JUSTIN NORDEN: Welcome back to the Stanford Healthcare AI

podcast.

We're so excited to be joined by Kimberly Powell, who's

been leading healthcare efforts at NVIDIA for almost

the last two decades.

We'll get into a lot more of what that means in a second.

But unless you're living under a rock,

NVIDIA is the largest company ever today, over $4 trillion

in market cap.

And we're just so thrilled to have Kimberly joining us today.

KIMBERLY POWELL: Thanks so much for having me, Justin and Matt.

I'm excited to be here.

JUSTIN NORDEN: I'll signpost a few

of the topics we'll go through.

We'll talk through some of the recent AI developments.

We'll talk through a few of the different things

we're seeing around the world on chat with EHR.

And then excited to get into some

of the robotics angle, which we have yet

to talk about whatsoever.

But I wanted to start with this familiar topic around search

and how people are starting to use these tools.

Matt, can you tell us what we're looking at here?

MATT LUNGREN: Well, yeah, I mean, it's interesting.

I think a lot of my clinician colleagues

are also seeing similar trends where, again, there's

that-- and we've talked about this on the show before,

but this is one of those threads that I think

will continue to come up, which is, how are our patients using

the models?

And it's almost like you're starting

to see a lot more sophistication on the part of the patient

in terms of having had a conversation with one

of the frontier models and their health data,

or their diagnosis or their treatment or whatever.

And it's actually been, in some cases,

really refreshing feeling like I can

have a conversation with my patient

and they have a lot more background

and frankly, have been spending a lot of time

working with the models, explaining some of the concepts.

But at the same time, I think this is a good way

to look at maybe that trend from a different perspective,

because I think historically it was always--

you go to a search engine, you start Q&Aing that,

then you go through a bunch of links,

you go down some rabbit hole.

And then I feel like a lot of times, at least

historically in the pre-GPT days,

you're spending a lot of time in the clinic visits

explaining away some of those problematic rabbit holes

because there's no model behind it.

And maybe they didn't have the context or maybe even

the literacy, frankly, to navigate

some of these really complicated sites

or informational resources, et cetera.

And so I guess this trend--

listen, is this definitely because of the models?

Hard to say.

It's kind of an interesting coincidence if it isn't.

I think more and more folks--

I don't know if Kim--

I mean, I asked this all the time to my friends too,

are you putting your data in?

Are you putting your loved ones?

Because I literally do that with my own health data and it's

fascinating.

And I learned some things, too.

KIMBERLY POWELL: Yeah.

I mean, listen, it has some similarities.

I watched the show where you guys

talked a lot about when you had the power of the internet,

both as patient and clinician, how it had to change

the way that you work.

Talk the patients out of all the things that they read.

What is so different now, though,

with these large language models is the introduction

of reasoning.

It's not just the memorization of anything

that might have happened with some adverse effect of a drug.

It's listening to you.

It's remembering things about you.

In fact, just this weekend, in our own family situation,

it asks you questions and says like, oh, well, you

were up in Maine.

Is it possible that you were bit by a tick

and that maybe you should screen for Lyme disease?

So the fact that it can reason, the fact

that it has memory and context about you as a patient,

I think that's what's going to be kind

of game changing about this.

So hopefully the models are going to help not have to do

so much explaining away and actually empower the patient

a lot to just all that context that we take for granted

as patients, what is such powerful information

for doctors, but you can't always remember it when

you're in that 30-minute visit and you're feeling nervous

and anxious and all these other things,

chat can help maintain all of that memory for you.

So I think it's going to be really stunning,

a stunning experience.

JUSTIN NORDEN: Totally agree.

And I think what's also fascinating is

when you compare this--

and obviously there's many, many things going on here,

and it's hard to isolate one thing.

But when you compare the health queries to everything else,

the health queries are the most down.

And that's fascinating to me.

And we were actually laughing before we came on.

I still use the Google search like, oh, my gosh,

what a Luddite.

I'm not using technology for everything.

But I do use search for something still.

I know I shouldn't, but I'm used to it.

