Stanford OnlineStanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Applications, AI in Life Sciences
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
Chapter summaries
Setting the stage: guests from Anthropic and Chai, and the class plan
Apoorv Agrawal introduces the session’s two guests and frames the format: short guest presentations followed by Q&A. He previews each guest’s background and why their work is relevant to AI in life sciences.
- •Apoorv Agrawal introduces Eric Kauderer-Abrams (Anthropic) and Josh (Chai Discovery).
- •The class will hear about the guests’ work first, then transition to audience Q&A.
- •The guests represent complementary perspectives: life sciences tooling and AI model/product development.
Josh’s path: from pre-med track and family of doctors to coding and biotech
Josh explains that he did the pre-med coursework and originally expected to become a doctor, influenced by a family of doctors. He shifted toward coding because software can scale broadly, and later recognized drugs as another way to “scale medicine.”
- •Josh says he planned on medicine and took pre-med classes, coming from a family of doctors.
- •He learned to code partly to differentiate from older siblings already in medical school.
- •Programming felt powerful because software can be distributed and scaled widely.
- •He later realized biotech and drugs also scale medical impact, linking his interests in coding and biology.
Chai Discovery’s mission: a CAD suite for molecules and “zero-shot” drug design ambition
Josh describes Chai as building a computer-aided design suite for molecules, with a focus on designing antibody molecules. He outlines a long-term hypothesis: much lab work will move onto computers and, eventually, molecules could be designed “ready for patients” directly from computation.
- •Chai is building computer-aided design for molecules, starting with antibodies.
- •Josh frames today’s drug creation as a trial-and-error process from hit finding through property optimization.
- •The long-term bet is “zero-shot” design of patient-ready molecules from the computer.
- •Chai chose a partnering model with the ecosystem rather than trying to make most drugs in-house.
- •Josh notes that large pharma companies already use Chai’s models for drug design.
Eric’s journey and Anthropic’s life sciences vision: accelerating end-to-end R&D with Claude
Eric recounts moving from math and physics into AI and then into biology, including experiences starting biotech and medtech companies. He explains Anthropic’s goal: train and productize Claude to accelerate life sciences work end-to-end, from basic research through clinical and regulatory processes.
- •Eric describes his trajectory from physics/AI into biology, including running a molecular diagnostics company.
- •He says a step-change in model usefulness convinced him AI could accelerate life sciences R&D end-to-end.
- •Anthropic’s scope includes basic research, drug development, clinical design, regulatory work, and program strategy.
- •He emphasizes both model training and product integration into life-science workflows.
- •He mentions an upcoming life-sciences-oriented interface conceptually like “Claude Code for Bio,” tailored to scientific visualization and tool connectivity.
Drug development walkthrough: targets, modalities, preclinical design, IND, clinical phases, and timelines
Prompted by Apoorv Agrawal, Eric lays out a start-to-finish view of bringing a drug to market and where time and cost accumulate. He argues bottlenecks are distributed across multiple stages, not confined to a single phase like clinical trials or design.
- •Eric states end-to-end drug development often takes about 10–15 years on average, with faster exceptions.
- •He describes target selection as the real start of the clock, followed by modality choice and drug design in the preclinical phase.
- •He notes a preclinical target-to-clinic timeline of about four years historically and argues it’s a major opportunity for compression.
- •He outlines IND enabling human trials, then phases 1 (safety), 2 (efficacy signal), and 3 (validation for approval), plus manufacturing transfer work.
- •He highlights “target crowding,” saying only about 30 net new targets enter the clinical phase per year despite thousands of potential targets.
Where Chai fits: linking molecule quality to easier clinical trials and “outer loop” iteration
Josh explains Chai’s focus within the broader pipeline: generating better molecular matter that modulates disease, which can improve downstream trial feasibility. He argues AI can reconnect discovery and development by producing more potent medicines, increasing effect sizes and simplifying clinical validation.
- •Josh reframes drug development as finding molecular matter that changes disease outcomes, then proving it in patients.
- •He argues more potent molecules can increase effect size, which can make trials easier (e.g., fewer subjects or clearer outcomes).
- •He draws an analogy to code generation evolving from iterative debugging to zero-shot app creation.
- •Chai focuses on molecular generation as a leverage point, while expecting reinforcement from faster iteration cycles enabled by broader AI tooling.
- •Josh emphasizes improved medicines and broader coverage, noting many diseases lack approved medicines.
“Why now?”: better AI, more data, greater willingness to experiment, and geopolitical pressure
Josh and Apoorv discuss why this moment is different, beyond long-standing human interest in medicine. Josh points to improved AI scalability, larger compute, new datasets, and increased organizational openness—plus competitive pressure he observes in pharma conversations.
- •Josh says AI has improved via more scalable architectures, more compute, and new datasets.
- •He highlights increasing willingness in pharma to experiment with AI tools.
- •He describes conference conversations clustering around “AI and China,” and claims China is outrunning the U.S. in drug discovery efficiency and speed.
- •He argues AI can be a democratizing force because “China cannot run faster than AI.”
- •Apoorv presses for sizing impact: whether timelines shrink modestly or dramatically.
