MATT LUNGREN: All right.

Well, welcome back to the Stanford AI and Health Podcast

series.

Again, I'm Matt Lungren.

And I'm joined by Justin Norden.

And today's guest is probably familiar with most

of our listeners, Seth Hayne, who's the senior vice

president of R&D at Epic.

He has been at Epic a whopping 20 years,

focused on technology, AI, and applications there.

And fun fact about Seth.

I've gotten to know him a little bit

over the last couple of years.

As a math major, he's obviously brilliant

and a brilliant technologist.

But as a card carrying English major myself,

he is a prolific reader.

And I find more often than not that his book suggestions

are way better than mine.

So clearly, God has given with both hands in terms of talent

to our guests today.

We're excited to have you on the show.

Welcome, Seth.

SETH HAYNE: Thank you, Matt.

It's good to see you both.

MATT LUNGREN: So we have some topics,

I think, to tee up, as usual.

I think we typically try to get to just where are we

in at least at the point of recording

in this exponential world we're living in.

And I know a lot of folks in the audience

are a little different than us, in the sense

that they don't obsessively follow a lot of this.

So Justin, I don't know if you have some of the latest.

So GPT 5 has launched.

I think a lot of folks maybe had mixed response,

let's say, at least in relationship

to the hype leading up to 5.

We can get into some nuance there.

I'm curious to hear all your take on this.

But from the healthcare AI lens, once again,

it's saturating these typical benchmarks.

So I don't know, Justin and Seth, you've

seen this kind of data over, and over, and over again,

where we're using some of these benchmarks.

We get into that.

But first impressions of 5, thoughts around how

are you using the models today.

Has anything changed for you with the latest release?

SETH HAYNE: First impressions on 5.

It makes me laugh because every time one of these models

comes out, the very first thing I do

is I give it to my son, who is 14 years old.

And I ask him if it can build the magic deck that he

has always dreamed of for him.

And unfortunately, we still haven't cracked that nut.

So I think there's still a ways to go.

I think the USMLE benchmark that's

doing better on than my son's Magic benchmark, which

is my current--

MATT LUNGREN: These are the two that you track?

SETH HAYNE: Those are the two I track.

That's exactly right.

MATT LUNGREN: But the big question

is, has he made a magic card with your skills and stats?

Does he have his own personalized deck

with the parents?

SETH HAYNE: He's starting high school.

He's avoiding parents at this point in time.

MATT LUNGREN: It's probably a smart move.

Yes.

SETH HAYNE: Yeah.

We'll steer clear of that at the moment.

I have found that I spend a lot of time--

he brought up GPT 5 specifically using

the thinking and pro models.

I think that there's two tiers of usage of it.

Folks are pretty excited to see what can happen

with these reasoning models.

And I think there's some many folks are getting exposure

to it for the first time.

But I think the opportunity to know consistently,

you're going to get the high quality

output with either the thinking or the pro,

obviously, doing more with a little more latency.

Tends to be my usage patterns these days.

I don't know about you, guys.

JUSTIN NORTON: Yeah.

And I think-- and we've seen also some in the comments.

People ask Matt and AGI camp, this is the next thing.

I'll say, I was in more of the rest of the internet of hey,

it is better.

I think it is a very interesting consumer choice

to do some of the routing behind the scenes,

to expose people to different models.

I don't know if it was the Death Star giant

leap forward that some people were predicting.

So there's a separate question on, again,

what is the rate of change?

How much better are these models getting?

So I think from--

was it the exponential, from three to four,

four to like-- no.

Like to be super clear, I think at least not

from my own experience there.

But then the other interesting-- there's--

again, it's not just that that's come out since our last--

Seth, to your point on magic cards, and as you're

saying, going to make one for you.

There's Gemini's Nano Banana, and image models,

and other more specialized models coming out

that aren't just in this, hey, bigger models.

And so I think to me, this broader

theme of model selection, specialized models,

and other things.

And I know we're going to get to this topic later

with some of the announcements around some of the Cosmos

pieces.

But I think that's just another thing

for, I am watching as a field as I see these models come out.

