Patrick Hsu

1. Key Themes

The GLP-1 Blueprint Is Reshaping Biopharma's Capital Structure

Hsu argues the GLP-1 class has shifted who funds and directs biomedical innovation. Corporate R&D, not government, is now ascendant, and the money is concentrated in one horizontal product. Patrick Hsu: "Eli Lilly, for example, is, you know, the largest biotech company in the world... It's valued about $1 trillion. The, you know, the GLP-1 revenue is going to be something like $90 billion this year, and they're going to scale their internal R&D budget linearly or along with their GLP-1 revenue" 00:13:53. He adds that "Eli Lilly spends more money on R&D than the entire country of Germany" 00:14:24 and that by the early-to-mid 2030s it "will surpass the United States" in R&D spend if revenue keeps ramping. The investable implication is that GLP-1 created a blueprint: "small disease where you can have very clear mechanism, bigger disease, then everybody. So you could go after narcolepsy, right, then you can go after sleep, then you can go after wakefulness and attention and nootropics and enhancements" 00:19:08. Hsu also notes that "a lot of companies are explicitly trying to attack this playbook."

Existence Proofs Collapse Discovery Timelines

A recurring non-obvious claim is that most of the work in innovation is believing something is possible. Patrick Hsu: "Just knowing that there's alpha and that someone is making progress in this general area, you don't need to know a lot of specific details. That's sufficient" 00:17:39. He cites the four-minute mile, the nuclear bomb, and quant trading, plus the frontier AI labs converging on the same Millennium Prize problems. He concludes: "Just knowing that this is possible creates the existence proof that gives people and markets somehow the courage and changes something about the plumbing of the universe to make it possible" 00:18:25. For investors, this means a visible first success (like GLP-1) should be treated as a signal that fast-followers in adjacent categories are about to appear.

Biology's Bottleneck Is Data Shape, Not Just Data Volume

Hsu says scaling alone won't fix bio-AI, because the data types the models need don't exist yet. Patrick Hsu: "If you ask a machine learning person, 'What type of biological data do you like desperately want?'... time series data, right? And biology does not tend to have great time series data" 00:34:19. Today's single-cell methods kill the cell, so you only learn about similar cells. Arc's Felix Horns has engineered protein nano-cages that bud out of living cells carrying RNA: "It's sort of like sending out a FedEx package, you know, every few hours" 00:36:52. Hsu sees a "fundamental need for biotechnology invention that creates that shape of data to train that type of model" 00:35:39. He also frames the coming decade as "a discipline of context maxing" 00:31:03, because agents lack the context and evals to be token-maxed on biology the way they can be on verifiable math.

Perturbation Data Is the Labeled-Action Data of Biology

Hsu's central methodological bet is that perturbation data, where you know what you changed, is far more valuable than observational data. He draws on General Intuition's world models: "They would have a keylogger where they could record every sort of keypress. So they would be able to map the action space into the frames" 00:58:06. His translation: "perturbational data in biology is having the labeled actions for what happens before and after, as opposed to the majority of observational data that we have today. And that's been in many ways our fundamental bet across our institute initiatives" 00:58:57. This is why Arc is building what Hsu describes as the largest repository of perturbed single cells, "an internet-scale data set of perturbed single cells" 00:38:13.

Lab-in-the-Loop and Ground-Truth Evals Are How Virtual Cells Earn Trust

Hsu describes an active learning loop: "The model will make predictions zero-shot, you can test those in the lab, feed them back into the model, and then improve the model's predictions in multi-round loops" 01:02:01. He proposes a Nobel-grade eval: "Can a virtual cell model rediscover Yamanaka's reprogramming factors, right?" 01:04:41. His logic is that "you need to be able to recapitulate the most important discoveries of human-labeled biology in order to trust the machine ones" 00:26:06. Arc runs a public Virtual Cell Challenge, which has attracted "over 1,000 teams" and scores on held-out in-house data "so there's like no hacking of it" 01:06:39.