Converging trends and timeline compression: LLMs, specialized foundation models, and high-throughput biology data
Eric attributes “why now” primarily to rapid large language model progress, complemented by modality-specific foundation models and expanding biological data generation. He argues there’s line of sight to large reductions in total development time, including meaningful improvements in clinical trial efficiency.
- •Eric emphasizes LLMs enabling the “outer loop” of scientific iteration, complementing domain models (e.g., for molecule design).
- •He highlights increased biological measurement capacity (e.g., sequencing and other high-throughput assays) as fuel for model improvement.
- •Eric suggests a plausible path to reduce full drug development timelines into the five-year range as an upper bound, with potential for a few years in some cases.
- •He notes clinical trials create lower bounds, but argues endpoints can be accelerated via better proxy measurements and assay-driven endpoints.
- •He adds that larger drug effect sizes can reduce enrollment needs and speed trials.
Where value accrues: why AI tools might capture more value than past life-science tooling—and what changes for startups
Apoorv asks why tool platforms might capture more value in biology, given historical value concentration in drug revenues. Josh argues tools become more valuable as they materially increase success probability and lower cost/timelines, and he discusses how improved tooling could reshape biotech formation and strategy.
- •Josh claims many existing drugs are “primitive,” and improved tools should matter more as they raise success probability and quality.
- •He argues that if tools improve probability of success, reduce cost, and shorten timelines, toolmakers should capture meaningful value.
- •He contrasts scalable tools with less scalable end products and calls current lab tools (e.g., traditional experimental systems) the real “competitors.”
- •Josh discusses potential democratization: easier target insight and faster programs could enable more startups, but biotech will evolve rather than disappear.
- •He notes second-order effects are hard to predict because progress feels exponential.
Big unsolved problems: expanding beyond antibodies, scaling target discovery, and building more sophisticated medicines
Apoorv asks both guests what remains hardest at the frontier. Eric emphasizes extending AI-enabled engineering rigor to additional drug modalities and solving scalable target discovery, while Josh adds a push toward more complex, controllable medicines enabled by finer molecular understanding.
- •Eric says AI success in antibody development needs to expand to small molecules and other emerging modalities.
- •He frames progress as pushing the frontier of “drugable” targets and turning modalities into engineering disciplines.
- •Eric highlights the need for scalable discovery of many more high-quality targets to avoid continued crowding on a small set.
- •He cites approaches like virtual cell/cell perturbation modeling and genetics plus health-record correlations as active directions.
- •Josh argues more sophisticated medicines become possible as models enable deeper molecular-level understanding (including protein folding in their work).
Commercially massive opportunities: muscle-mass targets and sleep as “consumer medicine” scale
Apoorv asks for examples of potential next blockbuster-level programs. Eric points to drugs increasing lean muscle mass as potentially strong commercially, while Josh highlights sleep as a broad lever that could begin with disorders and expand outward.
- •Eric says emerging targets around increasing lean muscle mass could be commercially significant, while noting medical use cases as well.
- •Josh jokes internally about employee-only access if such a drug existed, illustrating perceived demand.
- •Josh argues sleep-related therapies could be huge because sleep affects many disorders.
- •He compares the trajectory to obesity drugs moving from initial indications into broader populations.
- •He suggests this could mark a shift toward large-market “consumer medicine” dynamics.
Bullish vs. skeptical: pharma as capital aggregators, Jevons paradox for lab work, and caution on tool-selling models
Apoorv asks each guest for what they’re bullish on and what they’re skeptical about. Josh argues improved tools can make pharma more effective and increase demand for experiments, while warning that non-frontier software approaches may be displaced; Eric highlights AI-native wet-lab execution as a key long area and expresses skepticism about relying on selling tools to pharma as a business model.
- •Josh is bullish on pharmaceuticals becoming more effective reinvestors of capital into drug pipelines as tools improve.
- •Josh predicts a Jevons-paradox-like increase in lab work because better experiment design increases experiment ROI.
- •Josh is skeptical of purely software approaches that are not frontier, expecting powerful AI to overtake older computational methods.
- •Eric is bullish on scalable, AI-native wet-lab execution models, including AI interfacing via CROs and eventually direct instrument integration.
- •Eric says tool-selling-to-pharma business models can be difficult, and highlights “pipeline in a person” as barriers to running drug programs fall.
Audience Q&A: why development gets less attention, scaling training runs, and limits of “single-shot” biology
Audience questions probe why AI discussion skews toward discovery over development, what bigger training runs might unlock, and whether “single-shot” approaches can work in deep-stack domains like biology. Eric points to operational complexity and longer feedback loops in development, while both speakers discuss capability unlocks, iteration speed, and uncertainty about when (not if) ambitious goals become feasible.
- •Eric says discovery opportunities are easier to name, while development is less glamorous, more operational, and harder to tie directly to model intelligence.
- •He lists development opportunities: smarter patient recruitment and site selection, automation of trial administration/monitoring, and improving effect size to raise success odds.
- •On scaling training, Eric describes incremental improvements plus “capabilities turning on” at certain scales, giving an example of models beginning to understand proteins in a later generation.
- •A question challenges whether deep-stack tasks can be single-shot; Josh argues the frontier keeps moving with scaling laws, more data, and more compute, though outcomes aren’t guaranteed.
- •Josh adds that discovery gets more attention partly because feedback loops (design and test in weeks) are shorter than clinical development loops.
