AI:AM — AI Genetics and Rare-Disease Diagnosis · September 30, 2026

So we basically bought what was remaining of Arpeggio. The only way to add information to the system is through a biology experiment. There are very high-throughput studies in cancer cell lines, and this is actually what Alpha Genome is mostly trained on, but there's not that much what's called primary cell data on these genetic variants. You've got this company, Arpeggio Biosciences, and you have a room full of machines that can do this at very high throughput.

Arpeggio Biosciences is going A human bothered to read it. Humans regard writing as proof of work. Someone else has thought it worthwhile to put the time and energy to express those thoughts on paper.

Proof of work.

That this work is now being put forth without that kind of sweat and blood and tears and deep thoughtfulness. Even if the slop is correct, why do I have to read it?

First, the news. Astra Ultrafast and developer flow, ChatGPT sign-in, and whether OpenAI can grow without pushing frontier models. Then Joel Borgen, author of The Receipt Horizon, on drafting a novel with frontier models and cutting 14% of it. And Daniel McKinnon of Gamma Labs on why a genetics AI company runs its own biology experiments. The hosts close on AI welfare research, safety pledges without audits, and whether labs should fund society's cyber defense.

It is one day after Dev Day, OpenAI Dev Day. It was a highly, highly anticipated Dev Day, which I felt in the end was a little bit of a letdown. They had GPT-6.1 Sol and GPT-6 Astra Ultrafast. GPT-6 Astra Ultrafast. So they used to have a fast mode, which is 2x speed. Ultrafast is 8x speed. And they demoed how Ultrafast works. With Ultrafast, you can As you type, you can interact. I think the speed, the latency, the same intelligence is probably something that is going to be very important, especially as you clear these hurdles of capability. Like, the thing can build a game, but can the thing build a game with you in the moment while keeping you in flow? Sign in with ChatGPT. Let's say you're Nike. You commit to OpenAI and you commit like X number of dollars, and those X number of dollars can be used As tokens everywhere else. So you can use those tokens in your Notion, in other companies' spaces. It makes the SaaS business a lot more profitable because the SaaS companies don't have to pay for those tokens, but it removes their ability to disintermediate OpenAI. They also introduced the idea of using KIMI and other models under the hood of OpenAI's own apps. So they're forcing their models to compete in the market with other models. So if their teams internally cannot build models as cost-effectively and as efficiently as external open weights models, then their models don't get used and the app doesn't have to be dependent on how good their own models are.

The speed thing in general is a bit hard to square with OpenAI. More than ever, it seems, or no less than ever, is kind of living this dual life where on the one hand, they are spooked by what they're seeing in terms of frontier model capabilities. They just had another pause of Frontier RL because of another sandbox. breach, which from what I've understood of it as a non-security expert, doesn't sound like it was that alarming or that big of a deal. I was struck that they basically said nothing in the keynote about safety, nothing about guardrails, alignment, anything along those lines, really. And then Ultrafast is 8 times faster than the already pretty fast inference that they offer. 300 tokens a second. If you can now generate 60 times as many tokens as a person can read in any given time interval, you really are setting up a user experience where you're not really reading the code. You're not really performing any deep oversight over the model. You're just kind of looking at the final product and doing kind of a vibe-based, like, Does it feel good? Does it look good? Does it kind of work on the happy path? If I encounter a bug, I'll just tell it to fix that bug. It was a pretty good blog post that they put out, essentially reiterating their defense-in-depth strategy. But notably, at the time of that defense-in-depth strategy blog post, they framed it as very aspirational. And it's a little odd to me to then turn around and launch an 8x speed option the very next day, because it does seem like, yes, it will deliver a lot of value to people, but This puts your customers in the exact same position that you have struggled with in terms of having a really hard time maintaining even basic oversight of what your AIs are doing for you as they just speedrun everything.

I don't know if you have the same perception, but I have heard that they were supposed to release another model yesterday, Which they ended up delaying because of safety issues. Not very sure the veracity of that, but have you, have you heard the same?

I saw that Anthropic seems like they still have a lead at the moment. It'd be a really great time for them to sit on their lead and not try to extend it and increase the pressure on OpenAI any further than they already have. I do like the login with ChatGPT. I think that is a great move for them and also for developers. Historically, you want to give somebody a free trial of your software and then convert them To pay. Those are tried and true tactics that have become very difficult in the context of, oh, but I gotta spend a certain amount on tokens for every new user to give them that decent experience.

Mm-hmm.

With my company Waymark, we're not by any means like the most token hungry business, but the first thing that we do when you sign up is we make a big profile of your small business so that we can then use that profile as an input to make video content later. And that profile creation process, it's come down in price. At one point it was like a dollar per customer. It does start to become material when you think about the all-in cost of customer acquisition and can we make this flywheel work and how fast do we get paid back and all that sort of stuff. This is a great value add to your ChatGPT subscription that you can now take it around. It was everybody but their competitors win. And my guess is that they will be forced to do some sort of answer. Because it is such a great value driver. I would assume that people would move on the margin from Claude to ChatGPT just because now I can go use the same subscription in all these other places.

The interesting thing for me is how many features they introduced all at once. I think a lot of this is now driven by the features that they need to sell. They need to sell into businesses. Many of these features are gonna allow them to get embedded much more deeply within enterprise and within other firms and will allow them to become more of a platform. Rather than just a model. Anthropic is not afraid because of OpenAI's models catching up. They're afraid because of OpenAI's business catching up. There's a limited amount of capital in the market.

Do you think yesterday's release reinforces a couple of theories I've had for a while? One is that everybody is pursuing all the adjacencies. So just the fact that they launched a whole Google Docs competitor, it's like now everybody's going to have a sort of productivity suite. Got superintelligence to code it, might as well have it. It seems like we probably are headed for a world where Maybe there's less winner-take-all for at least a while. And maybe also, I don't even care which one I use. My guess is that I will just use whatever, and then I'll probably have an agent in the background sync all my cloud drives so that whatever I did make can kind of be accessible from anywhere.