From Reading and Editing Biology to Writing It

Hsu traces his career arc from tools (CRISPR) to models (Evo). The tool bottleneck has moved from editing to design. Patrick Hsu: "A lot of the work that we've been doing in my lab lately has been developing large-scale genome editors beyond CRISPR, known as bridge RNAs and bridge recombinases, that allows to change up to a million bases in the genome at a time. The problem is once you start to be able to... you quickly run out of ideas about what to do, cuz we only know how to change things at the individual gene level" 01:14:07. The frontier is therefore models that predict how many genes interact. Hsu says, "We barely speak the foreign language of the genome or the cell" 01:15:38. As a proof point, he cites work from Arc and Stanford colleagues: "the first time that AI's been used to create entire viable organisms" 00:00:43.

Chronic, Multisystem Disease and Crowd Data Are the Next Medical Frontier

Hsu argues medicine is structured around acute, single-organ problems, which leaves long COVID, autoimmunity, and interoception-linked disease underserved. Patrick Hsu: "Medicine is really structured around treating single injuries, like acute care, or single organ diseases" 01:21:36. He sees a path in "scraping Reddit data" from niche patient communities and training models on anecdotal data to "design like very clear, like interventional, smaller-end trials" 01:22:25. He compares it to gym "bro science" that became standard without RCTs, and says that with the right oversight, the large population self-experimenting with peptides could generate "tremendous value" if blood profiling were done before and after 01:23:07. He also flags a coming Arc discovery on the gut-body-brain axis, "a profound interoception modulator" 01:30:37, which has gone unstudied because it is multisystem.

The Peptide Opportunity Is Gated by Chemistry, Which AI Can Attack

Hsu explains why peptides excite everyone but have few successes beyond GLP-1: "There are a tremendous number of bioactive peptides... Most of them don't have drug-like properties. So they're released spatially in very specific places locally... and they degrade really rapidly in circulation" 01:32:59. The known levers are lipidation ("attaching a fatty molecule to a peptide can really increase its half-life") and blocking protease cleavage sites 01:34:45. AI adds receptor prediction and, eventually, virtual-cell functional fingerprints per peptide. Sleep and wakefulness are cited as the example target class, with Hsu being "much less predicated on the drug class and much more on the product" 01:36:53.

Institutional Design: Arc as an Edison Lab Between Invention and Scaling

Hsu sees a structural gap: academia does invention, industry does scaling, and nothing bridges them. He says biotech companies "tend to get frozen in time relative to the time of their founding" 00:40:00. Arc is built to cover "the invention bucket, the engineering, and then the scaling bucket" 00:37:40, with the hope that "this will be an Edison lab for many companies over time" 00:47:38. Scientists are prefunded with flexible money to avoid "the Gordian knot of incentives that currently entangle grant funding success, single-author credit, and high-impact publications" 00:52:01. Hsu says Arc has just two flagship problems (Alzheimer's drug targets and virtual cell models): "We're not trying to solve 50 problems" 00:48:56.

2. Contrarian Perspectives

Research Taste Is the Appreciating Asset as Models Get Better

Most people assume AI commoditizes scientific intuition. Hsu says the opposite. Patrick Hsu: "I would say research taste and intuition is only increasing in importance, in relative importance. Like I would buy call options on it if I could" 00:00:04. His reasoning is that "biology is so high-dimensional, and the models today need so much inductive bias" 00:30:31, and the agents lack the evals and context to run unattended. He adds, "I really struggle between the modes of being long human and long AGI" 00:30:00.

The Smartest Person in the Room Isn't the Best Scientist

Hsu rejects the genius narrative of his own career. Patrick Hsu: "I grew up the dumb son. And so, I always had to try really hard" 00:05:22. In his PhD class, "none of them would have said that I'm the smartest. I'm certainly not, right? But I think I was one of the hardest working" 00:06:07. He pairs this with an unusual psychological formula: "deep insecurity plus sheer megalomania kind of have to meld together in weird complicated ways" 00:10:02. The corollary is that being "too smart" can hurt: "It's very easy to be too smart, right? ... You can talk yourself out of any idea" 00:55:35.

Don't Model Toxicity with a Digital Twin Until You Can Model a Single Cell

The industry hype around digital twins and drug-toxicity prediction is premature in Hsu's view. Patrick Hsu: "If you can't even model what happens in a single cell, how can you model multicellular systems and then kind of organismal... things like whether or not this drug is going to be toxic, right? You need to be able to model every organ and how they respond and interact across them" 00:32:31. He concludes that this "only works when you have something that's a whole, complete world model of biology."

