Ep 27: Oscar Co-Founder Mario Schlosser on How LLMs Can Be Used in Healthcare Today and The Path to AI Doctors

Oscar Health is a $4B public healthcare company, providing healthcare insurance to nearly 1 million members. Oscar is at the forefront of AI adoption, continuously developing new AI use cases in healthcare. On this week’s episode of Unsupervised Learning, we sat down with Oscar Health Co-Founder, former CEO, and now President of Technology Mario Schlosser to talk about where AI will have the biggest impact in healthcare, top AI use cases at Oscar today, AI adoption challenges Oscar is facing, and limitations of GPT-4 in healthcare. Mario also shared his takes on open-source vs. off-the-shelf vs. healthcare-specific LLMs, and why can't we have robot doctors today.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Jacob Effron?

  I love it. The most powerful brain model we've ever built still can't understand how the world health insurance works.

• 0:05Speaker 01

  So maybe we shouldn't piss off our smartest robots by using it for health insurance tasks because that's what's going to make them kill us.

• 0:13Jacob Effron?

  This is Unsupervised Learning, Red Points AI Podcast. I'm Jacob Effron, and today we had an awesome episode with Mario Schlosser, the CTO and co-founder of Oscar Health. Oscar's a $3 billion public health insurance company. They've been at the forefront of innovating in tech and healthcare for the past decade. And so Mario just has a fascinating seat and lens toward all things happening in AI and healthcare. Oscar is both an insurer. They provide care to patients. They're experimenting with GPT-4 in pretty much any way you can in healthcare. I think they were one of the first people to get OpenAI to actually be regulatorily able to work in healthcare. And Mario just had some fascinating insights on what it means to actually implement these models in the ground, in the weeds, for real use cases, as well as some broader kind of zoom out looks on the future of healthcare and AI. I think folks are really going to enjoy this conversation.

• 0:57Jacob Effron?

  Well, Mario, thanks so much for coming on Unsupervised Learning. Really excited to have you.

• 1:01Speaker 01

  Great to be here. Thanks for having me.

• 1:02Jacob Effron?

  Yeah, you guys have been pretty prolific in publishing a lot of the AI work you're doing. And so I feel like it will be, it'll be fun to dive into a lot of threads from, from, from some of the things you published on Medium and Notion and whatnot.

• 1:12Speaker 01

  Yeah, no, looking forward to it for sure. Let's see if I can represent it verbally as much as in written form.

• 1:19Jacob Effron?

  I certainly believe.

• 1:20Speaker 01

  And LLM should do that. That's like very similar to the LLM restrictions. You know, they can, they can write quite well and they can't speak as well just yet. So.

• 1:26Jacob Effron?

  I guess I'd love to just start at the highest level. You know, you're one of the first kind of healthcare guests we've had on the podcast. So I'll just kind of ask you, in the next decade, you know, where do you think AI is going to have the biggest impact in healthcare?

• 1:37Speaker 01

  So one of the things that I think LLMs are very good at is the, is going from informal language to formal language and the other way around. From literally, you know, conversation to summarization, from comments to writing code and GitHub copilot, whatever else. And I actually really think that healthcare might be the single industry where you have the most of both. We have the most formal language, ICD-10 codes, you know, CPT codes, utilization management guidelines, and all kinds of other stuff that is very, very formalized, very regulated, it's very structured, but also lots of really human language, right? Any conversation between the provider and the physician and the, and the patients

• 2:19Speaker 01

  and any electronic medical record notes and things like that. And that is one of the reasons why healthcare, I think, had comparatively little algorithmic coverage, I would say, in the last, whatever, fifteen years or so, right? There isn't really the equivalent of page rank in healthcare. I don't know what that would be, you know? I mean, there's all kinds of models, of course, in healthcare, predictive models that predict when you go to the ER or whatever else, but they're logistic regressions and stuff like that. And they always sort of ended at that, at that surface of where you go from formal language into informal language, in a sense. Now we have L L M's and they are incredibly good at exactly that threshold, go between formal and informal and back and forth again.

• 3:00Speaker 01

  And so I am just looking for all the use cases that are right at that intersection. And for Oscar, and then I think for the healthcare system overall, that initially, I think, means that a lot of the use cases are in the, on the administrative sides. So when we pay claims, for example, we run a claim through a very, very big rule base, which is really just all rules based, and we can create a very long trace of which rules got applied and which order and which ones failed and which ones fired and all that kind of stuff. And you can look at that and as an expert, you can make sense of it and say, okay, this was why the claim paid, didn't get paid. But it's hard to do for any kind of layperson, look at this thing. And L M can now look at this thing and say, okay, let me translate this from this very

• 3:41Speaker 01

  formal language of trace from the claim system into informal language, which is what the hell even happened here. And there's a lot of use cases like that. So I think that is going to be the next five years of healthcare. All this administrative stuff has to become real time, has to become sort of like, you know, bidirectional very easily, has to become much more transparent and all that stuff. And I think if we do that right, then we will have a healthcare system where you go to the doctor, you will know right away what something will cost you, you will be able to right away say, you know, whether there's a better alternative than the thing you're about to do as a patient or the thing that the doctor's about to tell you. And there's no more issues with weird claims getting denied and authorizations are coming through and things like that.

• 4:25Speaker 01

  I think that's kind of the next five years, if you will. Of course, then the big price has to be, how much can we replace the caregivers and the clinical intelligence? With machine intelligence. And that really means eventually, you know, bringing down the cost of doctor visits by a factor of 10 or a factor of 100 or whatever else and some, you know, replacing specialists with AI and things like that. That of course runs into all the issues of hallucinations and safety and LMS insulting people and biases that are baked into the trading manuals and everything else. And I, but I, that obviously has to be the goal as well, that we eventually get there. For Oscar internally, that's exactly how we're sequencing it.

• 5:08Speaker 01

  We have of the top four use cases right now, I would say three are administrative in nature and one is clinical. And the clinical one, we don't have a lot of expectations for the kind of breakthrough in the next three months, but for the administrative ones, we are like, all right, by mid February, we're going to solve this and then we got to go move on to the next one and just much more, much more D than R in the R and D.

• 5:29Jacob Effron?

  It's a really interesting point. I feel like the ability for LMS to, yeah, to go to different levels of sophistication, depending on the end user or end audience, you know, one makes a ton of sense for people to finally be able to understand their insurance bills or where their claim sit or all these things that are probably, you know, you have to speak one language to, you know, internally and to regulators and then everything, and then another language to your end user of an insurance product. And then you can obviously imagine on the clinical side, you know, there's probably a level of a conversation the doctor wants to talk to another doctor about as they're referring a patient, but then they also want to communicate the visit itself to the patient. And so the ability to kind of transform the same set of data into, you know, different information for different audiences, I think is super compelling.

• 6:10Speaker 01

  Another thing we certainly realize when you compare these use cases is that in the administrative use cases, the healthcare system collectively has already spent a lot of time to make the inputs and the outputs very clean. Okay, if you pay a claim, it's going to get paid based on the stuff that's on the claim. There's a clearinghouse, there's a format for it, an 834 and whatever format, an EDI format and all that stuff. And you largely decide based on what's on these things. Okay. And the output is very clear as well. Are you going to pay it? Are you not going to pay it? Are you adjusting it or whatever else? That is not the case in clinical use cases. And one of the things we realize when we even do something like summarizing a conversation

• 6:51Speaker 01

  between an Oscar provider and an Oscar member in a virtual primary care setting, for example, right? Oscar runs a medical group, it's about 150 caregivers, providers, nurses and so on in it. We record all those conversations, we have an LLM summarize those and generate the medical record from them and all that kind of stuff. What we realize there is that the biggest difference between the LLM-written summaries and the human-written summaries is that the human-written summaries contain subtle contextual knowledge that the LLM cannot possibly have had, because the provider might, I don't know, look at a previous conversation, whatever, even remember a previous conversation that he might have had with that particular patient that is not showing up in the medical records, you know?

