The hottest running app has nothing to do with speed | E2303

24:38Lon Harris:Is that about right? Yeah, about$60 a year. Perfect. Yeah. So I went through this with comms. So that means if you get one in five people to convert, you know, five times 12, you hit that$60. And you said they last for 17 months, which means on average they make you$90 or$85. And then maybe they tell a friend about it if it has an internal feature launch. So you can maybe add a factor of like one in five, add a friend, which you divide the 17 months by five. You get another three months. And each month costs$5. You got an extra$15 in value. So there's all kinds of return on ad spend, ROAS, and cost per install.

25:21Lon Harris:And it's a really interesting science. And there are funds that can help you. I went through all this with Calm, FitPod. We have a great company called ToneBass that does music, Musician that does music, Steezy that does dance. And it just becomes really hard to get this right. but if you do get it right, you can have an incredible flywheel and build an incredible brand like Calm and Fitbot did in ToneBass. Steezy didn't work out exactly for it. That was a harder one to make work, dance. But yeah, continued success, Luis. And thank you so much for sharing all your secrets. Yeah, thanks, Luis. Good to have you.

25:59Lon Harris:Continued success. I'll see you when I'm down under. Sounds good. I'm joined for that one. I'm coming along on the Australia trip. Yes, you are. Well, you know what I'd like to do is there's four cities there. Perth, Sydney, Melbourne. What's the other one that always competes for startups? Brisbane. Brisbane. So there's like four centers of excellence. So what I want to try to do is get two or three or four of them to join forces to bring my stack to Australia. Yeah. So I want to fire up this week in Startups Australia again. Mark Pesci used to do it for me. We did like 12 seasons. Many years ago, we started doing that.

26:41Lon Harris:Yeah. So it'd be great to get that fired up again, to bring Foundry University there and to bring the launch accelerator there. And then my vision for it would be to get those three cities to collaborate, chop up the cost of doing this. Oh, sure. And then rotate it. So Foundry University is in Perth, then it's in Sydney, then it's in Australia, then it's in Brisbane. And it just rotates. Canberra too, maybe? Canberra also? Whoever wants to chip in to get the flywheel going. And I just want to have an excuse to go there, frankly, what my family wants or wants. Hey, everybody. Welcome back to Twist.

27:14This is Alex. Now, AI is having a moment. People are mad about data centers. People are mad about Anthropic. People that like Anthropic are mad at OpenAI. Space X AI is suddenly a hyperscaler. Job loss is either here or never coming. And AI regulation is becoming a battlefield. Are you tired of all the negativity? Well, something that many AI believers love to trot out is that AI is going to cure cancer, bro. And the thing is, maybe. That's why I wanted to get Alice Zhang from Verge Labs on the show to tell us about the state of using AI to discover new drugs to tackle our most intractable species-level diseases and maladies.

27:49So please join me in welcoming to the show. It's Alice. Hey, how you doing? Good. Thank you for having me on the show, Alex. I'm so glad you're here. We're also talking to you mere days after the company rebranded from Verge Genomics, the name that I've always known it under, to Verge Labs. So one, congratulations on the rebrand. And two, from a very high level, why was this the right moment to kind of change the name of the company and redirect in a new direction? So we started 10 years ago, really with the mission that drug discovery could really be turned from a guess and check problem to really a prediction problem, and that the missing piece was really missing data.

28:26So we, over the last decade, have built one of the field's largest brain data sets directly from patients, over 12 ,000 human brains and 6 ,000 patients. And we initially used that to develop our own drugs. and we went through that experience which was really useful but that experience really taught us the importance of an even kind of more valuable problem which is when students develop the drug how do you actually predict what patient will respond to that drug which is something we did not originally foresee so the first thing is that we learned this really hard lesson about this very valuable problem and we also had the data sets that were necessary to really solve that problem.

29:08The second is that the architectures in AI have finally gotten to a point where they can actually solve one of the key challenges that actually prevented us from solving that problem in the first place, which is the kind of incomplete and fragmented state of most patient data sets. So it was really the kind of intersection of the fact that our data actually achieved the scale we needed and the multiple architectures were advancing to the point where we could actually see this much larger opportunity, which is instead of just buying a lottery ticket and developing a drug ourselves, can we actually make a better machine that sells those lottery tickets?