Specialized Models May Be Tool Calls, or May Be Eaten by Frontier Models, and Nobody Knows

Against the confident "bitter lesson" narrative, Hsu is genuinely agnostic. He cites Periodic's open-source material-science model, which beat a frontier model at half the inference cost, though the frontier model with max thinking came close: "Are all kind of domain-specific or data-specific models going to just become tool calls for these sort of, you know, kind of super agents, or do you have to take that data and add it into training these big models?... That's playing out in real time" 01:09:06. His strategy is to care about the capability existing regardless of which architecture wins.

Heart Disease and Infectious Disease Could Both Effectively Disappear

Hsu makes a bold epidemiological claim. Patrick Hsu: "100 years ago, infectious disease was maybe the number one killer, right? Today, that's heart disease. We may see both of those go away" 00:00:43. The mechanism is programmable RNA interference with a quarterly injection replacing the daily pill, including multiplexed targeting of "both high LDL and high triglycerides" 01:17:10. That leaves "cancer, with Alzheimer's disease, and then... a really long tail of chronic autoimmune disorders" 01:18:28.

3. Companies Identified

Arc Institute

Nonprofit research institute co-founded by Hsu, partnered with Stanford, UCSF, and Berkeley. It builds bio foundation models (Evo) and virtual cell models, and focuses on Alzheimer's drug targets. Mentioned as the vehicle for institutional experimentation. Patrick Hsu: "This will be an Edison lab for many companies over time, right, on the basis of technology or insights that are much more well validated and ready to spin out at the right time" 00:47:38. On spinouts: "There's a number of other companies that are being incubated at Arc that will... become public to the world soon" 01:28:59.

Eli Lilly

Pharma company with a roughly $1 trillion valuation and the leader in GLP-1s. Mentioned as the clearest example of corporate capital displacing government as the engine of biomedical innovation. Patrick Hsu: "Eli Lilly spends more money on R&D than the entire country of Germany" 00:14:24.

Thrive Capital

Venture firm where Hsu is a venture partner. Mentioned for backing Isomorphic's spinout. Jackson Dahl: "Thrive spinning out Isomorphic from Google, with a crazy amount of capital" 01:27:09.

Isomorphic Labs

AI drug discovery company spun out of Google, cited as the heavily capitalized end of the nonprofit-to-commercial spectrum. Mentioned only by Jackson Dahl in the context of the Thrive backing above.

Notion

Collaborative workspace and presenting partner of the podcast. Hsu is on its board. Jackson Dahl: "Notion plays a Switzerland role. It is friendly with every AI, in that it integrates with whatever model, whatever tool, whatever agent you're using" 00:03:07.

General Intuition

Company building world models trained on gaming video with action labels, spun out of Medal. Mentioned as an analogy for why labeled actions (perturbations) make models far more efficient. Patrick Hsu: "They can take real videos, just normal videos, like in real life, and then they can pass it through this model that's frames in, actions out and then create an action space that would have created this video" 00:58:06. The company was discussed as Pim's company.

Medal

Gaming clip platform from which General Intuition spun out. Source of the screen-recording-plus-keylogger dataset 00:58:06.

Periodic Labs

Materials science AI startup. Mentioned for releasing mid-trained open-source models on proprietary lab data that outperform a frontier model at half the inference cost. Patrick Hsu: "Their own custom model is state-of-the-art, and it's half the inference cost" 01:08:36.

Arrowhead Pharmaceuticals

RNAi company. Mentioned for multiplexed RNAi targeting LDL and triglycerides at once. Patrick Hsu: "Arrowhead has been creating multiplexed RNAi that can target two things at the same time. So, both high LDL and high triglycerides" 01:17:10.

Alpha School

AI-assisted school model where students complete academics in about two hours a day. Mentioned as evidence that learning can be compressed with the right scaffolding. Patrick Hsu: "For just two hours a day, you can get done all of your academics" 00:08:14.

Math Academy

AI-assisted math training platform. Mentioned for cognitive scaffolding that lets students "speedrun a lot of things that normally take people many, many years into something that's 10% of the time or less" 00:08:14.

Scale AI / Mercor / Surge

Data-labeling and expert-data companies. Mentioned as the template for a potential "Scale or Mercor or Surge for bio data." Patrick Hsu: "What does the Scale or Mercor or Surge for bio data actually look like?" 00:31:51.

