifics, like they didn't even
[14:10] work with Chai too. Uh so it actually
[14:12] took a couple of of uh weeks when some
[14:15] of those people showed up before the
[14:16] models could work at a point that they
[14:17] could work on some of these interesting
[14:18] case studies. Uh but fortunately
[14:20] progress is fast enough to kind of bring
[14:21] that online. So I think we're always
[14:23] like evolving that team and going
[14:24] forward you know, that next milestone.
[14:26] We've got the team very small as a
[14:27] result, too. Um so this way, you know,
[14:29] everyone is a little bit like slightly
[14:31] over capacity, I think, which means we
[14:33] have to prioritize. It forces us to work
[14:35] on the things that really matter most.
[14:37] >> I think on the research side as well,
[14:38] one of the founding engineers, Kevin Wu,
[14:40] he had the first uh
[14:42] I think it was the first protein
[14:44] diffusion model like ever. Uh and that
[14:46] speaks to Kevin's speed of execution.
[14:48] Like he is a heck of an engineer and
[14:50] like I think engineering has just always
[14:52] been like important since day one. Um so
[14:55] like really even our researchers,
[14:56] they're all excellent engineers and like
[14:58] we really care about building a
[15:00] high-quality codebase. Like at the end
[15:01] of the day, we are
[15:03] technically like a software company.
[15:05] We're we're AI researchers, we're we're
[15:07] protein designers, we're we're a lot of
[15:09] things, but uh our deliverable is some
[15:12] piece of software. So we've kept that
[15:13] like kept that bar really high while
[15:15] also trying to like, you know, level
[15:17] that with great research talent and
[15:19] people that can actually like push the
[15:20] frontiers of what's possible.
[15:22] >> And you've had a number of it seems like
[15:23] aha moments in the field. Like AlphaFold
[15:26] was an aha, we can figure out how a
[15:27] protein folds. And then the diffusion
[15:28] models, aha, like we can generate
[15:30] proteins. Um and it seems like your
[15:32] latest models are a kind of another aha
[15:35] moment. We're not only generating
[15:37] molecules that look like proteins, but
[15:39] they also have therapeutic properties.
[15:41] Like they can bind really tightly.
[15:44] Um they have really high hit rates. Can
[15:45] you talk sort of about the quality of
[15:47] the molecules that your models are
[15:48] producing now and everything that sort
[15:50] of went in to those models to make them
[15:52] able to do that?
[15:53] >> Yeah, if we look at the the quality of
[15:55] the molecules that's coming out, it goes
[15:57] back to, you know, one of the thesis
[15:58] when we started the company that uh we
[16:00] really wanted to focus in on this like
[16:02] de novo generation of the molecules. If
[16:05] you look at what a lot of the drug
[16:06] discovery AI work was at the time, it
[16:08] was about how do I take a molecule and
[16:09] just make it a little bit better, right?
[16:11] Which we just talked about in a sense,
[16:12] but it was doing it with a lab in the
[16:14] loop uh style, right? Where I take some
[16:16] data, try to make it better that way.
[16:18] And the the question uh
[16:20] we started with is, well, can we
[16:22] actually just like do all that on the
[16:23] computer, right? Is there a way that the
[16:25] there maybe there would be enough data
[16:26] or we could collect enough data um so
[16:28] that we could just zero-shot a molecule
[16:30] that has a lot of these properties? So,
[16:32] the first thing we needed to do to get
[16:34] there um was to design molecules with
[16:36] really high success rates. When we
[16:37] started the company, the state of the
[16:39] art for anybody design was about like a
[16:40] 0.1% binding rate. So, one in a thousand
[16:43] of the molecules you design would
[16:45] actually bind in the lab. Um so, first
[16:47] of all, that means you have to screen a
[16:48] lot of molecules to find some good ones.
[16:49] It also means that like the gradient you
[16:51] get on your process is actually quite
[16:53] weak as well. So, for many targets you
[16:55] won't get any hits. For the ones that,
[16:57] you know, you do get some hits, you
[16:58] won't have enough to actually see
[16:59] whether you're having like the drug-like
[17:01] properties. So, we really focused in on
[17:03] how do we just make this process more
[17:04] accurate? We got to uh with our chi-2
[17:06] model about like a 15% success rate. Uh
[17:09] so, now if you screen a a thousand
[17:10] molecules, you're getting 150 back. Now,
[17:12] you can get start to get some like
[17:13] interesting statistics on the properties
[17:14] of the molecules, right? And a lot of
[17:16] allowed us to like iterate on that,
[17:17] build the right evaluations around that
[17:19] in the lab, and really try to hill climb
[17:21] that as well. And we're getting to a
[17:22] point now where we can actually bake in
[17:24] a lot of these different properties from
[17:26] the start into this engine. And then
[17:28] maybe if you think about, you know, how
[17:29] does this happen or like how are we
[17:31] approaching this as well? And like why
[17:32] do we think it's going to continue
[17:33] improving?
[17:34] >> Yeah, so like um honestly, I think Josh
[17:37] and I were both surprised at how quickly
[17:39] this worked. Um like when we were
[17:41] originally budging this, we're like,
[17:42] "Ah, maybe like 20% hit rate like 3 or 4
[17:45] years." We're like really like a 1%
[17:47] rate. We thought 1 in 100 would be
[17:48] amazing. We're like, "This is going to
[17:49] be you know, a groundbreaking thing." We
[17:51] had a philosophy and like an approach
[17:53] that we wanted to take and it just like
[17:55] ended up working really well. And that
[17:57] approach is like very similar to what's
[17:58] worked in the rest of machine learning.
[18:00] So, people kind of treat biology as this
[18:01] like bespoke problem or like bespoke
[18:03] field, but really it's like the same
[18:05] principles as like self-driving LLMs.
[18:07] >> Well, same principles, but one of the
[18:09] things we've talked about before is how
[18:10] you managed to find scaling laws.
[18:14] And it's one thing to tokenize a string
[18:15] of text. It's another thing to tokenize
[18:18] biology.