• 7:35Speaker 01

  And so when the inputs and outputs are less clean, it's an unfair playing field for the LLM, because how the hell would the LLM know that, you know, you talked to this guy 10 days ago, whatever? I mean, you know, it's sort of in the ether, but you put it in there, but you don't often have it in there right from the get-go. And so to some degree, in improving of LLM's performance in some of these more wide-open use cases, many of which are clinical use cases, will also have to not just come with improving the LLM itself, but improving the horizon of knowledge of the LLM, so to speak, and all the stuff you put into it.

• 8:11Jacob Effron?

  That's so interesting. For some of these administrative use cases, because of how it has to be structured, you have all that, you have the same information, the same playing field, but, you know, but the point is, you know, maybe there's a hallway conversation or a text message or something like on the clinical side that the, you know, that you can't feed in this context. That's really interesting.

• 8:28Speaker 01

  I mean, we used to, just to make another example there, actually, we, so Oscar has these concierge teams, or now we call them care teams, where we have a number of people and they focus on a particular block of members, okay? The way we organize them, and we found this out many years ago, is that we said we should have them focus in certain geographical areas. And the reason was that if you if you live in Manhattan and or in Brooklyn, whatever else, your way of getting around the city and getting to physicians, it's just going to be a different conversation if you live in Phoenix, Arizona, whatever else. So are you going to drive a car in Phoenix? You're not going to drive the subway in Manhattan or the bus, whatever else. And so that contextual knowledge, like what is the weather today in Manhattan?

• 9:09Speaker 01

  And is it okay for that guy you're talking to to go to the bus stop and take a bus downtown or whatever else? Is we acknowledge that the agent probably has, that the that the L M probably doesn't have, unless you give it to the L M, of course. And that, of course, then goes back to, okay, how can you give way more context to the L M that we currently generally do in these in these in these setups?

• 9:29Jacob Effron?

  Taking a step back, I'm curious, like, obviously, as Oscar, you, I think you said at this really interesting place where you're obviously an insurance company. You're, you know, kind of renowned for your consumer experience on the insurance side. You also have, like you said, a provider group, you have care managers. There's like so many ways I could imagine you guys using AI internally. And I think even on your Notion page, you've gone through a lot of the potential ones. You you mentioned kind of three administrative use cases, one clinical, like, I think our listeners would love to understand a bit more about just how you guys are leveraging AI today.

• 9:54Speaker 01

  Yeah, so I would almost say the so we're an insurance company, right? Which means we we sign up members and then they pay us, if you will, a subscription fee. That's how I like to put it. So we like sell them a package of healthcare services effectively. And then everything that they do in healthcare will have to get paid for by Oscar, whether it's a cancer surgery or a drug or, you know, some medical equipment or something, a wheelchair or something like that. And so that means we have really three levers that determine our financial outcomes, which is always a good thing to start with when you think about running a company, right? At some point, the finances finances have to have to work. One is, can we grow faster? Can we retain people in a better way?

• 10:39Speaker 01

  And we have a bunch of applications in that category. Okay. The second one is, how efficiently do we run the operations? And that's all the administrative use cases I mentioned before. And the third one is really all the stuff on the clinical side. Can we somehow reduce your medical costs and improve your outcomes, right? Those sort of go hand in hands. And so what can we do there? And to give you a bit of a taste of how this of how the financials work there, which is honestly one of the reasons why I originally started Oscar, because I thought it is just so juicy and interesting to be able to be an insurance company and try to affect these these financial outcomes. One of the reasons that Oscar went into the Affordable Care Act market is so that we have

• 11:20Speaker 01

  a bit more control over who stays with us and who leaves us as a member. Typical insurance companies would usually sign up employers as opposed to individuals. And so, you know, employers don't leave and join Oscar because or an insurance company because the experience, great or not, right, there's all kinds of other considerations. Right. But you as an end user, you might really stick around if we gave you a great experience, if you like the service, you like the conversation and all that kind of stuff with us, right? So that's in the gross and retention buckets. But then on the admin side and the clinical sides, eighty five percent typically of insurance companies' costs are medical costs. And so of a hundred dollars we get, eighty five dollars will go right out the door throughout

• 12:01Speaker 01

  the year for drugs, cancer surgeries, all that kind of stuff. And then fifteen dollars can go to what's admin and even commissions and marketing, whatever would be those fifteen dollars as well. And so in some senses, way easier to, you know, automate an internal process to, you know, save a little bit of money on the on the on the administrative overheads. But it's way more fun and powerful and better for the healthcare system to get at the eighty five dollars that constitute medical costs, right? A little bit of of savings there goes a very, very long way towards improving margins. And margins, by the way, of insurance companies tend to be between two and five percent. Right? So you you you save a tiny bit on the eighty five percent and you've you have doubled your margins suddenly, right?

• 12:45Speaker 01

  That's really the way the way that this works in a very powerful way. So in all of these three buckets, we try to see how can how can AI work, help us and how can LMS work? I can give you one for each, if you will. On the growth and retention sides, I would say we've done the most on the retention sides. Okay? So the way. Our business works is that we have essentially six weeks at the end of each year, where members can re-sign up or newly sign up for the insurance for the entirety of the next year, which is an insane sort of Christmas business type of thing as if you compare to retail, right? Like, we have that plus like, times 10, basically. Nothing, everything that does not happen these six weeks is basically just not very relevant.

• 13:29Speaker 01

  And that means we better land the play in a really good way in those six weeks. And so on the retention side, for example, what we've tried to do a bunch there is to just run a lot of outbound campaigns, reminding members in a very personalized way of why we think they should like Oscar, why we did something good for them in the preceding 12 months, you know? And there are things that we find that have a lot to do with personas of members. For example, and honestly, I can go one level down, but two levels down, I don't know, and I don't think anybody else knows either. But we found, for example, that reminding people of getting colorectal cancer screenings had worked particularly well from an ethnicity point of view with Asian Americans in our case.

• 14:13Speaker 01

  You know, so interesting. There is an interesting segment now that we can identify and we can go after more directly and then say, okay, we reminded you of getting your colorectal cancer screening. You went, you did that, that was helpful to you, that felt like a kind of good preventative care step to take, and you should be happy about that. That Oscar reminds you of these things, can get you to the right doctor at the right time. And that's why you have a health plan. That's why you have Oscar, because we like to have that conversation with you. There's an interesting difference between chronically ill members of Oscar and generally healthy members of Oscar. You know, one of those segments responds much better to messaging around empathy and the other segment responds much better to messaging around just convenience, you know?

• 14:58Speaker 01

  And so if you are chronically ill, you are already dealing with your illness to some degree. You already know how to use the healthcare system for the most part. And so then cutting out some superfluous steps for you goes a lot farther than if we told a generally healthy member, hey, you can go to the doctor more quickly. You know, that resonates less and the sort of like empathy, nice message resonates better with them. People like you and I are probably generally healthy and just need a little bit of talking to to go back to the doctor, you know, for the right reasons. And so those are all the things where that are very much in this domain of. You know, language matters, how we phrase things matters, and how we sort of read your mind a little bit, if you will, through the conversations you've had with us matters.

• 15:41Jacob Effron?

  Has like G P T four out of the box worked for that? Or like, have you had to tweak? Like, it just kind of works. Yeah.