OpenAI

Cited as an example of a nonprofit-to-commercial structure. Patrick Hsu: "The OpenAI Foundation is the largest foundation in the world, owns 25, 26% of the company of the PBC" 01:27:53.

Runway, Genie, World Labs

Generative video and world model companies. Mentioned as examples of the "holy shit moments" Hsu wants biology to have. Patrick Hsu: "When you're watching progressions of image or video gen models... from Runway, or Genie, or World Labs, or whatever... every generation, the outputs of these things gets crazy and crazier" 01:10:35.

Google (Gemini for Argon / Antigravity)

Sponsor segment. Gemini for Argon is positioned for long-horizon work like code migrations, and the host used it to build an interactive compendium from 60 episode transcripts 00:50:02.

Regeneron / Intel ISEF

Regeneron now sponsors the high school science fair (ISEF, formerly Intel ISEF) where Hsu met Nobel laureates, a formative experience 00:23:23.

Evo 1 and Evo 2

Bio foundation models from Arc, Evo 2 being "the biggest bio-model basically ever created." Hsu: "They're only getting bigger" 01:00:17.

Fast Grants

COVID-era rapid research funding initiative that informed Arc's design. The survey result Dahl read from Arc's first blog post: "nearly 80% of Fast Grant recipients said they would change their research program, a lot if all of their existing funding could be spent in an unconstrained fashion" 00:52:01.

University of California

Mentioned for holding valuable patents on engineered strawberries, evidence that plant genome editing is commercially real. Patrick Hsu: "I think many of the University of California's most valuable patents are for engineered strawberries" 01:19:48.

4. People Identified

Co-founder and core investigator at Arc Institute, Stanford assistant professor of pathology, Thrive Capital venture partner, and Notion board member. Subject of the episode. Pioneer in CRISPR and bridge recombinases. See his self-description: "In some sense, I work three jobs, right?" 00:11:31.

Feng Zhang

Hsu's PhD advisor, CRISPR pioneer. Cited for teaching simplification as a research skill. Patrick Hsu: "One of the most important things I learned from my PhD advisor was how to simplify. Biology is really complicated, and most projects and teams get mired in endless complexity" 00:55:35, adding that "Feng has such a unique talent in distilling" 00:56:06.

Shinya Yamanaka

Nobel laureate who discovered the four reprogramming factors that turn any cell into a pluripotent stem cell. Hsu's proposed ultimate eval: "Can a virtual cell model rediscover Yamanaka's reprogramming factors?" 01:04:41.

Felix Horns

Arc and Stanford faculty who invented engineered protein nanoparticles that bud from cells and export RNA. Patrick Hsu: "One of our brilliant faculty, Felix Horns... has invented these nanoparticles that can be budded out of cells" 00:36:05.

Isha Jain

Arc and UCSF investigator who developed small molecules that tune oxygen binding to red blood cells. Patrick Hsu: "So you can do like altitude therapy in a pill, right? That's like, you know, insane. And she's using it to treat neurodegenerative diseases, to treat cancers" 01:28:39.

Hani Goodarzi

Arc investigator training models for designing RNA medicines. Patrick Hsu: "Hani Goodarzi at Arc has been training all kinds of models for designing RNA medicines" 01:28:59.

Roger Bannister

First person to break the four-minute mile, used as the example of existence proofs triggering rapid replication 00:17:39.

Usain Bolt

Cited by Dahl as another example of rivals closing in once a record is shown to be achievable 00:20:09.

Dudley Herschbach and Richard Roberts

Nobel laureates Hsu met at Intel ISEF. Their encounter taught him that "research is an even playing field, kind of by definition. No matter how great or accomplished you are, we all have the same shot" 00:00:04.

Hsu's Brother

Hsu's brother graduated from college at 15 and started a Stanford PhD, an example of the accelerated-education environment Hsu grew up in 00:21:39.

Hsu's Grandfather

His grandfather's Alzheimer's at age 11 was the origin of his biology career: "That was deeply moving for me, and it created the idea of decline and death became extremely visible and tangible" 00:22:30.

