AI Creates New Viruses, Building AI for Healthcare & Moats in Healthcare IT | Lightwork

0:00New viruses designed by AI sounds scary. My sense is a lot of people are freaking out that there is this idea that anyone in their garage can now create a virus. In this case, you needed a wet lab and trained researcher. It's not that AI designed and created a virus. It's that AI helped design a roadmap for creation of that and then scientists were in the loop. The healthcare Super Bowl, as you said. Health IT Super Bowl, yeah. This is the time to be bold and to be ambitious, and it's truly an energizing moment.

0:36We're back with another big week, this time a healthcare and science edition. Scientists built 16 viruses designed entirely by AI, not found in nature, and they're aimed squarely at the drug-resistant bacteria that our antibiotics can't kill. Then we chat through Vertical Healthcare AI and all the exciting opportunities for founders who want to build in this space. Breaking it down with me is Dr. Brenton Farnoli, who is a physician and partner on our health tech team. And he's sitting in while Josh is out. Let's get into it. Well, Brenton, I am so excited to have you on here because especially given the news from this week, I feel very excited that there is a real doctor and someone who has studied all of this in front of me.

1:22I did a video recently on the headline that scientists create 16 new viruses that have never been found in nature before. And it caused a lot of buzz. Were you able to check that out? Yeah. New York Times, Wire. There's a lot of press and even more on social media buzzing about this headline grabber. Yes. New viruses designed by AI. Sounds scary. It does sound very scary. It was really interesting because when I had read it from the tech sphere, it was positioned as this is really groundbreaking for what we can do in the antibiotic space, for some of the foundations of this idea that we can have like N of 1 biologics.

2:02But then, you know, I think the flip side, the more scary fearmonger side is that this could potentially turn into a world where anyone is creating any sort of virus. So I thought it might be helpful to actually demystify what exactly happened. I don't know if you wanted to talk through a little bit more about what exactly happened with these 16 new viruses. Yeah, let's chat about it together. Yeah. So I believe it was a paper in Science Journal that came out from some Stanford and Ark Institute researchers. And they had used, really think about it as a genomic or genetic language model as opposed to a large language model.

2:43Genes themselves are almost like a language. you have AGCT as the core nucleotides, and they use the genomes from a certain type of virus called a bacteriophage and used, I think, a couple million of those genomes to train a model to then design specific bacteriophages. Now, the key part here is they use the data set of bacteriophage genomes. And now bacteria phage, the name is bacteria, which is bacteria, and phage, which comes from, I believe, ancient Greek, which is to devour, to eat. So these viruses do one thing, they infect only bacteria, and they devour or eat, and therefore kill those bacteria.

3:29And the reference I always go back to is actually in Project Hail Mary. I don't know if you've read the book, but the whole premise is that they're going after these astrophages, this idea that there are these sunlight-eating bacteria out in space. And I think if anyone's seen them, you can understand what these bacteriophages do. Right. And so the researchers were able to use this, and they designed a number of bacteriophage genomes, some of which created new variants of viruses, and that's really a first of what we've seen. And so that is both exciting because bacteriophages actually play an important role in medicine, in the world and ecosystem, and also a bit troubling in terms of this proof of concept of being able to design viruses with AI.

4:24Well, what other types of viruses might be able to be designed? And so in this case, you had researchers following really good practices and with good intent. But I think some of the buzz is the question of, well, what if, what about when, and how might we think about this in terms of biosecurity, biosafety, and things of that nature. And there's a lot I want to get into around how we build guardrails around this. I think where people are concerned are not necessarily that tomorrow someone is going to release a bunch of pathogens that are going to kill us all, but rather the capability jump that we believe will happen and continue to happen given the inflection points that we're seeing with the underlying models.