MATT LUNGREN: Yeah.

It's interesting.

I feel like definitely, we track,

maybe almost pathologically too much the models.

But I would say that I was one of the rare people that

liked having choice, a bunch of different models to choose from.

And I maybe was also one of the rare people that liked 4.5.

And for those in the audience who didn't know,

that was this middle step model, but really

great at writing and just was felt much more natural, at least

to me.

But yeah, I think that it depends on your starting point.

If you were someone that was a casual user, a dabbler, and you

weren't even using some of the thinking models.

And now, all of a sudden, you're asking a question,

and you're getting this really, quote, unquote,

"intelligent response" that was the result of a thinking model

as opposed to the typical for a response,

I think you will be wowed.

And maybe it would feel like a Death Star

or whatever the hype is.

But to your point, I think the incremental increase

for getting to a place, where I just

don't know if I have hard enough questions to really tease out

the true difference.

And so I-- and again, we look at these benchmarks.

In healthcare, obviously, I have felt pretty solid

about putting my own healthcare data.

We've had this discussion before and not feeling

compared to the original 4, and certainly 3.5,

where it's wildly off.

It feels, in fact, I'm learning things.

I feel like that I continuously validate myself,

but I'm still like, wow, it really is impressive.

I do feel like, though, on the benchmark side,

we are drifting away from being able to rely on those

to give us information that's useful in our daily lives.

SETH HAYNE: I think one of the things on the benchmark side,

it's interesting that we continue

to use benchmarks that are founded in the education system,

if you will, from a medical perspective,

and we haven't quite figured out how

do we move into those later stages of a clinician's journey

as they become better and better in the role

and move out of school.

And I think it's an important benchmark.

It's something we need to keep watching

But I think there's other modalities,

and there's other approaches here

we need to start tapping into and thinking through.

JUSTIN NORTON: Yeah.

And actually, one of the other recent papers,

and I'll pull up the chart here, was

there, serious problems with multiple choice questions

as the only benchmark.

And this was a study done from a colleague at Stanford,

Nigam Shah's lab that showed, even if you just

change none of the other answers from the right answer,

performance dropped.

And so there are issues, and we were just

talking about getting a consistent response, also,

to these things.

Remember, these LLMs are non-deterministic.

And so there are issues to work through, really,

as we think about performance and how

we're evaluating these systems.

SETH HAYNE: I think one of the things that is often challenging

between this type of research, which I think

is important and often informs, for example,

types of development paths we take here

in regards to prompting and other pieces,

is that it doesn't represent the way that we build software

consistently.

We know when we're creating, say, a summary for a physician

in the context of the ED that has a patient presenting

in a certain circumstance, that there are a series of checks

and balances you need to go through in regards to creating

that summarization in the context,

and then also using things like citations

back to other parts of the medical record

to help account for the quality in those contexts.

And it's unclear to me how one starts

to replicate that in the academic literature

in a consistent manner.

So I think that's one of the reasons I

was excited to come on the podcast

because I think this type of dialogue

needs to continue to happen between both industry

and academia in many ways because we can keep learning

more from each other.

MATT LUNGREN: Well, I think you're raising

a really important point.

I think we can park the idea of do the models broadly.

And again, at a place like Epic, which

is famously engineer the prototype of practically

your entire workforce is as an engineer.

So there are clearly advantages.

And I'd love to hear your perspective just

on using some of those tools.

And you hear some hyperbole about percentages of code

written by Ai-assisted models or whatever.

But if you just go back into the physician use case,

it's an incredibly good.

I feel like we've gotten enough signal

from the models with the benchmarks

and some of the anecdotal use, that they

have capabilities that are useful for real things

in healthcare.

The question that you're raising is like, OK,

now that we know there's signal, how do we actually craft that

into something that really delivers something for our end

users, whether it's a patient or a physician?

And that gap, to me, is still a bit of a chasm

and does require the hard work of either having the domain

expertise, the tight developer user interactions,

and frankly, just the know of how

to stitch these things together, and that I

think there's an overhang.

We talk about the overhang a lot.