29:47And that's really what drove the shift is this kind of culmination of all of the above. I've never heard a startup founder come on the show and say, you know, we're doing now is we're selling the lottery tickets instead of scratching them ourselves. You can invest right here. No, I really appreciate that summary. I now want to go through that in a bit slower pace to let people know what's changed and how technology has kind of brought you to this point. So I think it's a very important story. So in the earlier days of Verge Genomics, you guys were working on Converge One, which actually helped you select a candidate drug that you then took into testing, if I'm right.

30:19And I'm curious about the process to getting Converge One built and how surprised or not surprised were you when you took it to the real world with this drug you put together, the results didn't quite match what you were hoping for.

30:31Lon Harris:Most AI tools are adding friction, not making your life simpler. And it's another tab to switch to. And maybe you forget to even do it. It's arduous. What you really want is one system that's going to make you more efficient and save your time every single time you do work. That's why I love Superhuman Go from the amazing team behind Grammarly, which I have insisted all my team members use since day one. Now it's an AI chat that lives on the side of your browser. It's always there. Maybe you're drafting an email mid-meeting. It goes and helps you finish it without switching apps. Maybe you got a 40 email thread to get through before that call.

31:07Lon Harris:It's going to summarize it for you in seconds without losing your place. No new tabs, no starting from scratch, no context switching. Superhuman Go has the context of everything you're working on. It works inside the tools and sites you already use. Your inbox, your docs, your browser. Maybe you're doing social media all day long like me. It's part of my job. You can try many of Superhuman Go's features for free. Find out more. Superhuman.com. That's superhuman.com. Yeah, so what we built Converge originally is what we call, it's called a target discovery engine. So it's how do you actually find the proteins to go after that cause disease and then design drugs around them?

31:46So to do that, we started accumulating this very large data set, which is that instead of starting with a mouse or a cell, which is how most researchers start, we asked, why not actually go directly to the source, which is the brain? for neurological diseases because that's where it happens. And so we started sequencing these brains. We paired them with multimodal data, like their clinical records, how they progressed in the disease. Alice, can I ask a question about that? Just because I'm really curious. My brain's inside of my skull and hasn't ever, to my knowledge, left. So when you're talking about getting brain samples, number of patients, number of brains, how much tissue are you getting?

32:24Are these from politely living people? Are these the recently deceased? I don't know. And I just thought I'd ask for everyone out there who's curious. So they're from deceased patients. This is why it's actually it's so hard is because, you know, in cancer, the problem of how do we actually find the right patient and match them to the right drug has been partly solved because you can actually take a tumor, right, from a living person. You can profile it and analyze it, and then you can match it to the therapy that you want that patient to be on. In the brain, you know, you can't take a brain from a living person, right, in neurological disease.

32:58And so you can only take it from autopsy patients. And so that is what we have done is that we've partnered with more than 24 different tissue banks, hospitals, academic centers across the world that have thousands and thousands of patient brains from people that have passed away from disease and donated their bodies for research. And then we've built an end-to-end infrastructure that can actually ingest these samples, quality We control them, dissect them for our data consistency, quality, and traceability. And then we essentially digitize them, which means that we sequence them. So we capture the behavior of all 30 ,000 genes in the genome at multiple levels from the DNA to RNA to protein.

33:41Okay, that's super cool. But I think when you guys were working on Converge 1, the first iteration of this engine, there was a mismatch between the samples of data that you could collect from the, I guess, tissue banks of the world and maybe the brain of someone who had a particular disease you were going after trying to fix. And there was a bit of a gap between the two? Yeah, right now you can only get brains from deceased individuals. And one of the challenges is that when you actually go into clinical trials, you're actually going into a living person. So how do you actually measure what's happening in that person's brain, which is the really only window into what is happening with disease?

34:20And so what we've developed in the last year is a world model of disease that can ingest all this brain tissue that we've collected, combine that with patient data from living patients, and essentially create what we call a virtual biopsy of the brain. So that's essentially a reconstructed picture of what's happening in your brain that can be built from just a single blood draw. And so that allows us in a living patient to actually say, hey, what stage is your disease at? And how might you actually respond to a given therapy? So with the information you have from the deceased and these tissue banks and some information about living patients, you can kind of bring the two halves together using AI, which is what's changed since you started the company, and therefore kind of close the gap using, I guess, the power of generative AI.

