
We're building agents for autonomous science. Closed-loop, faster iterations, more discovery, less time. Massive human scientist + AI scientist collaboration.

Autonomous science. Founder and CEO at Potato (@readysetpotato). Former neuro at Brown, NIH, UCSD.

We're building agents for autonomous science. Closed-loop, faster iterations, more discovery, less time. Massive human scientist + AI scientist collaboration.

we aren’t close to seeing what’s possible when frontier LLMs are trained on bio data at scale I think there’s (rightly) consensus that the models can’t keep getting better at bio without better data. BUT - I think we’re underestimating the gains left on the table by scaling compute + training on the existing data we have available today at some point gains will saturate - the question is when. imo we’re further off that many think
Claude was trained on protein sequences, at scale. We gave 20 language models the start of little-studied bacterial proteins and asked for the next amino acids. Only the latest Claude models predict them. Not GPT, not Gemini, not seven open-weight models.

I want one
Gene therapies can be delivered in a protein shell called a capsid. Mythos 5.2 found a new way to predict a whole capsid’s structure. In our tests, it accurately modeled 28 of the 34 capsids it attempted. Biologists could use it to design capsids faster and more cheaply.

Wow two years ago. My heart is melting
Found this by my son’s backpack. These are the moments

Everyone’s building autonomous labs. Does anyone know what data the frontier labs need?

I’m actively looking to match ambitious scientists with funding. Through some combination of my writing, podcasts, in-person events, scientific research and other activities that my friend @markwbudde would call “hypernetworking,” I’ve managed to meet a lot of wealthy individuals interested in funding ambitious scientists. I’ve had my own work funded this way, and though it’s not sustainable enough to replace the work of larger funding bodies of the NIH, for the right kind of person it is just what they need. I am writing this post because I am helping one of these groups – Fulcrum Science – find projects worth giving money to. While I’m pretty flexible, I’m particularly interested in hearing from people who are actively working on open sourced LLMs for frontier biology. Plenty of groups have shown what is possible with frontier, closed-source LLMs. I’ve also seen it in my own work (which I hope to talk about soon). I also think that the absolute frontier of intelligence isn’t required to do interesting things in science. I’d like to help match an existing lab – academic or industry – with a decent sum of funding to showcase what open sourced models can do in biology. Pitch me what you’d do. Please note, I am not looking to hear about any projects that involve Claude or ChatGPT for any reasoning, experimental analysis, or embodied automation stuff. Open source only! I’m also open to hearing about other things within biology. If you have an idea you have struggled to get funded, and you have a good reason for me to believe you, please reach out. My email is eryneym [at] gmail.com. You can also DM me on Twitter.

Been comparing notes with researchers at labs and data vendors, The new data money at frontier labs is going off-screen. Coding data they can mostly make in-house now. Wet-lab runs, robot hours and a real company's inbox they still have to buy Anthropic alone paid data vendors $757M in the first 9 months of this year. Here's where it's going: 1/Bio is the hardest to fake Anthropic built its own wet lab and bought Coefficient Bio for ~$400M in stock OpenAI ran GPT-5 through 36,000 experiments in Ginkgo's cloud lab, and got 173 scientists to write a bio benchmark. Best model: 36% What they're really buying is bench time 2/Robots are growing fastest OpenAI reportedly has ~100 contract teleoperators working in shifts around the clock @xdofai sells to 20 customers, several of them frontier labs. Mecka hit a $100M run rate in June video is easy now. Touch and force are the gap 3/Then the white-collar stuff @HalluminateAI builds PE due diligence environments for 4 of the top 5 US closed labs. Best model gets 51% Google paid $10M for bankrupt Spirit Airlines' emails and Teams messages, and outbid Mercor A dead company's Slack has a price now 4/Even evals turned into a purchase. Anthropic and Accenture each plan $1B+ over 5 years for evaluators embedded at Anthropic So the scarce thing is access: a bench, a robot, a real company's history

