By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Sequoia CapitalChai Discovery's Bitter Lesson: Drug Design Is Another Scaling Problem
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
Jump to publisher timestampsChapter summaries
Making drug discovery feel like engineering, not serendipity
The conversation opens with Chai Discovery’s ambition to shift drug creation from trial-and-error “discovery” toward a more intentional, engineering-like “design” process. A co-founder compares the desired abstraction layers in biology to the way software engineering and code-generation tools enable rapid iteration.
- •Chai’s goal is to industrialize and systematize parts of drug discovery that have historically been serendipitous
- •The co-founders want biology to have clearer abstraction layers, similar to modern software engineering
- •The current process is described as heavy trial and error, with tools not yet enabling fast iteration
- •Chai positions itself as applying foundation-model-style thinking to biology and molecules
From wet-lab “verification” to a design-first paradigm
Pat Grady asks whether the boundary between what can be engineered versus what must be tested is moving. A co-founder reframes the question: the lab remains essential for verification, but the bigger shift is from searching for needles in haystacks to specifying what you want and having models generate it.
- •The lab is framed as a verification mechanism, analogous to evaluation in AI
- •Traditional “drug discovery” is portrayed as massive screening to find rare hits
- •The “dream state” is specifying desired molecular properties up front and generating candidates accordingly
- •A co-founder suggests lab work may increase because improved ROI could drive more experimentation
Protein AI progress: folding breakthroughs to design and diffusion-era prompting
The episode traces a timeline of advances in protein modeling, starting with the impact of deep learning on protein folding competitions and culminating in diffusion models that can generate structures and sequences together. The speakers describe early distance-prediction approaches, quality improvements over time, and the emergence of more realistic “prompts” with constraints like binding locations.
- •Protein folding competitions are cited as a benchmark that saw big performance jumps starting around 2018 and again around 2020
- •Early deep learning approaches predicted inter-amino-acid distances and used a renderer to produce protein structures
- •Progress is compared to image generation improving from blurry to high-definition outputs
- •The field expanded from folding to designing sequences that fit structures, moving toward drug design
- •Diffusion models enabled generating structure and sequence simultaneously and supporting more realistic constraint-based prompting
Why start the company in 2024: antibodies become tractable
Asked what made 2024 the right moment, a co-founder points to improving antibody-related structure prediction and a shift in beliefs about feasibility. The discussion emphasizes that antibodies were long seen as too hard due to limited data, and that designing therapeutic antibodies depends on first being able to predict their structures.
- •Early conversations about antibody–antigen structure prediction helped crystallize the company’s timing
- •A co-founder says many believed antibody folding was too hard due to insufficient structural data
- •Antibody design is framed as the “holy grail” because it aligns with what much of the drug industry wants
- •A key premise: if you can’t predict what an antibody looks like, you can’t design one
- •The speakers argue biology can be simpler and more interconnected than it first appears, as many entities reduce to amino-acid sequences
Founders’ backgrounds and the “DNA/protein as language” motivation
Pat asks about backgrounds and how the necessary disciplines fit together. One co-founder describes early excitement about AI and biology, time at OpenAI working on GPT models and scaling laws, and the motivation to apply similar methods to DNA and protein; the conversation also highlights that the company builds models from scratch rather than fine-tuning general LLMs.
- •A co-founder recounts starting at OpenAI and working on GPT-1, GPT-2, and scaling laws
- •Motivation is framed as: if models can learn human languages, why not DNA and protein?
- •The company focuses on choosing architectures, compute configurations, and tasks that make progress possible
- •The speakers emphasize they build their models from the ground up rather than fine-tuning a general model
- •The timing is linked to early evidence that previously “too hard” antibody problems were starting to work
Diffusion models intuition: iterative “make it slightly better” generation
Pat asks for intuition on why diffusion models work. A co-founder explains diffusion as a multi-step denoising process that gives the model “more time to think,” contrasting it with variational autoencoders that try to compress complex inputs into a simple latent distribution all at once.
- •Diffusion is described as a multi-step process rather than one-shot generation
- •VAEs are characterized as trying to compress proteins into fuzzy Gaussian vectors, which was difficult for the domain
- •The diffusion approach is framed as learning repeated small improvements: fixing a “slightly broken” protein step by step
- •This iterative decomposition is presented as particularly effective for biology
- •Pat connects the idea to iterative prompting games with ChatGPT, while the speakers note the analogy is imperfect
Building a cross-disciplinary “Avengers” team while staying small
Sonali Singh asks how Chai assembles a team across AI, biology, chemistry, and product. A co-founder describes a staged hiring approach—starting AI-heavy, then adding antibody engineers, product builders, security, and scaling/infrastructure talent as the models and deployments demanded it—while intentionally keeping the team small to force prioritization.
- •The founding team skewed toward AI researchers, with required roles expanding as milestones changed
- •As models progressed, Chai brought in antibody engineering and scientific expertise for specific campaigns
- •Product interfaces are emphasized as necessary to make powerful models usable in practice
- •Security and GPU/infrastructure “hacking” become important as deployments and scale increase
- •The company keeps the team small so everyone is stretched and forced to prioritize what matters
From 0.1% binding to 15%: hit rates, de novo design, and scaling laws
Sonali asks about the quality of molecules and what made high hit rates possible. A co-founder explains a focus on de novo generation and dramatically improving binding success rates, moving from a baseline where only a tiny fraction of designed antibodies bind to a regime where results support faster iteration, stronger evaluation, and incorporating more properties from the start.