But for health queries, I'll never touch Google first.

It will definitely go to a model to go give more context,

ask more questions, go down a rabbit hole.

And so it does feel different, at least for me personally,

from a health care standpoint compared

to other things that I haven't fully changed my workflow yet

on.

But for health questions, it's by far the first place for me

to go.

KIMBERLY POWELL: Absolutely.

MATT LUNGREN: Yeah.

And I think part of it, too, is what

I would like to see hopefully in the future is a correlation also

to another chart that shows interest in preventative health.

Routine screenings are up.

I would love to see that this lead to a carryover effect

to potentially, I guess, health literacy

writ large in this democratization of, I guess,

the care management journey, because ultimately we

can't really experience any of the true outcomes

until this causes behavior change.

And as we know, that's hard to do in a 15-minute clinic visit

or whatever.

And you often wait for these episodic crises of care

as opposed to, hey, I'd love to just talk about--

let's plan out the next six months, two years, whatever.

I'm optimistic though, that just that literacy and education

will lead to some of those other behaviors.

And I guess we'll have to stay tuned to find out.

JUSTIN NORDEN: So I mean, actually, Kim, you

brought up context and asking for more information.

And were you traveling in Maine has

kind of laughing, thinking through-- this

is like asking our infectious disease consult where these are

the clinicians that spend more time with the patient,

go far deeper, and pull all these things together.

One of the interesting things we're also

starting to see now in a few different places

is people trying to bring more context

to models working in systems.

And so the examples I'm starting to touch on

are these different with EHR pieces

where, hey, rather than having the patient or the clinician

have to feed everything to that subsequent model,

can we bring some of this data almost directly to the models

to query in different ways?

And so there's a few studies that have come out here

over the past.

And actually, Matt, yeah, what did you want to highlight here?

MATT LUNGREN: Well, I think this is

that chat EHR that I think kind of made quite a few headlines

from the work out of Stanford.

But the basic gist of it is systems are setting up--

well, so first of all, systems are

seeing the trend where clinicians

and other folks in the healthcare workforce

are realizing that there's some real advantages to using

the models for the things that they do.

That's obviously not necessarily a compliant--

there's a compliance risk.

There's all the things we've talked about in prior episodes

around that problem of folks just using the OpenAPI or web

app on their phone, for example, in the health setting.

So then health systems are then setting up alternative ways

to access frontier models that are more of safe, compliant,

secure.

But that still requires cutting and pasting,

and to your point, Justin, like bringing the context over

from the records or whatever or having to put more information.

Now it's like, well, why don't we just connect it--

and in a lot of cases, it's a SMART on FHIR,

just connected directly into EHR,

pull in the data that's relevant.

And I can ask summarization questions or maybe best tests

or potentially reference guidelines

or whatever those things are.

I think that it does raise an interesting point

because there's a couple articles-- well,

the one on the right is from a group out of Europe

where there's additional regulatory considerations that,

in this case, required them to leverage local models.

And I think we are seeing, when we

talk about the prior frontier and the capabilities

of the models, that models are getting smaller,

models are getting faster, models are getting cheaper,

and models are increasingly run in local environments.

And so Kim, I mean, clearly NVIDIA

is leading the world in so many areas,

but I think one of the big ones is efficient

running of frontier models, distillation of the models,

on edge deployments, and all the advantages there, too,

that I think--

particularly in healthcare, where there's so much

concern around privacy and obviously restrictions

around use.

This could be a really powerful use case.

I don't know if you're seeing a similar thing where folks

are like, hey, I'd love to deploy it locally in this area,

but I need the horsepower to run it

or I need to connect it a certain way.

I don't know if you have a sense of that trend,

but it's definitely something we're seeing.

KIMBERLY POWELL: Yeah.

I mean, we're interested to enable these models to run

everywhere you can imagine.

I mean, you want it to run inside your car.

You would want it to run inside your operating room, right,

Matt?

You don't want to have to ask a nurse or yourself disengage

a surgery ever to maybe pull up a prior piece of information,

whether that be an image that you took of the patient

prior to the surgery or any potential comorbidities

and things that might get tricky when your surgery is getting

more lengthy.