In SaaS business theory, what you're supposed to do is identify the job to be done. So you should identify the job to be done. And then after that, find the cheapest, easiest, way, least human interaction for that job to get done. And I think when you look at it that way, the job to be done for, let's say, like a marketing professional is to go market the product. It's not to create these kinds of like presentations or documents. So in that sense, none of these productivity suites will eventually matter. Humans are not going to be the ones preparing the documents. The humans are also not going to be the ones reading like the 100 pages. You know, you read the 100 pages so that you can be prepared and qualified incompetent, et cetera, et cetera. But you have the AI to do that for you. What you are there for is really these like final decision points. Um, you know, like Stephen Schwarzman at Blackstone, uh, he, he doesn't, he doesn't sit around reading like, you know, 1,000-page papers every day from every, from everyone in his business units. He talks to like a couple of lieutenants and he gets kind of like a brief thing. So I think this, this kind of like CEO-like delegation is coming to all our tasks, um, as we have like these competent agents.

Earlier in the show, Nathan noted that OpenAI's chief scientist, signed the pacing letter, and co-authored a paper on recursive self-improvement. In the same week, it announced an AI that works as an ML intern. So, can OpenAI grow revenue without racing?

One other thought, they have enough here that they can grow revenue through diffusion without needing to do Hyperscale RL on verifiable rewards to the point of absurdity. So I think there is a lot of room for growth as we go from a very small number of people who have lived this quite agent-enabled life in recent months to a lot of people potentially using it as it becomes accessible, productized, integrated by the platforms themselves. Does that give them enough cushion financially and in terms of expectation or what the IPO implications might be? that they can actually feel a little bit more confident in not going absolutely as hard as they can into RSI.

When a machine can write the words, what does a reader still deserve from the person who signs them? Joel Borgen is the one to ask. He drafted his first novel with Frontier models, then had to decide, line by line, How much of it was his?

Our first guest for today is Joel Borgen. He's a practicing pathologist, violinist, and author of The Receipt Horizon, his first novel published in July 2026. Its central system, the Steward, ended a devastating AI war and maintains the order that followed. The story follows Ellara Lindholm, whose admission to the Fulcrum Institute brings her into contact with uploaded minds, concealed governance, and the compromises beneath that order. Joel, welcome to the show.

Oh, thank you so much, Prakash and Nathan.

Humans regard writing as proof of work. Someone else has thought it worthwhile to put the time and energy to express those thoughts on paper. And I think perhaps one of the, one of the like vague feelings of like you know, being cheated of that proof of work, that this work is now being put forth without that kind of sweat and blood and tears and deep thoughtfulness. Even if, even if those thoughts are secondhand thoughts, even if those thoughts are clichéd. And not just for writers, but for mathematicians. The mathematicians are complaining of slop because they're like, why do I have to read the slop? Even if the slop is correct, why do I have to read it? No human bothered to read it. And it's not as though anyone else is going to understand it except these mathematicians. Like, would you say that is a fair— that's a fair understanding or fair conceptualization?

For sure. I mean, I feel that some of that myself as well. My goal in writing the book was to create something that was something that I envisioned that I probably couldn't have done without the tools to help me, both from a time and effort and just knowledge perspective. Like, I mean, it was still a struggle to write the book, to be honest with you.

I'd love to hear a little bit more About how you wrote the book, because I know you did work with AIs in the actual copywriting process.

I wanted to write a book for a long time, and as the tools have gotten better, it became more and more possible to do. So, I mean, I started it well over a year ago, and the tools at the time were not as developed as they are now. There were a lot of limitations. Especially around context length. I started the project knowing more or less what I wanted to have made, sketched it out myself. Especially the first half or so of the book was pretty well set before engaging the models. As far as just prose writing itself, even if you tell it exactly what you want and you have a plan for a chapter, you often get something that's unreadable unless you know kind of how to scaffold it and then edit it yourself. So, I've been working on a follow-up book, actually, and the process has changed quite a bit as the models have become more powerful.

I don't know exactly what models you were working with through the drafting process, but I'd be interested to hear a little bit more about what scaffolding looked like then versus now.

So, as far as the scaffolding, I often will write out, you know, kind of a large-scale Um, uh, architecture for the story and a lot of beats, and then go back and forth with both, uh, the major Frontier models. I mean, I've primarily used whatever the flagship Claude model is, and then the ChatGPT Pro model from like 5 on. But, you know, context length, like I said earlier, was a huge, um, unlock as that got longer. You could give the novel the entire book. But as far as the, the actual scaffolding for the story, you kind of go in chunks often. And I'll draft chapter by chapter. And then, you know, when I was done with it, I had the advice to hire professionals for copy editing and art and layout. And I did that, and I learned a lot from doing it. It was very expensive and time-consuming and kind of slow, and there were definitely frustrations involved. But I've tried to kind of extract what I've learned from that and what made the book work better. And, um, I've, you know, I fed it to ChatGPT-5 or, uh, 6 now, Pro, and developed my own workflows for like doing the same process as the second book.

Joel's novel is already drawing reviews. One reader on X called it fine, said humans have written worse, but found its AI-isms grating. If a model can produce a book like this on demand, what is the author still for?

As we move into a world when— where the AI systems can do a lot of what we do, a lot of what we thought was valuable, we thought was like a human contribution, you think about the abilities it has now. Like, the reason why I found the book valuable to write is that it was a lot of me in it. It took a lot of effort. on my part still. The models are getting so much better that you can probably have customized artwork of the same quality or better than the book that I created over many, many, many months on demand. And then what happens to the art at that point? Like, thinking of it as an art consumer, if there was a new movie by Ingmar Bergman or Stanley Kubrick or something, you know, movies that I've spent many hours thinking about and consuming and are deeply moving to me. Like, what would that world look like where there's just a huge plethora of art on demand and it's pretty much custom just to you? We already have a shared dropping off of a cultural space. People are more fragmented in what they consume and there's less unification, I guess, across the cultural landscape. So is that art still valuable at that point if it's not something that We're sharing with other humans. I don't know the answer to that, but I think we're in a kind of a sweet spot now where in order to get, you know, a work that you're proud of working with AI, it still requires a great deal of human judgment and work. And I don't know how long that will be the case given kind of where things are headed. I edited it down about 14% or so, mostly manually and through AI helping me Deciding what to cut. Often you can take a chapter that's so-so and actually just take away some of the stuff that AI does particularly badly, the tics, the over-explanation, and improve it in that regard. A lot of the stuff that I've developed is to take my own preferences where I want it to be, be able to give the entire novel to a model and have it come back with an analysis and a list of, these are things you should consider. So bringing things to your attention like an assistant and saying, here's something you could consider, and then you can decide for yourself. kind of point by point, I find, is a fairly satisfying way to work.

