Stanford MS&E435 Economics of the AI Supercycle ｜ Applications, AI in Life Sciences — Annotated Lecture Notes & Key Takeaways

Course Information

Institution: Stanford

Original Course: Stanford MS&E435 Economics of the AI Supercycle ｜ Applications, AI in Life Sciences

Instructor Bio: This session is led by **Apoorv Agrawal**, Adjunct Lecturer in Management Science and Engineering at Stanford University and Partner at Altimeter Capital. Apoorv Agrawal holds a Bachelor’s degree in Computer Science from the National University of Singapore and an MBA from Stanford University. At Altimeter Capital, he has led investments in AI-enabled life sciences and biotech companies, bringing specialized perspective on AI’s economic impact in healthcare and biomedical research.

Course Description: This closing lecture analyzes the economics of AI in life sciences and healthcare. It covers AI-driven efficiency gains in drug discovery, genomic sequencing, and clinical diagnostics, evaluates the R&D return on investment for biotech AI platforms, and discusses how regulatory environments shape commercialization pathways. The session concludes with a forward-looking view of how AI will accelerate innovation across the life sciences industry.

Two Guests, Two Competing Visions

 

This session of MS&E435 featured Eric Abrams (Anthropic, heading up biology and life sciences) and Josh (Shai Discovery, founder). The setup was deliberately contrasted: Eric represents the general-purpose AI lab perspective — get biology to "run on Claude." Josh represents the specialized startup route — build a CAD suite for molecules. Both are chasing the same prize: moving drug discovery from a painstaking lab process toward something more like software engineering.
 

A useful framing from the professor's intro: Josh's earlier work on ESM-1 now accounts for roughly half of all citations in a major slice of the field. That's a striking claim about the reach of a single foundational model — and a reminder that in AI for biology, a model released years ago can still be the gravitational center everyone builds on.
 

Josh's Path: From Pre-Med to "Code as Medicine"

 

Josh's personal story is more winding than the polished founder bio suggests. He actually took all the pre-medical classes and comes from a family of doctors — older siblings already in med school while he was a kid. The twist: he learned to code in high school and discovered biotech late. His reasoning is genuinely interesting and worth sitting with:

Code is infinitely scalable — write once, distribute widely. Drugs, it turns out, have the same property: a single molecule can treat millions.

That's the core insight that pulled him from medicine into computational biology. He liked coding more than the lab, but sees them as linked rather than opposed. I found this a refreshingly honest origin story — not "I always knew I'd disrupt drug discovery," but a teenager noticing that both code and medicine could scale impact.
 

One small confusion:​ the professor says Josh "was not a pre-med major," but Josh corrects this — he did take the pre-med coursework. It's a minor factual slip, but it matters because the "scenic road" framing depends on him having deliberately stepped off the med-school track, when really he supplemented it with programming. Not a big deal, but worth noting if you're trying to reconstruct the actual narrative.
 

Shai Discovery: A CAD Suite for Molecules

 

Here's the heart of Josh's talk, and it's the densest part of the lecture. Shai's pitch has three layers:
 

Layer 1 — The near-term vision.​ About half of all approved drugs today are antibodies. The current process is trial-and-error: find an initial "hit," then make a huge number of modifications to get the properties a drug needs (binding affinity, stability, manufacturability, safety). Shai wants to let you design those antibody molecules on a computer instead.
 

Layer 2 — The big bet ("zero-shot").​ The long-term goal is to generate molecules that are "ready for patients right out of the computer" — no iterative lab optimization required. This is what Josh calls zero-shot design: the model produces a candidate that works on the first try.
 

Layer 3 — The company's origin story.​ When they started ~2.5 years ago, everyone told them they were crazy. Two objections: (a) the technology won't work, and (b) even if it does, shouldn't you be making your own drugs rather than selling tools?
 

This third layer is where the lecture gets economically interesting — and where I think the argument is weakest. Josh clearly believes tooling will win, but the "shouldn't you just make the drug yourself?" objection is not really answered head-on. It's the classic platform-versus-product tension, and in life sciences the platform bet has a mixed track record (Benchling, 10x Genomics, etc., come up as Anthropic partners in Eric's intro — tellingly, those are also tooling companies, which may be why the partnership makes sense).
 

What I wanted more of:​ an example. Josh describes the workflow abstractly — hit → modifications → properties — but never walks through a concrete molecule or a before-and-after comparison. Even a toy example ("here's a naïve antibody, here's what our model changed, here's the predicted improvement") would make the "CAD suite" analogy click. Without it, the phrase risks sounding like a slogan rather than a describable product.
 

Eric Abrams: Getting Biology to Run on Claude

 

Eric's portion (as introduced) is shorter in the transcript we have, but the setup is clear. His job at Anthropic is to convince the biology world to run on Claude — the general-purpose frontier model, not a domain-specific one. Anthropic launched "Claude for Life Sciences" last October and has partnerships with Benchling, 10x Genomics, and Novo.
 