You just want to be able to talk to something that's

listening ambiently and provide you

that very relevant, very important real-time information.

So having models be able to just live on

edge it's a necessary condition.

But it's just like, models run in the car,

but you're going to phone home to more capable models,

potentially, that need to do other type of tasks

that run in the cloud.

And so we absolutely see a complete hybrid.

And being an accelerated computing company,

we're just always been passionate, focused,

and taking the most important computing technology

and fitting it into the tiniest compute footprint, one,

so it's accessible, two, so it can, as you say,

run on the edge.

Some of the most important and I would say,

risky applications are these edge applications

because there's usually life at risk.

And if it's life at risk, you need

to have the models to be so close to the humans

that it's interacting with.

So I think there's reason that we have to be really focused

on taking this amazing capability

and putting it in tiny packages and running it everywhere.

JUSTIN NORDEN: Yeah, I totally agree.

And it's interesting, thinking about autonomous cars,

other places where really life is at risk.

Things really have moved to the edge.

If I look back though at where we're at in healthcare today,

I would say, we're early.

I mean, one, we don't have autonomous surgeries going on

or anything like that yet.

Most of our problems we have stop.

So difference just for people.

If you're driving a car, you can't just slam on the brake

if you lose internet connection.

That's real bad.

For a chatbot with EHR function, yes, it's not good

if you don't have that data, but you're not

in the middle hopefully of an operation or something where you

were depending on this data.

I guess, Kim, where do you see this starting?

Because if I grade us right now, we're still very, very early.

So where do you see--

because you're going to see the literal frontier of this.

Where do you see starting of things

getting pulled to the edge and moving that way in healthcare?

KIMBERLY POWELL: Yeah.

And let's think about what kind of agents

there are or use of utility of agents in healthcare.

We think of digital agents, they could be vision agents,

they could be ambient listening agents,

they could be information retrieval agents.

And as you say it, we're connected with--

it's approaching 5,000 AI startup companies in healthcare

alone here at NVIDIA.

So we do see quite a bit.

But vision, as Fei-Fei at Stanford pioneered for us

or helped pioneer for the whole world, vision is a huge aspect.

And I think we're not yet taking full advantage of it

in the health space, but it's going

to be an absolute necessary condition.

And then the ability now with being

able to do speech recognition, we're

doing a crazy amount of awesome research

with fantastic companies like Abridge,

where you need to know exactly who's speaking, who they are,

the context of that person, because you're

digitizing that language and you're going to turn it

into some kind of action.

These tokens of speech turn into some kind of task

that we have to end up going to do.

So you need to know if the patient said that,

if the nurse said that, if the doctor said that.

And so all of that has to be really, really looked after.

But I think that vision systems are coming into play now

because there's a lot of fantastic efficiency

that you can have just by being able to track things

in the hospital or timing of surgeries,

or being able to see the flow of patients within hospitals.

And then connecting a lot of that vision,

now you know that the trend of vision language models

that have reasoning, in fact--

this is the vision systems of the past where

they might have been a little tricky for nursing stations

because they would have overalerted perhaps

and caused actually potentially more work on the staff than you

would want.

I believe that these reasoning systems and chain of thought,

where you can really attach the agents to how that hospital runs

in very particular ways, we're going

to be able to really reduce some of that overhead that

came with the first generation of vision systems, for example.

And then digital agents that then combine into, as I said--

we were talking very early on why are

the health searches going down.

Because you can just have a conversation with your phone

now as a patient and you're getting

prompted and things like that.

And so the fact that we can now do that ambient listening

in a health setting--

oh, my gosh.

I get so excited about the possibilities

because it's not just the thing that was written down.

It's all that other context that's

being digitized, both by the doctors and the nurses exchange

and the patient exchange.

All of that can facilitate just a massive amount of opportunity.

So vision systems, ambient listening systems, and then

the ability to attach them to what is otherwise really messy

healthcare IT systems.

We're in a whole new category of possibilities now.

World is the oyster.