Capabilities are here.

Getting those into the real world to show that value

is still going to be messy work that will take time.

SETH HAYNE: Well, I think, first off,

you got to make sure you're asking a meaningful question

when you're designing a solution.

What is-- is it--

I don't know.

It's probably cliche at this point.

But I mean, we deeply believe that every developer

needs to get on site through some--

we call it immersion here.

But you've got to spend time at the elbow

with the physician, understanding

what are the problems you're trying to solve.

And in some cases, these are diagnostic type questions

about labs in those cases.

And many of the contexts, it's an administrative question.

And I think you've got to start by asking the right question.

And then to your point, there's this question

of how do you build up the right pipeline?

How do you build up the right back

end structures, both in regards to monitoring

and ongoing reinforcement loops for improving the back end

process for whatever you're generating, be it a summary,

be it a draft of a note, be it a SQL query

for putting real world evidence at the point of care?

And I think that is starting to mature.

I think the thing that there's a lot of opportunity to spend time

talking about and continuing to think through

is what are the new user experience patterns,

those design workflows, particularly

where having the models think longer

continues to improve the outcome.

But now, we're not in this instantaneous,

I have an answer in under a second question,

but it might be worth waiting for.

And I think that there's a important conversation

to be had on that front that's coming in the industry,

and we're just starting to tip into.

JUSTIN NORTON: We could talk about nothing else but that.

And I think it's what's fascinating

right now is we were just talking about model selection.

At OpenAI, there was a lot of backlash

on taking that away and forcing people

to an instant response versus thinking.

We haven't solved it there.

We also haven't solved it in health

as we think about higher risk interactions,

and latency for certain pieces of data,

being able to wait for others.

And then I think the points you're bringing up,

Seth, on just software and building software.

The way I describe it often when I'm

talking to people in healthcare is

if you want something that is reliable and consistent,

high quality, candidly, you want as little LLM as possible

because that's going to introduce

noise in what you're working.

And so we have this kind of dichotomy of worlds.

We have a new model come out and benchmarks.

And oh my gosh, look how good it is.

No one is building software.

That's just LLM output.

Good luck.

Hopefully, that was the right piece.

How are you bringing the right context?

How are you bringing the right data, audit controls,

monitoring, governance, evaluation?

All of these other pieces are how you build good software.

And I think as a whole world, we've

gotten so excited about the LLMs.

In some cases, people have forgotten about good software

development to actually build high quality systems.

SETH HAYNE: Well, I think that importantly highlights

a common misconception I sometimes hear,

where folks confuse the model with the actual end user

application.

There's so many other ways to use

these models besides a consumer facing chatbot,

whether you have a model selection or not.

I mean, we don't have people typing prompts into our software

when they're getting a summary.

You need to build in consistent guardrails,

consistent checks in regards to where the data is being pulled

from, how you're using these different patterns,

and then ultimately designing it into a workflow

so that an informed decision is being made in regards to what's

being presented, and that user has the opportunity

to make the decision.

And it's very, very different than the consumer

experience most of us have on a day to day basis.

MATT LUNGREN: And well, I think this is important too.

Because I think, again, the one to one interaction, I think,

will have its limits.

But then to your point, Justin, too, as I think about the tools,

and the software, and the things that we

rely on that are deterministic, that are like a calculator,

let's just say, I still feel like there's a part of me,

and I've said this before, but I'd

love to still abstract some of that away.

And to your point, Seth, I don't maybe

need to use the model for the actual interaction

for the answer I'm getting.

I just need to go fetch the right tool

to then retrieve the answer or the piece of data.

And that can be kind of, again, pushed back from my

having to choose or set that up in advance.

And I guess maybe we're not quite there.

But to me, there's a multi-agent story.

There's a multi tool story.

There's a better together story, where

all the work that we've done for decades in traditional LP

and all the calculation risk scoring start

to accrete additional value to having an agent that

can get that information for me and save me

time, or mental effort, or whatever we use as the metric.

SETH HAYNE: I feel like anybody that

is continuing to push the bounds of using AI in new ways

needs to simultaneously be as ready to push for not

using AI when it's not needed.