35:13Yeah, and that's what the power of what these models have brought in the last few years is if you look at traditional deep learning or machine learning models, they've really had required every patient to have every single measurement. So you have to have the brain, the blood, the clinical treatment data all in one, but that's not how it happens in the real world. In the real world, you might have a patient that goes into a clinical trial, and you might have a different patient that gives their blood, and then you might have yet a different patient that donates their brain tissue. And the power of these transformer-based architectures is that it allows you to actually piece together missing data and infer missing data from what you have.

35:53So you can start creating a unified representation of what a patient looks like and start filling in missing data modalities. Now, you guys said that brain tissue is the lidar of neuroscience, applying the kind of world models we've heard about from self-driving companies like Wave and I think also Wabi and so forth are working on that. And you think that brain tissue is going to help your world model have high fidelity and high accuracy. Are people out there trying to build similar world models for similar tasks without using actual brain tissue as part of the data grounding for that work? Yeah, so we are using the very kind of same models that some of the self-driving cars are because it allows you to not just pattern match based on observational data, like it doesn't just pattern match how their previous driving scenarios happened, but it can predict a new person in the road.

36:47And similarly, that's what we're doing with our world models. There are world models in oncology because that's a much easier space to get data. In fact, that's kind of a pattern you see in the space that AI companies get just simply built because of where it's easiest to get data set. But we've kind of taken the opposite approach is we've actually asked what's the biggest problem right now? And then how do we actually do the hard work of collecting the right data? So with neuroscience, most of the data, it's not that there are people building world models with the proxy data. It's just that that's where most of the data is today.

37:23So it's probably tempting to go there first. But the issue with the proxy data, and when I say proxy data, I mean things like blood, brain imaging, spinal fluid that can easily be collected from a living person, is that they're all just downstream consequences of the disease. They're like shadows of the disease. So in order to really understand what is happening in the disease, you need to go into the brain where it's happening. And so the reason it's a bit like LiDAR is it's like thinking about self-driving. If you were to build a model only on just camera data alone, like kind of Tesla has, you have limited information.

37:59But we saw that when Waymo integrated cameras with LiDAR, which was an actual direct reading of 3D depth, that could vastly increase the speed at which they could get accuracy in self-driving. And so in a very similar way, that's why I say brain tissue is like the lidar of neuroscience in that it's just the molecular ground truth of disease. And for the model to work, you need that anchor to be able to anchor the relationships between blood, between your brain images and to actually what's happening in the brain. OK, so some people are spinning up drug discovery companies using AI and they're going to where there's a lot of data because everyone knows if you can bring a lot of data in, you can fine tune a model.

38:36You can therefore do a lot of work with it. but you know honestly Alice if everyone's going to go just to where there's easy data it seems like they're all going to be competing kind of along the same vector whereas you guys having done uh years of data collection that's special and unique will have a different approach okay that makes good sense to me now when it comes to world models for this work I'm a little bit confused because when I think about a world model in the self-driving context I almost imagine like a video game if you will like a place where there's you know physics and people moving around and interactions and so forth.

39:07When you're doing world models for brains, what does that look like? Or does it actually look like anything? Or is it just code? So it's what a world model looks like is that, so in the self-driving world, instead of pattern matching on a previous scenario, it creates an internal representation of how the world works, you know, so that it can anticipate new scenarios. So similarly, you know, instead of a road, Our road is essentially the patient or the human, exactly. But what we do is that we take all of these inputs ranging from your genetics, from your blood, your brain images, and your brain tissue, and we fuse those into a single internal representation of each patient.

39:54So actually, each patient is represented essentially as a 512-dimensional vector. Oh, okay. So this boils down to a series of numbers in a list. Yeah, exactly. It's a lot like the kind of current large language model architectures. Not to be a total brat, but vectors are, I think, one-dimensional tensors. Do you actually use vectors or do you use higher dimensional tensors? So the actual model architecture is at the kind of core, it's a transformer in the same vein as ChatGPT, Cloud, and other LLMs. So it leverages the flexibility of those transformers, but it has several innovations that are unique in the bio.

