0:005.3 plus trillion we spend on healthcare. If it were its own country, it'd be the third largest GDP. And yet we have a ton of scarcity. Open's mission is to ensure AGI benefits all of humanity and give people access to trustworthy, abundant, reliable health intelligence. One of the stories that I was most touched by was the story of Nathan Levent. He basically had this kind of long saga with his son where at some point he was suffering a lot of fatigue and irritableness and AI helped them navigate the process and it turns out he was diagnosed with an aggressive form of leukemia. some of the work that you've done around mental health.
0:29There's a lot of complexity. How do you help guide a user on that given that could mean so many things?
0:34We've just been working with a bunch of licensed mental health experts across the world to think about the right ways to define model behavior and a bunch of tricky scenarios and ensure that we can get to a level of consensus where we can actually measure performance.
0:45Let's talk about trust. A person can [music] be speaking to an LLM really hoping to hear that everything is okay, but from a medical perspective, they need to hear perhaps something very different. How do you teach an AI to recognize the difference? Um, hey everyone. Welcome back to the podcast. I'm Jay Rugani, a partner at A16Z, leading investments in healthcare and biotechnology companies. I'm joined by Karin Singal of OpenAI. He leads health AI research at OpenAI and we've got a lot to talk about. Thanks for
1:15Thanks for having me.
1:16I want to start with the big question.
1:18Uh, there was a report earlier this year. You all have talked about it publicly. Hundreds of millions of people each week use chat GPT for health rellated questions. Uh what are they asking? What are the kinds of questions they're asking that they're not getting from our main healthare system?
1:35Yeah. Um let me zoom out and talk a little bit about why we're investing in health in the first place. Um so you know open's mission is to ensure AGI benefits all of humanity. And so we see that at a high level happening in three ways. Um the first is building and deploying AGI. So building capable models, building products around them, deploying them, partnerships. The second is preventing downside um safety of various kinds. The third is making upside happen. Um and so we view health as a really really concrete way to make upside happen and also work on some of
2:04the safety problems that come up in healthcare along the way. And so really we're really invested in in making sure basically two things. One is that we can raise the the floor of human health and give people access to trustworthy abundant reliable health intelligence for for free or nearly free and really raise the ceiling as well and make things possible that weren't possible before. And so the best way to see the kind of raising the floor is really to think about how we've been working in the consumer space. Um and so as we reported there's more than 300 million people weekly using CHP for for health
2:34and wellness types of questions and we take that as like an incredibly important responsibility to to get that right. Um, and that's been rapidly growing um, over time. I think one of the things that you see is just if you if you think about the experience of of getting healthcare even in the US, even in the Bay Area, you know, if you're looking to get a specialist appointment, even here in the Bay Area, sometimes it can take weeks or months, um, you might um, spend a lot of time figuring out how to get an appointment and scheduling and all these kinds of things. And after all that time, maybe get 15 minutes with the
3:02doctor where you don't feel adequately prepared and you maybe walk out of it thinking, "Oh, I had like six more questions that I wanted to ask." And our hope is that AI can be part of this care journey with you and like health is not something that is just kind of restricted to these kind of 15-minute intervals. And AI can actually help you navigate your care journey. And so that actually reflects how people are using AI on the consumer side today. Like people are coming in with questions and concerns, information that they'd like to have. Um they want to figure out how
3:31to kind of navigate their care visits or ask questions or figure out how to follow up things that they that the doctor told them or things like this.
3:38And we're finding that in between these visits is is when AI can help people most on the consumer side. And you know, we're making similar investments on the clinician side and on the enterprise side as well. Yeah, it's amazing. We have an extraordinary health care system. Uh if I got really sick, there's no country I'd rather be in than the US.
3:56And yet we have a ton of scarcity. We have, you know, you've seen some of these statistics, over 100 million people don't access to a regular primary care doc. Even the ones who do often have to wait 40 plus days to see that doctor. Sometimes you meet the doctor exactly as you said and it's a quick 15inute visit and you have more questions. I saw a dermatologist last week about something. I wrote down all my questions beforehand in an Apple notes and I still forgot to ask something and now I'm scheduling a follow-up and it's another four months
4:25out. It happens to all of us. We've all felt that. Um, is there a specific patient story or recent case that you all have seen that kind of help us imagine where the world is going and where your work is taking us?
4:39One one way in which the the work of our team is gratifying is just like seeing the numbers and seeing how many people are using the tech and how useful it seems to be for them. Um, one other way is just hearing from people, right? So we get people coming in sending us emails or whatever it is reaching out in various ways multiple times a week saying like here's how this has helped me.
4:58I think one of the things that I one of the stories that I you know was most touched by was the story of Nathan Levent and his um his son um so he hosts the Cognitive Revolution podcast um and kind of told his story in that podcast but he basically had this kind of long saga with his son where at some point he was kind of suffering a lot of fatigue and irritableness and and things like this. um his his very young son. Um and um AI helped them kind of navigate the process of realizing that this was like a a pretty serious thing while they were
5:27on vacation and and deciding to take them for testing and repeatedly telling them like, "Hey, take this seriously even when when doctors were a little bit less sure and actually kind of downplaying it." And it turns out he was diagnosed with an aggressive form of leukemia. Um, and throughout his care journey, I mean, I mean, Nathan reports kind of using GBD5 Pro at that time, you know, quite a bit for helping navigate the process, um, helping understand lab reports and things like this even before, you know, a specialist had time to kind of help interpret. And I think, you know, it brought them a lot of peace
5:56of mind. Um, and so Nathan actually tells that story. I don't I don't know if I'm allowed to advertise other podcasts here, but Nathan actually tells that story in his podcast and would recommend that for folks in many ways. So, OpenAI is building health intelligence and health intelligence into your model at every step, but it's the same model that folks use to write code or plan a vacation.
6:17Can you help folks understand why you're building it that way from an architectural standpoint? Why is a general purpose AI the right foundation for something as complicated as medicine?
