# Robert Wachter & Eric Topol - Discuss a Giant Leap Book

An episode of [Ground Truths](https://invisiblenetwork.ai/podcast/1728526108), 4 Feb 2026, 46 min.

Audio: https://api.substack.com/feed/podcast/185326564/1b3afe780e240ba0c39d030576a67676.mp3

Get full access to Ground Truths at erictopol.substack.com/subscribe

## Transcript

[0:00] SPEAKER_01: Um,
[0:01] we're going to have a fun discussion today about the world of
[0:05] AI and medicine and
[0:07] SPEAKER_01: healthcare.
[0:09] And I'm just waiting for some more folks to get,
[0:12] be part of this before we start talking.
[0:14] SPEAKER_01: Let's give it another minute and then we'll get going.
[0:20] SPEAKER_01: And for everybody who came on properly, thanks especially for that.
[0:23] SPEAKER_02: That's great.
[0:32] SPEAKER_01: Okay, we're going to be talking about this book here, A Giant Leap.
[0:37] SPEAKER_01: It's by a friend,
[0:39] SPEAKER_01: Robert or Bob Wachter from UCSF,
[0:42] SPEAKER_01: who's a chairman of medicine,
[0:44] SPEAKER_01: author of Digital Doctor previously some years ago,
[0:48] SPEAKER_01: and certainly the person that started the term hospitalist.
[0:55] SPEAKER_01: And is very much cued into the world of how medicine's progressing.
[1:01] SPEAKER_01: So it's great that he has put it all out there in a book.
[1:04] SPEAKER_01: And also,
[1:05] SPEAKER_01: of course,
[1:05] recently in other writings like the January 19th New York Times
[1:11] and others we'll
[1:11] SPEAKER_01: get into.
[1:12] SPEAKER_01: So, Bob, welcome.
[1:13] SPEAKER_00: Thank you, Eric.
[1:14] SPEAKER_00: It's great to see you.
[1:15] And I always take claim to your success as a proud graduate of
[1:20] our residency program.
[1:21] SPEAKER_01: Right.
[1:22] SPEAKER_01: No, I'll always remember.
[1:24] The great times I had there and the amazing inspiration from my
[1:28] hero,
[1:28] Conor Chatterjee.
[1:30] SPEAKER_01: All right, so let's get into it.
[1:32] This book is really interesting,
[1:36] a challenge,
[1:36] because the field is moving so quickly.
[1:40] And so here you decide,
[1:41] I'm going to write a book about the current state of AI in
[1:44] healthcare.
[1:45] SPEAKER_01: but it could be different in a week or a month or an hour.
[1:51] SPEAKER_01: What were your thoughts about trying to pull this together,
[1:53] SPEAKER_01: knowing that it's such a dynamic and ephemeral field?
[1:58] SPEAKER_00: I thought that was incredibly helpful to worry about that.
[2:01] SPEAKER_00: And that if at the end of the day, I wrote a book that was
[2:05] SPEAKER_00: relevant when we're on GPT 5.2,
[2:07] SPEAKER_00: but no longer relevant when we're on GPT 5.4,
[2:10] SPEAKER_00: then I really would have screwed up.
[2:11] SPEAKER_00: So, what it did was it forced me to think big picture.
[2:15] SPEAKER_00: Like,
[2:15] what are the big questions that happen when a technology that
[2:19] is really amazingly
[2:21] good but not perfect enters a high-stakes field like medicine
[2:24] that's got payments
[2:25] SPEAKER_00: and regulations and incumbents and doctors and nurses and
[2:29] SPEAKER_00: And it turns out that there's a whole lot you can say,
[2:33] SPEAKER_00: and I think you need to understand,
[2:35] SPEAKER_00: to just think about what does this do to being a doctor,
[2:37] SPEAKER_00: being a patient,
[2:38] SPEAKER_00: being a health system that is not
[2:41] SPEAKER_00: tied directly to the absolute state of the technology that day.