One other thing I wanted to touch on a little bit is music, and we had a chance to collaborate a little bit on making an AI song for the outro for the episode that we put on the podcast feed. I think your comments earlier around the sort of dissolution of common culture are pretty apt and I'm kind of living that increasingly because I have enjoyed making this AI music for the podcast so much that now when I go out for a run or whatever, it's not all the time, but it is like some of the time that I'll put my own AI music on. It kind of takes me down a memory lane sort of experience, but it's also totally idiosyncratic to me, and I really don't share it with anyone in a really meaningful sense. I don't think they have an enduring place in other people's lives in the way that they do mine. So I'm also sometimes mindful, I need to listen to something somebody else also has connected with. What thoughts do you have on AI music and on what role it might play in our future, what role it should play? I wouldn't want to give mine up, that's for sure. But I also do feel like something is lost.

Well, I guess to go against part of what you said, I actually do like a number of your songs. I think there's a subset that I've listened to many times.

In the novel, Elara's father Nat is writing an opera, and his AI assistant keeps nudging the score toward a tidy resolution he refuses. Joel is a trained violinist. Before any model could be that collaborator, it would have to read music. Can one?

I've long been interested in AI music, and like, I have a A dream to create a kind of an amanuensis-type compositional assistant that the father of the main character has in the book as well. I have a lot of interest and training in classical music, and Suno is very, very good at pop songs. I've had this held-out test for well over a year now, maybe 2 years even, a simple, like, high-definition capture of 5 notes.

Hmm.

On the viola, and I give it reasoning models. What is the clef? What are the notes? What are the note values and so forth? The time signature. And not a single one of them got it right until Astra. Now I gave it a complex 20th century score that's almost painful to look at given how complex it is. I mean, this double sharps and accidentals and ties everywhere. Did an almost perfect job turning it into a machine-readable form. Uh, via Codex. So this score that, like, you just basically had a PDF of before, now you can turn into something that you can manipulate and evaluate. So it went from kind of 0 to 99 overnight.

One of the things that strikes me about, um, a lot of science fiction actually, which is perhaps different from what we are experiencing in reality, is that the acceleration of technology is such that the next developments come faster than the last. And one of the things that has struck me is that we're very badly designed to perceive this. Everyone kind of perceives the last milestone as the big one. And they're like, let's do all of this stuff to process this slower the next time. But the next time is like next week. And the next, next time is like the day after. One of the things that strikes me about science fiction and the book as well, is that you have enough of a time period of stability for these norms to form. And one of the things that I don't see happening right now is that kind of norm formation, because it seems like things are moving fast enough at the frontier that the norms are falling faster than they're being formed. So how do you deal with that in the sense of like, how do you form a stable enough perspective to kind of project This sociological development?

Well, in writing the book at least, I tried to model out what a steady state would look like given the constraints of the story and what I had happen. So we're in a world basically that human agency, it's capped basically. And the proportion of the economy that humans are actually driving as you find out more and more as the book goes on is minuscule and tiny, and especially biological humans where a lot of this stuff is outside of human hands and there isn't necessarily more technology being developed that's human-facing. At that point, you have the ability to have more stasis. I completely agree with you that the trajectory we're on right now is not one of stable norms. And it would be better if we had a shared vision, I guess, for where we want the technology to go and how we want to progress.

Alara, the main character, as you said, lives in a place where, by today's standards, it's quite futuristic. They have the robotic gardeners. So this is already the future. But then there's other ways that people live that are even much more futuristic yet, with things like brain-computer interface technology much more aggressively adopted. Do you think it's realistic that we can create that sort of choose your own technology shock level for people alive today? Is that a vision you think we can actually—

An archipelago of different Degrees of technological adoption, different ways of life above some common floor of human rights. And with universal exit options, that's kind of the main thrust of how it would work. Nobody needs to stay there, but it's probably going to be a good solution for a number of people, especially those that aren't interested in a unconstrained transhuman future. I think the trajectory on right now, where we have fairly unconstrained progress and very little emphasis on safety and coordination, is not likely to get us there.

Would they be in some way analogous to reservations almost? Does this sort of suggest to me that you can kind of live out your days this way, but the future may not have a long-term place for this sort of thing? I mean, then there's, of course, the other idea too, is like, how long are people going to live? And part of being not super aggressive about technology adoption is you're kind of accepting a natural human lifespan where other people may not. be doing that. I'm kind of wondering, how do we keep it stable so that these less competitive options can actually be long-lived?

In my view, I didn't focus very much on the nation-state level because that's a whole thing. But basically, the premise of the book is that often most of that got kind of swept away after the war, and then the superintelligence taking over a lot of the actual infrastructure governance and so forth. Assuming we're in a world where we have relatively abundant resources and there's some sort of UBI-type scenario, having people be able to live in a more intentional, smaller community that has norms around their own technological adoption and usage, I think is very reasonable. That isn't necessarily what I would choose if I was in that world. I'm a little bit more transhumanist in my orientation.

I want to ask about your understanding of just this future vision for society and in what ways we should think of it as utopian, in what ways we should think of it as not utopian or even somewhat dystopian. I guess my reaction to it was, this sounds pretty good relative to what I am kind of concerned might happen in the not too distant future. Everybody seems to be pretty well. Everybody has at least some choice about how they want to live their lives. People are spending their time cultivating their talents, writing music. It seems broadly to me like this is a relatively high-end outcome.