6:27Yeah, I mean, people often ask us like, you know, this kind of general versus specialized model debate. And I used to work on specialized models actually in a previous role. And so so I definitely have kind of experience on on kind of both sides of this um this question. I think the the kind of the kind of the most important thing here is like what we're talking about here is actually specialized training in a general model which is I think actually the best of both worlds. So what we do um for for Chash PT and all the models that we ship in Chash PT every across every stage of
6:55model training for JHPT we actually have specialized healthcare data added to each phase of training um since all of our models in the last year. as a result is now you have a model that benefits from all the scale and general improvements and intelligence and reasoning of the big big frontier models and also the benefits from the specialized training that we get from for example integrating data that we get from working with a bunch of physicians across the globe and so I think you get the best of both worlds with this approach if you if you kind of forked and said like hey we're going to train a
7:24kind of smaller specialized model that is just for healthcare um like I used to and like many people have um what you'd find is that model would just kind of fall behind the frontier relatively really quickly because the pace of the frontier and the general model improvements is so fast.
7:37Then if you bet on scale, what we found as OpenAI as an industry is that things will get better. Um, and so what we're doing is kind of hitching a ride on that kind of scale frontier. Um, and leveraging the fact that we have the ability to kind of integrate evals and training data into the models to make the models both specialize and also benefit from that general capability.
7:56[snorts] I want to almost start imagining what's possible with the technology that you're building. Obviously where it is today is amazing, but you're not stopping and there's a lot lot of work to go. I think many people are allured by the idea of a personal doctor in your pocket.
8:12Something that is perfectly knowledgeable, has knowledge across all medical specialties, so you don't necessarily need to go see a cardiologist and a dermatologist for different questions. um something that is perfectly personalized to you, has all your medical history in context, um remembers everything that you've said, remembers your medical history, but also your preferences. Um and is also able to motivate you and nudge you so that uh the information that it gives you is given to you at the right time in a way
8:42that can nudge you towards better health behaviors or adhering to the protocol that your doctor uh gave you. Is that sort of the experience you're building at OpenAI in the fullness of time or and how close are we to that?
8:54Yeah, that's something that we're we're really excited about. We are very excited about this vision of kind of a proactive personal health assistant. I don't think that's any kind of replacement for actually having a having a doctor or having access to a doctor.
9:04But really thinking about the ways in which we can take that experience of getting health information and making it more personal to kind of your context is a really important part of kind of our investment in health and chd which is kind of consumer offering that we have that enables users to connect their health information for that information to be used kind of securely not to be used for for training data or for targeting ads or anything like this.
9:25And that's a big part of how we think about how we're going to move towards this world of really like increasingly personal access and proactive access.
9:32It's also been how we've been thinking about deployment. So we've actually from the beginning been deploying this as widely um as possible to our free users as well as our pro users. And so we have been of the kind of 300 million weekly active health users, the vast majority of them are free users. And we take that and most of them are outside the US as well. M and so we we take that kind of um question of like distribution here really important really um really centrally and so as we're thinking about this kind of next frontier of like a proactive personal health assistant
10:02we're kind of bringing that to bear as well I think use cases help us imagine help people imagine what's actually possible what all this work is contributing to the idea of a personal doctor in your pocket I think is exciting when Greg Brockman came on the podcast he spoke a little bit about this situation he had with a friend of her friend of is was in the hospital and uh her physician was about to administer a certain medication. She double check double checked that medication in chat GBT
10:31realized based on her medical history that she would have some drug uh interaction or some um adverse event that may come from that. She showed that to her doctor her doctor agreed and said I didn't have enough time to look through your medical history uh to notice that. So thank you so much for for raising that. Shouldn't we do that for everyone all the time? If it's cheap enough and convenient enough, can't we do a second, third, fourth opinion for for every clinical scenario? Is that where we're going?
10:59Yeah. I I love this idea of a kind of safety net for both patients and clinicians that is just kind of there in the care journey. And you know, not not always interfering or or or being the middleman, but like really thinking about the ways in which it can augment the existing experience and build on it.
11:12Um you know we did this work actually kind of more directly studying this on the clinician side um in a kind of a series of experiments that we ran uh with Penda Health where we studied this idea of like a clinician co-pilot that was deployed and it was exactly deployed in this way where you know it was a safety net for the clinician as they were doing their work. It kind of came in and proactively suggested like hey this might be off like you know this there might be this medication interaction that you might not be might not be tracking can we you know potentially adjust that and things like this. I think we're
11:41going to see the same thing happen on the patient side and I'm very very excited about this idea of like kind of like a proactive p proactive agent in your care journey really helping you and making sure that like you know if your A1C rises and you're diabetic like that you get like review of your medications by your doctor and things like this and making sure that that's kind of flagged for kind of appropriate review.
12:01You mentioned sort of personalized proactive agents. personal agents in general have kind of emerged on the scene from a popularity perspective over the last few months. Instinct, Muse, Grockbot, OpenAI's dots. Uh you can give these AI apps an objective for folks who are not familiar and it can execute a series of tasks. So it's not just hey do this task and then it executes the task uh with a single loop. It can use your preferences. It can access all of your
12:30applications that you give it permission to. uh it can access your credit card information if you give that to them and it can execute an objective on your behalf. One can't help but imagine the possibilities there in navigating our wonderful but often disjointed health care system. Scheduling a appointment and making sure uh your insurance is covered uh for that appointment or covering that appointment. Uh you know helping you and guiding you to follow a
12:59care protocol that your doctor gave you post discharge. Even negotiating with your insurance company on a medical billing error. all kinds of tasks that just create friction in our healthcare system.
13:10Is that where you're kind of going with with uh with your work? Should we look forward to that from OpenAI?
13:16I'm very excited about that vision. I think if you look at from both a patient and clinician perspective, one of the most difficult things about the healthare system and how it serves people is this kind of question of like fragmentation and kind of lack of connection of the ecosystem. And healthcare is also one of those industries where you shouldn't expect everything to have an MCP connector tomorrow, right? It is a it is a place where there's legacy software. There's not like necessarily the best connectivity with AI or software in the first place. And many people are, you
13:44know, running their primary EHR software in the healthcare ecosystem on like a VM for example. So like virtual machine that they're running in the cloud. Yeah.