5:09that i think is is worthy of discussion but just to even you know demystify and clarify some things bacteriophages scary word not necessarily bad they're very common it sounds like in bad if you're in this case e coli the bacteria e coli because those are uh were were targeted but yeah okay and in this case my understanding is they were only going after e coli so there's no world based off of what even the data that they fine-tune i know that i think the base model is trained on, it sounds like, 9.3 trillion nucleotides. So all organisms spanning bacteria, plants, animals, humans, and viruses.

5:47But for this specific model, EVO2, they had only trained on bacteriophages. I've seen a lot of comments that are like, this might be able to jump and evolve into something that could infect humans. Is that even feasible? yeah it's not not really the case here but i think the question is more um if the scientists were using different sets of data that they were if they were able to access and they had other intents than what could what could be um the interesting thing about bacteria fish is it's actually quite a simple genome and so to your point on well with ai improving exponentially more complicated genomes and generative a generative genomics could could be at hold But this is really kind of the first example.

6:32And when we talk about what this could unlock, I think the exciting part is this idea that we could develop almost more variants of bacteriophages given what's happening in the antibiotic phase. So I think generally a trend line that we're seeing is that there is worryingly a lot more antibiotic resistance. And so tell me a little bit more about that. Yeah, so bacteriophages, you know, have the potential and, you know, in instances have been shown to play a role in, you know, limiting the spread of bacteria and whether that's in agriculture or for human types of purposes. And obviously viruses in general can be quite scary and bad, but they also can be used for medical applications, whether that's for gene therapy.

7:24Typically, the vector of delivering that gene therapy is a virus. And so being able to design and control that better can always be helpful. And I also go back, I think an important point of this research is it's not that AI designed and created a virus. It's that AI helped design a roadmap for creation of that. And then scientists were in the loop doing the work and doing the synthesizing. And so that in itself also is a guardrail of there's a human in the loop of these parts of things. But I wanted to also note that. Yeah, I think that's really interesting because I do think there was a lot of the assumption that AI had basically one shot at a virus, but when in reality it was immensely helpful in the design process.

8:12But ultimately, in order to manufacture and produce something, it has to go through an actual lab. So I guess there's a question of like, you can do the actual DNA design, but then my understanding is like the manufacturing part, there's a lot of guardrails. There's only a small number of companies you can actually bring this virus to market and manufacturing. Like, I think the information part is easy, but the actual production part, it sounds like, is quite difficult. Is there a lot of regulation in guardrails right now? I think basically what I'm getting at is that my sense is a lot of people are freaking out that there is this idea that people can, you know, anyone in their garage can now create a virus.

8:51Well, in this case, you needed a wet lab and trained researchers and things of that nature. That's really where I think there's opportunity to bring in some of this regulation because that still remains the bottleneck. I guess as you think about this space evolving and, again, these capabilities evolving, where do you feel like safety measures need to be put in place? Yeah, I think it's a broader discussion, right, as you're talking about frontier intelligence. And I think you've seen a couple themes come through, which is one, even the most advanced frontier models in biology and in cyber and some other areas actually revert back to less advanced models when queried on such topics.

9:31And so some of these labs have been putting in guardrails on their own models. And the other is, when is there a human in the loop versus AI going through completely? And so I think those are two areas that are, I think, part of a broader discussion with AI and intelligence and certainly is a specific example here with the design of AI-generated design of viruses. Do you think people should be worried with this or what are you thinking in terms of the timeline? I mean, I think, again, zooming out, it's similar to the broader AI conversation, which there's so much hope and excitement, but there is also trepidation in questions.

10:13And so, you know, there's both bring both that tension, I think, can live live together in one. And this is this is probably one of many stories that will pop up over the coming weeks, especially as these proof of concepts turn into the hypotheticals of, well, what if this when? And I think it's conversations like this and others that hopefully can catalyze the discussion so that we are prepared and thoughtful around, well, what if, when we've played those scenarios out across key stakeholders. And while I have you here, I think there's been a lot of discussion around questions of building verticalized AI solutions in a world where big labs, big tech, incumbents are all building in overlapping spaces.