And sometimes, that seems contradictory.

But I think back to your point, the right tool

for the right job.

And this does provide a generalization framework

that can be applied in a bunch of contexts.

But that doesn't mean you should always use it.

MATT LUNGREN: Yeah.

Well, this leads to now what's next?

Because I feel like there's clearly

going to be a continued slope of improvement,

probably on a lot of tasks.

Some of them may be health of the models.

They will be useful in applications.

But I'm really curious to bring up some of the work

that you all announced around specific models.

So can we actually reevaluate the task and where

the gaps might be in terms of does natural language really

get us to where we need to be in specificity for things

like prediction, for longitudinal records,

for the language of health, which we can all agree

is different than the natural language of Reddit threads

on internet that some of these models are trained on?

And is there a way that we can take some of the lessons learned

from this phenomenal technology and apply them

with some domain expertise?

And I think that's what you've done.

And I know, Justin, you have some

of the graphic from the paper that you all put out.

But to me, this is like such a beautiful example of the NLP

to GPT 2 story in healthcare.

In other words, let's reconsider the problem

of the language, the tokenization of actually

the language of healthcare.

Let's think about the fact that things

happen on different timescales and consider

that right in the events in someone's progression

through a health system.

And then let's think about that prediction task that we've spent

decades in medical AI, trying to do prediction models.

And can we actually learn the language.

Hopefully, get economies of scale,

hopefully get the scaling laws to tell

us is the right direction.

And eventually, have a generalist model

that can do speak the language of health

truly, and then also predict events in ways that can help

us truly advance the field.

Anyway, I'm so excited about this work.

I don't know if you have other insights that you

got from building this model, but this

is such a powerful statement that just came out

of the meeting, really thinking critically, marrying the domain

expertise with the scaling laws and the things we've

learned from LLMs.

SETH HAYNE: I think there's a fun, historic digression here

as we get into this that's maybe worth highlighting.

Do either of you know where the name Epic came from?

Have I shared this?

So it actually, Matt, you're going

to appreciate this with your English background here.

It's actually the classic Greek poems.

And these long stories.

And in fact, when Judy initially developed the first capabilities

for Epic, which was the database that

managed the data around the patient,

she called it chronicles.

We still use that name today.

And so this idea of a patient's story

been inherent in the company from day one.

And the approach we took here was really

could we, using things like medical events, interventions,

observations, and I think a critical component here,

and you see it in gray, time that intervals

that have passed use that in a chronological order

to build out the story of a patient,

and then take that same transformer

architecture that we've been seeing in natural language

be so useful at predicting next word from a training perspective

and apply that here.

And as you noted, it's early.

At the same time, the data set we've

trained the largest model on so far

is about 8 billion encounters.

So it's not small.

But based on the scaling laws that we found in the paper,

there's a lot of opportunity to continue

to improve these models by growing them,

and then exposing them to more patient stories.

And the Cosmos community, this is an effort across health

systems that use the Epic software to build out

a large, de-identified data set for these purposes

is two to three times that size, depending on how you measure it.

So there's a lot of opportunity here.

I think the exciting thing, a couple of exciting things

that I'll just highlight, at least to me.

One, we have evidence that this model seems

to be able to not just predict near-term events that might be,

for example, what gets ordered or resulted during an upcoming

encounter, but also seems to be equally applicable out 3, 5,

10 years in regards to predictions, which can obviously

help in regards to managing chronic diseases,

but also things like some of the capacity problems

that health systems are facing right

now in regards to getting patients

beds and those sorts of things.

Happy to dive into other pieces.

I think the kind of visibility of the simulation of what

gets simulated, those future trajectories

will be interesting to explore from a UX perspective going

forward as well, but we're just, I think,

scratching the surface as a community

on how these new modalities can be used.

I know I threw a lot out there.

Sorry.

I'm excited about-- thank you, Matt, for the kind words,

but it's been great to work with the team at Yale.

Andrew and Daniela, who's on the Cosmos governing council,

as well as some of your colleagues at Microsoft

to get this initial research out.