• 15:47Speaker 01

  I think it works well for this stuff. I mean, sort of, you know, extracting from customer service conversations, I mean, certainly what the issues are the patient came with or whatever else, or, you know, a little bit persona, character, whatever, that stuff you can extract quite nicely, quite easily. Yeah, that that definitely works. It works even in small ways, like, we, I think, have, I think we have ethnicity information for about sixty five, seventy percent of our members. It just, you know, we are able to collect it in the onboarding, or we get it from one of these state exchanges, we get it from the federal government, whatever, something like that. Okay. And as I said before, that's an important matching mechanism to get you to the right doctor, have the right conversation with you, all that kind of stuff.

• 16:31Speaker 01

  How to fill the remaining forty percent is a great L L M use case, you know, looking at the name perhaps, and, you know, looking at the, you know, other sort of the conversation in, you know, just even detecting the language that the last conversation was in. Those are all these kind of things that were definitely possible with normal machine learning models before. But they definitely get easier by just sending one message to G P T four and say, all right, which language was this? Was this Tagalog? Or was this Kizwaheli, whatever else? You know, it just knows this stuff. And so that's, yes, useful for lots of things out of the, straight out of the box, for sure. So that's growth and retention.

• 17:08Speaker 01

  I'd say in the admin space, that's probably the area we've published the most about. Yeah. And current top of mind use cases are, we've been running this call summarization in parallel with our care guides taking notes on customer service calls, and we're increasingly phasing out the manual note taking, and just kind of completely letting the L L M do it, the summarization, note taking of the call, and not even having the, and telling the care guides, you can stop taking notes, you know, the L L M will do the whole thing for you. We've launched large test summarization in the Oskar Medical Group, we've launched.

• 17:45Speaker 01

  Secure messaging, medical records generation in the Oscar Medical Group as well. We've launched these kind of claims explainers to the internal care guides as well. Those are all things, you know, I would say they probably save us on the order of a few cents. PMPM, those are the magic four letters in healthcare. It stands for per member per month. Okay, so Oscar probably, I mean, to fill in these numbers from earlier, we probably have a PMPM revenue premium of about $600 per member per month that we get, okay, per member. And then 85% of that goes, as I mentioned before, to the, to medical costs and 15% goes to admin,

• 18:26Speaker 01

  whatever. So, you know, you can do the math, one or two cents PMPM does not like shoot the lights out, but you get 10 of those and it starts getting more interesting, you know? That's definitely a bit more, that's we add up the savings really. So that's on the admin side. And on the clinical side, I'd say earliest, those are the earliest use cases, but the biggest one there right now is the, I would call it talking to the medical records.

• 18:53Jacob Effron?

  For the doctor to talk to their own medical records or for the patient?

• 18:56Speaker 01

  Yeah, for the doctor, for the caregiver, for the customer service agent, almost for anybody inside of Oscar, you know?

• 19:00Jacob Effron?

  I mean, I think a lot of our listeners are building general AI applications and they're probably like a bit scared of healthcare, right? I mean, there's all these regulations, there's, you know, HIPAA, there's BAAs that people want to sign. I think it would be really interesting, like, can you talk a bit about like those requirements and how they kind of manifest themselves when you're like just implementing these models off the shelf, like if you're using a GPT-4 or Mistral or Claude, like, you know, have, have people built around these requirements? And what like restrictions does that mean for, you know, maybe there's a customer support company out there that's like, oh, we could do this for healthcare, but healthcare sounds scary.

• 19:34Speaker 01

  Yeah. And I mean, those are real things. And so. I always say that I think we were lucky that we were a highly regulated insurance company because we had to take this stuff seriously from the very beginning and I think it made it easier later on to, I don't know, not blow up because our financials were in shitty states, whatever, you know, or because we weren't audited properly and stuff like that. So I mean, it made it easier to go public, for example. So, you know, 10 out of 10 recommends to be highly regulated. Well, you know, partly. So HIPAA is the biggest one. It's the biggest constraint in these things.

• 20:08Jacob Effron?

  Right. And HIPAA is around just basically like not being able to share patient-specific information, right?

• 20:12Speaker 01

  Yeah, we could not. They had to sign a so-called BAA, Business Social Agreements, and they weren't signing those at that time. And healthcare organizations had to go to Microsoft and Azure, and I think they were starting to sign them around whatever February, March timeframe or so. And I think Oscar was the first organization to sign a BAA with OpenAI directly. So, great claim to fame. And, but it's, but you know, it's not that big of a deal. Like Anthropic will now sign a BAA, you know, OpenAI will obviously will do it now.

• 20:44Speaker 01

  You know, you know, Miss Charlie, you can run yourself, so you don't have to sign a BAA, right? So, so there, so you can definitely, you can, you can deal with it. It's just the thing you gotta watch out for. The place where it comes into play today is, you know, waiting for Google's Gemini Ultra, for example. And we're sitting on Google Cloud Platform on the analytics side, and we have AWS on the, on the production side. So we can have both of these at the same time. And we are always testing all of the new LLMs that come out from either of those, of those area, from, from those guys. And, and we really mostly have OpenAI right now. And in OpenAI's case, we wrote our own GRPC service that interfaces with them. And then we have internally, internally, we essentially expose all the models,

• 21:26Speaker 01

  but it runs to a defiance pipeline to them effectively. That lets us audit better who's using it, what are they sending there, all that kind of stuff. Where we see HIPAA come against us today is when any of these guys publish a new model. Like when Google now puts out Gemini Ultra, for example, it will not be under whatever release thing that automatically makes it part of the HIPAA agreement we have with them. That will always come later. So we have to always have, we essentially have test data that we internally generate. It's that's more or less synthetic. It's some mix of like anonymized and synthetic that we will use for these new models. Because we know we can't just send, you know, medical records there that are real medical records.

• 22:11Speaker 01

  But once that's sort of like a distance of, I'd say, three to four months typically. You know, I would expect Gemini Ultra to come out and then we can start using it in early access. And then, I don't know, three months later, they will put it in broader release, whoever, still kind of guarded, but in broader release for people like us. And then we can start using it for everything and then sort of it becomes public, you know? So I think that's actually mostly fine.

• 22:32Jacob Effron?

  That's really interesting. Basically, like you can test it, you want to test out the model anyway to see if it's good. So you'll test it out while you're waiting for them to get, you know, kind of be willing to sign that. When you're running like Mistral or an open source model, like what, you know, I'm sure there's companies out there that are doing that and they're like, well, I guess we could, you know, since we're the entity itself, like we could try and go sell into healthcare. Like what do they actually need to put in place in terms of checks and balances to make sure like, you know, they're ready to go talk to a hospital about, you know, doing, helping them with like customer service calls or, you know, patient messaging or whatever it is?

• 23:04Speaker 01

  It's a huge question and I don't think there's a great answer. So hospitals will make you go through security reviews and policy reviews and so on. And they're basically long checklists. And I'm sure you can ask honestly a hospital what the checklist is. You know what, actually, let me check with my legal people. If this is a service Oscar could do to society, take one of the recent checklists we fit out and just publish it. So you can see what hospitals generally would want. But there are also some, there's some certifications that if you have them, it will be much easier. Like high trust is one of those, you know? If you're high trust certified, then then a bunch of these things fall away. But I would look at those things as final hurdles to get over. I don't think they are the ones that really make the make a difference for you.