Cell-type and pathway-specific excitation and inhibition of neural circuits has massive potential. This is inspiring!!
A major milestone for Chemogenetics. DREADD technology demonstrates therapeutic efficacy in humans. A new study published today in NEJM demonstrates preliminary safety and therapeutic benefit from DREADD-based chemogenetic treatment of refractory epilepsy, marking an extraordinary milestone for neuroscience and precision medicine. nejm.org/doi/full/10.1056/NE… The approach combines targeted AAV gene delivery at the epileptogenic locus with very low, subthreshold doses of clozapine to activate engineered DREADD receptors and modulate neuronal activity. Congratulations to Bryan Roth, whose pioneering discovery of DREADDs made this field possible, and to the investigators of the NEJM study for bringing this technology to patients. This achievement has a particularly meaningful connection to our work published in Science Magazine in 2017. science.org/doi/10.1126/scie… We showed that clozapine, rather than CNO itself, is the active mediator of DREADD activation in vivo and proposed that subthreshold doses of clozapine could be used to activate DREADDs in humans, anticipating a practical path toward clinical translation. At the time, many in the scientific community viewed these findings primarily as relevant to the design of preclinical DREADD experiments. We saw a much broader opportunity, translating Chemogenetics into human therapeutics. Nearly a decade later, that vision is now becoming a reality. Our 2017 study was supported by the research funding through the National Institutes of Health (NIH) and The National Institute on Drug Abuse (NIDA), a prime example of how fundamental basic research can ultimately enable transformative clinical applications. This is an exciting beginning, but larger, well-controlled studies are essential to establish the magnitude and durability of therapeutic benefit and long-term safety. Chemogenetics is moving beyond an experimental neuroscience tool toward a potentially transformative neuroscience platform. We are only beginning to realize its potential.

Whoa this is a huge deal!!
Dreadds work in humans nejm.org/doi/full/10.1056/NE…

AI agents enter the wet lab. In new @Nature N&V article, I discuss recent advances in agents running automated labs and ideas on how to make these systems more reliable. nature.com/articles/d41586-0…

Elon’s talking about drug discovery. What a world
Eli Lilly has some banger drugs coming out!

Out and about recently. Randomly stumbled across someone with a Nobel in their backpack

Neurotech is Humanity’s Next Frontier! I’ve been investing in neurotech since 2021 and working with teams building BCIs, NeuroAI, and connectomics. It now feels like AI did in ~2019. We are entering an extraordinary era: neurotech already restores functions, it can enhance our capabilities, and will evolve what it means to be human. PLC Neurotech is the fund backing the extraordinary founders and defining companies of this booming industry.