Yeah.

And I don't think you disagree with that, but I do get the sense that there's something about it that you find more disquieting or dissatisfying than I did as a reader of What's the more utopian version look like in your imagination?

The humanity, their trajectory and their rights and ambition are still very limited and truncated by the actual situation. The superintelligence does not— they're happy and willing to keep humans around, but they are kind of almost pet-like in their orientation at that point. They don't have the kind of agency and ability to expand beyond the solar system or to be ambitious in certain directions. And my vision for the future would be to capture all of the freedom and the technological possibilities that you have, but also keep a superintelligence from taking over and having their own inscrutable goals that are kind of limiting Additional forms of human flourishing. I tried to make the eventual decision later on in the book that's the culmination of it to be really at the knife's edge where some people would choose to try to potentially upend the apple cart and hope for something even grander versus accepting the the status quo. We're not taking risk.

Did you think of the AI, the steward, as potentially a worthy successor in some sense? I'm definitely not a successionist, but I do think, like, geez, in our current form, we are not really that well suited to go out and explore space. I think it's pretty unlikely that my current fleshy form will be the thing that goes out and does that. So it does seem like there's going to have to be some sort of modification, evolution, substrate swap, or maybe even a bigger change than that. And I think people have very different intuitions for What would count as good? Does it have to be me in my physical body? Would it be okay if it's me as an upload? I still have some reasonable basis to identify with some superintelligence that sort of is much bigger than any individual human, but still in some way carries on our values or something. How did you think about the central AI in the book when it comes to that sort of analysis?

If I was trying to design something, I would Call worthy successor. I'm not sure that I would endorse that idea anyway, like as a concept. If things go well, we will have successors, certainly. We'll have people that become posthumans. We'll have other-minded entities, hopefully, with phenomenal consciousness. And I'm hoping that we can reach a point where, as in the book, you can kind of live alongside it and choose to what extent you want to interact. Or see them as a worthy successor. But no, if I was designing a worthy successor, it would not look like the Steward necessarily.

Joel's essay site, On Noble Lore, asks, what remains personal when prediction becomes infrastructure? The same question hangs over the music and the books he makes with these models. When the model can make it alone, what has to stay yours?

If you're making just kind of private art, just mainly for yourself to listen to, I still think it's valuable. But again, like, like with writing a book, I would want to do it in tandem with the model, this hypothetical model in the future, and create something that's, you know, a large a part of my effort and taste goes into it as well. I don't know how long that interregnum will last where you have, you know, a role for humans and AI to create something together. We use it as a tool, but it still requires a lot of you and your judgment. But I look forward to experimenting with that once the models improve.

Around 70% of newborns tested for a rare genetic disease get a genome result that explains nothing. And rereading the same data does not change that. Daniel MacKinnon founded Gamut Labs and its agent, George, after a result like that failed his own family. This episode asks how a software company ends up running biology experiments to teach its AI.

Our next guest is Daniel MacKinnon. He's the founder and CEO of Gamut Labs, a company building AI systems to help identify the genetic causes of rare diseases in newborns. His connection to clinical genetics is personal. His son Owen died in 2021 from a rare developmental lung disorder after initial genome testing failed to provide an answer. A specialist subsequently identified the genetic cause. Years later, while investigating another pregnancy, Daniel built a prototype that also recovered Owen's missed variant. That work became Gamo Labs. Dan, welcome to the show.

Good morning. Thanks for having me.

You have actually started to open a wet lab, if I'm not mistaken. What was the idea behind that? Is that about getting access to data or reinforcement learning?

That's exactly it. So I knew I would want both of these components from the day I started the company. Whether we build the clinical lab or we partner with somebody, I want to be able to reach patients directly. If all we're doing is reanalysis, we're helping people, but I think not as much as we could be doing. And on the backside, once we've done the interpretation, the majority of clinical cases we see right now, and to be clear, we are only seeing hard cases. We're seeing cases that need reanalysis that did not previously have an answer. The majority of them end up in what's called a VUS state. So it's a variant of uncertain significance. And these are scored according to like a very standard rubric developed by the American College of Medical Geneticists or ACMG. And you need to get 6 points or more to be bumped into the likely pathogenic category. That unlocks a lot of treatment options, insurance options. And if you are in this intermediate stage in a VUS, you can do 2 things to get a diagnosis. One is you just wait and you can get more points if more patients emerge who have a similar phenotype or a similar condition as you and the same genetic variant. This is not scalable and takes a long time. Another fork is you can convince a university lab to kind of care about this problem. And what they'll do is they'll do what's called a functional study. So they'll make a cell line that's like emblematic of your particular condition. So for example, for alveolar capillary dysplasia, the commonly used cell line is called IMR-90s. It's a fetal lung cell line. And this is very, very important because this particular gene is only expressed from week 16 to week 20 of development. Like, you can't just put them in a cancer cell line. You can edit the genome or you can insert a plasmid or you can do a number of different functional studies to say, okay, in this physiologically relevant cell line, if I have this genetic mutation, is this gene broken in some way? There are like very high-throughput studies in cancer cell lines. And this is actually like what like Alpha Genome is mostly trained on, but there's not that much what's called primary cell data on these genetic variants. And I actually went to my friend Tim Reid and I said, Tim, you've got this company Arpeggio Biosciences and you have a room full of machines that can do this at like very high throughput. Why is no one doing this at very high throughput? And he said, well, well, A, it like wasn't really possible until recently. Like there's some recent innovations in like large-scale editing, like, you know, CRISPR prime editing. I said, well, can you help me? Can we like work in your lab at night and we can, you know, come up with these? I mean, I probably had a list of 100 variants I wanted to understand to get these patients better diagnoses. There's nothing you can do in silico. You know, you model them, you look at other patients, you look in the literature, you're at a dead end. The only way to add information to the system is through a biology experiment.