• 23:49Speaker 01

  And at the end of the day, what makes a difference for you at the end of the day is whether you can get a hospital to trust you and your solutions. And that will not be outsourcable to high trusts. And that is very difficult. I'm not going to lie. I think. Health systems and insurance companies are not good at rapidly prototyping things. And they're not good at, you know, doing that rapidly, you know? I mean, really, anyway, the ones who are are then not good at following up on the rapid prototyping. So you just would spend endless amount of cycles in that process. And so it is one of the fallacies, I think, in health tech is that you spend too much time on your products

• 24:34Speaker 01

  and too little time on this kind of enterprise sales process. The best products in healthcare unfortunately do not win. It really still is the best enterprise sales processes that win. And it makes way more sense for you as a founder to spend time in conversation with whoever's on the other ends. And then, you know, kicking the tires together with them and things like that. Then for you to tweak your model for the tense time and hope that somehow that's going to make a difference for the for the sale, you know? So that's a little bit the disillusionary message there. I think I would say, well, one of the things we did recently is we were part of a consortium of health systems

• 25:14Speaker 01

  and some insurance companies that wrote a document about principles for AI and the use of it in healthcare. There's a website as well.

• 25:23Jacob Effron?

  Yeah, we'll link to that in the show notes.

• 25:25Speaker 01

  You know, I helped write that document and a whole bunch of other very smart people in healthcare helped write it. So, you know, folks from UC San Diego and Boston Children's and HCA and some, you know, various payers and so on. And one of the things that we put in there, and I was excited about how excited everybody else was about this, is a lot of positivity around AI solutions. And really saying AI and generally AI in particular can democratize analytics in an organization, you know? More people can use it and it's almost our duty as leaders of these organizations to get more people into contacts with GPT-4, Mistral, whatever else. And we put that in there with exactly that purpose so that the flywheel can eventually start spinning faster.

• 26:07Jacob Effron?

  Well, you talk about how, I mean, I guess like, you know, hospitals and traditional payers may be a bit slow in trying a lot of these things out. Obviously, you guys are super fast and nimble. And I feel like I've been trying all sorts of use cases. I'm curious if there's like an example of one where, you know, once you saw the power of these LLMs, you were like, wow, this would work so well for X. And then you built it and you're like, actually, okay, this doesn't, doesn't quite, not ready for prime time or not actually that valuable.

• 26:33Speaker 01

  Yeah. Well, I give I give an insanely simple example, which I think goes to a very basic limitation of L L M's, which is um call call summarization in in the following sense. If you take a bunch of customer service phone calls and you tell G P T four in the context window, um characterize each call and what it was about and give it a reason, like come up with reasons what these calls are about, your own taxonomy, and then give me how many were about reason A, reason B, reason C, and then sorts by reason, okay? So what you want in the output is these in these hundred phone calls, there were thirty one that talked about claims issues, there were twenty nine that talked about, you know, virtual primary cases or whatever else. G P T four will fail at that miserably.

• 27:19Speaker 01

  It is an incredibly weird, incredibly weird thing. And it will fail if you make it too big in a weird way. Uh, and so this is a very, very fundamental limitation of L L M's. Um, it because it almost like to put it naively runs out of layers. You know, these transformers, they have all these layers, right? And it's all the it's basically all the blocks and they get repeated many times. Um, and G P T four has actually haven't said how many layers it has, but probably like, I don't know, a hundred or whatever else, right? Um, and and on the simple language models like G P T two, whatever else, like really early language models, people have done all kinds of interesting interpretability studies where they've looked at when you take the sentence,

• 28:01Speaker 01

  the capital of France is blank and you let it fill in that token. Um, then you can see G P T four, I think, G P T two, I think, has twenty layers, whatever else, that by layer by layer ten, it has not identified the token that will be the final answer. But layer eleven has that token and it doesn't change anymore until all the way until layer twenty, okay? So there's very clearly a layer by layer processing going on that somehow at the very end leads to the synthesis of, yeah, it's Paris, you know? And that becomes the becomes the thing. And so effectively, when you have a do a task where it first has to process all these tokens and say, okay, here's call one is this reason, call two is that reason, call three is that reason.

• 28:45Speaker 01

  And in the same processing, it has to then also count them. It's sort of like maybe does the first identification until layer 10 and then needs another 10 years to count it.

• 28:56Jacob Effron?

  Have you been able to split that up just into different steps and different calls to the model to make it work?

• 29:00Speaker 01

  Yeah, so like chain of thoughts is that's the reason why chain of thought works. Right, chain of thought does not work because the LLM thinks whatever, that's not the case. Chain of thought works because effectively you chain LLMs to each other, right? You almost like expand the layer space, if you will, right? And you have the bottom like in the middle of the token generation really. But that is one of the big things you can do there. And so that I find super strange. And I think the reason I mentioned, I mean, this use case is easy to fix. You just exactly what you said, you just have it run several times, whatever, and you can refine it over time, all that kind of stuff. And you can have a taxonomy, you put in the taxonomy, you ask it to refine the taxonomy over time, whatever, whatever, that all works. But the reason I say this is because it's hard to sometimes see how composed your task

• 29:45Speaker 01

  is, right? This is an easily decomposable task, but the task of is this guy a smoker or not? How many subtasks are really under this task? That is somewhat hard to say, right? And so that is, I think, still for us, the big thing to think about. I'll give you one more example, which is a totally different flavor, but literally also something I looked at earlier today. So again, medical record extraction, okay? Medical record, you want to know certain things about it. For some bizarre reason, the question that is very important in deciding whether Oscar authorizes certain care for members, this is a process called utilization management. Okay, we look at medical records, we say, okay, Jacob should get a surgery or he's not allowed to get the surgery, right?

• 30:29Speaker 01

  That's ensures do this all the time. And we have to also do it. And I think there are good reasons why this is an important step in the process and all that kind of stuff.

• 30:39Speaker 01

  The way these things work is you effectively have questionnaires and you have to answer questions in more or less binary question answers. Okay, yes or no? And then they get mixed together in some kind of AND OR formulas, okay? And one of the questions that shows up quite a bit is, does this member have a post-traumatic injury? And so that is the single highest false positive question that an LLM has so far tried to answer for us. And I honestly don't really know entirely why. My speculation would be the following.

• 31:12Speaker 01

  There's also research that if you look at what G P T four really is, it's sort of a weird superposition of all kinds of different personalities, right? It's it's got the crazy personality, it's got the terrorist personality, it's got the, you know, the Harvard personality, it's got the whatever, like the gamer personality, all that stuff, right? And again, this is a very metaphorically speaking type of thing. And but it's sort of really what is in the weights, the weights of some weird position.

• 31:37Jacob Effron?

  Yeah, I guess as the whole Internet is, right? Like.

• 31:39Speaker 01

  Exactly right. And there's failure modes as well, as you probably see, right? You can put certain tokens in it, it suddenly flips over into some weird personality and whatever. So that seems to be going on. And my best guess would be that if the L M sees post-traumatic injury, it is enough of a concept that a lot of laypeople have some association with. Like, you and I would sort of have a general idea of what that is. Like, oh, yeah, shit, this guy broke his leg, he's got a post-traumatic in an accident, he's got a post-traumatic injury. However, that is a concept that is very well defined in the context of these utilization management decisions, you know, where that's not that same thing. And so you have to.

• 32:17Jacob Effron?

  I guess the training data for the model has all sorts of references to post-traumatic data that have nothing to do with your utilization management context.

• 32:23Speaker 01

  Exactly.

• 32:24Jacob Effron?

  Have you been able to then like, I mean, do you then have to just break it down into steps? Like, was there a, I mean, I don't know the utilization management criteria for post-traumatic injury, but like, is it, you know, was there a traumatic event, you know, get that marked, then like, was this injury related? Like, have you found a way to kind of sub-prompt it, I guess, into into answers or still a work in progress?