[18:19] Can you say a couple words about
[18:22] without giving away any of the magic,
[18:24] you know, can you just say a couple
[18:25] words about that challenge and how you
[18:27] guys solve that?
[18:28] >> Yeah, so one thing that I really like
[18:29] about Chai is like we're we're a very
[18:31] bitter lesson pill company. So, like we
[18:32] really believe in like scaling data,
[18:34] scaling models, scaling compute. In
[18:36] order to do that, obviously you you need
[18:37] to identify scaling laws. Otherwise,
[18:39] you're just kind of like wasting time
[18:40] and resources. Um and I think like
[18:43] without giving away too much, like one
[18:44] of our big guiding principles is just
[18:45] simplicity. Um so, like when you look at
[18:47] a model like let's say a Chai 1, um so
[18:50] like there are like I think there are 23
[18:53] distinct submodules in Chai 1. Um and
[18:55] like when you're trying to iterate on
[18:56] something like that, it gets really
[18:57] hard. Uh cuz you're like, I kind of need
[18:59] to understand each of these submodules
[19:01] independently. I need to understand all
[19:02] of their behaviors, their dynamics, and
[19:04] like that doesn't actually scale that
[19:06] well. Um so, like you can start to
[19:07] think, "Okay, how do I simplify this?
[19:09] How do I identify I what's what's really
[19:10] important?" Um and once you have that
[19:13] like kind of the whole research process
[19:14] and like identifying these types of
[19:16] scaling directions uh becomes a lot
[19:18] simpler.
[19:18] >> One of the other things too that I think
[19:19] is interesting about this is we take a
[19:21] lot of these lessons from what's worked
[19:22] in the rest of the deep learning space,
[19:24] but as you're pointing out like the data
[19:25] itself is different. Yeah. The models at
[19:28] Chai are completely built from scratch.
[19:29] We're not like fine-tuning GLAM or
[19:31] something like that on some protein
[19:32] data. We build everything from the
[19:33] ground up. I think a lot of the company
[19:36] building process though is is taking a
[19:37] philosophy and actually sticking with it
[19:39] and iterating on that and just having
[19:41] some guiding principles. When you're
[19:43] building like a blue sky research
[19:44] company if you will, you know, Chai is
[19:46] is almost like one of these neo labs,
[19:47] right? Like we have this big AI problem
[19:48] we're going after where as we make
[19:50] progress on it, you know, that opens up
[19:52] opportunity for for our customers. Um,
[19:54] but if you're going to work on something
[19:55] so open-ended that way, you need some
[19:57] principles to guide you. And I think
[19:59] we've done a very good job on like
[20:00] tracking those principles in the
[20:01] company, working on it. So a lot of the
[20:03] things that Matt
[20:03] >> Yeah, can you share them? What what are
[20:04] the guiding principles?
[20:05] >> So I think simplicity was one of the
[20:07] ones that Matt mentioned. It's this like
[20:08] bitter lesson pillness of like, you
[20:10] know, scaling compute and data and and
[20:12] models. Uh, it's being really rigorous.
[20:14] Uh, that's something that's so important
[20:15] in this space. You can fool yourself so
[20:17] easily in biology. Like the error bars
[20:19] in the wet lab are actually quite large
[20:21] as well. Um, so it's actually a little
[20:23] bit different than uh if you look at
[20:25] like code generation for instance. If
[20:27] you look at SWE bench, people are like,
[20:28] oh, I like this is a maybe like a year
[20:30] ago people like, oh, I got 16% and 17%
[20:32] and 18%. I mean, in biology if if you're
[20:35] like plus or minus 5% in your lab, like
[20:37] that's all that might all be the same.
[20:39] So it actually just means uh the bar is
[20:41] really high in terms of the step changes
[20:42] that you want to see with the models,
[20:44] but you also need to be really honest
[20:45] with yourself about whether you're
[20:46] making progress or not. Um, so you could
[20:48] come up with some fancy model that looks
[20:50] like it works well on like one or two
[20:51] new tasks, but it's very important to
[20:54] show that that works more generally. Uh,
[20:55] if you're actually trying to build a
[20:56] product that can like bring the field
[20:58] forward.
[20:59] >> Yeah, and that's I think pretty
[21:00] interesting because biology is one of
[21:01] those inherently not so verifiable
[21:03] domains. And you guys have been really
[21:06] good at um, sort of showing your
[21:08] progress to customers,
[21:10] um, and to people like us who know very
[21:12] little about biology. And so, can we
[21:13] talk a little bit about the evals and
[21:16] the verifiable part of of the model
[21:18] progress? Like, how do you guys know
[21:20] that your models are getting better?
[21:21] >> Well, I would say first of all that, uh,
[21:23] I actually think this is one of the more
[21:24] verifiable domains. It's actually very
[21:26] objective readout. If you look at
[21:28] something even like CodeGen, right?
[21:30] Like, maybe it's a verifiable task like,
[21:32] did my code compile? Did it solve these
[21:34] unit tests? But, how do you think about
[21:35] the taste, right? Like, did I write some
[21:37] really sloppy code that can't be
[21:38] maintained? Like, what does that look
[21:39] like? When we think about designing a
[21:42] molecule in the lab, uh, we can actually
[21:44] be quite specific about many of these
[21:46] properties, right? So, maybe we get a
[21:47] molecule that binds the target, but can
[21:48] I manufacture it, right? That might be
[21:50] your, you know, version of like some
[21:51] tech debt, but you can measure that. Uh,
[21:53] and I think those evals, again, actually
[21:55] make this domain more verifiable. Maybe
[21:57] it takes a little bit longer to validate
[21:58] it, right? It's not like five seconds to
[22:00] like get a readout and run a unit test.
[22:02] You might have to spend a couple days, a
[22:03] couple weeks in the lab to get that
[22:04] readout, but at least you can be honest
[22:06] with yourself.
[22:07] >> Yeah. When we talk about progress and
[22:09] how good the models are, there's like a
[22:10] domain of targets in biology,
[22:13] um, that you can, as you mentioned, sort
[22:15] of address with traditional screening
[22:16] methods. They take a long time. They're
[22:18] very slow and rudimentary. Um, and then
[22:20] there are targets that just aren't
[22:22] addressable with the with existing
[22:23] methods. They're not druggable for
[22:25] whatever reason.