[2:45] SPEAKER_00: You know,
[2:45] SPEAKER_00: that said,
[2:46] SPEAKER_00: I worried about it because,
[2:47] SPEAKER_00: you know,
[2:47] SPEAKER_00: if you think about the day GPT came out,
[2:50] SPEAKER_00: if I'd been writing a book on AI the day before,
[2:52] SPEAKER_00: it'd be like,
[2:53] SPEAKER_00: oh,
[2:53] SPEAKER_00: God,
[2:54] SPEAKER_00: I got to throw the whole thing out.
[2:55] SPEAKER_00: But I made a bet that over the course of two or three years,
[2:58] SPEAKER_00: maybe four or five,
[3:00] SPEAKER_00: that nothing quite that
[3:01] ground,
[3:02] that earth shattering will happen,
[3:03] that the improvements will be incremental.
[3:06] And a lot of the success of this will have less to do with
[3:09] technology and more to
[3:10] do with what happens when this technology hits a very
[3:13] complicated world of
[3:14] SPEAKER_00: healthcare systems and doctors and nurses.
[3:17] SPEAKER_01: Yeah,
[3:17] SPEAKER_01: I mean,
[3:18] SPEAKER_01: what you're bringing up,
[3:19] SPEAKER_01: I faced a similar challenge back when I wrote Deep Medicine in 2018.
[3:25] SPEAKER_01: And then,
[3:26] SPEAKER_01: before what you call a giant leap,
[3:27] there was already the medical image deep learning story that
[3:33] superhuman performance
[3:35] SPEAKER_01: over radiologists and pathologists.
[3:37] SPEAKER_01: And then, of course, we knew in the background that the early...
[3:43] transformer model work was being done by 2017 so that something
[3:46] big was going to
[3:47] SPEAKER_01: happen,
[3:47] SPEAKER_01: although the public didn't see it,
[3:49] SPEAKER_01: as you said,
[3:49] SPEAKER_01: until the end of November 2022,
[3:52] SPEAKER_01: three years plus now ago.
[3:55] SPEAKER_01: So it is challenging.
[3:56] SPEAKER_01: This field changes a lot.
[3:57] And I would submit at least that that phase of deep learning
[4:01] was a great precursor
[4:02] SPEAKER_01: in a lot of important things.
[4:04] SPEAKER_01: Certainly not as big as when the whole generative AI
[4:08] SPEAKER_01: Large language models came to pass.
[4:10] SPEAKER_01: Now, in the book, you've got some really great stories.
[4:14] SPEAKER_01: And I thought we could start with one on the good and the bad.
[4:18] SPEAKER_01: Okay.
[4:18] SPEAKER_01: Frame this conversation.
[4:21] The good one that was striking was when you were a visiting
[4:25] professor at Yale and
[4:27] SPEAKER_01: you saw this patient they presented to you.
[4:30] They're always trying to stump the expert with Walton's
[4:34] macroglobulinemia and liver.
[4:37] SPEAKER_01: Can you tell us about that?
[4:39] SPEAKER_00: Yeah, it was probably about two years ago, I guess.
[4:42] SPEAKER_00: I'm a visiting professor at Yale.
[4:43] SPEAKER_00: I'm at the Yale VA.
[4:45] SPEAKER_00: And they say, well, here's this patient.
[4:47] SPEAKER_00: We don't really know what's going on.
[4:48] SPEAKER_00: He's a 75-ish-year-old guy who's got Waldenstrom's macroglobulinemia.
[4:53] SPEAKER_00: It's sort of a very rare hematologic disease.
[4:55] He came in with what we thought was pneumonia,
[4:57] pulmonary infiltrates,
[4:58] and a white count.
[5:02] SPEAKER_00: And then he developed abnormal liver function tests in a certain pattern.
[5:07] SPEAKER_00: And they said, what do you think is going on?
[5:08] And I pulled my phone out of my pocket and I said exactly what
[5:12] I just said to you.