• 32:44Speaker 01

  I'd say yes and no, it's certainly work in progress. And one of the things we've been doing there is to say, is to actually have the L M create a sort of self-consistency questionnaire. So we say, okay, post-traumatic stress injury is the following, give it a few shot examples is always good, right? Give a few examples. And then have a pre-processing step where you tell the L M, I want you to tell me 30 different ways that this will show up in the medical Okay, so you prompt the LM for its own knowledge, basically. And then we use those thirty ways and we put them back in the prompts. And we then have those kind of get executed in parallel. And then you take those together and you have it sort of rates rate the thirty different independence reviews of those different different methodologies.

• 33:26Speaker 01

  And that gets you that gets you another ten, twenty percentage points in in sort of accuracy, roughly.

• 33:31Jacob Effron?

  Andrew, so it gives you like thirty potential ways that would show up in the medical record. You then have it evaluate those thirty ways to see if they're actually good answers. And then based on the output of that, you put you keep putting that back in the prompt at the start.

• 33:42Speaker 01

  Exactly right, exactly right. Exactly, and you have to do it independently. You can't do it in one pass, but otherwise you get the token, whatever pollution, if you will. Cross pollution, pollination, pollution. But independently, it does work, yeah.

• 33:55Jacob Effron?

  That's super cool. I mean, one, you know, I think like, yeah, the specifics are always, I feel like, the most interesting parts of these conversations. And like, one other one I was fascinated by is I think you guys, you know, built this claims assistant product. And it's like, you know, if I were to describe what it is, I think it's basically like, you guys internally have all sorts of complex logic around where a claim sits. You know, is it going to get approved, like all that stuff? And it's used in very insurance-y type language. And then I, as a member, probably want to be like, where's my claim? Like, is it good, you know, what's going on? And so I think natural LM use case, I think you you guys worked on basically, how do we take where that claim is in our own internal systems, complex language, and make that intuitive to the end user? And I would have, like, on the outside, you know, like many problems, it feels like, oh, maybe GPT-4 out of the box could just solve that. And I feel like that was far from the case here.

• 34:39Jacob Effron?

  And so I think it would be interesting for our listeners, like, maybe talk a little bit about like, why that was such a complex problem and like what you had to to do on the model side to make it actually work.

• 34:47Speaker 01

  Yeah, no, it was it's fascinating. And I would say it's still fascinating that the part of a use case we have, I think, more or less cracked, is the is the simple denial question, which is basically why did a claim get denied? The part we have not fully cracked is the actual. Why did a claim get paid eleven dollars and not ten dollars? Okay, that's that's yet another step of complexity. So we're still working on that part. But yeah, exactly. You you said it very well. We have we built our own claim system. Um, it's a giant rule base and, um, you know, it's very well audited. It can go forward and backward in time and all that kind of stuff you need for running these kind of big systems in that regard. And we get a we have our own domain specific languages in the claim system. We have a language called layer cake. Don't ask me why it's called that way.

• 35:32Speaker 01

  It's all some bakery related, uh, you know, metaphors. We've got like we've got Muppet Show, we've got animals and and bakery related stuff. Those are like the engineering team code works now. So anyhow, um, so layer cake, you get essentially a very, very long list of which rule got was fired and what happens in the rule. Okay. Um, you know, which contract got invoked, which with provider, um, under which, you know, exception, all that kind of stuff. It's it's, I think, a thousand lines of logic, whatever else for any given claim. When you just look at it ends to ends and a lot of superfluous stuff and redundant stuff, whatever else, but you get that's what comes out of the system.

• 36:14Speaker 01

  If you put that in the G P D four, first of all, you literally have a you have a context window issue. We started doing this, um, you know, in the summer last year, whatever. Uh, I think the high the biggest context window was eight thousand tokens and it just literally didn't fit in there. Okay, so that was not that big of a deal. Now it's thirty two K and and and whatever. Um, but a version of this problem you will always have. There's always some like abstraction level type of thing you have to think about with regards to the context window. Um, then the other thing is that the longer this this got, the less G P D four knew what was going on. Uh, it was good for shorter decision trees for longer ones.

• 36:50Jacob Effron?

  I love it. The most powerful, like, you know, brain model we've ever built still can't understand how the world health insurance works. I feel like there's something to tell you about that.

• 36:57Speaker 01

  That is a beautiful and very terrible thoughts. You're right. I didn't I never thought about that, but you're exactly right. Or maybe, um, we shouldn't piss off our smartest robots by using it for health insurance tasks. Is that's what's gonna make them kill us.

• 37:14Jacob Effron?

  Fair, fair. We'll give it some more fun, more straightforward tasks.

• 37:17Speaker 01

  Poetry writing, whatever, you know, that bounds out, like, kind of, you know.

• 37:20Jacob Effron?

  Shakespeare rap songs, you know, whatever.

• 37:23Speaker 01

  Exactly right. So, um, yeah, so it failed at all this stuff. And so what we basically realized that we had to, we had to give it these prompts, these traces, but we had to give it the traces at a certain level of hierarchy. And so we said, okay, we'll give you everything, but we don't give you the sub procedure calls. Like whenever it stepped down to some kind of a function internally, we, we just said, okay, it's now calling the function of determined geography or like determined gender or something like that, you know? And, and we didn't give it the output of the function. And then we told G P D four, now it's your job to find out which of these functions you want to double click on and, and where you want more detail.

• 37:58Jacob Effron?

  Interesting.

• 37:59Speaker 01

  It with a sub trace under that particular call. And that ended up being working very well.

• 38:04Jacob Effron?

  You basically had to focus it on like, you have such a mass amount of information that you basically had to figure out how to prompt it to focus on some subset of that information.

• 38:11Speaker 01

  Exactly right. Exactly right. And, and so in the denial case, the reason why the denial case is easier is because you can generally, you can generally find quite well which function call it was that made the claim get denied. You know, that function call might have had, might have had other functions calls in it. But at least you know where to look. Like there's 99 function calls and the 100th one fails. All right. That is the one you want to double click on and go down into and then whatever.

• 38:37Jacob Effron?

  I guess one, one thing I'm curious about from just hearing you talk about these use cases, it feels like you guys have gotten really smart on like different kind of prompting strategies and ways of, to what extent do you feel like you're picking this up from like the literature, um, picking it up from like informal conversations with other, you know, tech leaders that are trying this out? Or just like trying a bunch of stuff and it's like, oh, that, that, that bucket happened to work.

• 38:56Speaker 01

  Yeah, it's a great question. So I would say honestly, it's 90% trying stuff out. Um, it's, um, so I do a lot of paper reading, just honestly not. Because I'm very good at it, just because I like it, so. And I I take notes on all the papers I read and I put them all online. I mean, that's all in this notion side. Now it's actually in Squarespace, so.

• 39:16Jacob Effron?

  Oh, nice.

• 39:16Speaker 01

  It looks nicer. So, but people should take a look at that hi, Oscar dot A I. Um, we upload new papers there all the time and so and and the ones I I have a very big bias towards reading papers about interpretability. Um, and these weird failure cases, because I think they tell you the most about how these things work. Um, but the ones I always send to the team are the ones about prompting strategies and what I would call systems design, right? This is really systems design. Um, it's almost like less about the prompts, I would say. It's more about the. As you yourself said, that how you chain them together, when, when, what's the first call? What's the second call? What's the third call? That stuff I think is the most powerful. And so there's a good amount of.

• 39:56Speaker 01

  So I think this, the, the, the idea of, of prompting G P T four to generate its own way of looking at the medical records. In the literature would be sort of in this self-consistency area. That's a sort of chain of thought disciple, if you will. Um, and that was really researched quite well there, but I think the guy at Osko did it. Um, didn't even get it out of a paper, just discovered it independently. I'm like, hey, this is this, this is what that is called, you know? Um, the guy.