11:01Since you spent a lot of time in the healthcare AI world, you figured it might be helpful for a lot of the founders and operators who are tuning in to get a better sense of how you're thinking about the space. I guess for starters, as an investor looking at broadly verticalized healthcare, are there any mental models as you look about the white spaces and opportunity spaces, given that there are all these players all kind of hovering around similar spaces? Yes, certainly. And, you know, health care isn't one monolith, right? You have life science companies, payers, employers who pay for a lot of health care, and even provider groups.

11:36There are large academic medical centers, smaller health systems, community practices. And there are pockets, small and large, of opportunity within all of those with different approaches. I would say we've certainly seen a lot of uptake in adoption at the earliest phases of this AI era in the provider space, which actually wasn't always the case, right? Health systems often quite slow to adopt new technology, private practices and independent practices as well. that has certainly flipped as you've seen the scribing explode in use as a use case, as you've seen voice AI for scheduling. And so these areas that had been left behind by the SaaS era and really hadn't adopted as fully as other industries, you're having this leapfrog moment in that area and great founders and companies kind of dashing to create opportunity there.

12:38And at the same time, this has not gone unnoticed, as you mentioned, by the big labs and incumbents like in Epic. And they have their UGM, their big meeting coming up right now. The healthcare Super Bowl, as you said. Health IT Super Bowl, yeah. And so here's also always this question of, well, how does this insertion grow over time and expand as others take notice who might have their own platform or distribution and other things in that regard? I want to zero in a little bit into the first part. I want to also definitely get into how to think about players like Epic also with massive distribution playing in the space.

13:21But with the first part, I think what you said around this idea that almost the dynamics have flipped where you are seeing a lot of uptake and adoption inscribing in some of the intake parts where historically SaaS is ignored. What dynamics do you feel like have made this environment possible for startups to start building this space? Is it the AI part? Is it, you know, what's happening broadly with, you know, the idea that, like, you know, I think all software is being challenged, all existing paradigms around software is being challenged. What do you feel like makes this space so ripe right now?

14:00one is the delta between where things were and this absolutely delightful we've heard uh when demos are being given kind of this is like a magic show a reaction and so that wow factor uh you know in venture you often talk about well is this product 10x better and people say 10x and all this and you actually now have have examples of that and so that is resonating whereas before it was, you know, a little bit better type of thing. And the other is, you know, on a physician myself, you know, providing care has become, you know, quite challenging. There's insurance burdens. There is, you know, bottom line pressure.

14:40And so the buyers are really looking for ROI on their investments because their bottom line is pressured. And so that also creates an urgency, right? Necessity is the mother of invention. So you have that makeup there. And then the last thing I would say is some of these are able to scale quickly by bottoms-up adoption. And so physicians latching on to AI of their choice and not waiting for the administrators to get around to selecting a vendor and contracting and things of that nature. Yeah, I think that tension is really interesting because the buyer is making decisions for the end user. And I think oftentimes they have different preferences.

15:26Have you seen? And you see that in the EHRs all the time. The NPS from the users often quite low, but the buyers, you know, they like how it documents and gets the billing done and does these types of things. And so if you're a founder building in this space, how do you even think about all these stakeholders? Because I think oftentimes it almost feels like you have to trade, or it's a trade-off between the two, but it sounds like some of these newer solutions are actually, if anything, seeing massive adoption from bottoms up, but then also satisfying buyers as well. Exactly. Yeah. There's multiple approaches right now, right?

15:57Top down, bottoms up, and different segments, right? So you have some founders attacking the independent practices, and there's multiple specialties there with different bespoke needs that often a large incumbent or another large player can't quite satisfy in a really focused, focused, directed way. And then I think on the second part, this is a really hard, I think for any founder in this ecosystem, it is one of the most exciting times to build with just what capabilities there are that have been unlocked. At the same time, it means that everybody has access to some of these same capabilities.