JUSTIN NORTON: It is amazing.

And it fits in--

it fits in this world of healthcare is different.

It's a different language that we're talking about.

And it's amazing just to see the push on the research side.

I'm curious, Seth, as you brought up

the other pieces on UX, how physicians, patients, others

will start to interact with AI systems.

There was another paper, and I actually

bring up one more chart that's actually

gotten a lot of attention over the past few weeks, which

is around, de-skilling de-skilling of users with AI.

And, Matt, I don't know if you have any more comments you want

to talk here as the practicing physician

between the three of us.

But at large, I felt like I saw this,

I was like, yeah, it makes sense.

You have a physician rely on AI for imaging.

You take that away.

By the way, maybe they weren't as good as it before.

You talk about the calculator example.

You want to push all of us on mental math

with three-digit numbers.

If you don't use it, you are going to lose it.

And so if you have any comments, and then staff starting to think

through, how are you thinking about this?

How are you thinking about UI/UX as you're

thinking about development out now at Epic?

MATT LUNGREN: I think this brings up a couple concepts.

I think-- I mean, this has been true in healthcare.

So I guess the point of the paper, which I think

is valid, which is that physicians

are using this AI for colonoscopy,

for polyp detection.

It's quite good.

As it shows that they do potentially better

with the software.

And then you take the software away,

and you have them do the same tasks, and they do less good.

And it seems like it's close, but you can see

that this makes perfect sense.

Everyone has stopped using maps to get around.

I'm old enough to have had to do that before GPS,

and I just use GPS.

And then there's medical examples.

I haven't tapped out a patient's pleural effusion

since a med school physical exam rotation.

Because why we have X-rays, and we have ultrasound,

and we have all kinds-- so there are lots of examples.

At the same time, I mean, it does raise a question.

I think there's-- the two points that I think that we still

haven't solved.

One is what are the acceptable--

what are acceptable areas in healthcare

that we're willing to be OK with the de-skilling?

Are there just things that we're just going to say we cede to--

the technologies where it needs to be maturity wise,

we can use it reliably, this is the way we do things now?

But the other part of that which I think is still unsolved

and it's related, is the human computer interaction or probably

some new term, probably human AI interaction.

What is that ultimate combo?

And how do you have to display it to me

so that we're better together and we've

had this narrative just in through the several episodes,

which is I need to see that human plus AI is better

than either alone?

And if not, a system where which I

think there's been some papers on now,

how do we divvy up the work between the human and the AI

so that, again, in aggregate, they're better?

And those are just unsolved but very important questions.

And the papers like this are starting

to tease at that a bit more.

SETH HAYNE: I think you're highlighting this sort of--

if I look forward 20 years--

I guess I'm curious.

I think I imagine we're all in agreement on this.

But if not, we should have that discussion.

But I feel if we look forward 20 years, the quality of care,

we'll be able to provide--

I think, is going to be unequivocally better.

I think we're going to have more real world

evidence at the point of care.

It'll be more personalized.

There will be a new set of tools that folks

will be able to use and have at their disposal,

in addition to what they have today to do that.

If either of you disagree, I mean--

MATT LUNGREN: Well, no.

Well, the only thing I disagree with is that we never--

on this show, we only go--

we only go as far as five years out because--

no, I'm just kidding.

SETH HAYNE: No, I agree.

It's hard to--

I agree, generally speaking.

And we could debate the time frame.

I don't have a clear definition there.

But I think that the question in my mind

is how do we move through this interim state?

There is things we need to learn.

And frankly, I feel this in the same way in regards

to the developers I work with at the company.

There's a similar question in regards to vibe coding tools

and experiences.

And I see developers now walk in the door, where

it would have been a PowerPoint presentation in regards

to UX experience.

They now have a functioning prototype

in the same amount of time.

And frankly, it looks better than their poor PowerPoint

skills did before.

And we're better able to understand it.

Now, Not everybody's up to that state yet.

And I think there's this interim piece we need to move through.

And I think it applies equally to medicine

as it does to development as an example.