40:30The first is that each data actually gets its own encoder, each data type. So it's multimodal. And that maps it to the shared kind of mathematical space. So blood, brain, and genetics can all kind of live in the same kind of mathematical language. And then we fuse all the data layers into a single unified vector that represents each patient. And the way you think about it, it's essentially like a patient fingerprint. Okay. Right? And then the last thing we do is we use what we call contrastive alignment. So this is actually a new architecture. We use a form of it that's a new architecture that's only been developed in the last 18 months, which is called contrastive multimodal learning.

41:06So unlike your kind of classic contrastive alignment, which, you know, image models often use, and that only keeps two types of the kind of data that two types of data agree on. Ours keeps three things, which is, you know, what the blood knows on its own, what the brain knows on its own. And then what's the synergy between both that combines them? And in biology, the synergy is huge because it's where kind of real signal hides, where no kind of single measurement can capture. And so lastly, once we have that fingerprint, then we freeze it and we can build a bunch of task heads on top of it that answers specific biological questions like who is going to respond to this drug?

41:46What does their brain look like? And what is cool is that this form of training, because we are using masking, actually starts to learn tasks that it was never explicitly trained on. Can you explain masking for me in that context? It's a bit similar to kind of how AI does masking, right? Which is that, so in large language models, AIs do masking by actually hiding a word and then predicting what that word is. For us, we have all types of data, blood, genetics, brain. And what we do is that we can hide one type of data and the model trains by learning what data type is missing and how to fill that in.

42:25So as a result, it can start learning tasks that it wasn't taught. So we have seen that our own model with high accuracy can actually accurately reconstruct brain activity from blood alone. And that's actually not a task that it was asked to do. It's just a simply emergent property of this training task. I love AI. It always finds some new way to delight me and make me excited about the world. Okay. So you went from the first iteration of the company, Verge Genomics. We're going to identify candidate drugs and test them and bring them to market. And now you realize that your technology is probably a better tool for other people to go out there and do the very expensive guessing and trials work.

43:04It makes a lot of sense to me. Who is the target customer for this new iteration of Verge? So it's really anyone that's developing a drug, right? It's the whole pharmaceutical business. Companies can work with us essentially three ways. They can first come to us with a specific problem and we can run our targets against it. We can also directly license insights or targets that we've already found, or we can license the data and models directly. So, for example, if you're a company with a phase two drug in schizophrenia and you're like, holy cow, this drug is behaving differently in every patient, you can come to us and we can help you pick out which patients to roll in your next trial that actually respond to your drug and let you design a much smaller and cheaper clinical trial.

43:52that's so many ways to make money and the companies that you're going to have as customers are uh famously large and frankly quite wealthy which is good for you guys um do you charge for this on like a per case basis it sounds a little bit custom on the pricing side if that makes sense um so we have you know we've done two major partnerships already actually with eli lily and um astrazeneca alexion um those are target discovery partnerships or if you're more traditionally structured. So it's in those cases, it was a 25 to 42 million upfront with then milestones that total up to anywhere between 700 to$800 million each as a very traditional therapeutic structure.

44:34Now we've opened up new platform models as well that allow you to engage with it more kind of how you might used to be engaging with kind of a direct model license, Right. So, you know, companies like, you know, in the space like Chai and Noatech have done kind of these multi-year licenses to pharma companies. We also work with smaller biotechs as well in a kind of platform as a service format where they have actually a specific question. They can come to us and we can kind of answer on a on a question by question basis. the deals you're talking about uh back when you raised your series b in uh 2000 i think it was late 21 you said that the company had announced a 706 million dollar partnership with lily to quote develop new treatments for oh hell oh als there you go using its platform so how did that contract go did the milestones come in because one thing i noticed alice is that you guys haven't raised money in a while which is fine but also may imply that there was some revenue along the way We did.

45:32Well, we haven't announced publicly any additional funding, but we have done those actually a few major deals and we've raised some unannounced funding in between. Those partnerships are also did provide some milestones. So Lily actually in 2024 announced that they actually optioned two of those targets into their internal ALS pipeline. So it's actually the first AI-derived targets that were actually internalized into their ALS pipeline, which we're quite proud of. And one thing that was actually really quite striking from that partnership was going into the partnership, Lily had said to us, you know, even if 20 % of these targets validate in the lab, that would far surpass our expectations.