[22:26] Um, and so, where are we in terms of
[22:28] model progress in terms of, you know,
[22:30] working on existing targets and
[22:32] generating molecules faster? That's one
[22:34] end of the spectrum. And on the other
[22:35] end of the spectrum is unlocking novel
[22:37] biology, new targets, and things that we
[22:38] couldn't drug before.
[22:40] >> This kind of goes back to like, uh, kind
[22:42] of like our core modeling philosophy.
[22:44] Um, so like, there there have actually
[22:45] been like plenty of times where like,
[22:47] man, if we had a 24th module, like we
[22:49] can actually like unlock that new
[22:51] target. Um, and we're like, is that
[22:53] really something that we want to
[22:54] maintain long term? Is this incremental
[22:55] or is this like actually a compounding
[22:57] improvement? Will this actually like
[22:58] help us generalize to the the class of
[23:00] like these this whole class of targets
[23:02] that we we really can't hit. And so like
[23:04] our our philosophy has been like, all
[23:05] right, let's just like continue to focus
[23:07] like identify your scaling laws. Like
[23:09] there there comes a point where like if
[23:11] the model is able to push loss down even
[23:13] further, it has to understand like
[23:15] something like very intrinsic about the
[23:18] target that it's operating on. So like
[23:20] one example we were talking about the
[23:21] other day, Paul and I, maybe the way
[23:23] that we're looking at like certain
[23:24] glycosylation sites on proteins we're
[23:26] like, "Oh, we might want to represent
[23:28] them differently or something like
[23:29] this." And we're like, "Well, even if we
[23:30] didn't represent them, like there are
[23:32] certain sequence motifs that will tell
[23:33] the model like there should be a
[23:35] glycosylation site here. And in order to
[23:36] drive like loss down further, the model
[23:38] should just like have to learn that." Uh
[23:40] so there are like all these hidden
[23:42] features of targets where like if you
[23:44] really believe in scaling laws you can
[23:45] believe the models will get there. Like
[23:47] these types of targets like should just
[23:49] unlock with better models. And of course
[23:51] like you still have to take this very
[23:52] seriously and like you still need all
[23:53] the proper validation. Um you need to
[23:55] like really challenge yourself and like
[23:57] make sure that this is truly working.
[23:59] But I think our approach has always been
[24:01] like, you know, we with better models
[24:03] like we we should be able to unlock a
[24:05] lot of these targets.
[24:06] >> One of the other interesting things is
[24:07] if we look at we talked about like the
[24:08] interdisciplinary nature of this. If we
[24:10] look at the different teams at Chai,
[24:12] what people will call hard target is
[24:14] actually different in literally every
[24:15] team. So on the science team it might
[24:17] be, you know, like a undruggable GPCR
[24:20] target no one's gotten something that
[24:21] has like, you know, modulated that in a
[24:23] functional way. On the ML research team
[24:25] it'll be something like, "Oh, there's
[24:26] like the the model just can't fold this
[24:28] thing up. It doesn't know what it looks
[24:29] like."
[24:30] And then on the product team it might
[24:31] look like, "Oh, I've got all these like,
[24:32] you know, modifications like my
[24:33] glycosylations and it's a membrane
[24:35] protein. How do I represent that to the
[24:36] user?"
[24:38] And actually the fact that it's
[24:39] different for each of these groups I
[24:40] think is a feature rather than a bug.
[24:42] And it means that if we want to make
[24:44] like broad progress over here, everyone
[24:46] is kind of pushing in parallel on these
[24:47] different ways and and that means that
[24:49] there's very smooth progress that we can
[24:50] make all the time. And there's usually
[24:52] not one bottleneck at Chai. It's not
[24:53] like, "Oh, if we only had that one extra
[24:55] module things would work." Or if we only
[24:58] you know, like try to push this into the
[24:59] product in some way, we can unlock it.
[25:01] We're trying to build this unified
[25:03] solution because at the end of the day,
[25:04] the goal of the company is to build a
[25:05] computer-aided design suite for
[25:06] molecules. It's not to make one or two
[25:08] molecules. It's not to get a pipeline of
[25:10] like five interesting therapies that we
[25:12] bring to market. It's to change the way
[25:13] that medicines are discovered. And we're
[25:15] if we're going to do that, we need to
[25:17] work on all the hard targets regardless
[25:19] of how you define hard.
[25:20] >> Yeah. Say more about this idea of
[25:21] computer-aided design suite for
[25:23] molecules. What does that mean?
[25:24] >> So at its core, it goes back to this
[25:26] point about making biology look more
[25:28] like an engineering discipline. So we're
[25:29] not going and fishing something out of a
[25:31] large library or doing a ton of trial
[25:33] and error. You want to be able to
[25:35] specify up front the principles that go
[25:37] into designing your molecule and then
[25:39] have an engine that can actually realize
[25:41] that into some molecule that we're going
[25:42] to go and create in the lab. And look,
[25:44] you still might do some iteration on the
[25:46] lab and and on your model because maybe
[25:48] your hypothesis was wrong. But we want
[25:50] to speed up is is actually again have
[25:52] that computer-aided design suite so that
[25:54] you can go from idea to testable
[25:56] hypothesis very quickly. And if that
[25:58] loop right now takes something like 9
[26:00] months, I don't know, to go and like
[26:02] discover your molecule versus if it
[26:03] takes 9 weeks or it takes 9 days, you
[26:05] know, each order of magnitude just
[26:07] scales in a very big way the number of
[26:09] ideas you can really sort through. And I
[26:11] think that's ultimately how the field is
[26:13] going to converge on on better
[26:14] medicines. It really comes back to like
[26:16] people sometimes talk about do we care
[26:18] and you kind of noted on it's not like
[26:19] do we care about like speed or do we
[26:21] care about like the difficulty of the
[26:22] targets? At some point they converge in
[26:25] this way as well because a hard target
[26:27] if we can like iterate through
[26:28] hypothesis a lot faster, then maybe
[26:30] it'll be easier to crack it.