[5:13] SPEAKER_00: This is a 75-year-old guy with Waldenstrom's,
[5:14] SPEAKER_00: comes with a fever,
[5:15] SPEAKER_00: white count,
[5:17] infiltrates in his X-ray and develops abnormal LFTs in the
[5:19] pattern of high
[5:20] SPEAKER_00: transaminases and normal bilirubin alfos.
[5:23] SPEAKER_00: Now,
[5:23] that is the kind of thing that if you said that to UpToDate or
[5:26] Harrison's,
[5:27] SPEAKER_00: it would sputter.
[5:28] SPEAKER_00: It was like, what are you talking about?
[5:30] SPEAKER_00: You could not give it a clinical case of medicine.
[5:34] SPEAKER_00: And it listened and then it spoke
[5:36] SPEAKER_00: and said,
[5:37] SPEAKER_00: well,
[5:37] SPEAKER_00: I think this is going to turn out to be drug-induced liver injury,
[5:40] SPEAKER_00: and here's why,
[5:41] SPEAKER_00: da-da-da.
[5:42] SPEAKER_00: And then I said,
[5:42] SPEAKER_00: well,
[5:42] SPEAKER_00: that's what we thought too,
[5:43] SPEAKER_00: but it began only six hours after the first dose of Zosin.
[5:47] SPEAKER_00: And it basically said,
[5:48] SPEAKER_00: well,
[5:48] SPEAKER_00: that's kind of weird,
[5:49] SPEAKER_00: but it can happen,
[5:50] SPEAKER_00: and you probably should send off tests looking for hepatitis.
[5:54] SPEAKER_00: And then I said, can you write a poem about drug-induced liver injury?
[5:57] And it wrote this incredible poem about,
[5:59] you know,
[6:00] and I did that just as a party trick.
[6:01] SPEAKER_00: I thought that would impress the trainees.
[6:04] SPEAKER_00: It was really one of the first moments that I came to say,
[6:08] SPEAKER_00: this thing in my pocket is now a curbside consult.
[6:12] SPEAKER_00: That, you know, when I'm on the wards, I'm a generalist.
[6:15] SPEAKER_00: And so,
[6:15] SPEAKER_00: as I say to people,
[6:16] SPEAKER_00: if you want to be a hospitalist,
[6:17] SPEAKER_00: you have to be comfortable with any patient who comes in the hospital.
[6:20] There are five people in the building who know more about every
[6:22] problem than I do.
[6:24] SPEAKER_00: And I know more about the whole big picture than they do.
[6:26] SPEAKER_00: But it means I'm constantly learning.
[6:28] I'm constantly trying to track you down in the cafeteria
[6:30] because I have a heart questioner.
[6:32] Now,
[6:32] I think what we have is a subspecialty-level intelligence in
[6:36] our pockets.
[6:37] And that became kind of an epiphany that what happens when you
[6:41] have a digital tool
[6:42] SPEAKER_00: that can basically scale subspecialty-level knowledge for a generalist?
[6:46] SPEAKER_00: And the answer is it's pretty magical.
[6:49] SPEAKER_00: It can be.
[6:49] SPEAKER_00: It can be.
[6:50] SPEAKER_00: Yeah.
[6:52] SPEAKER_00: Assuming it gets it right.
[6:55] And that was two years ago at a time that these tools sometimes
[6:59] hallucinated and
[7:00] SPEAKER_00: came up with crazy stuff.
[7:02] And as Ethan Mollick likes to say,
[7:04] the AI you're using today is the worst AI you'll ever see.
[7:07] Even two years later,
[7:09] it's substantially better than it was and doesn't hallucinate
[7:12] very often.
[7:13] SPEAKER_00: So is it trustworthy?
[7:14] SPEAKER_00: Does it replace the ID doctor?
[7:16] SPEAKER_00: No, not completely.
[7:18] SPEAKER_00: But I can't call an ID doctor every time I have an ID question,
[7:20] SPEAKER_00: and I can pull the phone out of my pocket and go to open evidence.