And that's why we see just the startup community and the AI

applications that are really domain-specific, really

vertical for health being wildly successful.

100 to 200 million annual recurring revenue startup

companies in the matter of 12 to 24 months

is absolutely unheard of in this industry.

They used to be fraught with barriers

to entry, whether that was computational footprint,

whether that was safety considerations,

whether that was the ability for clinicians or nurses

to adopt it.

Here the adoption is human language, speaking to a phone.

They're all familiar with it.

There's no learning a new UI.

The UI is speech or the UI is a camera just watching over me.

So the barriers are really coming down super fast.

That's what's super exciting.

MATT LUNGREN: Yeah.

I mean, this is so true in terms of the multimodal aspect

because we always say there's so much that you can do.

We've shown this in prior episodes of different data

around the performance of these models

and traditional text-based tasks.

And there's a lot still to do.

But when you get into multimodal, to your point,

that unlocks an entirely new surface area.

There's so much more information that we process just

as human clinicians with our eyes and our senses

and we know that there's information

being transmitted that may not be captured in language.

And do we have the infrastructure

of the low latency systems?

I think we're getting there in terms

of the computational capabilities.

But then also, to use the voice example

just as a different modality, just on the voice biomarkers,

we're seeing startups that can take just the audio data

and have a very high prediction rate

for neurodegenerative disease or depression or other things.

And it just adds these layers of more opportunity for us.

And I think some of it you might ask an old timer clinician who

could probably walk in a room and tell

within two minutes what's going on maybe

before anyone speaks a word.

Because there's something, to your point about the vision

aspect, to the sound that--

anyway, I'm just as optimistic as you are that we're

scratching the surface.

And then I think on your point about the alerts,

I think to me this is where that abstraction layer with an agent

that understands context you can interact with it almost like,

quote unquote, "like a human," but can feed or use those tools

as independent systems in the context of the situation

and help filter out some of the potentially false positives

or alert fatigue that ends up going on when you have,

like you said, dozens of these running and they're just like

being triggered by a simple event as opposed to a contextual

awareness.

I think there's a lot of work there

that I'm seeing that's really exciting.

Again, that taps into the idea of agents.

But you all have really thought a lot about agents.

And I think what I like about the approach NVIDIA has taken

is it doesn't just stop with the digital world.

You really are pushing into physical world, I think.

And I don't really know of another tech

company that's thought quite as much about the robotic space.

And then, by the way, can we create a developer ecosystem

where folks can build on that no matter what the vertical?

But obviously, I'm biased.

I think healthcare robotics is going to explode

in the next five years.

And I don't know what you're seeing,

but I just feel like there's so much opportunity

to marry the context, the multimodal with now interactions

in the physical world.

And that could be as simple as supply chain and back office

type things, but it also be literally in the operating room.

I don't know where this is going to go.

I don't know if you have a sense,

but it seems really exciting.

KIMBERLY POWELL: Yeah, we're more than excited about it.

I mean, a lot of the stuff that we were just talking about,

you could just go so far as to say every hospital

is going to be a robot.

It is a 3D.

It's a 3D space.

It's a physical space.

You have to understand the 3D world to do all of the things

that you just described, to really have the appropriate

context.

The distance between things matters, the speed at which

things are traveling matters.

The size and volume of things matters.

And so having that physical understanding of the world

is what will enable the future of what

I'd say an embodiment of an AI hospital.

And it's going to be a necessary condition in the future state

of surgical robotics.

If you think about this as a multi-scale problem,

on the one hand, you want to have your hospital be

a robot so that efficiency-- like you

were talking about, a lot of back office

and a lot of efficiency where obviously

is absolutely necessary.

We're tens of millions of healthcare professionals

short of the demand of healthcare.

So we have to do something, even at the hospital environment

level, to offload any kind of non-clinical work.

It's just a necessary condition for us

to be able to serve the population.

But then you get into, OK, like we were talking

about in the doctor's office.

That office is kind of a robot in itself

listening, observing, capturing.

And then moving into a more intense environment

where you have to actually receive treatment.

