Last week, OpenAI acquired Orchestra, the operational platform I’ve been using to build my company’s drug program (congrats Ashoka, Shaan, and Team!). On Tuesday, Anthropic released several vignettes of how Claude helps teams accelerate drug development. In the blog post, Anthropic noted its acquisition of Mithril Medicines, a virtual, AI-leveraged single person biotech led by Cody Tranbarger (congrats Cody!), former EiR at Atlas Venture.
Why are the frontier AI labs making moves into biotech?
Pharmaceuticals are a large market. Eli Lilly’s tirzepatide will do over $100B in sales next year. But as I’ve explained, I think the primary motivations aren’t financial.
Anthropic CEO Dario Amodei’s Machines of Loving Grace leads with “Biology and Health” as a headline positive application of advancing the AI frontier. OpenAI CEO Sam Altman has backed several biotechs including Retro Bio. Both labs have released agentic AI products for scientists (Rosalind and Claude Science). Selling tools to help Pharma discover, develop, and market drugs is a durable business. But, I suspect motivations beyond greed and beneficence drive these acquisitions: the frontier labs are using drug discovery and development to prove out new ways of human-AI collaboration to improve physical-world productivity.
Drugs are durable gifts
Pharmaceutical innovation is slow
Drug development is rocket science
Most R&D is more like managing a drug project than a codebase
The medicine cabinet of progress
Today, much of medicine is service provision. The problem with providing services is that, perhaps until recently, scaling expert services provision required producing more experts. As physicians retire, society must continually invest in the supply of new physicians. But, that’s not true for drugs. Once a drug has been discovered and developed, its benefits are indefinite. Anyone who’s had a major surgery today benefits from the early 20th century discoveries of anesthetics and antibiotics.
Over time the pharmaceutical industry has progressed from developing new treatments for common bacterial infections, to complex enzyme replacement therapies for rare diseases, and functional cures for devastating conditions like Trikafta for Cystic Fibrosis or Harvoni for Hepatitis C. While drugs for CF and Hep C remain expensive, in the near future generic competition will make provision of these therapies essentially costless. Decades of work from academics and industry transformed HIV from a death sentence into a chronic condition manageable with a twice-yearly injection. That’s not going away. If one wanted to have frontier AI produce a durable gift to humanity, a slate of new drugs - antibiotics, antivirals, and rare disease cures - would be among the most valuable contributions frontier AI can make to benefit humanity.
Developing drugs takes time
I see many computer scientists frustrated by the pace of progress in biomedicine. It’s understandable. Next month is the fourth anniversary of ChatGPT’s release.
Since 2022, we’ve seen LLMs go from struggling to count the “r”s in “strawberry” to proposing novel solutions to Millennium Prize Problems.
Eli Lilly’s next-gen GLP-1 triple-agonist retatrutide entered Phase 1 in 2019. Lilly finished Phase 3 studies in 2026 and plans to submit a BLA in early 2027. In advance of its release, eager consumers are buying counterfeit tirzepatide labeled as retatrutide from Chinese ex-fentanyl dealers. Lilly’s first legal commercial sale of aetatrutide will happen in late 2027/early 2028.
The timescale of drug innovation is necessarily long. If the endpoint is patients losing weight, not developing Alzheimer’s, or not getting cancer, you need time. Patients will want to know that the drugs they take are safe and effective. Even when efficacy comes quickly, you can’t dose twelve men for one month to get a year of chronic safety data. We have discovered biomarkers predictive of efficacy. I hope AI tools help us discover and develop more! But, good biomarkers like LDL-C, IOP, and BMD rely on correlations with outcomes data established over years.
Can frontier AI accelerate drug discovery and development? Both labs would say yes. I agree. Yet, to date, much of the enthusiasm around AI impacting biomedicine has focused on neat demos of “designing” “binders”, enabling data “analysis” slop, or processing mountains of low-content static data. None of these limit progress in biomedicine.
Biochemists in the audience will be familiar with the concept of a “rate-limiting step” in a metabolic pathway. The process can’t proceed faster than the slowest enzyme in the pathway. What is the rate limiting step in drug discovery and development?
I’d argue the rate-limiting step in drug discovery is the validation of therapeutic hypotheses. Animal models of disease range from noisy, but weakly predictive to outright misleading. Imagine having Claude/Codex write your application, but with bad, hard-coded tests, so you get to enjoy unexpected behavior live. That’s the problem of drug discovery. Except, the validation loop takes three to ten years.
In drug development, the rate limiting step is again… validation. If tomorrow Claude spits out the chemical structure for a cure for Alzheimer’s, we’d know if it works in about a decade. Seven years if we’re fast and optimistic. Unfortunately, we’re rate-limited by disease progression. Pray it works in the 5xFAD Alzheimer’s mouse or you may never pass Go.