46:17And we actually found at the end of that partnership that 83 % of those targets actually validated in wet lab experiments. So that kind of far surpassed even our own internal expectations and starts to create this kind of surplus bullpen of targets that we can continue licensing. And by targets, we're talking about ideas for drugs that might solve. OK, cool. Sorry. It's actually like, what are the proteins to go after with a drug that might cause disease? The target proteins to go after to help either reduce or resolve ALS in this case. Yeah, exactly. Yeah. Okay. So you guys are focused on the brain, which I think is fantastic because I'm a big fan of having my brain and working in all those good things.

46:56And also I would like to live for a long time with my mental faculties. But I'm curious about the idea of taking in people's information, tissue samples, and applying AI to them. Does that work in a similar way, for example, in my liver? Or is this more of a system that is set up because the brain works a certain way and it wouldn't be applicable to other organs in my body? Yeah, absolutely. And there are other companies that are doing something similar in cancer. The reason it is such a big problem and so hard in the brain, though, is because the brain is the hardest organ to access. So pretty much in any other disease, in cancer, in fact, it's standard of care to get your tumor kind of taken out, to get it analyzed.

47:40In IBD, you often do that. Most tissues, you can actually go and take a sample of that tissue and the patient can continue living. With a brain, you simply can't do that. So being able to accurately reconstruct what's happening in the brain has been one of the field's longest standing challenges. And it's why I think, you know, neuroscience is long behind cancer by 10, 20 years. And it's really honestly probably the biggest driver of mortalities in the next generation as we all get older. It will really be Alzheimer's disease and dementias. Yeah. No, I mean, I'm at the age now when my parents are in their mid-70s and you start to have thoughts and fears about how they're going to do and what we can do for them and how to care for them.

48:26This is very apropos to things that are near and dear to my heart. One thing, though, that I've heard from basically every AI-ish CEO that I've spoken to, and I include you in that bucket, of course, is that if they have more compute and they have more data, they can do a much better job over time. It's kind of a standard kind of path that direction. Does that same relationship apply to the second version of Converge and also understanding which proteins in the brain we want to go after? Or is there a limit that is different from other applications of AI in that context? So actually what we have found so far, we actually have not deliberately chased parameter counts yet because the biggest gains we've seen have actually come from scaling data and modalities.

49:12so it's not about making the AI bigger it's about feeding it the right pair of data but you kind of contract like something interesting between kind of text models and bio models are of course in text models scaling just works you have this kind of everyone believes that if you just make it bigger and it gets better but you know the reason why that is in text and I don't think most people realize why is because when you're actually training a text model you're predicting the next word in a sentence. And that task inherently forces the model to learn everything about reasoning. For example, reasoning, code, tone, everything.

49:49But when you interact with the real world like biology, self-driving, robotics, it's much, much harder. Because first of all, there's no single task where predicting the next thing can teach you whether or not a drug works, in which patients, whether it be toxic. Most biological data are actually proxies, So they're kind of shadows of what's happening. And most biological data is observational, but you're actually wanting to ask counterfactual questions like, what if I take this drug, what will happen? And kind of analogous is kind of self-driving again, because, you know, Waymo didn't solve self-driving by collecting just simply more and more camera footage.

50:25they fused sensors cameras lidar radar maps etc and so biology is the same you really need to fuse modalities rather than just scaling one and but it's even harder because you don't have a perfect geometric representation of the world like lidar does biology doesn't have that kind of same sensor so the takeaway in biology is that scale really only matters when it's pointed in the right data in the right direction so it's not to say that scaling doesn't matter but it's i think in the beginning, the gains will come from kind of combining the right data sets and scaling the right data. So then would a major unlock for the company then being able to access more brain tissue samples to expand your underlying data?

51:08Yeah. And that's what we're doing, not just more brain tissue, but more modalities. So on the roadmap for us next is bringing in imaging, bringing in proteomics, bringing in even longitudinal data so that we can not only predict a snapshot of the brain but we can actually create a virtual model of the patient in time where we can actually run forward each person see when they'll get the disease how the disease will ah wait no i don't like no wait a minute you everything you've set up at this point has been fantastic but then you just told me you're gonna tell me what i'm gonna die and i i don't i don't know alice if i'm on board for that more knowledge is power is it though i sometimes ignorance really is bliss uh okay but if if i'm being serious if you were to tell me you are at risk of getting alzheimer's or whatever dementia early then i presume that i could take at least some steps to limit that risk and manage it okay that that makes a lot of sense and in alzheimer's disease a lot people think it's actually not even just finding the right drug it's actually being able to intervene early enough to change your trajectory so that becomes even more important but i want to to get back to the data points.