That could be the patient room where you might have

lots of monitoring systems.

And then all the way into where it's very, very, very

complicated into the operating room.

So the operating room itself is a robot.

The surgical device and all of the other devices surrounding

that patient is going to be robotic.

And then the human, we are going to want

to simulate every individual patient's anatomy and understand

its essential physicality and functionality.

And that is essentially living in a simulation environment.

So at the atomic scale so that we can

teach these robots how to work.

And so we're super excited about the physical AI space.

We've created what we call the three-computer platform, which

is truly accelerating this.

There's a reason why robots kind of were steady state in not

being able to be super useful.

And that's because we didn't have the third computer, which

is essentially simulation and understanding

the physical world.

And all physical things in the future

will be born in a computer first.

You won't make the robot then try

to teach it in the real world because as you can see,

it takes decades to try to generate enough training data

or write enough rules in software

for them to be robust enough to be in our physical environment.

But once you can introduce this new concept of physical AI,

which there are these new class of models called

world foundation models--

again, Fei-Fei Li and the team at Stanford doing amazing

work there.

NVIDIA doing amazing work with our Cosmos models.

You can essentially create millions, billions,

trillions, if you like, scenarios, in which case

you can start to train these otherwise physical things.

Could be a surgical robot, could be a cobot.

It could be just any kind of medical device

that will be working with patients.

And so that third computer has been the missing link.

And it is here now with world foundation models

and the ability to do very physically accurate

digital twin environments.

So synthetically generating information and then

being able to physically accurately

represent it in a computer such that it

obeys the laws of physics.

So we can train robots in very real time.

And everything from every medical device,

whether that's an ultrasound machine

is going to be autonomous in the future.

You're going to be entering autonomous X-ray rooms

and diagnostic suites, diagnostic imaging suites all

the way through to--

in fact, the very first, [CLEARS THROAT] excuse me,

FDA approved autonomous robotic task just came through

at our GTC conference with a company called Moon Surgical.

This is a surgical assistant robot and they pioneered--

by watching the tools, the scope actually follows and moves.

Otherwise, a surgical technician, a surgeon

is talking to a technician to do that motion.

And it's on obviously robotic arms,

so it has just extreme precision.

And that's the first time through AI vision systems

that they've actually allowed the robot

to actuate in a surgical environment.

So we're moving now.

We're cooking in this area.

And I think we're going to see a lot of very, very

rapid advancements in the operating

room, in the patient room, and I would

say in the hospital as a whole.

MATT LUNGREN: Yeah, so much there I'm excited about.

I think that the idea of digital twins-- and I

honestly should do a tip of the cap to another space

that we don't talk about as much on the show, which

is manipulating the chemical environment,

the molecular environment for drug discovery,

which is a whole other topic altogether.

But to your point, this simulation idea

is fascinating to me because again, we

encounter all these edge cases all the time.

In fact, that's how you can separate the good proceduralists

from the inexperienced proceduralists.

It's having that muscle memory, seeing the edge cases,

being aware of that and having that assistant the whole time.

I've been pushing my team here occasionally--

I haven't gotten any takers, but to literally do

like-- you remember, Tesla had the Dojo

where they would simulate millions of different outcomes.

I always thought we could have a hospital Dojo and we just let

everything play out and just see what happens in all these

different.

And my running joke, my dad joke version of this

is that in 99 out of 100 scenarios,

the clinicians rise up against the administrators

and take over the hospital.

I say that with tongue in cheek.

But I do feel like there's an opportunity

to take real-world data, possibly physical environments

and to your point, simulate this with the world model

and probably come up with the optimal care.

That is like the quintessential precision medicine.

KIMBERLY POWELL: Yeah, absolutely.

And this is well within reach, Matt.

At our GPU technology conference this year,

we announced a platform called Isaac for Healthcare.

And this is all of the tools necessary.

It's digital assets that are already

pre-made for you, hospital beds, operating tables, lamps

that you can start to build out these digital twins.