A few things stand out about this approach versus Josh's:

• Generalist vs. specialist.​ Eric is betting that a broad model, with the right integrations and safety posture, can serve biology well. Josh is betting that molecules need purpose-built representations and tooling.
• Distribution strategy.​ Anthropic's partnership list is essentially "embed Claude into the existing lab software stack" — Benchling (lab data management), 10x Genomics (single-cell instrumentation). That's a very different go-to-market than Shai building an end-to-end design environment.
• The "scarred memory" detail.​ Eric apparently took his final EE PhD exam in this very course's room and is "still traumatized" from a device physics final. Charming, but also a useful reminder: the people building these tools often have serious technical training, not just business backgrounds.

 

A limitation worth flagging:​ the transcript cuts off (marked <truncated>), so we don't get Eric's full presentation or the Q&A. That's a real gap — Eric's section feels underdeveloped relative to Josh's, and we miss the back-and-forth where the two visions would presumably clash. Any conclusions drawn here should be read as provisional on the truncated portion.
 

The Central Argument: Lab Work Moves to the Computer

 

If there's one thesis tying the whole lecture together, it's this — and it appears in both talks: a large share of today's lab work is going to migrate onto the computer. Josh says it explicitly as a "medium-term" goal. Eric's entire role ("get biology to run on Claude") presupposes it.
 

Why might this be true? A few reasons floated, directly or indirectly:

1. Scalability.​ Josh's original insight — code scales; molecules, it turns out, share that property. Once designed digitally, a drug can be manufactured and distributed widely.
2. The ESM-1 citation fact.​ If foundational biological models become as broadly useful as that citation share suggests, the "lab" starts to mean "training and inference infrastructure."
3. Economic pressure.​ Drug discovery is famously expensive and slow (a point the lecturer alludes to via the trial-and-error description). If computation can compress that loop, the incentive is enormous.

 

But here's where I think a listener should pump the brakes. The lecture presents the migration as a near-inevitability, and the evidence is largely aspirational​ (Josh's "big goal," the zero-shot dream) plus anecdotal​ (one model's citation count). What's notably missing:
 

• A concrete timeline. "Medium term" is undefined; "long term" is explicitly deferred.
• A discussion of why zero-shot might fail.​ Biological systems have context-dependencies (tissue environment, immune response, delivery mechanics) that aren't fully captured by molecular structure alone. Josh acknowledges the goal is ambitious, but the professor doesn't press him on specific failure modes.
• Validation data.​ We hear no numbers on how Shai's designed molecules perform in actual wet-lab assays versus conventionally discovered ones. Without that, "CAD suite" remains a vision rather than a demonstrated category.

 

This is the section I'd most like to revisit after reading the Q&A — presumably a student asked exactly these questions.
 

What's Strong, What's Unconvincing

 

Strongest points:

• The code-as-scaling-analogy for medicine is genuinely illuminating and explains a real career pivot without resorting to hype.
• The "half of approved drugs are antibodies" statistic gives the audience a concrete foothold — it explains why antibody design is the right first target.
• The professor's two-object objection (tech won't work / should you just make drugs yourself) is a genuinely useful framework for evaluating any AI-biology startup, not just Shai.

 

Weakest points:

• No worked example​ of the CAD workflow. Abstract descriptions of "hit → optimize properties" don't show the listener what's actually being automated.
• The zero-shot claim is presented but not defended.​ "Someday we'll generate patient-ready molecules from a computer" is a bold prediction to state without detailing the remaining scientific hurdles.
• Truncation.​ We lose Eric's counterweight and the Q&A, so the lecture leans heavily on Josh's optimism. A balanced note would need that second half.

 

Questions Left Hanging

 

A few things I'd want to ask in the room:

• For Josh: When a Shai-designed molecule goes to the lab, what fraction of candidates hit the target versus conventional discovery? Without a comparison, "design" is just a word.
• For Eric: Does Claude for Life Sciences compete with specialized models like ESM-family models, or rely on them? The partnership model suggests collaboration, but the underlying tension (generalist vs. specialist) is unresolved.
• For both: Who owns the data — and the IP — when a foundation model designs a drug? This comes up implicitly through the Anthropic + Benchling/10x Genomics partnerships, but isn't addressed directly.

 

Bottom Line

 

The lecture is at its best when it lets the two speakers embody a real strategic choice in AI for life sciences: general-purpose model as platform​ (Eric/Anthropic) versus specialized molecular design tool​ (Josh/Shai). Both agree the lab is moving toward the computer; they disagree on what that computer should run. That's the genuinely useful takeaway, and it's one you can apply well beyond this single talk.
 

Just don't mistake the vision for proof. The strongest claims — zero-shot patient-ready drugs, the full migration of lab work onto silicon — are stated as goals, not results. Treat them as hypotheses to test against the validation data that, in the portion we have, hasn't yet arrived.
 

Content Disclaimer:

This article is for general reference only and does not constitute professional R&D guidance, production process advice or quality certification. All material performance data has specific test premises; readers should verify parameters against actual equipment and working conditions.