• 40:20Jacob Effron?

  There's all this like great research being done within companies and I wonder how that ultimately gets, you know, it's not formal academic work, but it's like obviously incredibly helpful to the ecosystem. And I wonder how that ultimately gets ends up getting shared across companies, because there's probably an engineer in so many of these places just reinventing, you know, similar to.

• 40:36Speaker 01

  I agree. I think we should all share this stuff. I mean, I think, um, this is why we put that the, the, the A I side out there, right? And and I ask other people all the time as well and and so. I mean, there is a good amount of. Learning, I think, going on in companies, exactly right, and we should all learn more about it. I mean, I'm always. Surprised with how. Little we still know about this stuff. I mean, it's, it's just, it is mind-boggling, you know? I had this thing the other day. This is a fun one. I, I, I wrote about this on Twitter as well. Um, you know, OpenAI has these custom G P Ts now where you can, you know, you, you can build your own G P T, whatever, right? And you can upload some files to it as well, whatever. And so you can upload Python files, like it connects you to Python code and it can be files you upload it as well, which I find pretty cool. It can call functions as well, but I didn't want to have mine have a function call.

• 41:21Speaker 01

  And so I try to sort of like. Make my custom GPT follow a fairly formalized state machine, okay? Like literally go from state zero to state one to state two, whatever, and back to state one, all that stuff. And as is very typical with LLMs, it would constantly lose track of that thing. It would just be like in state two and would go like invent a state or whatever else. And so the way I then I was able to get it done then is to in the Python codes, tell the thing and now you have to go back to state zero, you know, which is an unbelievably weird programming paradigm, you know?

• 41:54Jacob Effron?

  Totally.

• 41:55Speaker 01

  Not only generating code, but generating language.

• 41:57Jacob Effron?

  Yeah, some natural instructions in there.

• 41:59Speaker 01

  Yeah, it's the strangest thing. And so anyhow, it's just a novel intelligence we're talking to there. And anyhow, I think that stuff is really best done by sharing it and discussing. So I'm glad you are getting people to talk about it.

• 42:11Jacob Effron?

  No, it's really, it's super interesting. I mean, I guess another thing I'm curious about is like you guys have obviously been really thoughtful in how you've chosen models. And I think you said, you know, it sounds like GPT four out of the box ends up being the most powerful one. There's obviously been this movement toward like healthcare specific models. And I think Google has like MetLM, which maybe, I don't know if you guys have played around with. But I'm curious, like if you're a betting person going forward, like, do you think you end up for a lot of your use cases staying on like a general purpose model? Or like using some of these more healthcare specific models that are trained just on healthcare data? And maybe don't have like the, you know, for post-traumatic injury, maybe you're just trained on like medical record data on that versus, you know, whatever bloggers on the internet think.

• 42:49Speaker 01

  What we've consistently found to date is that when you go and specialize in a particular area, you just lose alignments, you know? The model just will not follow your instructions anymore. And the simplest thing is always the famous, you know, your answer needs to be in JSON. That's simple. I mean, that sentence is just one of the most terrifying sentences for LLM users, right? Because you want your output to be formalized and all that kind of stuff. And GPT four has a switch now to create JSON as opposed to just natural language. And you tell MedPound two to put out stuff in JSON and it just forgets it half the time, at least half the time. And it's a very simple example for losing alignments. You just will not get it to follow your instructions anymore.

• 43:31Speaker 01

  And so I don't think, I personally don't think that's a general purpose model that would be specialized would be more useful than a general purpose model that is just bigger. Because of that issue. However, it's entirely possible that in the future, we're able to. De-link the symbolic processing, if you will, the planning and take that out of, you know, the transformer logic and have a transformer, just do the contents, if you will, right? That's sort of like the dream of every AI designer, that you have a separate reasoning system and then like a content system where else? And we don't know how to do that right now in GPT-4 or anybody, anything else. But once you get to that point, yeah, sure, it'd be great to have a clinical model.

• 44:15Speaker 01

  Until then, I think the use the biggest model you can for the best reasoning and maybe do some RAG, you know, or retrieve up into generation, whatever. In the meantime, there's a recent paper that looked at the combination of RAG and fine tuning and the combination of both of these together. And they chose to be in the agriculture space and they basically showed that's, and I felt this interesting too, that RAG and fine tuning gets you independent improvement in performance. Okay, so you're better off doing both of them at the same time.

• 44:48Jacob Effron?

  Yeah, everybody wants it to be a binary thing. I feel like, you know, I guess, yeah, we would always ask like RAG or fine tuning. It's like, yeah, the reality is they're probably both helpful.

• 44:55Speaker 01

  Kind of do both, yeah.

• 44:57Jacob Effron?

  So another thing, I feel like a lot of companies are thinking about, and I think it's super relevant to you guys, is like how to structure AI teams. And so I imagine there's a lot of like basic work you guys do in AI that is very relevant across all of the AI products you ship. And then obviously, as you talk about the use cases, I'm sure you have like a claims team that builds products around members interacting with claims and a team that builds product for your providers. And, you know, I think there's this debate of like, do you embed the AI folks within these like individual product teams? Do you have some like horizontal AI team that sits across? How have you guys thought about like structuring that?

• 45:28Speaker 01

  That one I'm actually quite proud of, to be honest, because I think we have a good model going there. So this started with a hackathon, really. We do this quarterly hackathon at Oscar and it just basically means everybody can, you know, take a day or two, whatever, and just work on something, classical hackathon stuff. And the one with the most participation we ever had was at the beginning of last year and it was the first AI one we had, you know, we had GPT-4. That was before G P T four, I think G P T three and a half, whatever at the time. And and we had that direct link already. We had signed the B A A, I think with OpenAI already. And so people just got like the hack on stuff. And so it became so obvious to me, all these things you and I were talking about already just now, which is you've got to share this stuff.

• 46:11Speaker 01

  You got to talk to each other about what prompting does and all those things. And and that's where this hi Oscar dot A I site came out of. And so we formed the parts. It's a seven person parts. And it's got a got a two part managers. It's got a bunch of data scientists and got a bunch of engineers on it. And they have they have office hours every week. And everybody in the company who works on anything related to A I can come to them and say, this is my prompt. What do you guys think? Okay, what else have you seen? They have their own three priorities. Okay, so we always have essentially three projects that the part itself needs to finish and needs to do. And I think that's really important as well, because otherwise you just spend endless time on just research and, you know, kind of wavering

• 46:56Speaker 01

  and not getting not finishing anything off any anytime soon. And they also have a they also have a weekly hacking session, which is Monday nights, two hours, sometimes even longer, where anybody can come and the part definitely is there and I'm always there. And and anybody can can bring any idea that they want. And so we had an Oscar engineer bring his own trading system. He built like a little bots that helped you, you know, trade stock or whatever else, right? That was cool to see. And here's how he built it and stuff like that. We had last week, actually this week, we had the broker commissions team. You know, we use brokers to sell Oscar policies, whatever. Commissions is one of the other one of these things that has a lot of rules around it. And there's a language internally that that governs which who gets which commission paid and all that stuff.

• 47:41Speaker 01

  So they use that system to explain which commissions had gotten paid and which ones hadn't gotten paid. Sort of like that the also bought the claims bots effectively. They showed off their system and what they were doing there. And we had the utilization management to will earlier this week. And if we don't have anything, I usually show one of my things. I show the whatever picture generator I created last week or something like that. And but the point is we need to lower the bar, like to just make it insanely easy for people to come with something that they tried. Even if it didn't work, that's why I always also go and do something that I did because I feel like the more embarrassing I can make my stuff, the better because, you know, it lowers the bar for everybody else.