We're working feverishly, and we were just

at the Society of Robotic Surgery conference

announcing a whole bunch of new technology and workflows

for telesurgery, for being able to build an autonomous

ultrasound where you can use just a commercial off the shelf

arm and attach a mobile transducer to it.

You can speak to this robotic arm

and say, please go scan the liver

and it automatically scans the liver,

doing the vision system in the background that's

seeing liver lesions.

And all of this are now developer tools

available to the community, all open source.

And so we can build that environment.

And then you can take these Cosmos models,

and also we have group models, to be

able to create those corner cases.

We're working with clinicians to say

they did experience those corner cases, sometimes dramatically

for themselves, and their patients.

So you can build that into these foundation models

and make sure that it captures that preexisting experience,

as well as go hallucinate and dream up

a crazy thing that hasn't, thank God, happened

but could be physically feasible.

And so I think it's going to be just really exciting to see

how fast it's going to advance.

Just like we're seeing in the digital world,

we've not seen technology enter the health space

at this pace before.

I think in the physical aspect of what we're talking about,

we're going to feel a similar experience.

JUSTIN NORDEN: Wow.

So we're moving.

Totally agree with you, Kim also on how fast software is running.

And we haven't seen adoption like this

to come more on the surgical side, other places.

But help ground us how we get there from your perspective.

Obviously, the three of us are some

of the most optimistic people around where AI is going to go,

what's coming next, thinking about what's possible.

When I ground this in, let's say,

the perspective of a hospital CEO today

as they're looking at these tools, what do they see?

They see the financial impacts coming

from the one big beautiful bill coming down

and looking at serious cuts to staff, other resources,

ways to make ends meet.

I would say the broader context historically

is most of those CEOs who got there today,

the right answer for the past two or three decades

has been to ignore technology.

Every new shiny thing that's come along,

more or less the right answer has been to ignore it.

Just wait.

Let someone else try it out first.

Don't get distracted by blockchain,

first generation of AI systems that came out a decade ago.

It's been the right answer to ignore it.

How do you as you have these conversations start

to bridge for those people?

This is where we're going.

I know where you're coming from, where you haven't done--

why should they do something now to start

interacting with these tools?

What is the way either of you hook them in

to engage now with these topics?

KIMBERLY POWELL: Yeah, I mean, I think

we've also we have the great benefit of ChatGPT

being an interface the whole world could understand.

That's a great benefit.

Now you're speaking to it.

First you were typing.

That's still a pain in the butt for doctors and nurses.

Now you can speak to it.

I mean, just last night, I moved houses

and I had to unscrew something and I couldn't figure out

what bit to use.

I turned on the camera and I'm like, which bit do I use.

And he's like the third row down, two into the right.

And it was perfect.

And so it empowers me to do crazy things like that.

So I think we have the great benefit that this is just

so accessible to every human.

And I honestly believe that if you're a CEO of a health system,

you're going to start feeling it's unethical for me

to not help empowering my existing, overworked,

burnt out, potentially leaving the workforce

healthcare professionals with these tools.

They are demanding it and they're expecting it.

So in a way, it's almost unethical.

And secondly, there's no way they're

going to get out of the red if they don't find other ways

to treat more patients.

They can't hire the people, they can't hire more healthcare

professionals.

So they've got to find other ways where technology is

going to fill this great gap.

So there's fortunate things that are at the advantage,

and there's unfortunate things that

are also pushing in this way.

And then, as we said, the software

as a service technology of the last two decades

was tremendously challenging.

It was $3 billion in three years of installation

and then three more years of potential change management

to adopt these technologies of the past.

Today, when you want to engage a company like Abridge,

who has just a stunning clinical, conversational AI

domain system that is excellent, you

download an app on your phone and you ask your patient,

do you mind if I use this?

There's absolutely no learning curve.

So the excuses to ignore technology

have just completely diminished.

And the existential crisis that the healthcare system

is finding itself in is unethical and must demand it.

And so that's still very much in this digital world.

But the moment we start to see all this ability where you can--

you're not going to be able to hire the person

to check your patients in.

It's going to be a digital kiosk.

And why not?

Why shouldn't it be?