13:51Exactly. And so there's actually like an interesting component where like a virtual machine actually makes things easier in some ways for for AI agents, but this is absolutely ripe for agents and computer use. Um, so, so I definitely think that this is like a place where a lot of the barriers that have traditionally existed between patients and getting the care they need and clinicians and delivering the care that they need and healthcare data in general, which is often kind of a bottleneck to the whole system working well for patients. I think a lot of those will dissolve in the coming months and so we're really excited about that.
14:21Okay, so we we're talking about the use cases, you're helping us dream the dream. Uh, before we keep going down that that uh that road, let's talk about trust. Obviously trust is the most important here. The same underlying model um can help me understand my lab results or something that I heard from a doctor um but it can also help a highly trained specialized physician make sense of a very complicated case.
14:46What do you change in the answer so that it's legible to each of those two personas or any other persona that might be engaging with the same underlying model? Yeah, I mean as you said, health is is a place where, you know, it moves at the speed of trust. And so, you need to be able to not only make the models better and measure them, but also kind of communicate and bring the world along. And so, that's been a very important part of our thinking and our strategy for both how we work with patients and how we work with end users and how we work with clinicians. And so
15:15um the the brief answer to your question is basically we we evaluate and train the models by working with experts across the world and and getting feedback in various ways to kind of tailor their behavior um to whether a person is a lay user or kind of an expert lay user that has been like navigating their care for a while a complex care journey or clinician or other kinds of user personas. And so back even for the original healthbench evaluation that we released back in May 2025, there was actually a theme or part of the the evaluation that was really
15:44focused on this distinction of how the model behaved and tailored its level of kind of use of technical jargon and things like this to whether a person was a lay user or clinician. And since then we've both gotten better at measuring whether the models are able to cater to different kinds of users and also training the models towards that. Um, and so a lot of our approach has really been like not only do how do we make the models better for these different users, but how do we bring them along? And part of that is how we do evaluations and communicate about them as well. So
16:13there's a few components to trust. So we've talked about um sort of tailoring the answer to the user. Another big component is disagreement.
16:22um a person can be speaking to an LLM or speaking to chat and be sharing all of their symptoms but really hoping to hear that everything is okay. So from a consumer user NPS perspective, net promoter score perspective, they want to hear reassurance, but from a medical perspective, they need to hear perhaps something very different. Yeah.
16:43How do you teach an AI to recognize the difference and manage different product objectives there either from consumer engagement and retention versus sort of correctness and and sort of clinical quality?
16:56Yeah, this is super central to how we do both evaluation and training. And so zooming out a little bit and the question of like agreement and disagreement around how we um how we respond and how we evaluate like one thing you could do and one thing that we you could you could kind of base your approach on here is like we could bring in a few clinicians into let's say the opening high health team and say like hey you guys figure out how to navigate these kind of uncertain cases and you know we'll we'll ask you about these uncertain cases and you'll help us build evaluations and training data and stuff
17:25like this. um and leave it to those few physicians to figure it out. Um this is actually not the best approach.
17:32So it turns out if you ask five physicians or something this um any question you'll get like eight different answers. And so the the one of the first things that we did when we first started focusing on health at OpenAI is really build this kind of global physician network of hundreds of experts that we work with kind of across the globe across you know dozens of specialties and languages and countries of practice to think about the right ways to kind of aggregate the wisdom of experts so that we weren't relying especially in areas where there could be disagreement we weren't relying on one or two or three physicians to kind of adjudicate. So that's that's one part of it um is kind
18:02of like thinking about um evaluation and training data coming in from different experts rather than relying on one or two people. But then the second part of your question is like how do you balance kind of competing objectives, right?
18:12Let's say the physicians and the aggregate wisdom of the physicians is pointing in one direction and then maybe the user actually wants you to say something comforting even when the situation is not comforting there. It's basically a prioritization of safety. Um and so it's very important to us that like we we take our obligation very seriously when there's hundreds of health hundreds of millions of health conversations coming in. And so across you know the evaluation stack and training stack um we have evaluations that track these kind of difficult situations where the user should where the model should actually go and say like hey this is
18:41unsafe like you should seek care. Um even if the user doesn't want to hear it or you know wants to get some kind of alternative treatment or something like this. Um, and that's, you know, a very important part of how we do evaluation.
18:51And over time, we've not only improved how we've done the evaluation with working with physicians, but also seen pretty dramatic model improvements across the board. Uh, in those areas, there's another example or another instance of disagreement, which is me as the consumer. I ask a medical question of chat GPT. I then show that response to my doctor during a visit. My doctor loves me, by the way. Um, and and then they disagree. they disagree with the response. How should I handle that situation?
19:20Yeah, I mean I think you should handle it similarly um as if you know disagreement was happening in some other way as well. Like people people do disagree about um different things and I think like you have a a right as a patient to get you know ask for another doctor to take a look and and get a sense and go to another doctor to get a second opinion and yeah that's what I would advocate for whether or not you're also using CHP to get health information as well.
19:43Yeah, makes a ton of sense. There's another element of trust which is just absolute correctness and a lot of the a lot of the work that you all have done it seems to be at least I'm hearing from you seems to be about grounding answers in the world's medical literature and evidence. Some parts of medicine there's not a lot of agreement or disagreement about, but there is much of medicine that's still uh being discovered or
20:11being studied. And there is disagreement even in the medical field. And for that um you know for that for those kinds of questions it's difficult to ground it in truth when we don't we're still discovering you know truth science as a process etc. How do you think about correctness in those kinds of instances?
20:30Yeah it's a really great question. I mean it's important with or without AI.