Claude Science looks to me like a bid to own the economic version control system for life sciences. The logical endgame is Anthropic financing hundreds of drug programs in parallel for the biopharmas that leverage its harness. In return, Anthropic will ask for clinical data rights and use the data generated to 1) improve their ability to finance the next program and 2) eventually feed a God Model that allows them to create the holy grail: an “AI drug factory”. You can think of a drug program as an exercise in making a sequence of complex decisions under uncertainty: specifically, decisions around how to allocate incremental experimental capital to maximize expected value across safety, efficacy, commercialization, etc. A growing share of that experimental capital will be compute-mediated over time (e.g, Claude generated roughly 35,000 antibody designs in a month for Divergence Labs). But the most consequential increments will still go toward compiling therapeutic candidates in (very scarce and slow) human biological computers, which is where both the cost per bit and the consequentiality per bit of information are highest. Whoever owns the system that originates and integrates the most information upstream and downstream of those allocation decisions (in this case Anthropic) becomes the best suited to start making them (or at least gains a differentiated understanding of the relevant signals necessary to make those decisions well). Some examples of the experimental capital allocation decisions they are starting to help generate / facilitate and version: 1) Most importantly: Acquiring Mithril, where Cody Tranbarger (@prof_oak_) is massively levered by a swarm of agents that help him identify, evaluate, and ultimately select a much larger # of drug programs to run in parallel than is typically possible for a single drug hunter. He calls this a “Pipeline-in-a-Person” 2) Bristol Myers using Claude Science to analyze their proprietary data (guessing preclinical and clinical) to create optimal clinical study plans and weigh the study’s chance of success. If that program is actually executed, they should be able to calibrate how well those predictions turned out and in what sense they were right or wrong. 3) The Column Group spinning up Pleiades: a biopharma company with a mandate to “run many more—and more complex—drug programs than is usual for a small team.” Similar to the Formation Bio’s of the world. They are using Claude to evaluate target-disease pairs and find non-obvious, potentially valuable combinations. I believe Anthropic is running their own drug programs not to compete with pharma directly (yet), but instead to generate the learning signals that would allow them to become the single largest (and potentially best calibrated) underwriter / financier of drug programs. These financed programs would eventually create the data distributions necessary for them to create an AI drug factory (and start competing with pharma). Specifically, I could see Anthropic creating the first real instantiation of @AndrewWLo’s elegant Megafund model. TLDR on Megafund (proposed in Nature Biotechnology 14 years ago): any single drug program is a terrible bet on its own, but if one entity funds a large enough portfolio of them, the odds that a few hit become predictable enough to finance the whole thing with debt. Lo et al. proposed doing this by issuing “research-backed obligations” (bonds collateralized by the portfolio and its IP) alongside an equity tranche. Ofc, the pool of available capital that buys debt is far larger than the pool that buys biotech equity, therefore this would unlock capital from pensions, insurers, endowments, etc. Their simulations of $5–15 billion oncology portfolios estimated average annual returns of 8.9–11.4% for the equity and IMPORTANTLY 5–8% for the RBOs. This would be a fixed-income-like return that many new sources of capital would get excited about. For some reason this strategy never took off, and part of the reason is that it’s pretty hard to find a large basket of drug programs that’s sufficiently uncorrelated. But I also think it’s because there wasn’t really an incentive for any well capitalized player to push through all the schleps and actually unlock the capital sources to this end. I think Anthropic has that incentive. They have a structurally higher willingness to finance a drug program / make this structure real, because, by doing so, they would capture two return streams: 1) the potential economics of a blockbuster drug and 2) the widely amortizable value of what the associated series of experimental capital allocation decisions teach its models! In return for providing biopharmas cheap financing, Anthropic will ask for clinical data rights (including rights to learn from decision histories, negative results, protocols, etc.). Just selling tokens / the harness doesn’t give them the right to learn all of these things, but providing financing likely would. This would allow them to make more well-calibrated experimental capital allocation decisions over time and eventually feed the God Model that “cures all disease.” This would all also be quite charismatic: your retirement money funds the experiments that might let you enjoy a much longer retirement!
Over the past few months, Anthropic scientists have engaged in deep research partnerships with six organizations developing new medicines. We've also begun drug discovery efforts of our own. In doing so, we’ve learned much more about how Claude can support the many stages of drug development. In this blog post, we share some early results from these efforts, and describe how this work fits in with our wider ambitions in the life sciences. anthropic.com/research/early…

Dreams of AI ending aging and of AI killing everyone through a purely in-silico-designed pandemic both rely on this fallacy. It's understated here how much we lack ground truth data about what's happening inside of cells, or sensors to get that data. A single cell contains billions of protein molecules, across at least thousands of types, and at least trillions of interactions per second. Think we solved protein folding? We've made enormous progress predicting static protein structures. But these proteins constantly change their conformation and interact with one another, often in ways we can't observe or model accurately. Our best sensors capture a miniscule fraction of the activity in a living cell, orders of magnitude short of what's happening. We actually lack the ability to perceive what's going on that would be required to build such a bottoms up simulation. Let alone to scale this up to an organism or a global population. We know we have an expanding universe not because of breakthroughs in theory or computation - but because Edwin Hubble got new data by looking through the Hooker telescope. Possibly the greatest mind, Einstein, was wrong in his model of the universe because he lacked the right instrument to bring him the right data. That's where we are in biology. Our sensors are still so crude. No amount of pure computation or AI can build a perfect model. We need an instrument revolution in biology, not just a computational or intelligence revolution. AI can generate hypotheses, but without sufficient ground truth data, we can't know whether our simulations are modeling reality.
If you had an exponential amount of compute you could solve simulate biology exactly and wouldn't need to run any experiments. Some of you will think "of course, but that's an unrealistic assumption" and others will think "no, you need experiments even with that much compute. That's just How It's Done". You would of course need evaluations once you have a solution, but you don't in principle need experiments to find it: Experiments are a way to save compute, dating from a time when it was much more expensive. And I think the presentation is exponential but real systems aren't irreducibly exponential. AlphaFold worked, after all.