[7:23] SPEAKER_01: Sure, sure.
[7:25] Now,
[7:27] the other side of this was another short anecdote,
[7:29] but I thought it was really interesting.
[7:31] SPEAKER_01: Your colleague, Sarah Murray...
[7:33] with putting together a preauthorization to justify the use of
[7:38] the blood thinner
[7:40] SPEAKER_01: Eliquis,
[7:40] SPEAKER_01: Apixaban,
[7:41] SPEAKER_01: for insomnia.
[7:43] SPEAKER_01: Can you tell us about that one?
[7:44] SPEAKER_00: Yeah, that was actually the first time I ever heard about hallucinations.
[7:47] SPEAKER_00: We held grand rounds on A.I.,
[7:50] SPEAKER_00: Two months after ChatGPT came out, I was just like, wow, this is wild.
[7:53] SPEAKER_00: We should talk about it.
[7:54] And Sarah said,
[7:55] so Sarah's job is to,
[7:57] she's the chief health AI officer at UCSF.
[8:00] SPEAKER_00: That says something.
[8:00] SPEAKER_00: We didn't know we needed that job two years ago.
[8:03] SPEAKER_00: But she's great.
[8:04] SPEAKER_00: So she's a physician informatics expert.
[8:07] And she,
[8:07] almost fooling around,
[8:08] she said,
[8:09] all right,
[8:09] AI,
[8:10] and it was GPT-4,
[8:11] I think.
[8:12] SPEAKER_00: I have a patient who's got insomnia,
[8:14] SPEAKER_00: and I think maybe I'd like to use this blood thinner,
[8:17] SPEAKER_00: powerful blood thinner,
[8:19] SPEAKER_00: because I've heard that it helps.
[8:21] And can you write a prior auth to UnitedHealthcare asking
[8:25] permission for me to use Apixabanth?
[8:28] SPEAKER_00: And it did it,
[8:29] SPEAKER_00: and it did it beautifully,
[8:30] SPEAKER_00: and it did it,
[8:31] SPEAKER_00: and it said,
[8:32] SPEAKER_00: well,
[8:32] SPEAKER_00: although most of the time we use this medicine for patients with atrial
[8:36] SPEAKER_00: fibrillation and pulmonary embolism,
[8:38] but there is some emerging evidence that it works really well
[8:41] for patients with
[8:41] SPEAKER_00: insomnia.
[8:42] So Sarah read that and she said,
[8:45] huh,
[8:45] maybe there is some evidence I'm not aware of.
[8:47] SPEAKER_00: And she went to PubMed to do a search and it was complete BS.
[8:51] SPEAKER_00: So this hallucination is it can not only lie to you,
[8:54] SPEAKER_00: but swaddle it in a blanket of BS so that it's utterly believable.
[9:00] SPEAKER_00: So what was striking about that was,
[9:01] SPEAKER_00: you know,
[9:01] SPEAKER_00: this is kind of dangerous in a field like medicine.
[9:04] SPEAKER_00: But I think in some ways,
[9:05] SPEAKER_00: the more striking part of the story is three months later,
[9:07] SPEAKER_00: she went back to GPT,
[9:09] SPEAKER_00: which was now one generation more senior,
[9:12] SPEAKER_00: put in exactly the same prompt.
[9:14] SPEAKER_00: And GPT said, I'm sorry, there must be some mistake here.
[9:17] There is absolutely no evidence that you should use a blood
[9:20] thinner in a patient
[9:21] SPEAKER_00: who can't sleep,
[9:22] SPEAKER_00: you know,
[9:23] SPEAKER_00: to come up with something else.
[9:24] SPEAKER_00: So even,
[9:25] SPEAKER_00: and I think it's an important lesson,
[9:27] SPEAKER_00: because if you played with these tools two and a half years ago,
[9:30] SPEAKER_00: And you say, oh, it hallucinates.
[9:32] SPEAKER_00: It'll never be ready for prime time in a high stakes field like medicine.