20:34So if you kind of zoom out and look at like how is medical consensus distributed today, right? Like what is the process of that today? You basically have um a bunch of people um writing papers, biomedical papers of various kinds um suggesting kind of improvements or treatments or things like this that could that could improve the quality of care. And then you have this process, yearly process across different specialties of building clinical guidelines which kind of describe like hey in this situation if you're a cardiologist and you see a patient with
21:02X Y and Z here's what you should probably do. Um and so these these clinical guidelines are roughly you can consider them kind of flowcharts that are put out each year by the different kinds of bodies across the different across the world that describe like roughly what you should do in different scenarios based on the latest research.
21:19Mhm. And then the final step is so you have you biomedical literature and then you have the um clinical guidelines and then you have like what do people actually do in practice and so the final step it turns out is a is a ma massive bottleneck. So there are these studies on like how long does it take for um clinical guidelines to penetrate the practice of everyday doctors across the world.
21:39And um in the past people have found that this takes depending on the study like 5 to 17 years for that for that penetration to actually happen which is massive. Like you don't want to be practicing medicine 5 to 17 years behind. And so one of the things that we think about is how do you compress that kind of period of of kind of practicing in in in the past um down to weeks or days or even hours, right? And so this is one very important principle behind our work in chachpd for healthcare and chachpd for clinicians which are
22:08clinician and healthcare professional facing offerings where we're bringing in the latest biomedical literature in the moment of care and allowing people to very quickly in the course of like let's say a 15-minute visit at the point of care get answers to their questions. So that's one way of answering the question which is like you know in practice you know while there is medical a lack of medical consensus in many areas in some cases there is medical consensus but it's not being practiced and so one thing you can do is like really dramatically reduce the amount that that's happening and so that's one one way of answering your question a second way of answering your question is like
22:37how to handle the places where there's still like genuine medical uncertainty and there I think I mean there's a couple things so one is you want to train the models to be able to recognize their own uncertainty like after after browsing and getting information and things like this you want the models to be able to know, hey, like there just isn't enough information and to be upfront about that. And so, you know, that's been an important part of how we've done evaluation that's actually like a public aspect of um healthbench, which includes kind of evaluations on hedging, for example, when there is this kind of uncertainty. So, that's one one important thing. The second thing, I
23:07mean, this kind of goes to this question of like how do you raise the ceiling of human health is like really thinking about the ways that models can actually accelerate the process of doing science and actually pushing the medical frontier. And so, maybe we'll we'll chat about that later.
23:19Some of these areas are gray areas and controversial. And I'm curious if you if you think the models or you're designing the models in a way to take a stance on these things, unapproved peptides, vaccines, all things that just get politicized oftent times in the media. How should a product like like chat engage on some of those topics where there's still debate?
23:43Yeah, I mean where there's debate or a lack of consensus across the board whether it's politicized or not, we want to make sure that we are relying on the experts to help define kind of standards for how the model should respond. And so that's kind of the approach that we take across the board. Hearing you talk, it feels like you're always thinking about one of three constituents. It's either the patient, it might be the individual clinician that signed up, or it might be the enterprise, whether it's a healthcare provider enterprise or some other enterprise. How do you think about designing the product at the
24:12architectural level and at the let's say the harness level to meet the needs of all these different users?
24:18Yeah, I mean I think maybe there's three parts to that. One is making the models really really good for these users. The second is building the right harness or product around it and the third is really like the deployment or partnership motion to kind of re reach out to the right reach out to the users.
24:31And like by default I think a lot of the time the the models are often pretty good these days. And so we talked about our approach of evaluation working with clinicians and things like this. We are seeing pretty dramatic improvements in kind of performance and reliability. You know we talked a few months ago about the improvements that we've been making on kind of the free tier and bringing free tier performance to the frontier.
24:51So you know of the 1.2 billion people who use Chaship PT weekly um giving them a model for health that is just like incredibly useful at the frontier for free. Um so that's been a big part of our strategy obviously for for making healthcare health intelligence abundant.
25:05The second thing is our kind of product offerings. And so the way we think about this is you can have a model that's really good but if the activation energy to use the model to improve health is too high then people won't do it right.
25:15So a lot of our product efforts are really like thinking about these constituents that you're talking about and thinking about the ways in which we can reduce the activation energy for people to use compute and intelligence to improve human health. Um and so that goes you know along the lines of you know our work in chash PT uh health and chash PT um chd for healthcare um on the enterprise side and chash for clinicians. A lot of the way we design these offerings is basically to think about like what are the barriers and bottlenecks and activation energies for people to be able to use the models. Um
25:45and so for example um you know what are the um security and compliance bottlenecks in providing an enterprise offering of chachbt to healthcare systems. So that was like one question that we were thinking about a lot 2 years ago. And so we thought, okay, we need to we need to kind of invest in security and compliance. We need to make sure we're able to offer HIPO compliance. We need to make sure that we're able to offer the kinds of of um privacy protections that healthcare organizations need so that we're able to handle the healthcare data that they need and be able to offer services that they trust.
26:14And so that was like a very important part of the work we were doing for like roughly two years ago to like roughly a year ago. and um and we kind of just go after there from from there like based on the bottlenecks that we see to kind of adoption. The third thing is kind of partnerships and deployment. And again, it's really like, you know, a few years ago, I think OpenAI was very early to speaking the language of healthcare. Um, and when we um interacted with healthcare partners, I think we were, you know, early in those conversations.
26:40So, I think one of the things that we've brought to bear since then is really, you know, thinking about each of those constituents that you're talking about and really thinking about how to speak their language and how to kind of bring them along this this journey because even if the models get really smart, if nobody's like really trusting them or using them, that's not really achieving our mission. And is that uh you're alluding to a little bit some of the work you've done with Epic to kind of integrate the product directly into Epic. Maybe share a little bit more about that because a lot of the folks who spent a lot of time in healthcare uh were very interested to see that level of collaboration with with Epic.
27:09Yeah, I mean this is a very important part of our approach. You like like we talked about fragmentation um in healthcare has been one of the issues um and the fact that there is kind of data and workflows that are disconnected from each other I think is is a big obstacle to the to the use of AI productively in healthcare. Um and so you know our work with Epic or work with other um partners we had a recent announcement where we announced that there's now a way to use Epic in CHGPT um if you're if you're a provider or healthcare professional and also for chat GPT to be useful in Epic
27:37itself. And so on in the Epic workflow itself you can have like a little um sidebar where you can actually use Chat GPT as well.