• 48:22Speaker 01

  Like this, by the way, great CEO and founder rule, whenever you do karaoke, you have to sing as a founder. It's just knowing he's not gonna sing. Because, you know, if you sing, then everybody say, oh my God, you know, that guy can make an idiot out of himself. Now I can go and do it. And, and by the way, the other rule is also, it is absolute, you know, I, biases are a bad thing for sure. But it is absolutely the case that if you go and sing at a karaoke nights, and, and then you go and say, hey, Mario had a cool idea about healthcare. There's probably something in me that will wanna make, that will wanna make me listen to you. It's tiny bit more, because I'm like, oh yeah, that person took a risk and sang and entrepreneurial and you know, like, I wanna listen to the person.

• 49:02Jacob Effron?

  The next Oscar karaoke night's about to be wild, you know, now people listen to this, it's, everyone will be fighting over the mic.

• 49:07Speaker 01

  Probably, yeah. We haven't had one in a while, actually, post pandemic, but, but so that's how we set it up. We got this, they got the pod and the pod also has a list of all the other AI projects in the company. And just to kind of keep track of them and whatever else.

• 49:19Jacob Effron?

  Yeah.

• 49:19Speaker 01

  And I think that's a very nice mix of sort of, you know, you have something centralized, some things decentralized. And everybody knows that they can come to these guys and have ideas and whatnot. Works, I think, quite well.

• 49:29Jacob Effron?

  I love that. And I guess as you've been building these, you know, these features, I feel like we're so early in, in kind of having any sort of tooling around, you know, helping folks build them. Any like things that have come, you know, that you've seen like three, four times, you're like, God, I just wish there was something that helped us do X or Y, like, you know, I guess the shameless VC request for startups question.

• 49:51Speaker 01

  Excellent, yeah. I mean, I would definitely say the safety layer, right? Is there anything? Is there some way that you, your LM produces output, your model produces outputs? And you send it somewhere else and it verifies if that thing is not insulting.

• 50:05Jacob Effron?

  How do you guys solve that now?

• 50:07Speaker 01

  Honestly, right now with a human in the loop, I mean, any, any use case where, where there's danger of that happening, it just has, goes to a human right now. You know, that's the, the lab test summary goes to a doctor, the, um, the claims explanation doesn't go directly to the end user, goes to the care guide agents, you know, the kit to the, to the claims agents and whatever else. That can't stay that way. Obviously, if you want to have clinical chatbots, we have to have direct to consumer or G P D four, whatever G P D five outputs. But that's still a fairly big one, I would say, faster inference times, quite simply as well. I, I, it's sometimes I think. Why even, who needs G P D five when we could run G P D four a thousand times in parallel?

• 50:49Speaker 01

  Right? Like I do lots of, even with my kids, I try to get them to. Use get, I mean, I do it with them. Basically, I sit next to them and I say, let's do a game, get up co-pilots, you know, I now have, I have chat now, get a copilot chat, which is cool. I can chat with the L M as well. Um, and so, uh, whatever, let's do like a defender game or something like that together. And so in code generation, you can verify if the code works or not, if it's real code or not. And if you could generate a thousand listings in parallel. And run them all in parallel, like in a micro seconds, who needs a better L M? You just pick the best of the thousands outputs.

• 51:25Jacob Effron?

  I feel like G P D three point five pricing has come down. They're always slashing it and it's gotten to the point where it's, it's so cheap.

• 51:31Speaker 01

  Um, yeah.

• 51:31Jacob Effron?

  You know, you really could do that.

• 51:33Speaker 01

  It's, it's definitely, definitely from a cost point of view for sure. I think that one is just still in much inferior to G P D four from a quality point of view, you know, so.

• 51:40Jacob Effron?

  Well, then by the time five comes out, it'll be so much better that, uh, you know, who knows?

• 51:44Speaker 01

  I think that's probably true. Yeah. Who knows if we're, if we, anybody needs us still, you know, at that moment in time.

• 51:50Jacob Effron?

  Yeah, we shall see. Hopefully we have at least till seven or eight. Um, I, I guess like, you know, you probably, it's something, I mean, you think about all this stuff so much. I'm curious that if you weren't building Oscar and, you know, you were just starting a net new healthcare AI company, like. What do you think's like most interesting, you know, best commercial opportunity? Like where, where would you start?

• 52:09Speaker 01

  Does it have to be healthcare or can it be a generic?

• 52:12Jacob Effron?

  I know, I know. Everyone wants to flee healthcare once they've had to be been through it once, uh, but we'll, we'll, we'll keep you, uh, we'll keep you just because you know it so well.

• 52:19Speaker 01

  Yeah, yeah, sure.

• 52:19Jacob Effron?

  Where you see the opportunities?

• 52:21Speaker 01

  Yeah, that's a good question.

• 52:23Jacob Effron?

  But I feel like, does it have to be healthcare is the classic second time healthcare founder question.

• 52:27Speaker 01

  Yeah, I know. Because my answer in healthcare would have been, you know, pick a very, very, very obscure niche of healthcare, where, you know, you're solving some very particular issue for enough people who are not technical. Like, you know, regulatory filings composition, that actually might be a good one. I mean, I, you know, Oscar struggles with things like, how do we create the right kind of regulatory documentation without too much overheads? And not just even for healthcare regulators, but even for, you know, for SOX compliance and all that kind of stuff. You know, we, as a public company, we have to do a lot of work in that. And so, you know, it slows down engineering to some degree.

• 53:10Speaker 01

  It's also very important, obviously, to get it right because you want to have high quality output and you want to be able to make sure you pass these controls. But I always wonder if there is not an application of Genify in exactly that. Like, you know, regulatory reports have to be written, they're natural language, they're not just tables. Can you generate them more easily by having some LLM watch your data flow and generate them automatically? I think there's something there. In particular, actually, in the clinical space, you know, we produce clinical reports to, let's say, the NCQA, that's a sort of like, not a governmental organization, but a private organization that, however, is tasked with improving quality of health plans.

• 53:53Speaker 01

  And we file reports to all kinds of state regulators and state health departments. There's a lot of there that can probably get generated pretty automatically and would be totally fine to do in that way. That's probably where I would go.

• 54:06Jacob Effron?

  Super true. What about, like, I mean, one that comes up on our radar all the time is I feel like there was a million AI prior authorization companies that were launched last year, you know, obviously being the, you know, what the information you have to send to get approval from an insurer to get a certain kind of treatment. And, you know, I'm sure you guys think about this use case a fair amount. Like, is that, is that an interesting market or a good space?

• 54:27Speaker 01

  I would say yes, with a qualification that's, I feel like that is very, very close to the core competency of what an insurance company has to be able to do. And I don't understand really necessarily why people. Can't do it themselves better, including us. And so the the risk I think you run there is that you're trying to take a very neuralgic thing away from insurance company. And if you just take it away and you don't somehow make it very interactive and very, you know, platformizable, if you will, for that insurance company that you are sort of like taking this from. I think you're gonna have a low horizon of what you can get to, a low ceiling what you can get to, you know. This is really what insurance companies think that they need to be good at, clinical

• 55:11Speaker 01

  managements and and postage charge spending from the hospital and all that stuff. And I wouldn't say they are very good at it, but you're sort of, you know, catching them where it hurts, if you will. I would actually say the other thing is fraud, waste and abuse. That is still an industry that is that is very dominated by a lot of very old school players that make a lot of money still. And it strikes me as not clear why you need to overpay a bunch of fraud, waste and abuse companies on the on the backends.