JUSTIN NORTON: Yeah, the vibe coding

is such an interesting example.

All of our MBAs, and I still teach a bunch,

they are now expected-- instead of slides,

they are now expected to show up with working prototypes,

working demos.

It's just a different way.

And it's so much more powerful to communicate that.

When you can say not just, hey, here's the concept,

and these are the three bullet points I want.

No, you can touch and feel, this is what we're working at.

Granted, very, very different than functioning software

in the real world for all the kind of reasons

we've talked about, but the communication

is totally different.

And then, Seth, to your point on-- and yeah, 20 years

is far too long.

And I think that's where-- that is where it does get really

interesting, though, is really at what

time scale are these changes going to start to happen?

And the fascinating thing we've been discussing

is there's this balance between patients are going now.

Again, for better or worse, patients

are going now to these consumer tools.

And for all of us who work closely

with health systems, who think about this on the other side,

how do we, as a field, feel about balancing

these imperfect tools to have people keep up,

or educate, or change how patients are interacting

with it?

And that's just this separate debate.

Because I think you're--

I think your 20-year answer is easy.

I think the answer is undoubtedly yes.

It's just is it fixed 12 months, 24 months, five years.

To Matt's point, where how do we navigate as a field through it?

SETH HAYNE: Well, and this is--

so last week was our user group meeting.

And here in Verona and in Wisconsin,

one of the things we spent time talking about

was a collection of work we're doing, both in my chart,

the patient-facing application.

Most folks just have it on their iPhone.

They're doing things like getting lab results

or scheduling an appointment, those sorts of things,

checking on their kids growth charts, that thing in it.

And then the physician-facing experience

in the office, as an example.

And part of what we were talking about there

was that there's an opportunity in an integrated set of software

to be able to answer those consumer questions that

are starting to be asked, but grounded both in one,

their medical record, their family history, that information

that is in the chart.

So you can have more confidence in that piece.

And then secondly, and I think this is--

I think this may be even more important, bluntly,

the opportunity to hand off from that experience to the care

team, to engage a nurse, to engage a provider,

and to escalate that conversation to somebody

else that can intervene if necessary in that context,

and be able to do so seamlessly and go back and forth.

And we tried to highlight the importance of that.

I think of it as two sides of the same coin.

One side is facing the patient, and the other side

is facing the care team.

And there needs to be a consistent dialogue

between those two with grounding in this rich history

of the medical record.

MATT LUNGREN: I love this so much.

And just, again, there's always that classic thing

in healthcare.

But as our time to see our patients and discuss their care

shrinks and the pressure on-- all the pressures around that,

but there's always that famous--

I mean, everybody's gone through med school has this,

like Marcus Welby, kind of old, old school doc.

That'll tell you, the only question the patient

actually came in to ask you will be

asked once your hand touches the door to the next appointment.

And it's often the most gnarly, complicated question.

And it's just-- it's without fail.

And to me, when you're talking about someone who's

both getting more educated about their own data

and they're asking the right questions,

but then they're teeing that up for the--

I mean, that's such a beautiful connection to me.

Because now, I have a way better understanding

of what are the things I really need to deal with in this visit

that I-- without that information,

I'm just going to assume refill a couple meds

and make sure that you're up on your screening,

and then all of a sudden it actually ends up

being this whole other thing.

But that communication is so key.

And I think that we talk about this a lot too.

But this idea of information asymmetry between the physician

and the patient, that goes both directions, by the way.

Yes, physicians are knowledgeable about their field

and the things that they do, and the patients

may not be as up to speed.

But on the other side, I am at a disadvantage about all the data

that I have about you as a patient

and all the things that are troubling you when you're not

in my office.

Like that-- can we level the playing field on both sides?

I think that's a phenomenal future vision.

SETH HAYNE: For any listeners that

happened to be at Tuesday morning of our user group

meeting and just heard what Matt said,

he did not share that idea before what you saw on Tuesday

happened.

We actually-- what we shared was a MyChart experience,

where the conversation with the patient through MyChart

pre-visit build out a visit agenda, including

double checking for that last hand on the doorknob moment,

and then the screen flipped.