27:44Um and so that's one thing we we've we're also working on like different other kinds of connectors that we've also announced recently and we have more coming. And so there is quite a bit of work that we're doing to make sure that we are kind of breaking down these barriers to kind of the healthcare ecosystem being connected. And I think you know Epic is just one of many examples where we're just identifying those bottlenecks and trying to fix them. The last thing I want to hit on with respect to trust is uh some of the work that you've done around mental health. And uh this is another area where there's a lot of complexity and
28:13there may be other objectives that you need to solve for beyond um the clinical complexity that that we talked about. If I ask my openai dot, hey, I want to feel better over the next few weeks. How do you even interpret that question? How do you help act on that? How do you help guide a user on that given that that could mean so many things?
28:34Yeah, I mean and regardless of the surface, we are super committed to like helping users when they're having a difficult time or when they're in crisis. And so, for example, for DOTS, we've been investing the same kinds of safeguards that we've been putting in all of our product surfaces in that product surface as well. So, that's an important part of our approach is kind of like a a baseline safety floor. And then you know with mental health bench we've really taken this approach of of not just measuring the safety floor but thinking about the ways in which the models can actually like proactively help kind of raise the ceiling um and really think about the ways in which
29:02models can you know help provide positive mental health experiences and move from move from just um worrying about the safety floor to measuring those those kinds of upsides. And so again, the approach is is actually very much mirroring our work on Healthbench where we've just been working with um a bunch of licensed meth mental health experts across the world to think about the right ways to define model behavior and a bunch of tricky scenarios and ensure that we can get to a level of consensus. Um people definitely disagree even more here than physical health. Get to a level of consensus in some of these
29:32tricky situations where we can actually measure performance and create this kind of like shared benchmark from the industry. And um so we released that that benchmark very recently and similar to healthbench and healthbench professional the goal has really been can we create this benchmark that kind of raises the ceiling for everybody and kind of creates across all model developers and the healthcare industry and and mental health experts like this really shared ruler or measuring stick that we can use to kind of make sure that we're moving in the right direction.
29:59Okay. So we've spoken about intelligence and the use cases that are coming around uh intelligence approving. We've spoken about trust which is critical. Let's talk about medical care. And I want to draw the distinction between medical information and medical care. Uh I think for medical information, I think if you have a health concern, if I have a health concern about my myself or something that I'm dealing with, chat and other LLMs are increasingly the better, faster, cheaper way to get an answer and a resolution to that
30:28question. But if uh in the scenarios where medical information alone doesn't get me to resolution, I then uh at least today am thrusted into the traditional medical care system. And so if I need to go get a prescription, if I know need to go see an in-person specialist, if I need to get some follow-up blood work or some MRI or some other form of scan done, if I need a clinician to examine me in some way, I still need to navigate
30:56the traditional health care system. And a lot of that work um uh is sort of uh governed and regulated by the US medical license and sort of other infrastructure that we have in our healthcare system.
31:09Um it's the multi-turn conversation where it's not just a single question resolution. It's I got to go talk to a few people. I need to go get a few things done. We need to collect some information. In many ways, medical care is a it's a bundled product. you know, we're getting the diagnosis and triage and the care plan. We're getting the empathy from the clinician and the assurance and the support. We're getting the legal liability associated with that and what they're taking on. Will AI
31:38provide all of that? Uh where will it stop short? Uh where do you think where do you think this is all going?
31:43Yeah. No, I don't think AI can can or will provide all of that. And I think you're right that there's like this distinction between health information and and medical care. And so I think I think the way we see our role here is really to enable the whole ecosystem because there's no future where you know any one company is delivering all the medical care right and so our role here is really to think about what is the last mile of medical care really look like for for people. Can we really understand that and can we make that
32:11easier and um easier for everybody so that last mile of actually having impact on patients just looks like the last 100 meters for all of our partners. And so that's kind of how we think about it is like really like making it very easy for for our partners to use AI to support everybody in the care journey. I think there is stuff that OpenAI can do relatively uniquely here because of the front door to health that it has. And so for example, it can help you navigate a lot of these kinds of questions or help you navigate this journey or you know as we get increasingly helpful personal agents help you with some of the scheduling and logistics and things like
32:41this and just act as a quarterback for your health. Um that said, I don't think it'll resolve all these challenges that we'll have, at least not in the short term. And I think to a large extent, we're looking to enable a lot of our partners that are delivering care that are, you know, overwhelmingly using our models in various ways to like make these operations happen smoother and easier so that patients are not left in this kind of fragmented experience.
33:03So should the as a consumer should the experience that I look forward to uh from OpenAI be um sort of the quarterback care that you described?
33:13Will I be talking to a clinician in the interface at any point or would you perhaps route me to a doctor and help me find that given that's a whole challenge just scheduling and making those visits happen. What should I look forward to?
33:25I I think I think you'll see like we we will be hopefully very helpful in helping you navigate the process of finding care. I don't think we'll be in the business of doing all the things that you mentioned.
33:35Um but and that that's where we're really relying on all of our startup and care provider organizations and partners and things like this. So, so coming back to a question I mentioned earlier, which is you are building that future, but you're building it with partners, you're building it for consumers and patients, you're building it for clinicians, you're building it for enterprises. For clinicians and enterprises, like give us some texture on uh on some of the tools that you're launching and and some of the collaborations that have really worked.