• 55:41Jacob Effron?

  Makes total sense. I mean, you know, in your in your kind of like overview of AI use cases for Oscar, you talked about kind of interaction automation on the clinical side. And I feel like that I know our listeners are going to want us to ask about that, like that feels the big question in in AI healthcare, you know, will there be kind of this future AI doctor? And so I'm curious, like, you know, how you see that that kind of space evolving over over the next five, 10 years?

• 56:04Speaker 01

  I mean, I I have I see no reason whatsoever why we shouldn't be heading there. It's, you know, doctors are probably the the most, I, you know, don't quote me on this, actually, you can quote me, the most computerized profession we have in some sense. And not just computerized because they use computers, but their brains are the most computerized, right? It is very is regimented, it is it is all existing knowledge, it's a lot of inference based on complete data points or whatever, it's all very algorithmic. Why wouldn't L M be able to to map all that in at a very, very, very high quality? I I don't see the reason why that wouldn't be the case.

• 56:43Speaker 01

  When we look at what the practical issues are, I would maybe say there's two, well, there's three. Safety is obviously one of them, right? Right now, it's very difficult to get an L M to talk directly to an end user, so that that's for sure. But we talked about this already. The second one is is.

• 57:03Speaker 01

  Physical interact, what has to be physical, what has, what can be virtual? We did this research at some points that when we, we looked at all of our claims, so like a sample of claims a few years ago. And we had physicians go through all of them through the sample. And look at how many of these claims could have been done in a virtual setting, how many of them, how many of them would you have needed to be in person? And I think the in-person percentage was something like thirty five percent or so. Okay, so that means two thirds can go totally virtual. The problem was it was two thirds of claims, not two thirds of people. Okay, and so if you're diabetic, for example, there's going to be some amount of things you have to do in person, like, you know, foot exam and things like that. And I think that is the other big issue right now. When, when we look at our.

• 57:49Speaker 01

  Attribution to primary care physicians. So in other words, we have a member. And the member is, the member told us, this is my PCP. And we actually see in the claims, that guy is going to that PCP, primary care physician. Okay. You would think the percent of that prime members, primary care go to that PCP is like ninety nine percent of whatever else. It's not, I think something about twenty eight percent of those members will see another PCP in that same year. Which to me is mind bogglingly high. It just means that the, that the loyalty to specific physicians and even physician, specific physician groups, okay, you have to look at this in a group level. Is not that high, you know? And but it also, I think, is a reflection on the fact that.

• 58:32Speaker 01

  A lot happens as you go and need to do the one thing you have to do in person. And you immediately get sidetracked away from kind of doctor computer, if you will, you know, doctor digital. Yeah. And so as long as we can do lab tests virtually, as long as we can do sort of like your hands on virtually, whatever else. There will be this leakage that will keep clinical chatbots from, you know, be replacing physicians effectively or replacing the system. In a bigger way, I think that's a big holdup. And the third one is business model. It is just unfortunately the case that right now if you're a health system, and I don't mean this cynically at all, even though it's kind of makes you cynical a little bit, if you're a big health system, there's not a lot of incentive you have to switch towards lower costs, channels of care delivery, because that will mean the insurance companies will

• 59:17Speaker 01

  and the government will put pressure on you to reduce reimbursement costs, which of course will mean then you have to reduce capacity and all that kind of stuff. And so the insurers would be in a good position to deploy automated virtual primary care, but they mostly can't because they don't have the member engagement and all that kind of stuff. It's the conundrum of the U of healthcare system. And so anyhow, that's the third thing that I think we have to somehow cut through over time.

• 59:40Jacob Effron?

  Well, Mara, this has been a fascinating conversation. We always like to end our interviews with a quick fire round where we get your your quick thoughts on a few common questions. And so maybe to start, I'd love to hear just one thing you think's overhyped and one thing you think is underhyped in healthcare AI today.

• 59:53Speaker 01

  Well, okay. Overhyped, I mean, I think clinical chatbots generally overhyped.

• 1:00:02Speaker 01

  Underhyped, maybe voice outputs. Let's pick one thing, you know, I think you can get pretty far pretty quickly there. And so there's a lot more we can do there.

• 1:00:11Jacob Effron?

  As long as they're not talking about clinical things.

• 1:00:13Speaker 01

  Yeah, maybe, yeah.

• 1:00:17Jacob Effron?

  Which healthcare AI startup, you know, outside of Oscar, are you most excited about?

• 1:00:21Speaker 01

  Hmm. I mean, honestly, I would still say OpenAI. I hate to tell you, I just think

• 1:00:30Speaker 01

  I just think they're still building the best models, you know, I would also say having face, even though it's not AI necessarily, not healthcare either. But I just love the fact that they are, you know, they're creative, they, we have a contract with them, we use them for consulting purposes as well. Having one place where all the models collect is is just is cool.

• 1:00:48Jacob Effron?

  Well, I feel like there are so many threads in this conversation people are going to want to pull on. So I will leave the last word to you. Where can folks go to learn more about you, about the exciting AI work you're doing at Oscar? Floor is yours.

• 1:00:58Speaker 01

  They should go to hi. oscar.ai. That's our collection of papers of Oscar articles and insights and all that stuff. And we try to really be very open about insights we have and things we try. So if we crack that use case on the medical records extraction, for example, that'll be all on there. People can look it up and see and do it for themselves. I think they should follow me on Twitter as well, Mario TS, okay, at Mario TS. I try to post my own stuff there, my own explorations and certainly also all the Oscar stuff we're doing. I want to do a lot of stuff around games, by the way, so if people have any interest in helping me figure out those use cases in gaming, then I would love that.

• 1:01:36Jacob Effron?

  Are we going to get a gaming healthcare collab or is it completely outside?

• 1:01:39Speaker 01

  So, old idea, not done, but very, like, a long time on my mind is, take all the material on the Oscar Google Drive, all the presentations, all the documents, whatever else, run that through some kind of system that then generates an RPG, like a role-playing game, where you can be whoever you want an Oscar, like you can be a regulator or you can be the CEO or whatever else, and you have to sort of like relive the last 10 years. I think every company should have that, and I've had various attempts at it. Clearly, LLMs are the solution for it to actually do this, but it's not there yet. But if we're willing to help build that, then definitely, please come call me.

• 1:02:21Speaker 01

  The other one that I have a version of, but it's not working yet, is a game, it's like, you know, remember the Oregon Trail, this game way back in the days?

• 1:02:31Jacob Effron?

  Of course, classic.

• 1:02:32Speaker 01

  The Oregon Trail is like a basic program, literally the language basic, that's on GitHub as well, actually, you can look at that, it's pretty cool. And it's very simple rules, very, very simple sort of like formulas and rules, whatever else, and sort of like state machines of sorts. And an LLM that writes those games as you're playing it. So you start playing the game with a very simple mechanic, and it adds more mechanics as you play it, and it will add the mechanics so that the overall game economy always balances out.

• 1:03:00Jacob Effron?

  Interesting.

• 1:03:00Speaker 01

  That's another one, I have a very simple version of, but yeah, that's another one. But those things are on Twitter, for example, so just definitely follow us there. And yeah, just send an email or whatever, and DM, and looking forward to chatting more about this stuff.

• 1:03:14Jacob Effron?

  That's awesome. Well, Mario, thanks so much for coming on. Seriously, this has been a fascinating conversation.

• 1:03:18Speaker 01

  Yeah, thanks, Takeoff. Really fun to do, and keep me posted if you're seeing cool stuff out there, please.