You want to have your conversation with your doctor

recorded, both for the clinical outcome

and for the patient understanding,

and hopefully the preventative things

it's going to help and do.

You're going to want to have an agent tell you

what next appointment you need to make so that you can stay

on your course of care and hopefully reduce

the cost of the health care system long-term.

And so then we're going to be like, of course,

we're going to get more comfortable with cameras living

in the hospital, because it has all this contextual information

and it can catch and see things and do things

that wasn't possible before.

And I am a tech optimist here.

I am at NVIDIA for 20 years, but I am a patient and we all are.

And when I get a colonoscopy, I want an AI

looking at my colonoscopy.

When I get a mammogram, I want an AI looking at my mammogram.

If I'm having a surgery that Intuitive's da Vinci is expert

at, I'm going to want to go to a hospital that has that.

And so there's just so much push and pull

here that it's creating the conditions,

I think, where it just can't be ignored

and it shouldn't be ignored.

And the software developers of this era

really have everything they at their fingertips

to make the burden or the adoption of it

just supernatural.

And then wonderful companies that

are incumbents in this space like Epic,

they've opened the doors as well to allow for third party

integration so that there isn't this wall that you hit.

I might have a great technology, but I can't actually

get it to interface with the systems that

are the systems of record.

That is the operating system of the hospital.

And so that's a really important part,

is we have to recognize where these operating systems exist

and make sure that we're creating the conditions

where a whole ecosystem can thrive,

because that has been also some of the challenges in the past

is the interoperability, to use that fancy word the health care

industry loves of all of these different systems.

So that's how I think about it.

MATT LUNGREN: Yeah, I mean, that's

the trend that we're seeing.

It's the health literacy.

It's the ease of implementation, partially because of interop,

but partially because these systems are so--

I mean, they're accelerating the developers.

And then again, platforms, a lot of open source, a lot of folks

are converging there.

And then finally, I think the last two pieces to me

are, to your point, expectations on behalf of the patient,

because they're using the tools.

They have access like they've never had before.

And then the final piece, I think there's the--

I talked to health systems and they're

stunned at how many of their workforce

are interested or using the tools today.

But then also in the education side, the youngest doctors,

the newest generation, they're using it all the time, too.

And it's actually making them better

at learning the concepts, potentially even

patient interactions, having difficult conversations, which

is something we brought up before.

And this is probably the last chart we'll bring up.

This is from a paper that just looked at medical students

using an AI model that was meant to simulate a patient encounter,

and then how they ultimately did on their clinical tests,

like their evaluation.

Are they ready to have conversations

with patients, et cetera?

This is like one of those rites of passage

in your medical school training.

And on the bottom, you can see that the group that used AI

outperformed those that did not at all

or use the traditional curriculum.

But the other cool thing about this paper,

which, again, I think, Kim speaks to your idea of this

is democratizing.

Everyone has access to these.

The students, actually, 70 plus percent of them

weren't sure if they were working

with an AI or an instructor in some cases

with these chat interaction, which is like literally

the Turing test in medicine.

But I think to the point there's this familiarity,

there's these increasing competencies.

But as a learning tool-- and Justin knows this is one

of my soapboxes--

I don't think we're thinking about this

as often as we probably should.

At the ability to upskill folks in a variety of areas,

leveraging these models is not a question and answering machine,

but as a tutor.

I'm guilty of this too.

I've tried to do better at that.

And I think there's a lot more to come here

in terms of medical education getting transformed.

And then I think we've got a complete cycle where

things will continue to accelerate in the way

that you've outlined.

And I'm here for it.

KIMBERLY POWELL: It's awesome.

And I think the upskilling we have--

again, here we have tremendous opportunity.

I myself as a consumer, I was talking about how you can just

talk to your phone.

I mean, my whole way to work, I want

to learn about navigation systems for surgery.

And I'm thinking about the technology behind it.

And I just had a chat with Gemini all the way to work.

And I'm getting smarter on my way to work while I'm driving.

And I don't have to put myself in danger reading anything.

And it isn't a pre-recorded podcast

that could be two years old.