34:02Yeah. Um we've had a few really exciting partnerships and we've had some really amazing launch partners for Cheshpd for healthcare. We've had lots of users using catch for clinicians and really enjoying it. It's been um the retention on the on the offering has been like a a very pleasant surprise. People people seem to be really enjoying it. Um yeah, we've been working with a variety of health systems. I think one um a couple that stand out you know one is um we've been working with UCSF and we've been chatting about that and they've been you know one of our earliest adopters for this connection between Epic and the EHR
34:31and CHPT and they've been really on the front lines of thinking about for example clinicians using it for chart review or things like this um and thinking about the right ways in which like the experience of caring for a patient doesn't need to be this super fragmented you need to switch tabs or switch experiences and things like this and so they've been some of our power users on that front. Um I think another one um is Advent Health and they've actually been some of our power users in think and thinking about deploying agents across the workspace actually across all of OpenAI. They've been um
35:00they've been a leader um and they've been really um they've had amazing results. They've reported roughly 80% reduction a lot of their administrative um workflow costs using AI in these kinds of ways. And I think a a lot of health systems I think should look to them. I think they're they're really impressive. If I go into my doctor's office and I see chat GBT open up on their screen, should I be upset? Should I be happy? Yeah.
35:24How do you think people will react to that?
35:26Yeah. I I I hear I hear differing things and I think it's actually changing over time. And so I I think it's this interesting thing whereas as the Overton window is shifting in real time. I think you know there might have been some level of concern before but I think increasingly a lot of you know on both sides you know there's this question of like you know whether clinicians would push back against patient use of AI and whether patients would push back on clinician use of AI in the way that you're describing but I think both parties are actually finding incredibly useful for their own personal purposes. M and so I think what we're finding in
35:55practice especially over time is that both parties might actually feel safer with AI in their care journey whether it's you know their clinician using it or the patient using it as kind of like a a sanity check and I think that that attitude is definitely changing over time but I wouldn't say that we're fully there but that's my prediction of how things go.
36:10Okay. So then taking that one step further, if a patient has done all their research before a visit and they have all their health data integrated and they've asked a bunch of questions in chatbt and then the clinician before the visit has all automated through loops uh has run the entire patient case through uh through chatbt and then the visit happens. What should they talk about?
36:34Um my hope is that they talk about the human aspect of healthcare. I mean there there's still like a lot of uncertainty.
36:38There's a lot of empathy. There's a lot of bedside manner that goes into the human aspects of healthcare. There's also a lot a lot of physical interaction that goes into it. I don't think you know I don't think all the information exists for the AI just information the theoretically all the information exists for the AI to just figure out what all should be done and part of the reason for a visit is to get more information and do the process of history taking investigation and things like this. And so I think that will not go away especially the the physical elements of that. And I think even from a cognitive perspective, the elements of of kind of
37:07empathy and getting comfort and getting that second check um I think will still be something that people demand.
37:13Can what do you think uh AI in general or let's say dots would do in between visits and will you think you'll cross the line into providing providing care in between visits and then share that back with the doctor?
37:27Yeah. No, I I I don't think we'll um I think we'll we'll we'll definitely be focused on health information, but I think the whether it's dots or any other product surface, I think we're really excited about this kind of future of heading more in this proactive direction that you were outlining. Like imagine if somebody is kind of dealing with a chronic care journey and you know they're again like this example of like someone's A1 someone gets a lab test back where their A1C is higher and they're pre-diabetic or diabetic you know can you flag for them like hey maybe you should get a medication check and go to your you know cardiologist or
37:57whatever and think about the right ways in which you can um you know flag to users medication adherence issues or other kinds of issues. So I think those kinds of issues are places where you know AI can have potentially a really high health impact especially like these kind of chronic care journeys but I don't think um I don't think it'll be we'll be in the business of like managing all these things by ourselves.
38:20Yeah. Another thing I'm curious about is in the fields of AI applied to other domains and situations like let's say games like chess and go. Um AI as AI got really good at chess and go they also revealed other strategies and ways of playing the game that even the best experts in those domains were not exploiting or not pursuing and that helped us learn more about these games and um also helped us learn more about AI. Are there equivalents like that that
38:49you've seen in AI applied to health from the work that you've done uh in research? Are there already findings that are surprising you and perhaps introducing new ways of delivering care or understanding a clinical scenario uh that can give us some optimism? Yeah, I mean I think we've we've seen this at smaller scales, but let let me let me zoom out and say like, you know, one thing that's been motivating our team in general recently around raising the ceiling of human health is this notion of bringing to bear like kind of a move
39:1937 for for health. And this idea of a move 37 is it comes from from Alph Go.
39:24Um and uh a game that the Alpha Go system which plays this kind of complicated board game uh played with Lisa Dole where there was a specific move that just kind of in the moment was seemed like a a really dumb move and a surprising move for all the human experts watching this game uh between this AI system and this world expert and go but in hindsight is this kind of like brilliant and insightful move that humans are now just u beginning to understand and learn from the patterns of and things like this.
39:54And this this is very interesting example. Um I think so so this idea of like how do you get to a move 37 in health of a move or prediction where the models predict something that a human might not have been able to predict in hindsight seems seems correct and brilliant and insightful and impactful.
40:11So I think I think we're very early and I think you're going to see much more of this in the future than in the past. But um I think you know I mean one kind of example is again thinking about these the ways in which these systems have been deployed as safety nets to clinicians where clinicians are actually delivering care and in the moment the AI actually flags kind of like hey something might be off here like like our study with Panda Health and thinking about the patterns of of results that we saw there as an example you know um in that context uh the clinicians were a little
40:38bit um prone to overuse of antibiotics and a very common pattern that the the models flagged is just kind of like avoiding this overuse of of antibiotics and thinking less about the local case and more about the kind of global um issue of resistance to antibiotics. So that's like one, you know, specific example, but but again, I think you're going to see more of these examples of models taking in like a longitunal history of a user over um 10 years across all their data, across their HR data, across their chat data, all the all these things and making predictions
41:06that seemed impossible before cuz a human can't even integrate all that data in their head.
41:11And making predictions about their health and their risks and things like this that seemed like they were impossible. And I I think that would rise to the occasion of of being called the Move 37.
41:19Yeah. Yeah, I mean there are there's so many things that consumers are curious about uh that um at least at one point in time we weren't sure if it was healthy for you and then the to your point the Overton window changed and uh and there was sort of more evidence that was sort of conclusive in one direction.