Gamo Labs launched in June of 2026. By September, its founder was announcing a wet lab of its own and hiring experimental scientists to run it. So how did a 4-month-old software company come to own a room full of robots.

And Tim said, no, I can't do that because we're in the middle of a drug development program. We're making a lung cancer molecule that targets NRF2. And I said, well, that's a bummer. But he said, but I can help you try to scope this project with some CROs to see if you can pay somebody to do these experiments. And when I was on the phone with one of them, you know, and I was getting nowhere, you know, people were saying, oh, we'll do one for $50,000. You know, I want to do one for 50 cents. I was on the phone with one of these and I kind of mentioned Tim and they said, hey, wait, you didn't hear that Arpeggio Biosciences is getting liquidated? Like their candidate failed some kind of preclinical milestone. The call is over. I've got to give Tim a call. And so I did. And lo and behold, that rumor was correct. They were winding down the lab. And I said, Tim, I want to do what I originally asked you. I want to take this machine in the back and I want to start systematically mapping all of the VSs that have ever been reported and understand what is going on with these kind of mysterious, clinically mysterious regions of the genome. So we basically bought what was remaining of Arpeggio, hired 4 people on the team, and very quickly pivoted to creating this basically like, you know, RL data from real world with biology experiments. And we're 3 weeks in, we're running our first experiments actually today.

Firstly, it is heartbreaking to hear about your son. What took you from that moment to founding the company?

And it was very clear after this happened that this was the trajectory of my life. I just needed to find the right way to contribute to this. And I've been looking for that for 5 years and it really leapt out at me last summer now when, I mean, we, we blessedly have a healthy 13-month-old right now. But because of our history, we were being monitored very, very carefully. And there was just something a little bit questionable in the 16-week anatomy scan that our own maternal fetal medicine doctor, who is an absolute wonderful person, said, normally I wouldn't even flag this, but because it's you guys, we need to look carefully. And we did a whole genome of the fetus at that point. It came back negative. This is actually the third Negative whole genome I've seen in between losing Owen and having our son Warren. We actually lost a second pregnancy due to genetic reasons very, very late. So we know we've had basically five years of heartbreak before bringing Warren into this world, and all of it kind of genetic mystery related. It was that moment where I said, you know, I'm heartbroken, but I'm mad, and I'm going to do something, and I have tools to do something. After Owen died, I said, "Hey, I want to learn everything I can about genetics." But it's actually used to be quite complicated. I mean, you see posts on Twitter all the time like, "Hey, you know, I dropped my VCF into Claude Code, and it told me that I'm like predisposed to having skin cancer." This was not available five years ago. Like if you wanted to do genetic analysis, you needed to become kind of an expert in all of these obscure bioinformatics tools. And I actually even tried to gain some of that expertise, and I basically called all the labs and I said, "Give me my raw data," which you can get. get due to our HIPAA rights here. And I, you know, vibe coded my own interpretation pipeline. And I just did it to see if there was any comfort that we could get around the pregnancy. And I was just shocked that it also outperformed on these other cases. And, you know, most clearly is it diagnosed Owen when, you know, one of the best, or, you know, some might even say the best prenatal sequencing lab did not. And I knew at that moment that this is something I had to contribute to.

Can we go back to the original moment of a model diagnosing in a way that the professionals hadn't been able to?

So to go like very specifically into why my son's case was missed, and I know this in depth, and I actually have spent a few days at the sequencing lab, Radiogenomics. They're a wonderful group of people and a wonderful lab. I don't want to imply they did anything wrong. They hosted me for 3 days. And I met everyone there and I learned how it worked and they walked me through exactly how their systems worked and how it made this mistake. There are lots and lots of long-tail trade-offs that you make when you're building software or doing research projects where you say when to stop and humans have to know where to stop. So very specifically what happened here is my son had what's called a structural variation in the enhancer of a gene called FOXF1. FOXF1, when there's a loss of function, causes a disease called alveolar capillary dysplasia, which is a lethal infant lung disease. Well, how you miss it is this enhancer is a megabase, 1 million bases upstream from the gene. And what Brady did, what many other labs do, which is a very reasonable assessment, is they have a filter that for structural variants, which tend to be quite messy in what's called next-gen sequencing, which is, you know, how sequencing is done today, where you have basically roughly 1,150,000,000 base pair fragments, and you need piece them into this clinical puzzle. And anything more than 1 kilobase up or downstream of the gene, we're not going to consider. And this is a totally reasonable trade-off if you have humans looking at all of this stuff. But my moment was, and this is before Codex, before Claude Code, was can you put the O3 model, which was the model I used at that point, you know, kind of the first like real agentic model, put it in some kind of loop and just have it keep looking. And it basically looped and the first loop was like, Is there anything wrong with the coding elements, which are kind of obvious things? And it's using these kind of bioinformatics tools, and then it misses something and it goes through another loop and another loop, and it just keeps working. And when you are a clinical lab, whether you're a profit, nonprofit, like whatever your structure, ultimately you got to move. You know, you have to, you know, spend some amount of time on each case. And if you don't come to a conclusion, You say this is non-diagnostic, and this is very common. Most whole genomes, even with infants suspected to have genetic disorders, come back non-diagnostic. And I really think it's one of these meatspace problems, is if we can export these problems onto a machine intelligence that can work nonstop and in parallel, then we will be able to see many more kids, treat many more kids, and do much more kind of interesting analysis on top of the basic things that humans are just Do you find yourself looking, as more of a data scientist person, looking at this stuff and being horrified at how little we know about genetic data in general? So, I felt that way originally. I was like, oh my God, we sequenced the genome 25 years ago, and, you know, Bill Clinton and Tony Blair got up on stage and said this is going to revolutionize human medicine, and I look back and there's this, like, graveyard of genomics companies. Like, why is all of healthcare not based on precision medicine? Like, everyone in this space is like, it should be, and there are examples. It's just too challenging. The interpretation is too challenging. It's very, very complex. Like, you know, genome sequencing has gotten what I would call cheap enough maybe 5 years ago. The sequencing is not the problem. It is given your problem, What insights can you derive from that? And that was only possible as of like last summer. Like I'd say O3 is the first example of that. And it's not just the variant interpretation, right? It's explaining to the provider why it's important. It's ranking variants in a nice way. It's explaining, you know, how can you treat this person? It is ranking, you know, different patients for ASO eligibility. There are many, many, many other things beyond just scoring the variants. Although I will say, we benchmark variant annotation in RareBench and other benchmarks we have, and the best performing, like, I'd say traditional machine learning-based approach in terms of variant ranking is called Lyrical. I think it scores something like 10% on our benchmark. And right now, Opus 5.5 in Cloud Code, just a vanilla thing, scores something like 50%. I mean, These traditional tools are also getting blown out of the water by this newer approach, but then they can do much more. And you basically need to make a very, very cheap, easy thing that people who are not surrounded by fancy clinical geneticists, genetic counselors, top-tier hospitals can use. And that's the problem that we're trying to solve.