- •Chai’s early thesis focused on de novo molecule generation rather than incremental tweaks in a lab-in-the-loop workflow
- •A co-founder cites a starting state-of-the-art around 0.1% binding success for antibody design
- •They report reaching about a 15% success rate with a later model, enabling meaningful statistics and iteration
- •Higher hit rates are described as strengthening the “gradient” for improving the process because more targets yield usable hits
- •The approach is tied to broader machine learning principles, arguing biology is not fundamentally bespoke
The “Bitter Lesson” approach: simplify models, scale rigorously, and measure honestly
Pat asks about identifying scaling laws and the challenge of representing biology. The co-founders describe being “Bitter Lesson” oriented—scaling data, models, and compute—while also simplifying architectures so iteration scales, and maintaining rigor because wet-lab noise makes it easy to fool yourself.
- •A co-founder explicitly frames Chai as “Bitter Lesson”-oriented: scale data, models, and compute
- •Simplicity is a core principle; a co-founder notes a prior model had many submodules, making iteration difficult
- •They emphasize building models from scratch and relying on guiding principles for open-ended research
- •Rigorous evaluation is stressed because wet-lab measurements have large error bars
- •They argue progress must be demonstrated broadly, not just on one or two tasks, to support real products
Hard targets, unified progress, and the molecular CAD vision
Sonali asks about progress on known targets versus unlocking previously undruggable biology. A co-founder argues that instead of piling on ad hoc complexity, scaling better models should unlock intrinsic target features; the team also notes that “hard” means different things across science, ML, and product, which helps parallelize progress toward a unified computer-aided design suite for molecules.
- •The team resists adding one-off complexity for single targets, focusing on compounding improvements that generalize
- •A co-founder suggests that with scaling, models should learn hidden target features needed to reduce loss further
- •“Hard target” is defined differently by science, ML research, and product teams, which is framed as a feature that reduces single bottlenecks
- •The stated company goal is a unified computer-aided design suite for molecules, not a small pipeline of therapies
- •They describe accelerating the loop from idea to testable hypothesis, where order-of-magnitude speedups expand the number of ideas to explore
Future industry outlook: more intentional drugs and faster iteration unlocks more diseases
Pat asks for a future-back vision of the industry once molecular CAD is standard. A co-founder predicts higher-quality medicines, less resource waste from advancing “80% of the way there” molecules, and improved feasibility for difficult and uneconomical areas if iteration becomes cheaper and faster; another speaker emphasizes intentionality, specificity, and fewer negative interactions as modeling improves.
- •A co-founder expects medicine quality to rise as design becomes more intentional
- •They argue faster iteration could reduce inefficiency from advancing suboptimal molecules to meet timelines
- •Hard-to-treat diseases and smaller-population indications may become more feasible if development cycles compress
- •A speaker says the goal becomes being the “final answer,” not merely first- or best-in-class
- •They anticipate increasing ability to computationally model properties that today are not fully modeled
Platform vs pipeline: partnering model, pharma realities, and scaling advantage
Pat asks why Chai chose enabling the industry rather than developing its own drug pipeline. A co-founder explains they want to build infrastructure, keep investment focused on improving models (rather than funding clinical trials), and use partnerships as a forcing function for rigor and generalization; the discussion also covers how pharma partners evaluate claims and adopt quickly once convinced by data.
- •Chai’s choice is framed as building infrastructure for how future drugs will be discovered
- •A co-founder says the partnership model lets them keep allocating resources toward better models instead of clinical trials
- •Partnering is described as reducing self-deception because models must work across many real programs, not a few hand-picked ones
- •The speakers reject the notion that pharma can’t use AI, describing partners as rigorous and quick to adopt when data supports it
- •They discuss incentives in pharma to continuously reinvest in new technologies due to competition and the need for new drugs
Data sources and the compounding flywheel: structures, sequences, and lab exhaust
Pat asks where data comes from if it can’t simply be scraped from the internet. A co-founder points to a long-running structural protein database as gold-standard data, alongside massive sequence databases; the speakers describe how better models can convert sequences into predicted structures and how internal experiments generate additional data that feeds future iterations.
- •A major structural data source is described as a protein database built from decades of deposited lab-derived structures
- •Massive biological sequence databases are another key data source, described in token terms
- •The speakers describe a compounding loop: improved models predict more structures from sequences, expanding training data
- •They explain why sequence models can implicitly learn 3D structure: predicting amino acids requires understanding local 3D context
- •They add that in-house experiments produce “exhaust” data that further improves models once the base capabilities are strong enough
Staying ahead, scaling pain, culture, naming, and what’s next
Sonali asks about competition and staying at the frontier; the co-founders emphasize rigorous evaluations, rapidly locking in emerging capabilities into production, and competing primarily against established wet-lab baselines. In closing, they discuss what it’s like to work at Chai, the challenges of scaling infrastructure, why the company is named Chai, and what they’re excited about in the next 6–12 months.
- •They describe focusing on measurable results and evaluations, then hardening new capabilities for production shipping
- •Competition is framed as secondary to beating existing wet-lab protocols and expert human baselines
- •Team culture is described as highly mission-driven, with intense motivation driven by tangible results and partner impact
- •Scaling compute and keeping large training runs and clusters healthy is cited as a major operational challenge
- •They explain the name “Chai” as Chemistry and AI (and also an affection for chai tea), and express excitement about partner deployments and improved controllability of model outputs
Publisher timestamps
- 0:00Introduction
- 1:52From Discovery to Design
- 3:25Protein AI Breakthroughs Timeline
- 6:04Why Start in 2024
- 10:13Diffusion Models Intuition
- 11:41Building the Avengers Team
- 15:22Hit Rates and Scaling Laws
- 25:01Molecular CAD Vision
- 25:24Faster Design Loops
- 26:32Future Drug Discovery
- 28:37Platform Business Model
- 31:14Partnering Reality Check
- 33:44Data Flywheel Explained
- 37:16Staying Ahead at Scale
- 39:44Culture and What's Next