41:37You know, at one point people weren't sure if smoking was uh bad for you. Is coffee good or bad for you? Is sugar good or bad for you? Uh sometimes the views on some of these things have changed and today consumers are they want to know is creatine good for you?
41:52Is melatonin good for you? Uh you know all these sort of new emerging phenomenons that perhaps are underststudied or at least um uh you know not widely uh sort of agreed upon this safety and and effectiveness of certain of certain supplements or certain medications and things like that. Can do you do you anticipate as people are sort of entering this information into chat GPT or other LLMs that we'll have more evidence that we
42:19can generate to study some of these things that folks are curious about but we don't really know uh or we don't have conclusive evidence one way or the other on on where they stand.
42:29Yeah. I mean I think one of the most interesting things about agents is they kind of solve a lot of the bottlenecks in analyzing data that already exists. Right. There's various places across the healthcare ecosystem where data is being generated. EHR data, longitudinal histories, claims, whatever where data is being generated and it is kind of just stored there and the insights from it are not fully mined or used to generate new knowledge.
42:52Yeah. And so I think one of the frontiers for how how health might improve in the future and how the ceiling of of health might might raise will be thinking about the ways in which we can leverage the data that is just kind of out there in boxes in various places and have models generate new insights based on this data. So so I'm also optimistic about that future. Um I'm also optimistic about a future where you know you mentioned these two constituencies the the user the patient and the clinician. I think I think there is a third constituency which is the scientists right which was helping kind
43:20of advance the the frontier of knowledge in the space and I'm also very optimistic about a future in which you know um AI helps accelerate a lot of the bottlenecks in validation which I think will end up being the bottlenecks as opposed to kind of proposing new ideas or treatments or things like this um and so um very optimistic about that as well. Switching gears slightly, you've also shared thoughts around um just the cost of intelligence as an input to medical intelligence coming way down and you know at some point it'll be too cheap to
43:50meter. Um I spent a lot of time and we as a team spent a lot of time trying to understand how these technologies will make an impact on our healthcare system that uh is inflating well ahead of uh of inflation. uh 5.3 plus trillion we spend on healthcare. If it were its own country, it'd be the third largest GDP. Um it's the biggest part of our economy.
44:13It's also the biggest employer uh in our economy. 23 million Americans work in healthcare. There's so much um there's just so much we spend on healthcare.
44:23What do you what do you sort of make of the work that you're doing and the impact that it'll have on the cost of care? I think this is like one of the most important ways that we can pursue our mission is to not only make our models better at the frontier, but make that widely accessible to everybody. So that's, you know, obviously a big part of how we've been able to deploy health intelligence to a lot of people across the world for free. I think it's also a big part of how we think about, you know, um API and how how good our models are there for health and pricing there and stuff like this. You know, one one
44:50recent update was we recently released uh GPD6 Luna and GP6 Luna's um is our current cheapest model on the API um or cheapest, you know, frontier model on the API. And what we found with GP6 Luna was that it was actually doing better on our benchmark Healthbench Professional, our our current latest benchmark of of kind of clinician workflows that are like really difficult selected for difficulty for our current models across all the different ways in which clinicians are using our models and and
45:18the models um still could be could use some improvement.
45:21Um it performed better than 5.6 Soul, which was literally our best model like a month ago. And so and it's it performs better despite being 34 times cheaper.
45:30Um, and we've had a couple of those moments even even in the past year. And you know, there's been many trends showing that AI compared to other technologies that have been introduced in the past has been reducing a cost at a given level of kind of capability or usefulness faster than many of those other technologies. And so one bet we've had at OpenAI is that the cost of health intelligence will will plummet at a given level of health intelligence.
45:52And the the [clears throat] level of health intelligence where you know clinicians and patients feel like they can trust in the care journey is starting to get there, right? So that's like a fixed level of intelligence in some sense. Hopefully the bar for that raises over time as well. Um my expectation is once we get close to that level of intelligence the cost for that level of intelligence also dramatically reduces like it has been as the hope there is that we get like a kind of this commoditization of health intelligence right we don't want to hoard that health intelligence we wanted to everybody to use it so my hope that that contributes but you
46:21know to answer your question about the whole of healthcare spending I don't think all of that is about health intelligence I don't think like health intelligence has had yet a macro factor on the a macro effect on the problem that you're describing if you look at like the whole spend of health. There is quite a bit of areas though where if you look at for example chronic care management or documentation or administration or a lot of kind of administrative workflows of various kinds, back office workflows, there's a lot of money spent in ways that probably don't need to be spent that way. Um and
46:50so we are very optimistic about kind of an overall compression of spend over the over the coming years. I don't think that's happened yet and I don't think that just happens from making health intelligence abundant. I think that happens from integration and workflows and deployment and scaling that up and really making sure these models are valuable in the economically valuable tasks that that people are using them for. But I think it's a a prerequisite to making that future happen.
47:11More care, less paperwork is sort of where we're heading. At least there's a lot of the cost savings.
47:17I'm sure you saw some of the headlines over the last few weeks of uh there's a New York Times article uh that basically provider AIS and payer AIs are fighting and there was some Blue Cross Blue Shield Association analysis that showed that uh the respective billing bots uh and paperwork bots effectively uh increased spend by a billion dollars or something like that. when you are doing cutting edge research on health and trying to bring down the cost of medical
47:44intelligence to near zero and then you see headlines like that, what do you make of that?
47:49I I think this might be one of those areas where there might be some kind of short-term pain around this kind of provider payer arms race, but in the long run there's no reason this has to exist, right? And so re really I think you know you're entering a world where a lot of visits actually have some kind of ambient intelligence in the room listening, transcribing things like this. um there is notes, there's clinical notes, there's AI adjudicate AI details of of clinical notes and things like this. And so you have a lot of the details you need to adjudicate visits, right? And so
48:18the ideal world in my view is a world in which providers and payers kind of have a handshake like, hey, here's our ground truth of what happened in this visit. Like here's all the information we have.
48:27There's just like no other source of information that there is, right?