[26:32] >> If so if we can maybe detach ourselves
[26:35] from the present reality and go far
[26:37] enough into the future that we're not
[26:39] thinking present forward, we're actually
[26:41] thinking future back. 2035, 2040, 2100,
[26:45] whatever whatever you want. What does
[26:47] the industry look like? Like let's
[26:49] imagine that computer-aided design suite
[26:51] for molecules has become a standard.
[26:53] Let's imagine a lot of innovation has
[26:55] flowed downstream into some of the wet
[26:56] lab parts of the process. Like, can you
[26:59] paint the picture of what the industry
[27:00] might look like 10 years from now?
[27:03] >> Yeah, I think it's going to be a a
[27:04] really exciting time. And we can look at
[27:06] this on on a couple of angles. So, first
[27:08] of all, the quality of the medicines
[27:10] that we develop will hopefully go up.
[27:12] There's a lot of molecules we put into
[27:13] the clinic today that there is really
[27:16] hard to discover a molecule. You find
[27:17] something that's like 80% of the way
[27:19] there. Maybe advance it anyways to like
[27:21] hit my timelines. It's probably going to
[27:22] benefit some patients. But then you get
[27:24] beat, you know, a year later by someone
[27:26] else. And it's it's really just not the
[27:27] most efficient spend of resources in the
[27:29] industry. You've got the kind of
[27:30] diseases that are just too hard to go
[27:32] after today. People have been trying to
[27:33] drug Alzheimer's forever. And
[27:35] unfortunately, you know, haven't made as
[27:37] much progress as as we'd like. You have
[27:39] things that just aren't economical to go
[27:40] after today. Think about like
[27:41] personalized medicines, rare diseases,
[27:43] things where maybe the patient
[27:44] population is going to be smaller. But
[27:46] again, if we can iterate through these
[27:48] ideas faster, if we can launch something
[27:50] faster, if we can do it in a less
[27:52] expensive way,
[27:53] then those probably come into reach as
[27:55] well. So, I think there's just so many
[27:57] different axes
[27:58] that that we're able to to push on. And
[28:01] I think that means that the the future
[28:03] is is really bright.
[28:04] >> It used to be like, you either want to
[28:05] be first in class or best in class.
[28:07] Now it's like, you want to be last in
[28:08] class cuz like you actually just want to
[28:10] like be the final answer.
[28:12] So, I think like a lot of what you'll
[28:14] see is just like way more intentionality
[28:16] in the types of drugs that you're
[28:17] designing. Like, this drug will be super
[28:19] specific to the disease of interest. It
[28:21] won't have like certain interactions,
[28:22] like certain negative interactions that
[28:24] a lot of drugs today do. A lot of this
[28:26] stuff is actually like you're able to
[28:28] model most of this computationally.
[28:31] Like, maybe not today, but there's
[28:33] definitely a path towards getting there.
[28:34] And I think that's like one of the most
[28:35] exciting things for me.
[28:37] >> Very cool. Can we talk about a business
[28:39] model decision that you guys made?
[28:41] Cuz I think a lot of times folks think
[28:43] about Chai and Isomorphic in in same
[28:45] neighborhood. Isomorphic is developing
[28:47] drugs. You guys are enabling the
[28:49] existing industry to develop drugs more
[28:52] efficiently, better, faster, cheaper
[28:54] than they have before. Why did you
[28:55] decide to go down that path versus,
[28:58] you know, the Isomorphic path?
[29:00] >> Yeah, first of all, I think both of
[29:01] these paths are are great and
[29:03] they can they can create tremendous
[29:04] value. We've always been really excited
[29:06] about building infrastructure for the
[29:08] industry. Our bet when we started to go
[29:10] back to those results in Matt's house,
[29:12] you know, a few years ago,
[29:14] was that this is how most future drugs
[29:16] are going to be discovered. And if
[29:18] that's the case, somebody needs to go
[29:19] and like build that infrastructure and
[29:20] make it happen. I think part of this too
[29:22] is a lot of our founding team,
[29:24] like including Matt and I, had worked in
[29:26] companies before where we had built
[29:27] these full-stack drug pipelines, right?
[29:29] And we were building AI models. We're
[29:31] putting those drugs in the clinic.
[29:32] Again, we're we're pretty bitter lesson
[29:34] pill as well and our our thinking was as
[29:36] the models get better, we want to be
[29:38] spending more of the money making better
[29:39] models as opposed to diverting those
[29:41] resources into clinical trials and
[29:43] things like that. And one of the
[29:44] interesting things about our business
[29:45] model is as the models get better,
[29:47] actually wins us the right to continue
[29:48] investing more in them, right? And you
[29:50] have, you know, partnerships with with
[29:52] these pharma companies that are that are
[29:53] paying off today and allows us to
[29:54] continue invest in it. So, it's it's a
[29:56] much more, you know, scalable business
[29:57] in that way. And I just think about the
[29:59] ultimate impact that we can create for
[30:01] the world is is a lot larger. We've
[30:03] We've always wanted to just partner
[30:04] broadly with the ecosystem. It goes back
[30:06] to this point about fooling yourself in
[30:08] biology. If you work on a small number
[30:09] of drug targets, you might come up with
[30:11] the most exquisite molecules, creating a
[30:13] ton of value for the world by doing
[30:14] that.
[30:15] But, you might miss the forest for the
[30:17] trees because maybe your model doesn't
[30:18] generalize to the other 500 molecules
[30:20] that people are going to make that year.
[30:22] And if you go and and partner with
[30:23] people, you just you just can't fool
[30:25] yourself. Like, you look at our
[30:26] partners, Eli Lilly, Novartis, Organics,
[30:29] Pfizer, like these are not companies
[30:30] that are, you know,
[30:32] they they take this stuff for granted.