52:20So scaling parameters is not that important. Having the right data, very important. When it comes to text, you can scan books, right? It's a little bit easier to talk about than deceased people's brains. But is there a good pipeline of fresh deceased brains that you can, if you wanted to, access, collect more, and then expand your data sets over time as you learn more and tune your own models? I mean, that's really what we spent the last 10 years building is that end-to-end infrastructure. And it really took us 10 years. People always ask, you know, why aren't just big pharma companies doing this themselves?

52:53I mean, the real answer, it's not impossible, but it will just simply take a very long time and it's very hard. And so it's really the kind of unsexy blood, sweat, and tears that we put in over the last 10 years that have created the moat for us. And it's how we'll continue scaling these data sets. And what's exciting is that we are seeing scaling laws in our data, right? where then they're non-linear and increase as we add samples. And we haven't even started working on scaling the compute and the parameters yet. So there's still massive headroom for growth. Okay. So basically you've done all the hard work to have a pipeline of useful brain tissue samples.

53:29Other companies don't have that. So not only are you ahead of the game in your particular niche, but also you have a unique advantage of having more data. Okay. I want to spin the clock and look ahead a bit. Like not this year, not next year, but a couple of years down the road. I think some people have been impatient, incorrectly, but impatient with the pace of medical progress in the AI era. I think people have been seeing coding agents do so well and say, hey, why aren't we there with drug discovery and health yet? So if you could take a like a 50 percent confidence interval guess about where both Verge is and other companies in the bio AI space, where are we in five years?

54:07What have we unlocked? And are we going to feel that difference in our kind of lived medical reality? Yeah, I mean, I think even with some of the text models, right, that progress all happened very quickly. And there was also ongoing, you know, work that was going on behind the scenes that enabled it. I think with every technology, it's always a process of iteration and learning and facing setbacks and then learning from that. And then once things start clicking, right, progress gets made exponentially. Right now, I think that what's really exciting is just some of these transformer-based models and these world models are just performing in ways that we didn't expect.

54:50Even with us, we're starting to see, you know, performance on tasks like brain prediction directly from blood that it wasn't trained on that are far exceeding current clinical tools. We're seeing, you know, prediction of responders. And so what I see in five years is really I think AI will come into the pipeline at multiple points from multiple different models, right? I think you'll have models that are able to, you know, predict, hey, what patients will respond to what drugs. And I think the future vision for that is you can have a continuous monitoring of your health state, right? You can figure out, you know, when you're going to get disease, when you want to intervene.

55:29And that really brings us to a world of true personalized medicine where we're no longer just thinking of Alzheimer's disease as one disease, but we're thinking of hundreds of diseases where you might just have one form of a disease and you can really then seek a therapy that perfectly matches to the specific disease that you have as Alex or that I have as Alice. And that's really the way to start extending health span and age span is really by being able to address these chronic diseases. So when we sequenced the human genome, it cost like a bajillion dollars and took a while. Now we can do it for like$4 or something crazy.

56:07What you're describing to me sounds fantastic, but I'm curious about the price curve and if you think it's going to become something that is accessible to people, let's say on Medicaid versus with all of our friends and their concierge doctors will get it first, but will it make it down to the people that are less resourced? Well, so in terms of pricing, the thing I always think about is why are drugs so expensive now? it's expensive because it costs$5 billion on average all in to develop a single drug, right? And so that's reflected in the price. Why does it cost$5 billion? Actually, the vast majority of that$5 billion is getting spent on failures.

56:45It's because nine out of the 10 attempts fail at the last stage in the most expensive stage. So if you can actually be able to reduce that, even by a small amount, that has huge implications for how much is saved. And that ultimately is going to be the thing that drives down the cost of prices sustainably is actually being able to be much more efficient at how you develop drugs. So that's really what I see as the long-term solution is if you can perfectly with accuracy predict kind of which drug will succeed, it goes from 5 billion to really, you know, tens, tens of millions to really get a drug all the way through.