So, if AI doesn’t accelerate the rate limiting step - validation - how can AI accelerate drug discovery and development?
Well, I guess that’s partially what the frontier AI labs want to resolve. What I neglected to mention is that most drug candidates advance nowhere near their maximum theoretical rates. The average drug you take likely died at least once. Keytruda famously sat on the out-license list before Merck reactivated the program after BMS’s Opdivo success. As we learned during the 2020 pandemic, having committed capital for development and assurance the drug will sell can dramatically accelerate development timelines.
I think the labs are converging on the realization that the biotech industry already operates near the rate-limit for the process of creating next-in-class, or me-too/me-better drugs. These drugs are engineered from established templates, often with little thought. Yes, in the future AI “design” tools may shave weeks off the multi-year timeline while improving quality.
But, first-in-class drugs are a different question. To date, there’s been little focus on accelerating the discovery and development of transformative medicines. There, I think, the labs believe drug hunters can leverage artificial super-intelligence to find more first-in-class faster and cheaper. Here, iteration speed, intelligence, and risk management provided by highly-capable AI meaningfully improves the probability of success and financial returns for investors.
Developing drugs is rocket science
Drug development shares structural features with rocket science. Both drugs and rockets require enormous quantities of human effort and capital. Every detail matters. Teams of experts build rockets. Design-build-test cycles take months-to-years. Once a candidate is nominated, the fundamental constraints are locked. The ability to alter it and iterate narrows. You can spend years making progress, but the ultimate test (launch or landing) is binary. Blow-ups are hopefully infrequent, but not unexpected.
Now, for rockets, the constraints of the system are fairly well understood. One can model many of the constraints in the design phase before commissioning the build and test. In biotech, we aren’t that lucky.
The constraints of the biological system, and even the relevant components, are largely unclear. The surrogates we use to guide engineering of our molecules - animal models - are at best noisy proxies for the desired clinical outcomes. Once you have an animal model that validates a mechanism in an indication, you have an animal model that might predict activity in an indication - and nothing more.
From experience we derive rough rules (e.g. Lipinski’s Rule of 5, antibody developability) to guide our discovery work. Biotech is not so much shooting at a moving target as it is shooting in what you hope is the general direction of a target for which you have a rough outline. Ultimately, experience and expertise hold disproportionate value in the drug discovery and development process. This is why drug discovery has historically been a team effort. Success requires integrating expertise across multiple domains. It’s unlikely you want your in vitro assay specialist drafting Module 3 (CMC) of your IND. But, even for a team of experts, taking on the unknown is challenging. First-in-class drug discovery is as much an exercise in discovering the design constraints as it is in finding the right lead molecule.
So, how does one solve a difficult problem with limited understanding of scope? Learn quickly. One of SpaceX’s key innovations was vertical integration to accelerate iteration and capital efficiency. SpaceX extended the Pareto frontier for rocket engineering.
But biotech shouldn’t copy the mechanism. SpaceX vertically integrated because the rocket supply chain was thin, slow, and priced for government contracts. In contrast, the drug development supply chain is deep, competitive, and modular.
The AI-leveraged, virtual single-person biotech like Mithril extends the drug discovery Pareto frontier by leveraging unique biological insight while integrating across the existing drug discovery and development supply chain. Cheaper, faster learning guided by superintelligence working with expert Contract Research Organizations doing what they already do well1.
I’ve seen AI tools unblock our programs by enabling me to do things I wouldn’t have imagined possible two years ago. As AI capabilities have improved, we’ve seen our growth needs shrink while improving program quality and saving money. Today’s models can’t solve everything, but there’s a path to re-organizing today’s virtual biotech to fully leverage AI capabilities.
Efficient drug development is, in part, good program management. Whether you’re building drugs or rockets, you’re doing science on a timeline within budget. That’s where tools like Orchestra can help ensure programs remain on-track and under budget. Simple decisions made reflexively can have knock-on effects that transform success into failure. Effectively managing program context is difficult for any individual, human or AI. Presently most projects are set-up to be maximally expert legible. Part of Orchestra’s innovation was establishing an underlying data model that makes the process legible to AI agents. A country of geniuses in a data center working in parallel to manage your program and solve your problem may help!
Like rockets, building drugs takes capital. Now, capital efficient virtual biotechs aren’t new - although perhaps presently out of favor in our current $200M China-license “seed round” moment. But, if you need money to build drugs, you need investors. And investors need you to make them money. Simply put, for a fixed investment outlay, an ultra-lean AI-leveraged biotech like Mithril can advance a first-in-class program far more efficiently under uncertainty than before. Extending the drug discovery Pareto frontier means bigger multiples for investors.