Largely agree with these thoughts, and have some perhaps discordant points to add. - Solving biological diseases in the next few years will benefit mostly from the usefulness of the language models, substantially more than molecular ML. - Proponents of the bitter lesson for biological models (not LLM) aren't wrong, just (possibly pathologically) too early. We don't have enough data, and the data we can collect quickly is probably not the right kind (we may get lucky with biology being easy, but it's usually not). - I think just scaling models and public (passively accumulating) data is a cost ineffective direction (it's like scaling dense NNs on the internet). Today's model outputs are as good as GPT2. A single bioinformatician with linear regression, sometimes even without any additional data, can beat these models in many applications (people rarely benchmark AI advances against a simple Potts model). The most promising approaches for molecular models are innovating heavily on architecture (e.g. Baker Lab, Alphafold, etc). This may automate/speed up with LLMs. - Few companies (hopefully Dyno included) are betting that we've triangulated enough pieces from multiple experimental channels to predict (with scale) how things translate in humans. Having worked on translating in vitro things to primates, I'm very cautious with the idea that current models, trained on current public data, can do anything more than manipulating already working solutions, either explicitly or through implicit remixing of what has worked. It may still work, but the underlying bets offer no specific reason why there is any epistemic risk advantage other than being "made with AI". - Short term, there is more promise in combining molecular models with LLMs rather than scaling them. The two sources of lift in AI discovery elsewhere has been (1) Breaking Silos: AI synthesizes a bunch of knowledge/expertise that is rarely present in the same room (2) Data mining : AI pushes through volumes of data, with an intelligent eye, that humans or simple algorithms simply can't scale or adapt to. I'm curious what will come out of multimodal training of LLMs and biological data, which seems to be something the big players are pursuing. I think latent reasoning in the language (or better yet, + code) domain will reveal a lot more biology than scaling on nucleotides or proteins alone. - My view is that the next opportunity of combining lab and AI will be massive-scale hypothesis testing with LLMs. Currently a multiplexed assay generally tries to test a handful of ideas because the humans are bandwidth limited in designing properly powered experiments. LLM's can scale this very well. I have no doubt lots of good news will come on the impact of AI in biology in the next five years. I can't tell what paradigm will win, but my prediction is that it will not be the same paradigm imported from the LLM land (which will get a ton of funding anyways, because it's easy to pattern match for tech VCs).
There’s a really pronounced divide between people who are on the AI side of things and people who are on the biology side of things. Even among people working in AI, the divide persists based on where they’re focused. The AI side, adherents of scaling laws and the bitter lesson, believe that AI getting better means AI being more capable of novel discoveries in medicine (that go beyond discovering new drugs in well understood targets or speeding up dry lab work etc). The biology people think the idea of a bitter lesson in biology is kind of ridiculous and that regardless of AI’s capabilities it’s going to be at least 5-10 years until you could hope to reliably simulate human toxicology in silico (right now AI/ML can’t even reliably simulate in vivo studies like binding affinities or functional assays). Even the most AI pilled of them, who will list all the areas that AI can reduce the time spent on various areas, don’t think the concept of AI discovering novel targets or generating new understanding of causal biology is realistic in the near term. Most, in fact, will readily characterize the assumption that even an “ASI” level model could find cures hidden in our current understanding (without expanding that understanding with significantly more biological data) as patently false. This was a big part of why we wanted to experience it ourselves by using AI for a (very) small portion of what it can do in biology. It was both mind blowing and revealed some poignant insights into how different the problem of making AI better at something like coding is to the problems of biology. Even if it unlocked the ability to generate new molecules at a stunning pace, its simulation (in silico) of binding only had 49% accuracy. And if we wanted to improve that, we’d need to feed it back the data from the lab - data that took about 30x as long to get as it took to generate the molecules. And these were tests in a test tube (turn of phrase), imagine how long it would take if we wanted to improve the model with data that required recruiting humans and monitoring their reactions. Human biology is very complex, and adding a drug - especially a novel drug and doubly so for a novel drug acting on a novel mechanism - means mapping interactions that have nth-order-effects that we don’t have the data to predict. Virtual cells have made advancements enabled by AI, but the perturbations equate to a small fraction of the possible. This is before you even get into curing disease, especially uniquely human diseases. Brain function involves cellular identities that are unique to humans, meaning animal studies are of little use. How long does it take to understand what the side effects of chronic administration of a drug in humans are? AI can speed up analyzing the data, or designing the trial or even getting it approved…but it can’t make a 6 month trial complete in 3 months. I think its important to point out that significant advancements can be made in AI that contribute to biology, while also being realistic about how much it speeds up something as complex as curing a specific cancer or diseases of the brain.