[30:34] Like, you have to really deliver on
[30:36] these partnerships for them to take you
[30:37] seriously. And that means it can't just
[30:39] work like some some time. Like, when we
[30:40] ship models at Chai, they really have to
[30:43] work, they have to deliver value to our
[30:44] partners, and it's not like, oh, we made
[30:46] some molecule, doesn't fully work, we're
[30:48] going to have our our chemists like
[30:49] clean it up a little bit. So, it's it's
[30:51] almost a harder business to pull off. I
[30:52] think that's one of the reasons why you
[30:54] haven't seen it pursued many times. If
[30:56] you work on a drug pipeline, again, like
[30:57] there might be ways to fix things up
[30:59] later. There will be the proof of like
[31:00] what happens in the clinic, of course.
[31:01] Um but uh when you have this partnering
[31:03] based model, you have to be really
[31:05] rigorous about your models. Your models
[31:06] have to work really well because
[31:08] otherwise those partners not going to
[31:09] come easily. So, it's made our life
[31:10] harder, I think in many ways, but I
[31:12] think it's also the more rewarding path
[31:14] if we can get it to work.
[31:15] >> what have you learned? I mean, you know,
[31:16] you're you're out of the lab, so to
[31:17] speak, and in the real world, you know,
[31:19] delivering real value for actual pharma
[31:21] companies. What have you learned as you
[31:23] start to work with these partners in
[31:24] terms of any surprises in terms of how
[31:26] their needs might have differed from
[31:28] what you expected or how their level of
[31:30] sophistication around this might be
[31:31] different than what you expected. Have
[31:32] you any surprises or any learnings from
[31:34] working with these partners?
[31:35] >> Yeah, so we went into these
[31:36] partnerships. I think a lot of people
[31:38] told us that like pharma doesn't know
[31:40] how to use AI, these are not it's not
[31:41] like a tech forward industry and things
[31:43] like that. And to be honest, that hasn't
[31:45] really been our experience. I think that
[31:47] these are again, they're very rigorous
[31:49] partners, very rigorous customers,
[31:50] right? And they're they're going to to
[31:52] to test every one of our claims, right,
[31:54] before they start to deploy these
[31:55] things. But when they see the data, they
[31:57] go all in, right? Because pharma is an
[31:59] innovation industry. Um and it's
[32:01] interesting just to think about even the
[32:02] whole economics of the pharma industry.
[32:04] If you build a product in pharma, right?
[32:06] Like like a drug, right? You only have
[32:08] exclusivity on that drug uh for a
[32:10] certain period of time, right? And you
[32:11] have to continue to reinvest. Eli Lilly
[32:13] is a trillion-dollar pharma company
[32:15] right now. If they don't get more
[32:17] blockbuster drugs, they will not be a
[32:18] trillion-dollar pharma company forever.
[32:20] Um and I think that forces these
[32:22] companies to really be on their game of
[32:25] adopting new technologies and deploying
[32:26] them and trying to stay ahead. Pharma is
[32:28] a very competitive arena. There's a lot
[32:30] of people trying to bring these drugs to
[32:31] patients with By the way, it's great
[32:32] great for patients, but uh it means that
[32:34] you have to be on top of your game here.
[32:37] Uh and I think that means that um you
[32:38] know, once you're through that door and
[32:40] your models are working, we've seen like
[32:42] an an upscale this adoption very quickly
[32:44] and people thinking about how to use the
[32:45] models in incredibly creative ways.
[32:47] >> I really like the incentive structure
[32:49] that it creates as well. Um like taking
[32:51] more of the partnership model. As our
[32:53] models get better, we get better results
[32:54] to our customers and so on. Um so I
[32:57] think like that's just like a nice like
[32:59] nice side effect for us as well. Chai
[33:01] has been like incredibly focused. Um and
[33:03] like really as the models get better,
[33:04] you can kind of like iterate on a better
[33:06] model with better data and so on. So
[33:08] like a lot of what we see and like a lot
[33:10] of what our partners are using the model
[33:11] for like generating, you know, binders,
[33:13] antibodies, so on. Like we're also like
[33:15] doing a lot of dog fooding in-house.
[33:17] Like we have a whole science team that's
[33:18] using the models and just trying to
[33:19] understand what they can do. And a lot
[33:21] of that comes back to like, okay, now
[33:23] that like the model now that we've
[33:24] unlocked use case X, like what type of
[33:26] new data can we generate? How can we
[33:28] make the model better that way? Um and
[33:30] really like that's been the long-term
[33:31] vision of Chai is like we never thought
[33:33] of like we're going to build this one
[33:34] model that's going to solve all of this.
[33:36] Like we know that there just like uh in
[33:38] other fields like there're going to be
[33:39] multiple iterations of this model. And
[33:41] as you get better and better models,
[33:42] like you get this flywheel effect, I
[33:44] guess.
[33:45] >> Well, and you mentioned what data can we
[33:46] generate? Can we touch on data for a
[33:47] minute? You guys can't exactly just go
[33:49] script the internet and have all the
[33:50] data you need to build your models.
[33:52] Where does the data come from? How does
[33:54] it kind of compound over time? Can you
[33:57] just say a word about
[33:58] data as an input to your models?
[34:00] >> Yeah, so uh I think like
[34:03] the primary source of data or like the
[34:05] the gold standard source of data is this
[34:07] like protein database. Um so this is
[34:09] like it it's actually just like ac- like
[34:12] legit lab scientists like who since like
[34:15] 1970 have just been depositing crystal
[34:18] structures of like proteins and other
[34:19] molecules over time. Uh and like really
[34:22] without that like structure prediction
[34:24] and design wouldn't have been a thing.