16:38And so amidst labs, incumbents, how does a founder think about opportunities in these verticals? Is it important to keep up with what some of these incumbents and well resources labs are doing? How do you even advise a founder trying to navigate the space? It's a great question. And yeah, you have to assume over time, your insertion, there will be others like that for the reasons you mentioned. And so really the key is to insert at a strategic surface area and a point from which there you then can unlock more capabilities and turn more and more into a platform. And oftentimes we see that strategic surface area being at the moment of interaction with the patient, right?

17:23Time a patient calls in for an appointment, time a patient speaking with their physician or a time another physician is faxing a record to another physician. And so those moments of communication, typically a human receives them, documents them, puts them into the system of record, the EHR. But now you have this moment where maybe it's not the human recording that, but AI or an agent. And that creates a moment for the agent to then do more with that data beyond just putting it and compressing it into a system of record. And so that's really where you're seeing this tension between the EHR system of record and this kind of system of action layer, EHR moving into action and the action layer moving into becoming more and more systems of record.

18:14Yeah. And I think you actually see a common theme around that even in healthcare, but also in other circles, right? And I think what's interesting is when we interface with founders, there is this question of like, okay, I do ultimately want to build this venture scale business. I want to see a big outcome here and create a ton of value creation. But it is this process of, to your point, starting with a very interesting wedge of identifying the right wedge at the right time and being strategic about the ways that you build out that multi-product solution. It is not about day one going in and saying, I'm going to be everything for everyone and become this mega platform.

18:52A lot of it starts with some of these more strategic wedges. And I think you've seen this, you know, you've laid that out, what that looks like for healthcare. I think you see that in legal and finance. A lot of these tools kind of get a lot more traction starting with certain wedges that get a lot of that bottoms up adoption or have that very clear 10x value and demonstrate that and then go from there. When you think about what actually compounds into Advantage, how do you identify what that wedge? Is it starting with a place where there is that sort of very real data flywheel? Is it a place where there's a lot of workflow lock?

19:27I mean, I think all these are levers you can pull, but for healthcare, what are you seeing as these sort of levers? You fit some of the key ones. The other component often is the team. So you can have the same insertion in one team fixates there and another is able to see the broader vision, stitch things together and really have the buy in from their customer and the trust to get to get pulled beyond that. And so that ends up being a key piece. And those teams attract the talent, attract the capital. And that flywheel then expands. Yeah. I mean, it's hard. It's like you got to think deeply and like execute well on that insertion point.

20:09But you constantly think about like expansion and that multi-product vision. What do you look for in teams that can, you know, you think can execute on this? Yeah, so one is just really understanding the pain point and caring deeply about that. There's certainly examples of physicians in healthcare going after this and really understanding the user and the buyer and caring deeply about that and that really resonating. That's certainly one. Other is we're seeing more and more of these AI native technical founders who may not be the domain experts, but they are able to really use that as a benefit from first principles saying, well, how would we redesign this process, knowing what we know about AI in uncovering those pain points anew?

20:56You know, some founders actually, you know, spoken with, say, well, will you bring on someone who knows about this esoteric area of health care that you're trying to solve to the team? And they say, default, no. We want to really be first principled about it and using AI. So there's different approaches there. And depending on what you're tackling, I think different approaches working. Certainly as you're getting closer and closer to clinical care and patient care and the sensitivity and the complexity around that, often helpful to be default yes to bringing in domain expertise. But some of these other areas certainly can be shaken up with new approach.

21:37Yeah, and I definitely want to, you know, teasing a show that we're working on here. I think there's so much to also discuss about, you know, the team that we have at Lightseed where you have all of these kind of doctors turn investors. And I want to definitely dig into this idea of like doctors turn entrepreneurs, but also this idea of how do you build a great, you know, healthcare tech team. because I do think, to your point, it varies greatly depending on your insertion point, what part of the stack you're building in. I guess since you have so much visibility into how a lot of founders are attacking these spaces, are there any lessons or road bumps that you've commonly seen a lot of teams face?