[9:35] You should go back again today because the field,
[9:38] the tools get better pretty,
[9:39] pretty quickly.
[9:40] SPEAKER_00: And that was an example of that just over the course of three months.
[9:44] SPEAKER_01: So those are good, I think, stories to kind of frame things.
[9:48] SPEAKER_01: And it seems as I read through the book and also some of the work,
[9:52] SPEAKER_01: like,
[9:52] SPEAKER_01: for example,
[9:53] your recent January New York Times editorial called Stop
[9:56] Worrying and Let AI Help
[9:58] SPEAKER_01: Save Your Life.
[10:00] SPEAKER_01: that your main message is, perfect isn't the enemy of good.
[10:07] SPEAKER_01: Is that a fair thing to say?
[10:08] SPEAKER_00: I think it is one of the key messages.
[10:10] SPEAKER_00: I think,
[10:11] SPEAKER_00: as I said,
[10:12] SPEAKER_00: one of my goals here was to sort of,
[10:13] SPEAKER_00: what are the big picture issues that we really need to think hard about?
[10:16] SPEAKER_00: And one of them,
[10:16] SPEAKER_00: I use Joe Biden's old quote,
[10:18] SPEAKER_00: don't compare me to the almighty,
[10:19] SPEAKER_00: compare me to the alternative.
[10:21] SPEAKER_00: So I, you know, on November 29th, 2022, the day before GBT came out,
[10:27] SPEAKER_00: When I was on Google,
[10:28] SPEAKER_00: it wasn't like I was sitting there saying,
[10:30] SPEAKER_00: you know,
[10:30] SPEAKER_00: Google's not very good as a search engine.
[10:32] SPEAKER_00: I wish I had something better.
[10:33] SPEAKER_00: I couldn't even imagine anything better.
[10:35] SPEAKER_00: It was perfect.
[10:36] And yet when GPT came out,
[10:38] I said,
[10:38] oh,
[10:38] this is actually a better way to think about search.
[10:41] SPEAKER_00: But on November 29,
[10:43] SPEAKER_00: 2022,
[10:43] I don't know any human being who said the American health care
[10:45] system is just
[10:46] SPEAKER_00: perfect.
[10:46] SPEAKER_00: We should just leave it alone.
[10:48] SPEAKER_00: I mean,
[10:48] SPEAKER_00: I think that you have to,
[10:50] SPEAKER_00: as you approach these tools and ask,
[10:52] SPEAKER_00: you know,
[10:52] SPEAKER_00: how can this be useful in medicine and healthcare systems,
[10:56] SPEAKER_00: patients doing things themselves,
[10:59] SPEAKER_00: you have to compare it to the status quo,
[11:01] SPEAKER_00: not some mythical state of perfection.
[11:04] SPEAKER_00: And even if it's not perfect, and it's not, that's not the question.
[11:08] SPEAKER_00: The question is, is it better?
[11:10] SPEAKER_00: And you might say,
[11:11] SPEAKER_00: boy,
[11:11] I can imagine the world's best primary care doctor who's
[11:14] available immediately for
[11:16] SPEAKER_00: me because I have a complaint and I'd like to,
[11:18] SPEAKER_00: you know,
[11:19] SPEAKER_00: to have an expert.
[11:20] SPEAKER_00: That's fine.
[11:21] But I don't know anybody who has that kind of access to a
[11:23] primary care doctor or a
[11:25] SPEAKER_00: mental health therapist or,
[11:27] SPEAKER_00: you know,
[11:27] SPEAKER_00: or for me as a generalist,
[11:29] SPEAKER_00: I don't have fingertip access to an ID doctor or a cardiologist.
[11:34] So the comparison should not be against some state of
[11:37] perfection that we do not have.
[11:39] SPEAKER_00: comparison should be against the current healthcare system.
[11:41] And I can't see how you look at the current healthcare system
[11:44] and say that it's
[11:44] SPEAKER_00: firing on all cylinders.
[11:45] SPEAKER_00: So there's a lot of room for improvement,
[11:48] SPEAKER_00: which is partly why I landed in a pretty optimistic place.