Can we do better? Can we , there are efficiencies from running multiple programs in parallel. Aside from portfolio hedging, multiple programs often share costs. To date, running multiple programs has required a larger team. But, as AI progress gives drug hunters more leverage, we come closer to extending the Pareto frontier for first-in-class drug discovery again. I’m excited to see Cody (aka @Prof_Oak_) and Anthropic build out the Pipeline-in-a-Person.
Drug development as a lesson in tech diffusion
This is the most speculative part of my thesis. I have no inside information from the labs. These are speculations I’ve developed from using AI to run my own lean, AI-leveraged first-in-class drug program.
Software functionality is largely verifiable with little time. It appears part of pure math might be too. We see breakneck progress when modern RL techniques can work towards a well-specified reward with a large number of examples. But, as I discussed above, drug discovery is not. Drug development certainly is not.
So, is the economy at large more like software or drugs?
I’d argue most work falls somewhere in between. The problems of drug discovery and development - long verification loops, broad scattered context, a lot of judgement calls - recur across the economy. But, in drug discovery and development, tolerance for error and magnitude of uncertainty is amplified. If Claude and Codex can “solve” drug discovery and development, they can likely solve most business problems.
Drug discovery and development is particularly challenging because it substitutes surrogate QA/QC for software-like verifiability. How do you know you made the right molecule? Well, that’s not easy. You have a complex set of characterization assays and you ensure a consistent process. Ultimately, the process is the product. If the process breaks or the surrogate QA/QC fails, your drug candidate might be fine. Or not. You don’t know. So, do it again.
I suspect many engineering R&D problems look more like drug development than writing code. If you want AI to become broadly useful for R&D, you need to operate in environments with long verification loops and substantial uncertainty. CAD tools certainly aid many engineering disciplines at a micro and meso scale. But, if we want to leverage AI to bring online low-cost desalination to solve the American West’s water shortage, we’ll need AI tools that can process data from disparate sources, model the complex interactions of project dependencies, maintain provenance of inputs, and interact with long rosters of stakeholders. That sounds a lot like… drug development.
What about using AI to speed automotive development? Great! Frontier AI tools can likely model a relatively advanced electric vehicle in sufficient detail to accelerate automotive design today. What about manufacturing that car? Well, welcome to managing vendors with high precision, modeling hundreds of interdependent tasks, and integrating disparate human-legible data sources with millimeter-to-micron precision.
When AIs meet real-world long-loop verification, they will need different ways to manage information, model systems, and help humans make decisions. If AIs can accelerate the discovery and development of novel medicines, then lessons learned from building gifts to humanity may translate into AI-enhanced productivity across other complex physical sectors that appear recalcitrant to AI acceleration.
Realizing the productivity gains from AI
I suspect many organizations are optimized for maximum human legibility. To help AI agents integrate context, process information, and make decisions, we will need to alter how we organize our businesses.
For example, in the virtual biotech - a simple no-revenue business familiar to me - the companies are organized for maximum expert legibility. The connections between experts inform and guide decisions. But, are interacting information silos the optimal way to run an AI-leveraged single-person biotech? What if you have 90th percentile AI-agent experts responsible for each function? I don’t know, but I suspect alternative organizational structures will be necessary to unlock the full potential of the structure.
For example, in my experience, context rot across long document corpuses degrades LLM output quality. Breaking down literature or complex CMC documents into smaller pieces for LLM compaction works well, but when un-cited outputs are returned you have to be skeptical and verify the claims. So intelligence providers like Edison Scientific do a great job of making models cite their sources. Doing this iteratively burns through tokens. Can we have expert agents share context while maintaining provenance of claims? How do we preserve human legibility while maximizing AI-agent capability?
These are the questions we’ll have to solve to enable highly effective human-AI collaboration at the frontiers of R&D. I’m excited to have the chance to help solve these problems while bringing transformative first-in-class medicines to patients. Looking forward to the solutions people bring.
The biotech ecosystem is more like electronic hardware than rockets or nuclear plants. Electronics companies may design and engineer their product, but most inputs are off-the-shelf. There are many Contract Research Organizations (CROs) that specialize in single functions. “Assembling” a drug development program is similar to assembling a consumer hardware product from off-the-shelf CRO services. Here, trading money to save on CapEx and reduce error is often favorable for both parties relative to vertical integration. But, that depends heavily on the specific function and cost-structure of the program.
Great read! One fun thing to think about is that once all of these AI improvements have a decade to play out in the first-in-class space, I’d wager that there will be something that humans, via intuition or otherwise, were doing at the theoretical optimal rate. Maybe many things! I wonder what those are…