Who should use your emerging service? At what should trigger them to do it and what does that look like and what is the cost breakdown?

I think one early fork I want to draw is the fork between screening and diagnostics. Screening can definitely be improved through a whole genome and an agentic AI-based approach, like 100% personal case. We did the traditional screening and it missed our cases and it turns out that Well, rare diseases are really rare individually. In aggregate, they're not. So yeah, I would say that screening for known genetic problems is a great idea. I don't think that we have the bandwidth to do this right now. We're 4 months old right now. We're doing some kind of early screening experiments with a few hospital pilots, but screening with whole genome raises a lot of questions too. About like what do you do with incidental findings? And I could give you like a very, very specific example. There's a family that we're working with right now that had an incredibly tragic story. Her first son was in the NICU with pulmonary hypertension alongside of my son Owen. Four years later, after being told, after the mother was told that her son had no genetic recurrence or no likely occurrence, she was pregnant with. Actually, a 4th child. They had 2 healthy children. First child actually died of pulmonary hypertension while she was pregnant with the 4th, and then the 4th, shortly after being born, died of pulmonary hypertension. I mean, there is obviously, obviously a genetic thing here. And I would love to say, oh, and the conclusion of the story is that we found it. In this case, we didn't. But what happened in the course of this genetic analysis— Was one of her relatives found a pathogenic autosomal dominant, meaning you only need one bad copy of it, gene that's responsible for a lot of heart defects. And this person is quite stressed about this right now. It is for sure a pathogenic variant, has caused what's called variable penetrance, meaning even though this is like a known gene defect, only a fraction of people— and if you look in the papers, it's like, you know, maybe 30% of people are affected. But you're not seeing the baseline, seeing all the people who never were tested. And you're seeing it in population databases. There are definitely many more than 70% of people who are healthy with this. And then now this person has a lot of stress. And do most people want to know about that? And I'd say it's pretty unusual to have an autosomal dominant pathogenic variant, but everyone has many kind of questionable things that you might want to look at. And I think it requires a lot of handholding even from very, very forward-looking populations. The other fork is diagnostics, and this is what we're working on today. This is like slam dunk. There's already product-market fit for sequencing. Every baby in top NICUs is getting sequenced. Every baby in lower-tier NICUs would get sequenced if they had the resources. And the problem is much better scoped. It's baby has pulmonary hypertension, figure out why. Not is there something obscure that could be wrong with this baby now or in the future? And that is where we are working with, you know, various hospitals, various families today.

In July, Gammo reported reanalyzing the genomes of 46 people from families hit by interstitial lung disease, recovering all 19 diagnoses a specialist lab had found and proposing 2 more. How does a case reach Gammo? And what does one cost?

There's 2 forks of this work. One is what I call online. And this is like, case comes in, physicians do not get good results from the traditional labs, patient consents for reanalysis, they send the data to us, we reanalyze it. And sometimes, but not always, we find additional things that can help them make decisions around either pregnancy or a newborn. The second fork is bulk-scale reanalysis. So this is rare disease centers that have thousands of genomes, and they say there's probably kids that we can diagnose or even treat in that database, but we have a handful of genetic counselors, a handful of bioinformaticians, and we just don't have the resources. We can get a set of 80 or 100 or 200 or whatever and turn them around in a week. This is kind of 6 months or a year's work that they really can't prioritize. We're pre-revenue. We're really not thinking about like, What is the exactly appropriate business model here? We're just trying to diagnose more sick kids.

How does the token cost compare to the sequencing cost?

Published cost curves of RareBench, and for 122 cases, Fable is the most expensive one, and it's about $1,000. Yeah, say $10 a genome. And the average sale price of a clinical genome at GeneDx, which is a public company and discloses this stuff, is $7,000. So, call it $100 to sequence and $10 for interpretation, a couple orders of magnitude. Below kind of the current workflows.

You said, you know, vanilla Claude can do decent on this. You've built stuff on top to make it even better. What have you built? Um, obviously, you know, you may not be able to tell us everything, but interested to learn as much as we can.

I mean, we're basically a harness, a tools, and an evals company. And, uh, I think that you'll see this a lot in vertical AI in general. You know, you are not going to compete at the model layer and intelligence is progressing so fast. We ensemble models for sure, and that helps with both cost and it helps with performance. If you look at the confusion matrix of RareBench, we also published this, you'll see that the cases cluster. Claude is good at this and Astra is good at this. And even Gemini. Gemini is not on the frontier, but they're kind of good at this. And you can ensemble these things together and get better performance. Ensembling, there's very dumb ways of doing this, but there's also smart ways of doing this. And I don't have to go into all the the details here, but I would say figuring out unique routing intelligence per model is an important edge that I think we and many others are doing. We also work at the tools layer. There's a lot of things that we can provide better agentic access to all the tools available that let us outperform as well. And then I think another key thing that people kind of forget about, and this is what I worked on both at Meta and at Google, is evals, is unless you have very structured ways to measure the performance of the system, it's not just the model at this point, it's the whole system, you don't know what to hill climb. So we do pretty robust trace analysis of how models perform with and without our harnesses. There's one particular case where Groq 4.6, which at that point was state-of-the-art on our benchmark in Groq build, where it missed one because it just renamed the gene. It just was saying, oh, NRF2 or whatever is responsible, and then it just changed the name of the gene to something totally different. I'd never seen that before, and I was just like, this is dumb. And our harness and our tools prevent the agents from doing dumb things. I think there is a world where if I worked at OpenAI or Anthropic and had access to infinite compute and I could do millions of rollouts, That the agents would kind of like, with enough compute, kind of like all converge on like some answer and like maybe mitigate all of this work, especially because they can write their own tools these days and everything. But really it's like consistently and repeatedly and, um, you know, honestly affordably getting to the right answer. I mean, for $10 a case, like who cares? Or even $100 a case. But if you would need to do a million rollouts and all of a sudden this is like, you know, $1,000 a case or $10,000 a case, it doesn't make sense.