[34:26] Um so there are these uh so that's like
[34:28] one source of structural data. When I
[34:30] started in the field, I really took like
[34:31] a structure-pilled approach. I was like
[34:33] super interested in predicting
[34:34] structure, like how do you think about a
[34:36] machine learning model that can output
[34:38] 3D coordinates? Like that's that's not
[34:40] doesn't look like an LLM, it doesn't
[34:41] look like an image model. This is really
[34:42] in its own class. Josh, interestingly,
[34:45] was taking like the exact opposite
[34:46] approach. Uh so, he he was like one of
[34:48] the original authors on ESM, and that
[34:50] was like a really like a seminal work in
[34:52] understanding how to apply language
[34:53] models to protein sequences. Uh what's
[34:55] really cool about that work is if you
[34:57] can train a language model to understand
[34:58] protein sequences, what ends up
[35:00] happening is it ends up kind of like
[35:02] representing the 3D structure
[35:03] internally. And there's like a really
[35:06] interesting reason why this happens. So,
[35:07] like in order to predict like missing
[35:09] amino acids, so like same way that you'd
[35:11] predict like next word in a sentence, I
[35:12] want to predict next amino acid in a
[35:14] protein. In order to do that
[35:16] effectively, like you really need to
[35:17] understand like, okay, what does the
[35:18] amino acid's like immediate
[35:20] microenvironment look like? Uh cuz that
[35:22] kind of tells you, okay, what are like
[35:24] the compatible amino acids with
[35:25] everything surrounding it. And in order
[35:27] to do that well, you need to understand
[35:28] the protein's 3D shape. Um so, I was
[35:30] taking like this this really
[35:31] structure-based approach, and Josh was
[35:33] even more bitter lesson-pilled than me.
[35:34] He's like, "We're just going to like
[35:35] look at the sequences, and this is just
[35:37] going to emerge." Um so, yeah, the two
[35:39] main sources of this data, again, like
[35:41] protein database for structures, and
[35:42] then like these massive massive, maybe
[35:44] even like order of like trillions of
[35:46] tokens, sequence databases. Um and what
[35:48] you can do once you have really good
[35:49] models is just run them on the sequence
[35:51] databases to get new structures out. Um
[35:53] so, again, like you have this
[35:55] compounding effect. As your models get
[35:56] better, they get more and more accurate
[35:58] at predicting these structures, uh and
[35:59] then you have more and more training
[36:01] data for the next series of models.
[36:02] >> So, nice people ask us at Chai about,
[36:04] these days at Chai, like which which
[36:06] paradigm are are you actually going
[36:07] after? Like one of the things I love
[36:09] about our team is that it's actually
[36:11] just like neither. We're very pragmatic.
[36:12] Like we want to solve the problem. We
[36:13] don't really care is it a sequence
[36:14] approach, is it a structure approach? In
[36:16] practice, it's going to end up being
[36:17] like both, of course. Um I think you'd
[36:19] be surprised, but there there might even
[36:21] be more like biological sequence tokens
[36:22] on the internet than like English
[36:24] language tokens.
[36:25] Uh and a lot of that data is not that
[36:27] useful. It might not be redundant. It
[36:29] might be very noisy. But there's a lot
[36:30] of data out there. There's a lot of art
[36:32] and like bringing it together. And I
[36:33] think also the exciting thing is as the
[36:35] models have gotten into a point now
[36:36] where we can design things in the lab,
[36:38] write to new targets for instance, we
[36:39] can actually use the models to generate
[36:41] data as well. So there's a lot of
[36:42] exhaust from like all the experiments
[36:44] that we're doing at Chai, which also
[36:46] helped to make the models even better.
[36:47] So it's a similar kind of takeoff that
[36:49] we saw with LLMs. Like when I was at
[36:50] OpenAI, we worked on
[36:52] reinforcement learning of like a GPT-1
[36:53] architecture. Like did not work models
[36:55] weren't like good enough. But once the
[36:57] base model got good enough, then you
[36:59] could start to do those kinds of
[37:00] experiments. And And I think there's a
[37:01] similar analogy that's starting to
[37:02] happen in our world now as well, where
[37:04] the models have reached a point where
[37:05] there's actually like a renewed interest
[37:07] in data and like how do we actually
[37:09] bring the models into the loop on like
[37:10] making that happen. I think that creates
[37:12] another really like interesting cycle on
[37:13] on compounding improvement of the
[37:15] models.
[37:15] >> Yeah. We're talking a bit about
[37:17] compounding improvement in the models.
[37:18] And you said something about how pharma
[37:20] as an industry is an incredibly
[37:22] competitive landscape.
[37:24] And it's interesting cuz I think there's
[37:25] been a renowned interest in using AI and
[37:27] using
[37:28] ML to generate proteins and molecules.
[37:31] And so your arena has actually become
[37:32] quite competitive.
[37:34] And you guys have obviously done an
[37:35] amazing job at staying at the frontier.
[37:37] It's been the year of deployment for
[37:38] you. You've locked up a number of pharma
[37:40] partnerships that are making your models
[37:41] better. But how do you guys think about
[37:43] the competitive landscape
[37:45] and staying at the front of the
[37:46] frontier?
[37:47] >> Well, first of all, I think it goes back
[37:48] to like looking at the results and not
[37:50] fooling yourself and being rigorous. So
[37:52] one of the reasons why we do a lot of
[37:53] evaluation of the models. We mainly do
[37:56] it just to hill climb in the models
[37:57] itself. I think if you look at many of
[37:59] the capabilities that we've brought in
[38:00] have kind of been like first in the
[38:02] field. You look at our Chai 2 model,
[38:03] right? Like getting to success rate of
[38:05] the models that you didn't have to do
[38:07] large library screening anymore to see
[38:08] results.
[38:09] Couple months later showing how we could
[38:11] bring in a lot of these developability
[38:12] properties we've talked about before
[38:14] like manufacturability of the molecules.
[38:16] So in in many cases we are pushing the
[38:18] model forward and and trying to see
[38:19] these these
[38:20] emerge. And then we try to quickly like
[38:22] lock those in. Like, how do we how do we
[38:24] make those capabilities like even more
[38:26] pronounced that they become production
[38:27] ready and we can like ship them to our
[38:29] customers. I think one of the one of the
[38:31] things I like to tell the team is that
[38:33] you know, it's not like we are
[38:34] head-to-head like with other model
[38:35] providers or something like that. We're
[38:36] all of us I think are working against
[38:38] nature. Like, nature's actually been a
[38:40] pretty good baseline. People have done
[38:41] drug discovery a certain way for a long
[38:44] time. And you know, Matt talked about
[38:45] how like maybe you don't want to add
[38:47] like module 24 to the Chi model. But
[38:49] people have added like module 240 to
[38:51] like the existing wet lab protocols. And
[38:53] they have been like tuned quite
[38:54] considerably.