It's the here and now.

It's reading all the papers that recently came out.

It's giving me market intelligence.

It's who are the companies working in this space.

And so you can upskill on complicated areas like

navigation systems and surgery in a 20-minute ride to work.

And it put me on a whole other learning path as I did that.

And to me, it's a joyous experience

because you're not afraid to ask stupid questions.

It's totally OK.

And then you just keep getting better and better.

So I think you're right.

I mean, this generation is going to be Gen AI native, not just

AI native.

It's like generative AI native where they

can watch the chain of thought.

Even watching the chain of thought is stunning.

We have vision language models that

are doing chain of thought for a radiologist.

This is how radiologists are trained to think through when

they're doing their study.

And it's really interesting to watch that.

Or you're, like you say, a veteran in the field.

You can actually embody your train of thought

into your own reasoning model.

So it can work just like that for you.

It's going to be a tremendous time.

And some people ask me like, well, what

does success look like for you.

You and I'm like, for me or for NVIDIA?

I'm like, if for me, it's when I as a patient

feel every day that the health system has changed for me.

And the fact that I could go see a doctor

and they are using Abridge, it's changing and it's changing fast.

Or if I can go find that colonoscopy center that

has Medtronic's GI Genius platform, this is good.

But hopefully you're not going to have to seek it out.

I think you guys were talking about AGI and medicine

a couple episodes ago.

And it's like hopefully actually as a patient, your level of care

and your ability to be a good patient is just

going through the roof like we've never seen before.

So that'd what success will look like for me.

JUSTIN NORDEN: Yeah, that's amazing.

And I think what's scary but exciting as well

is while these changes are happening so fast,

we do have an opportunity to use literally some of the same tools

to re-educate both our patients and clinicians for keeping up

with it and where it's going.

I guess, like you said, talking with a model

instead of listening to a podcast will make Matt and I out

of a job here on what we're doing.

But that's OK.

KIMBERLY POWELL: I love them.

But it's a very interactive, as I said.

It's hard to interrupt a podcast with a silly question,

but that's what you do in a learning environment.

You might raise your hand and ask the professor something

that you needed clarity on.

And so I'm oftentimes pausing a podcast

because I heard something I didn't understand,

hitting up Perplexity and then coming back in.

So we need all of it but the ability

to consume it is what I think is going to change the upskilling

factor, because you can consume so much more, so much more

complex information at such a rapid pace.

So upskilling is going to take massive effect into all of this.

MATT LUNGREN: Well, I mean, one of the reasons

we did this podcast, Kim, is so we

could ask dumb questions of very smart people like you.

And so we really appreciate you coming on and sharing everything

that's going on with NVIDIA.

We're just thrilled.

Again, I think we see a similar future,

and we'll be keeping very close tabs on the work

that you all are doing, particularly as we start

to move into this robotic space, which I'm personally fascinated

about.

Again, Justin and I don't get a lot of opportunities

to talk about it.

And I've learned something just talking to you today about that.

So I'm going to go read more.

Maybe I'll talk with the model about it.

But yeah, robotics is to me the next frontier

and I'm here for it.

KIMBERLY POWELL: Yeah, Matt, Justin, thank you so much.

I love when you say ask stupid questions.

You guys know this space incredibly well.

I'm just a student of the space constantly,

hoping that we can take some of the most advanced technologies

and bring them into this domain where

I believe it's going to have the most profound impact.

So I'm excited about that.

And I would love for you guys to talk

about another frontier, as you were saying,

Matt, which is AI for science, AI for biological science,

AI for the physical sciences, and actually

how just AI is going to transform science.

In a lot of the deep, personalized medicine space,

it's very much science.

Oncology and science are deeply similar in the way

that they work.

So it will be a really neat--

autonomous labs with super intelligent agents

is a tremendous future that I'm also super excited about.

JUSTIN NORDEN: Amazing.

We'll have to have you back to go deeper.

Kim, thank you so much.

KIMBERLY POWELL: Thank you guys so much.

Appreciate the opportunity.

It was a pleasure.