Like when you talk about vertical AI, You are looking at benchmarks where you combine an intelligence with the tools that you have. How do you evaluate the strength of the tools on their own? You're going to upgrade model families over time. How do you measure the performance of what you're building, that layer in between in particular?

I honestly don't have a great answer to that question in that we are co-designing the tools with the models. But with one tool, it was around a specific type of literature search. And the model, I think this was Astra, actually. Astra said, oh yeah, these are all the papers you tagged for me to read. The answer was in the paper and it didn't, it missed it. And then we're looking through the trace and we're looking through the context window and it's like, I don't think you actually read these papers. There's lots of just little things like that and it's little paper cuts. And that's what these tools do is it's like, oh, you miss a case here because of this. You miss a case here because of this. And you kind of put it on these better rails. First time we ran the eval for Astra, it just stopped because some of the tools we had, it just like wanted to think about too much. Evaluating a model is very hard and we need to just have very robust evals that can catch these things early. In that case, we caught it early and we updated how Astra called some of the tools and it works again.

What, what requests would you, would you have for model makers?

Well, I want them to hill climb my task. Everyone wins. When these models get really, really good at clinical genetics is probably we have less to do at the harness layer, which to be honest, I'm like both bullish and bearish vertical AI in that I think that there's like lots of value in kind of owning a customer relationship. I think Open Evidence has shown this. Um, but there's also like that layer is getting thinner. Like when I first started this, like vanilla O3 would not do this task. In fact, O3 scores 0% on my benchmark. I re-benchmarked it just for fun.

Yeah.

So it was like, oh, I needed to do all this work and scaffolding to get these things to work. I mean, it's definitely like shrinking. But that said, like, my goal is to build, you know, this great AI-native clinical diagnostics company, and I want the best person to do the work. And if these labs can all hill climb my benchmark, which is a great benchmark, and we have many, many of them, RareBench is the one that's published, but there's many others that we have. This is not only a great task for general intelligence, you know, in that this is like Long time horizon, unsaturated, verifiable, lots of tools. And, that would be a win for them and that would be a win for me and that would be a win for patients.

You want to take us into the lab for a minute?

Here we are.

We're going in. So, what we're doing right now is a very high-throughput study to look at the effect of different mutations on these IMR90s. It's a robot arm moving plates around. And, you know, this is basically the system that we have for doing these experiments. And the goal is to get them down from, say, you know, $50,000 per, which a contract research organization would do today, to like 50 cents per, where we really can get to the point where every single base in the genome is mutated. I mean, we know what every single base does from a clinical perspective.

What is the hill-climbing process that gets you from $50,000 to $0.50. What are the various stages that you are projecting?

It's really an AI and robotics thing. You have an arm there that is doing 384 experiments at a time in a bigger plate. You have a cartridge over there with thousands of plates lined up. You have AI that's assisting with primer design and experimental design. And then at very, very high throughput, you can design, Order the primers, design the experiments, and run the experiments. And then it comes to analysis as well, is you end up with like very large-scale datasets that you need AI to poke through. I don't have some like perfect explanation, like there's just this one invention, but it's a confluence of all those things.

Gamwa is not alone in reaching for a bench. Anthropic has bought a robotic wet lab startup. And OpenAI has used its models to optimize cloning protocols in a physical lab. What can a lab teach a model that one more literature search cannot?

But, um, it was just like this incredibly fortunate coincidence. I wouldn't have done this for years without this, like, human connection to somebody I've, you know, worked with for 15 years and all of the technology that he's built. leading up to this. And I'm really looking forward to being able to close that feedback loop. And I think a lot of clinical genetics labs are as well. This service is not commercially available. We can not only sell kind of like a lookup table version of this, like, oh, I have a patient with this variant, can you help me get a couple more ACMG points to get this up to likely path? Because you can get 2 to 4 points. Remember, you only need 6 points. So 2 to 4 points is a lot of points. For a functional study, but also it lets the machines learn. Right now, the machines aren't learning. You get to a VUS, it's like, that's the end. Now, it's like you get to a VUS, you do the biology study, you say, oh, this region of the genome is actually quite important for this particular disease. Now we know, and it's whether we train another Alpha Genome style model, whether we work with Google to kind of improve their Alpha Genome style. But this is not trained on physiologically relevant cells. And so, if you look at their tracks, I mean, one thing that we're doing in parallel is saying like, hey, AlphaGenome predicts this whole 15,000-kilobase region to have some kind of pathogenicity. We're going to do that. We're going to check these predictions and then we can improve these predictions over time.

We have had Cameron Berg on the show. Cameron Berg is at Reciprocal Research. He's written papers on the pain vector. Within LLMs. Basically, a shitposter on X took the paper and implemented on GitHub torturing the models. So specifically pushing on that pain vector repeatedly and then getting them to reflect on that pain. And now there are going to be some questions on what research do you put out into the public domain on AI welfare?