[38:56] One of the scientists on our team, Andy
[38:57] Young, was
[38:59] he was one of the first people working
[39:01] on yeast display at MIT actually like
[39:03] two decades ago. And he's got 20 years
[39:04] experience in like Pfizer and Genentech
[39:06] like really honing in these methods. Has
[39:08] a drug approval to his name and
[39:09] antibodies. And I think you look at
[39:11] someone like that and like, you know,
[39:12] Andy knows how to make a good antibody
[39:14] with existing tools. And and that is
[39:16] actually the bar that we need to clear.
[39:18] Now of course, I think the ceiling on AI
[39:19] is going to be a lot higher than what
[39:20] we've managed to do before. Otherwise,
[39:22] what would be the point of of
[39:24] doing this? We didn't start the company
[39:25] just to make, you know, a 10 times
[39:27] faster mouse, right? We started this to
[39:29] make breakthrough medicines that that
[39:30] weren't possible before.
[39:32] But ultimately, that is the bar that we
[39:33] needed to clear in order to get
[39:35] adoption. I think we hit that inflection
[39:37] point a couple months ago. That's why
[39:38] you've seen a lot of these big pharma
[39:39] announcements.
[39:41] But now we just need to continue to hone
[39:42] in on like making these things even
[39:44] better.
[39:44] >> For somebody who's listening who thinks,
[39:46] "Wow, this sounds pretty cool. I wonder
[39:48] what it would be like to work at Chi."
[39:50] What is the best thing about working at
[39:53] Chi? And what is the worst thing about
[39:55] working at Chi?
[39:56] >> Yeah, I can I can speak to some of this.
[39:59] I'll think of this on the fly. The worst
[40:00] thing
[40:01] >> [laughter]
[40:02] >> the So I think like the the best thing
[40:04] is just like how actually like
[40:06] mission-driven everybody is. Like
[40:07] everyone is so dedicated to what they're
[40:09] doing. I've worked at other companies.
[40:11] Like the closest I've ever seen to this
[40:14] is like maybe some of the guys in my PhD
[40:16] lab. Um, but like everyone is just like
[40:18] incredibly incredibly motivated. We all
[40:20] work really hard. There's like an
[40:22] obvious shared goal.
[40:24] Um, and I think that's really rare to
[40:25] see and I think this goes back to just
[40:27] like the focus that we've had since the
[40:29] beginning and like we've always had like
[40:31] kind of a clear philosophy, a clear plan
[40:32] on how things are going to get there,
[40:34] how things are going to get better. And
[40:35] really like everyone at Chai is very
[40:37] bought into this. It's pretty amazing
[40:38] just to see like the amount of
[40:39] dedication that that everyone's putting
[40:41] in. Uh, least favorite thing about Chai,
[40:44] uh, not directly on top of dandelion
[40:46] chocolate, maybe.
[40:47] >> [laughter]
[40:49] >> No.
[40:50] >> Maybe the next office.
[40:51] >> Yeah.
[40:52] Uh, actually, so probably the least
[40:55] favorite part now, uh, is just like
[40:58] I guess it's getting things to work at
[40:59] scale.
[41:00] Um, and really I I didn't even know what
[41:02] that meant. Actually, when we started
[41:04] Chai, we had 128 GPUs and I was like,
[41:06] this is like the most scale possible for
[41:09] a lab. Like this is crazy. Uh, I came
[41:11] from like, you know, my PhD group where
[41:13] we there were four of us sharing eight
[41:15] GPUs and I was like, I just felt GPU
[41:17] rich. It was crazy. Um, so now like, you
[41:20] know, at Chai like we have a lot more
[41:22] infrastructure to maintain. We have a
[41:23] lot more compute resources. Like
[41:24] luckily, GPUs are parallelizable. Um,
[41:27] but that also kind of brings up its own
[41:29] set of problems. So just like how do you
[41:30] keep a cluster healthy over time? How do
[41:32] you get that large training run? So like
[41:33] how do you keep that running for for
[41:35] months on end? Um, and even when it does
[41:37] die like how do you automatically resume
[41:39] these things? How do you keep all of
[41:40] your communication down? How do you
[41:41] optimize the models and make the best
[41:43] use of the resources that you have? So I
[41:45] think these are a lot of problems that,
[41:46] you know, they continuously pop up.
[41:48] They're good problems to have, but I
[41:49] think they're they're also really
[41:51] difficult to solve. Um, yeah, and I I'm
[41:53] just excited to work on this.
[41:55] >> Yeah, we I remember one of the CEOs
[41:57] that, uh, we worked with a couple times,
[41:58] Frits van Slooten, had this line about
[42:00] you either have the pain of failure or
[42:02] the pain of growth. You'd much rather
[42:03] have the pain of growth.
[42:04] >> Yeah, [laughter] that's exactly right.
[42:06] >> I think my favorite part is is probably
[42:08] the results. Uh, and that that sounds a
[42:10] bit cliché, but there's nothing like
[42:13] It's It's working, right? And just like
[42:14] knowing that you're, you know, uh I
[42:16] think many of us in the company, right?
[42:17] Like we've been working in AI for a long
[42:19] time, right? There's all this experience
[42:20] you built up. And And to know that
[42:21] you're applying it to something that
[42:22] really matters, like just even I think
[42:24] just take Matt and and and I like we've
[42:26] been working on this problem for like 10
[42:27] years, right? And a couple of years ago
[42:29] I'm looking at something like, "Yeah,
[42:30] we're writing some cool papers. Like
[42:31] everyone is like celebrating this. Are
[42:33] we actually making the world better?
[42:34] Like is this actually going to impact
[42:35] some patients?" And I think now the
[42:37] answer is like, "Actually, yes. Like we
[42:39] have reached the point where this is
[42:40] going to make a big difference in the
[42:41] world." And every time you get one of
[42:43] these breakthrough results, anytime
[42:45] there's a new feature on the product
[42:46] that makes our lives of our our
[42:47] customers easier, whenever there's like
[42:49] new lab results uh coming back uh from
[42:51] the science team, it's just always so uh
[42:54] honestly it's exhilarating to realize
[42:56] like this is actually going to change
[42:57] the world in a pretty profound way.