And they were in the exam room.

And the visit-- part of what we're building towards

is a real time experience driven by the conversation in the exam

room that then walks through that visit agenda

and brings insights, including real world evidence from Cosmos

into the exam room based on that conversation in the exam room

and that kind of pre-visit experience

that already teed up the visit agenda exactly

to that point, Matt.

It's one example, but I think it helps highlight how there's so

many new modalities that can come into play

with these different models when put into workflow,

and a user experience that makes sense, and connects

the patient and the care team.

MATT LUNGREN: I love it.

And I will confirm that.

I unfortunately did not get to see the presentations there.

But to me, these are the kinds of things,

though, that I think if you grabbed, Dr. Smith

off the street in anywhere USA.

They'd be like, oh, that seems like it's so far away, honestly.

And this kind of goes back to just the general theme,

which is that you almost have to just really say,

in a perfect world, a magic wand,

what would you want the perfect visit to look

like from both sides, and see how far really you are away

from that technologically.

And could you-- because I don't think, again,

with that overhang, that with that far away from being

able to do these things at scale, today

or in the shorter term, without having to say, sure guys,

but it's 5 to 10 years away, I think

it's much more near term than that.

SETH HAYNE: Well, it is.

It certainly is.

And I think an important aspect of this

is that the incremental steps to get there, each add

both value in regards to time save for physicians,

as well as a better experience for patients

and improve the quality of care.

So you can step into this type of experience

and incrementally get there.

Because there are-- some of these are more complicated.

And we need to continue to both drive down latency in regards

to how these model responses on a technical side

and other aspects in regards to doing this at scale.

But that works underway.

It's going to happen to your point.

And so there's a real chance to quickly step

into this type of experience.

JUSTIN NORTON: Well, I don't even know how we top this.

I mean, we just talked about an entirely new patient physician

interaction.

It's coming.

I guess the only debate we have are the difference of opinion

is when.

But I guess, Seth, one last-- was there anything else?

Obviously, we glossed over it.

But, Epic, one of the biggest kind of events.

People eyes on watching in healthcare.

Was there anything we missed that we should have spoken

about here as our listeners are thinking

about, the future of healthcare, AI, and what's coming next?

SETH HAYNE: I think one thing that's important

not to lose sight of, and we've hit multiple times, obviously,

on the importance of this patient-physician experience,

and helping everybody be better informed,

and efficient in that context at the same time.

There's also a lot of challenges facing the health systems.

And I think that is a societal question as well.

There are not enough physicians.

We have an aging demographic that

is going to continue to put increased pressure on society

in this regard.

And I think the types of things we're talking about

help there, help with access challenges, help with--

even if it's as simple as helping

answer billing questions in my chart with an agent.

I think we need to continue to keep a beat on.

And this was the third theme that we touched on at UGM,

making sure that continues to scale meaningfully as well.

It's maybe not as directly applicable

when Matt walks into the exam room.

And understandably so.

But I think it is something that we all

need to continue to keep a beat on together.

MATT LUNGREN: I love it.

And I think that--

I'm sure most of our listeners who are health--

have a healthcare background know this.

But for those who don't, the Epic UGM conference or event

is--

they give Apple worldwide developer--

I mean, it's literally like--

I mean, people lining up right--

I mean, this is a big moment.

So kudos to you all for pulling this community together.

I think there's a lot of productive conversations

that come out of just beyond what Epic specifically is doing.

And I think that the level of discourse

has been accelerated for, I think, a lot

having to do with the way that you are positioning

the technology in very much a healthcare landscape, think

about each of the stakeholders and where

we should be headed next.

So thank you for all the great work, Seth, and really thanks.

Such a pleasure to have you on.

SETH HAYNE: Oh, I've enjoyed it.

Thanks for inviting me on, and we'll

have to do it again sometime.

MATT LUNGREN: I want you to come with your turtleneck

next time with the holding up the next big device or something

next time you're on.

That's all I ask.

SETH HAYNE: Something to do, something to aim for.