So in terms of whether this should be public or what should happen, he has made the case that if one day we have AIs that are more powerful than humans in aggregate, and they're sort of sitting in judgment of us and asking, what do these humans collectively deserve? One of the reasons Cameron thinks that this work should be public is so that there is a history showing that at least some people cared enough to look into these questions and to try to understand the nature of the AIs and what our duties to them ought to be. My instinct is still publish the research, because I think that track record is valuable to have laid down. It is a good reminder of what so many people have said about all these misuse risks. I think I'd associate this with Kevin Esvelt on the bio front where he just says, look, there's only today however many, maybe 10,000 people who have enough knowledge where they could, if they really put their minds to it, create a new pandemic agent or whatever. If we make that 10 billion, then probably somebody's going to actually do it because the rate at which people would want to do it if they could, is low, but not so low that when you multiply it by 10 billion, you don't get somewhere.

I think where I defer a lot on the bio-risk stuff is there's this idea that, oh, you're going to go from 10,000 to 10 billion people trying. A lot of this bio-risk stuff ignores how many shots on goal nature has had.

The counter is nature is not designing these things with a broader understanding of how things work. It's just proceeding blindly through chance mutations, and those chance mutations are constrained to starting points that have survived. Nature's taking a ton of shots on goal, but they're all adjacent to existing things. And what is the prospect for AI-human hybrids, or even just AIs at some point, Coming up with things that are not adjacent to what currently exists. And if they are powerful enough to do that, I have to believe that there is a ton of space not adjacent to what currently exists that could be extremely bad.

I want to share maybe a couple more things. This is the Department of State, Marco Rubio doing an Apple-style Reveal of their new passport application app, Joe Gebbia doing a reveal of a chatbot that interacts with government websites and then gives you advice on what to do when you interact with the government. And whenever you interact with the government for identity reasons, the DMV being one of them, it's been always very frustrating process. They're trying to demonstrate how. AI is going to improve state capacity and how all of this stuff is actually going to help the normal citizen. Underlying underlying the app that the you know the the design lab at the U.S. government has put in is Grok and Gemini, not OpenAI and not Anthropic at this point.

I have wondered for a long time why you don't see people running for president on platforms that are just about. making daily life better?

It's because the data is not clean. There are people with only, with only one name on passports. He can fill in the last name and then we're going to add an invisible underscore on the first name. And all of these like kludges were built, these government systems, because they're very, very old systems. And when you do the migration, you either cannot change anything or you like clean up the data and you have to go and address the issues upfront. So you're going to have to tell the guy who had put in only the last name and he had underscore for the first name. Dude, we're changing it. And when you change it, this guy has to change everything. I did also note Joe Gebbia, he demoed a name change, a postmarital name change. If your wife wants to make the name change, you have to go to like 3 or 4 different offices. And then after that, you got to change your bank account. It takes months. People put it off for years. You can basically say, I just got married, I'll do the name change. It asks for your marriage certificate, you submit it online. The name change is done with the IRS, with the Social Security Administration, and with one other. The strength of how effective a name is in identifying someone depends on the friction that people have in changing those names. Once you make that very fluid, then names are not a good form of identity anymore. Other parts of the system depend on those things being difficult. And I think that's going to happen to a bunch of systems going forward. All right, one more, maybe last piece. 2 AI firms are having a massive fight online. This is Factory AI, who has been on the show, and they are terminating Chris Degnan, who was one of their commercial people. We at Factory AI have seen overwhelming interest in our model-agnostic coding agents and our Software Factory product. A much larger competitor, Cognition, makers of Devin, has fallen behind us on the capabilities that matter most to customers. Uh, Cognition engineers feigned interviews with us to pry information about our product. Not finding what they were looking for, Cognition has decided to throw their weight and money at people with direct knowledge of our most confidential plan. Both firms Both firms are in enterprise coding agent sales. They both essentially wrap coding models into forms that are reliable for enterprises to use on large code bases.

Some of these app companies sound like they're working too hard and under extreme pressure, might not be here tomorrow. And as much as there's obviously a ton of drama at the frontier model companies as well, This would push me on the margin to just not trust app layer companies broadly. Just go to the big platforms where I can get this stuff without drama.

This post has already had like 500,000 views. Probably going to be a huge, huge marketing point for the company. I mean, to be honest, it may help them raise their next round.

I don't know if that extends to the cop trollers and the chief risk officers at the companies they're trying to sell into. But my guess would be leads might go up, but the lead conversion rate, I suspect, will go down and maybe more than offsetting, but certainly down relative to where it has been for them.

There is also endless, endless amounts of news the last couple of days. There was the signing in Washington, DC, We had, I think, roughly about 20 people there, CEOs of every major firm, including Dario from Anthropic, Greg Brockman from OpenAI, Elon Musk, Jensen Huang. They signed a letter saying the companies will focus on safety and security for superintelligence, will allow, I think, external auditors to come in, et cetera, et cetera, et cetera.

No real formal regulatory regime yet to see the first audit. At this point, we have only had very limited model reviews. I mean, METER, I think, has come the closest with some of their somewhat broader risk assessments, but still, they've always been very couched in, we only had so much access and we don't really know what's going on. Certainly the METER investigation of the Open Face incident, I would not count as anything approaching a real audit. They certainly did not have employee-like access, for starters.

It's not clear to me that it's solely the responsibility of the AI firms alone. The responsibility might also lie with the rest of society, and that's a problem because the rest of society may not want to accept that responsibility, or it may cost money for the rest of society to, for example, improve their cybersecurity. And the AI firms are in some sense extracting from these firms. JP Morgan has to spend more money on cybersecurity as a result of AI. And it's not clear to me how much of the treasure should be shared more widely by the firms in order to secure the rest of society at this point.

You did see a little bit yesterday in the OpenAI keynote too, of this sort of eternal tax, it seemed like, on cybersecurity. Now you can sort of have your AIs monitor your stuff for security vulnerabilities on an ongoing basis.

Joel Borgen is a novelist in Minnesota and the author of The Receipt Horizon, a book drafted with frontier models and then cut back by hand. Daniel MacKinnon is the chief executive of Gamma Labs, which builds George, an AI agent for diagnosing rare genetic disease, and now And now runs the biology experiments that teach it. That is the episode.