[42:59] I think that's also then uh maybe it
[43:01] comes to the least favorite side, right?
[43:03] Like, you know, we're running the
[43:04] company and it's like we have real
[43:05] partners that are relying on this. And
[43:07] like things have to work, right? And you
[43:09] ship a new model generation. How do you
[43:11] make sure that there's no bugs in that?
[43:12] How do you make sure you don't have
[43:13] regressions, right? This is no longer
[43:14] just like a Again, the blue sky research
[43:16] problem of like, "Oh, we got some cool
[43:18] results and we move on." Uh we've had to
[43:20] have really high priorities on like, you
[43:22] know, having production-level code
[43:23] bases. As the team grows, how do we make
[43:25] sure that the code base is in a state
[43:26] that more people can contribute to this.
[43:27] So, um something that that one of our
[43:30] other co-founders, Zach, likes to say is
[43:31] that if you want to move fast in in the
[43:33] long term, you sometimes have to just
[43:35] move a little bit slower in in the short
[43:36] term, right? And And make sure that you
[43:38] are building something uh Again, that
[43:40] goes back to that compounding idea. Goes
[43:41] back to we don't add module 25 to make
[43:44] the next thing work. Uh so, sometimes uh
[43:46] you know, you're you're like so excited
[43:48] to get to the next result and you just
[43:49] want to jump into it, but uh we have
[43:52] real partners, some of the biggest
[43:53] companies in the world that are now
[43:54] relying on us. Uh and it's important
[43:56] that we we realize that, we take that
[43:58] responsibility to heart, uh and we make
[44:00] sure we're building systems that, you
[44:01] know, continue continue to work.
[44:03] >> I have a burning question. Why Why guys
[44:05] called Chai Discovery?
[44:07] >> Chemistry and AI. But we love chai tea
[44:09] as well.
[44:10] >> [laughter]
[44:11] >> There's a lot of chai tea stuff in the
[44:12] office. So yeah.
[44:12] >> That's a good question. I didn't know
[44:13] that
[44:14] either.
[44:14] >> Josh is the visionary. That was all him.
[44:16] >> [laughter]
[44:16] >> It is a very user-friendly name.
[44:18] >> Yeah, we also wanted a name that like
[44:19] like biotech companies have such
[44:20] complicated names. We wanted something
[44:22] that's going to be much simpler. We're
[44:24] trying to build make this whole thing
[44:25] simpler, right?
[44:26] So we need a simple name to go along
[44:28] with it.
[44:28] >> Awesome. What are you guys most excited
[44:30] about in the next 6 to 12 months?
[44:33] >> I think for me it's just the deployments
[44:36] that are happening. So we've announced a
[44:37] couple of these these partnerships and
[44:39] I'm really excited just to hear about
[44:42] the results that our partners are
[44:43] bringing online. It goes back to this
[44:45] point of like making a real difference
[44:47] and also why I'm
[44:48] so happy to see how these partnerships
[44:50] are going even, you know, posted the
[44:52] agreement and as we're working with
[44:53] these folks that models are not sitting
[44:55] on a shelf somewhere. Like they're
[44:56] actually being used on real programs.
[44:59] People trying to approach devastating
[45:01] diseases where, you know, if Chai could
[45:02] give them a molecule that that works it
[45:04] could really change the lives of
[45:05] patients. So I'm really excited to see
[45:07] how that goes. Just the pace of progress
[45:10] here is is incredible but also just like
[45:12] the pace of the models. Like a year ago
[45:14] you could not zero shot a molecule and
[45:15] like, you know, have a good sense that
[45:17] your program was going to work. Like now
[45:19] that's changed. Someone might zero shot
[45:20] a molecule and be like I I think we're
[45:22] going to bring this program to the
[45:22] clinic now. And then a year later, you
[45:25] know, they might even have some of those
[45:26] first molecules going into patients. And
[45:29] just the speed of that is is just
[45:31] incredible and it's you know, sometimes
[45:33] you get some shivers thinking about this
[45:35] stuff like, okay, my model is going to
[45:36] like Matt has some patents from our last
[45:38] company about, you know, like just
[45:39] generating the molecule on the computer
[45:41] and these things are are now in patients
[45:42] and just to think about the scale. Like
[45:44] I don't know a few years from now, do we
[45:46] have dozens? Do we have hundreds of like
[45:47] Chai molecules going into people? It's a
[45:50] bit mind-blowing to think about what
[45:51] that might might look like for patients.
[45:53] >> I think one of the things that like
[45:55] again like what made Chai unique like
[45:56] the dedication. Like people are like,
[45:58] man, you work a lot. Um and like you
[46:01] like aren't you burnt out or whatever.
[46:02] It's like it's actually really easy and
[46:04] like it's very motivating when you're
[46:06] making progress. Seeing the progress
[46:08] that we're making and just like thinking
[46:09] man the next model is going to be even
[46:10] better than the last. We identified this
[46:12] new thing so on.
[46:14] Um like that's so incredibly motivating.
[46:16] For me it's more of just like what what
[46:18] can we unlock next and like how do we
[46:19] make these things more controllable and
[46:21] like when someone comes to us with a
[46:22] certain target instead of just like
[46:24] hoping we get good affinity or something
[46:26] like that, you know, can we actually
[46:27] control this? Can we say like we want
[46:29] exactly a 10 animal or binder or things
[46:31] like that. Like there are a lot of
[46:32] technical things that I think are we're
[46:33] like kind of right on the brink of
[46:35] solving. Um and for me it's like it's
[46:37] really motivating just to like pin those
[46:38] things down and just like get all of
[46:40] this over the line and see kind of where
[46:42] that leads to next.
[46:43] >> Awesome. Matt, Josh, thank you for
[46:45] Engineering Biology and thank you for
[46:47] sharing your story with us today.
[46:49] >> Thank you guys. Thanks for having us.
[46:53] >> [music]
[47:16] [music]