22:20I think one is just keeping the long view here, which is at the end of the day, you're trying to help patients physicians um and doing that durably versus there's there's shortcuts you can take um there's arbitrage moments you can make uh to scale revenue in the short term quickly uh but these are multi multi-year efforts and again over time as we said others will start to crowd around and so if you aren't building for that long term with that long vision. And, you know, you need to end up really building for that long-term vision, having the tenacity, the courage, the passion, passion to do so.

23:00And we're, we're fortunate to work with, with many such, such founders. Yeah. It's an exciting time. I think maybe to a more pressing event, sounds like with the, you know, I keep calling it the healthcare IT Super Bowl, but given that it's coming up, I'm curious, just as someone who was tracking a lot of the developments and features that are being built in this space, Epic's obviously a very important player in the ecosystem. Do you have any predictions for what they might release there? Yeah. So it's interesting you said curious because one of their key innovations is called curiosity. Oh, okay.

23:31So I'm also curious about curiosity, where they use all of the data within Epic to train a model to predict for this patient what would be the best next action or clinical care. And so certainly last year they announced they were getting into scribing, which created much buzz. This year we'll see what they announce. But I think that area is one area in particular, how they're using their corpus of data to inform clinical decision support will be an interesting one to watch. Yeah. And to end, we've always used this as a space for people to better understand the type of startups that Lightspeed is spending time on, the ways that investors are thinking about the space.

24:20Do you have a request for startups or a space that you're particularly excited about right now? Well, that's a great question. You know, first and foremost, we're looking for just exceptional people to partner with in this moment. There's so much to do, whether it's, you know, AI native care delivery in the provider side, whether that is stemming the, you know, insurmountable cost increase in employer health care, whether that's leveraging robotics in the OR for manufacturing. robotics play a role throughout a lot of healthcare, all the way down to the data and models and infrastructure underneath that are healthcare specific.

25:02So there's a lot to do. And we're really excited to continue to work with such founders and meet with them, trade notes, be a resource. It's a very exciting time. Yeah. And we just did an episode of Ravi Raj on our team last week on just everything happening at the frontier. And it's just so exciting to be able to, you know, have folks on team imagining a world where, you know, we can have abundant life and like, you know, really expand our health stand and lifespan. And even just this idea of all these things that we're talking about converging world models for healthcare, but then at the same time, a lot of what our team is spending time on, to your point, very practical solutions around just cost saving and, you know, fixing some of the frictions in the healthcare systems that we have.

25:51And it is just exciting to see very practical solutions, but also these sort of sci-fi things that you could never imagine. I was working with Ravi Raj over the weekend on a sci-fi healthcare, you know, a healthcare frontier. So it required both of our, both of our team's heads coming together. It's amazing. We'll have to do this episode because I think when Ravi Raj broke down, And, you know, a couple of big themes were covering space, obviously, this idea of abundant energy, intelligence. But one big one was life, like this idea that, you know, we now have massive opportunities in the health care world.

26:25And what does that actually look like if we really think about the sci-fi future? Yeah, I think and I know we started the conversation around AI designing viruses within health care and the questions and concerns that might might raise. I think overall, as you look at the different industries where AI is being applied, healthcare seems to be one of those which is, you know, the most needed and can be the most beneficial. And as a result, you're seeing tremendous founders from outside of healthcare, whether they're from the large labs or otherwise, saying, I'm starting something new. And that something new, you know, increasingly is in healthcare, which is, I think, a very encouraging sign.

27:08Yeah. And if any of those folks want to build in this space, I think this is the time to be bold and to be ambitious. And I think it's truly an energizing moment. And I'm excited that we have great founders all excited about this space and feel free to reach out to our healthcare team. Yes, please do. Thank you so much for being here today. Hopefully we demystified some very scary headlines, but also some very helpful advice for founders building in this ecosystem. And I'm excited to see what solutions get built. I'm excited for more to come, more innovation. It's truly just one of the most exciting times.

27:47So thank you, Brenton. Thank you.

28:01ご視聴ありがとうございました