Tamer Wight, Charles Brecque, Tyler Cowen, Nabeel Qureshi

Past Dialectic guests cited in the sponsor segment for the idea of borrowing taste from those around you 00:52:01.

Pim

Associated by Jackson Dahl with General Intuition, which spun out of Medal 00:58:00.

5. Operating Insights

Prefund People, Not Projects, and Select Like a Founder

Arc's research model is to give flexible, long-term funding to scientists and select them as you would founders. Patrick Hsu: "You need to be able to convince and coax, and give good talks, and recruit people... You need to be the chief scientific officer, the chief marketing officer... You need to be able to, against your top-line revenue, decide how you're investing your resources into different areas" 00:26:06. Hsu says even a funded investigator "still has to do BD" 00:54:21. The takeaway for operators is that removing the fundraising constraint doesn't remove the need for allocation, recruiting, and storytelling skills, so screen for them explicitly.

Cap the Number of Flagship Problems at Two

Arc deliberately limits its ambitions. Patrick Hsu: "We ultimately have two flagship problems. We're not trying to solve 50 problems, right? We're trying to find drug targets for Alzheimer's. We're trying to make these virtual cell models" 00:48:56. Combined with the idea that interdisciplinary teams must align on "the big-picture vision" while "how we get from A to B is going to be weird and murky" 00:06:07, this is a template for running a research or product org with a clear north star and freedom in the path.

Make the Discovery Process Fun

A direct operating principle for high-failure environments. Patrick Hsu: "Science is really hard. It's such a long journey, and it's mostly failure. So, if you don't have a way of bringing the energy and the vibe and the vitality, it's pretty miserable. You need to figure out a way to make the discovery process fun because you're mostly eating shit" [01:00:00 range, from the episode intro]. Hsu frames boldness and risk-taking as "survival traits."

Use Public Competitions to Build an Ecosystem and Honest Benchmarks

The Virtual Cell Challenge gets an outsized community (1,000+ teams, many of them engineers outside bio-ML) to attack Arc's problem, while using held-out in-house data to prevent overfitting. Patrick Hsu: "It's an industry-wide problem, and kind of a holy grail in the field that we should have this public, open, and kind of transparent evaluation of all of the models that everyone's trying, so it's not just overfitted on the two or three specific evals that they're testing in their paper" 01:07:34. Operators can borrow the pattern: publish a benchmark with private test data, and let outsiders do the search for you.

Use "Business Genetics" as a Comparative Framework for Evaluating Deals

Hsu applies a scientist's pattern recognition to investing. Patrick Hsu: "I think a lot about business genetics... different types of companies have different genetic substrates and profiles and signatures, and they interact in weird and nonlinear ways to create emergent phenotypes, much like in cell biology" 01:25:28. He does "comparative genomics" across defense, robots, healthcare, RL environments, and cybersecurity, each with different markets, competitors, and players. The tactic is to systematically catalog the variables of each sector rather than applying one generic playbook.

6. Overlooked Insights

Verifiable Domains Get Unleashed Agents First, and Biology Isn't One Yet

Mentioned in passing, this explains where AI progress will be fastest and slowest. Patrick Hsu: "If you take something like math, right, where we're able to solve Millennium Prize problems today, right? We have this verifiable problem, and we can... unleash like 10,000 agents to simultaneously try to solve a problem from every possible angle. This is not possible in biology today" 00:31:03. The consequence is that value accrues to whoever builds the evals and context layers (the "Scale or Mercor or Surge for bio data") that convert biology into a semi-verifiable domain. That is a picks-and-shovels opportunity that most people looking at AI-for-bio are not underwriting.

Clinical Trial Regulation and Human Data May Be the Real Moat

Hsu drops a single line that reframes the entire bio-AI value chain. Patrick Hsu: "There are different types of tokens, if you will, for the models. There's like a protein structure token, there's maybe a DNA sequence token, there's maybe a cell perturbation token, we also need human tokens... the most valuable thing in the industry is data on how a human body will respond" 01:18:52. He then says innovations in "clinical trial, sort of regulations and enrollment and recruitment and end points in order to be able to test things faster in people" will be "an incredibly important trend." If design gets cheap through AI, the scarce, defensible asset becomes human response data and the legal infrastructure to generate it fast, which could favor companies building trial platforms, patient communities, and longitudinal data collection over those building yet another design model.