57:22And so you can see orders of magnitude reduction kind of then get pulled through to actual, you know, what the average consumer will see in and how much is paper drugs. So as we have better selection of possible drugs, we'll have a lower failure rate. Therefore, we'll spend less money spinning our wheels, spend less time wasting there. We can therefore offer better, more targeted drugs at a lower price point, keeping this in everyone's medicine cabinet, to use an analogy, I suppose. That's fantastically good news. I'm pretty excited about all this. Is there anything that like you're worried about that might not work out?

57:54Because this all feels like You have the tools, you have the technology, you have the data, and off to work you go. But are there any science risks left? I mean, I think the biggest thing I always like to warn people about is that people always like to be very reductionist about how they view technologies. Yes. They always like to say, oh, this drug has failed in clinical trials. AI doesn't work at all. Right. And rarely in the case of any transformational technology has the first attempt ever been the blockbuster success. In fact, actually, transformational technologies get built because people continue to learn from setbacks.

58:32they feed that back into their platforms and then they improve from those. And so I think the biggest risk is more of a human one, which is that we kind of lose interest in AI or in the application of AI in healthcare, just because we kind of face one step back and we generalize that about the promise of the whole technology. But I think that technology is built through iteration and transformation and kind of continued persistence. Yeah. Yeah, I hope that no one takes an early failure as indication that things don't work. I mean, if we believed that, we would never be in rockets, for example, as a species.

59:09Because if you go back to the early days of rockets, it wasn't exactly like they were coming out of the assembly line and going straight up. They were not. So it takes a lot of time. And in pharma, there's this tendency when you have a clinical trial failure to essentially just look away and just move on to the next thing. And that's why when we had our clinical trial, which didn't pan out, instead of looking away, We published the details and results in detail. We took all the data and we fed it back in the platform. And we said, hey, this taught us a really hard one lesson about what's important in this space.

59:39It gave us all the data to be able to address that challenge. Now let's feed it in, make the next version actually address what we missed, and then actually build an even better kind of tool on top of that. Well, you have me feeling both optimistic and excited because I'm starting to reach the age in which my body gets dings and scrapes and nicks and needs a little bit of help here and there. So I'm really glad that you're working on this problem and other companies are working on cancers and so forth because who doesn't want to live forever, Alice? All right. For folks who want to know more, it's no longer Verge Genomics.

1:00:11It's Verge Labs. What's the URL? And is there a job you want to shout out to the audience in case the right candidate is tuned in? Vergelabs.com. V-E-R-G-E, labs.com. And we are always looking for great AI research talent. So if you are interested in AI and biology, give us a shot. How hard is it to hire right now in that particular space? I know, it's crazy. No, I'm actually curious because I'm not sure if the people that are going to work for Anthropic are interested in the same problem space. So I'm kind of curious if your focus gives you access to talent that might otherwise be absorbed by the major labs.

1:00:47Yeah, it's kind of in the space. What is hard is finding the intersection of, you kind of have to ask, do you want AI? Do you want biology expertise? Because there's kind of folks from the Frontier AI labs, and then there are folks with kind of biology training that have developed foundation models. We kind of sit in between both. So it's actually more about finding the unicorns that are interested in both. So it's either people that have had deep Frontier AI experience that may have had a personal experience really with one of these diseases. And so actually what we find that once we find those individuals, it's actually quite easy to recruit them because there's such a strong mission alignment.

1:01:24And it's so kind of what we're doing is so differentiated from a lot of the other companies out there. But it's actually about finding those people that have both. If you're curious, founders, what people mean when they say mission, that is mission, not improving barbershop, CMS, phone call, cold outreach response rates. All right, Alice, an absolute treat. Please come back on in six or eight months when you have more news. I really want to keep track of what you're doing because I think it's fantastic. Thank you. Thank you so much, Alex.

1:02:14The Launch Accelerator invests$125 ,000 and connects you with 500-plus investors to help you raise your next round. Apply at launchaccelerator.co. If you're an accredited investor looking to gain access to quality deal flow, apply for Jason's Angel Syndicate at thesyndicate.com. We find two to three deals a month. And check out This Week in AI, Jason's experts-only roundtable with top AI founders and operators every week. Find it thisweekina.ai. Check out The Twist Ticker, our daily newsletter, at thisweekinstartups.com slash ticker. Follow the show on Instagram. Follow the show on x.com. This Week in Startups publishes three days a week, Monday, Wednesday, and Friday at 5 p.m.

1:02:59Central Time. You can submit an audio or video file question by emailing it to thisweekin.com.