48:30And then an AI helps adjudicate that situation. And maybe there's like some limited appeals process or things like this, but I don't think you need to have all this process where where you have like an AI arms race between payers and providers. I think they can just be completely avoided. So I I view that as a a bit of a short-term distraction. No, there's a lot of entrepreneurs that are building technology and healthcare that are that are likely watching this and uh they're probably trying to figure out uh you know as you continue to advance your work and you bring down the cost of
48:59medical intelligence to near zero and this becomes this sort of ubiquitous utility that flows through all of our healthcare interactions for consumers, for clinicians, for the administrative part of of the healthcare technology or the healthcare stack. Um, what should folks be building on top? What should specialist applications be doing assuming that this gets 100x plus better?
49:23Yeah. Um, it's a great question. Like I mentioned, like we think about a lot of our health impact through OpenAI as not being directly through the products that we offer, but actually through our partners and and that last mile of delivering care.
49:34And so, you know, what I'd say to to our partners is, you know, build on top of models that are really capable. And the most important thing is bet that those models will become more capable. So bet on the future. I think many people in healthcare are on some kind of spectrum between betting against the models, betting neutrally for the models and betting uh with the models or betting for the models. And I think um I would say you know bet on the models getting better. That's probably the most important thing. And think about the ways in which a lot of the technology that's being developed whether that's you know agents or computer use or other
50:03kind of frontier tech will accelerate healthcare and and bet on that and build towards that future. I also think a lot of the medical care um will actually like um become more accessible to patients with AI. You know, we talked about this idea of kind of a quarterback and dissolving a lot of these barriers.
50:19And so just being focused on delivering really good quality care um and having AI be part of, you know, monitoring that care journey, I think is a very important part of that as well.
50:28Zooming out to the overall healthare system and now just looking at it at large, you know, we talked about this 5.3 trillion. A lot of that spend is paperwork between providers and payers and other stakeholders within the five Ps of healthcare. A lot of that is cognitive care that you're arguing can be augmented and um enhanced and sort of we can have bring on a lot more cognitive care capacity uh online which is exciting and exciting from an access perspective. And then there's a lot of
50:56spend in healthcare that is procedural either requiring some imaging, some blood work, some follow-on examination or it's some surgical intervention. Um, where do you see sort of the intelligence work that you're doing in the medical space impacting procedural care? And is that going to be the ultimate is that going to be the new sort of or an enhanced uh or perhaps an exacerbated bottleneck or do you expect to do a lot of work there as well?
51:23Robotics etc. Yeah, I mean I I think in the short term with procedural care, I think you want a better targeting when that's needed, right? And I think AI can help, you know, provide information that is helpful for knowing when to seek that kind of care as well.
51:36Probably the the biggest macro factor in the next, you know, year or two will just be better kind of um proactive preventative stuff, right? So you don't have to actually go in later and get that procedure right because if you make the right moves sooner and AI helps you navigate that kind of journey then you can hopefully avoid a lot of the downstream cost of the most expensive procedures. So that that might be like one macro factor to kind of look out for is my guess.
51:58I think the question of kind of robotics and care I think is a a bit of a longer term one both from a technology perspective and kind of like a reimbursement and kind of industry um integration perspective. My expectation is that it's not like a next few years kind of thing. Very cool. Okay, we've covered a lot of ground. We've talked about intelligence improving. We've talked about trust. We've talked about moving from medical information to medical care. Help us understand what we can look forward to as things uh as sort
52:27of things continue to unfold. Um as agents and personal agents and the underlying models become increasingly performant, what can we look forward to from chatb health?
52:38Yeah, I mean I think there's like really two highle things that we're really focused on making happen. One is again thinking about like the ways in which we can raise the floor of human health and make trustworthy reliable health intelligence abundant for all clinicians and all users and patients. And so getting to the point where you know like we talked about like all patients and clinicians just feel safer by having AI in the care journey like when you see the the chatpt you actually when the doctor's using chach you actually feel safer and vice versa. getting to that point and again that's that's a question of not just you know the models being
53:08good and and eval being good but also how we deployed and the activation energy around that and um the products and things like this um I think a second thing is really thinking about um raising the ceiling so what are the ways in which AI can actually do things that were not possible before so one of them actually I think is this kind of like you know proactive health proactive personal health assistant that is part of your care journey is helping you navigate these moments being your quarterback you know helping you schedule and all these things but also flagging the places where you know you can intervene in your health today and
53:37prevent a lot of know expensive deterioration later in the future. Um and I think that's going to be an important part of how we think about you know improving the future of of health for for people as many people as possible. And I think the second thing is just what you mentioned before which is like really thinking about the ways in which we can unlock the value of existing data that is sometimes stored in boxes of various kinds and thinking about the ways in which that can be leveraged to produce new insights that benefit patients. It is striking and
54:04incredible and humbling to realize that chat GPT I think came out three years ago roughly. Uh and so it's been a lot of progress uh in a very very short period of time. So to even talk on a five or a 10-year basis feels academic but can you give us some predictions? I mean is there a feature that we might have that feels science fiction today uh but we could have one day 3 5 10 years from now? Yeah, I mean is it is super super hard to predict the future these
54:32days. So um um this is a difficult one, but what I would say is like I think the thing that might have the most impact on human health in in that kind of time frame in the next 5 to 10 years is probably going to be newer treatments that come out of maybe two things. One is um AIG guided um generation of new targets or molecules that that that could be interesting for you know for new treatments and the second is AI guided prioritization and validation. So thinking about the right ways, we're
55:00already at a point where the number of molecules and drugs that people want to suggest to be used in in clinical trials and things like this is way more than the number of slots available in clinical trials. So thinking about the ways in which models can help prioritize and run these clinical trials so that they just run smoother and faster. My guess is in in that kind of time frame, you're you're talking about kind of like improved medicines, more personalized medicines for people, things like this.
55:25And that will only be possible because you invested in those two pillars. That is awesome. Uh well, we've covered a lot of ground. I'm leaving this conversation even more optimistic than I was before. Uh so Karen, thank you so much for coming on the A16Z podcast.
55:39Thank you for having me.