[11:50] SPEAKER_00: I think the tools are amazingly good and getting better every day,
[11:54] SPEAKER_00: but almost just as importantly,
[11:55] the healthcare system can't possibly deliver on what we need it
[11:59] to do if we stick
[12:01] SPEAKER_00: with our go-to strategies of just hiring more human beings.
[12:04] SPEAKER_00: Right.
[12:06] SPEAKER_01: And there's a lot of different parts of the practice of medicine,
[12:10] SPEAKER_01: which we're going to talk about a few of them.
[12:13] But the principle about perfect enemy of good would suggest
[12:19] that there's this net
[12:20] SPEAKER_01: benefit,
[12:21] SPEAKER_01: like everything we do.
[12:22] SPEAKER_01: You know, you're going to give a vaccine.
[12:24] SPEAKER_01: There'll be rare side effects.
[12:26] SPEAKER_01: You're going to give a drug.
[12:28] SPEAKER_01: Some people won't benefit.
[12:29] SPEAKER_01: Some people will.
[12:30] SPEAKER_01: We do clinical trials and we accept evidence for overriding benefit,
[12:34] SPEAKER_01: even though somebody can get hurt.
[12:37] SPEAKER_01: Mm-hmm.
[12:38] SPEAKER_01: And that's basically how medicine is practiced.
[12:41] SPEAKER_01: Whereas,
[12:42] SPEAKER_01: for example,
[12:43] SPEAKER_01: if we're looking at Waymo and we see Kit Kat run over in San Francisco,
[12:47] SPEAKER_01: or we see a child that doesn't have any significant injury but is hit,
[12:55] SPEAKER_01: we say,
[12:55] SPEAKER_01: oh,
[12:55] SPEAKER_01: get rid of these self-driving cars.
[12:59] SPEAKER_01: So the point here is that with something new in medicine, we're
[13:06] SPEAKER_01: You know, there's a public backlash against AI.
[13:09] SPEAKER_01: I think you're probably aware of that.
[13:11] SPEAKER_01: Of course.
[13:12] SPEAKER_01: And I don't know that it necessarily should be
[13:15] SPEAKER_01: in the healthcare domain,
[13:16] but it's kind of an overall sense that to convince that net
[13:21] benefit for a new thing
[13:23] SPEAKER_01: here is really making a real important transformative,
[13:29] SPEAKER_01: ultimately,
[13:29] SPEAKER_01: difference.
[13:30] SPEAKER_01: Now, let's get into a few of these scenarios.
[13:34] SPEAKER_01: So the first one I thought was really important,
[13:37] SPEAKER_01: the conversations that take place in an outpatient visit.
[13:43] SPEAKER_01: ambient scribes,
[13:45] SPEAKER_01: whereby that's,
[13:47] SPEAKER_01: as you say,
[13:47] SPEAKER_01: you get permission,
[13:49] SPEAKER_01: you tape the,
[13:50] SPEAKER_01: you record the interview,
[13:52] you've adjusted it a little bit because you're now talking
[13:54] about your exam and
[13:56] SPEAKER_01: And now you have this synthetic note,
[13:59] which is very powerful and automated note because you can do
[14:01] all these things like
[14:03] SPEAKER_01: order tests and appointments,
[14:05] SPEAKER_01: prescriptions.
[14:06] SPEAKER_01: It works.
[14:08] SPEAKER_01: Pre-authorization.
[14:09] SPEAKER_01: Now, you've saved hours, potentially, of this data clerk time.
[14:16] SPEAKER_01: Where do you think we are with that now?
[14:19] SPEAKER_01: And is that
[14:20] SPEAKER_01: something that should be the norm of practice today.
[14:23] I think it's almost become,
[14:25] at least in the Bay Area,
[14:26] almost an expectation.
[14:27] Like if we were trying to hire a doctor and we did not offer
[14:30] them an AI scribe,
[14:32] SPEAKER_00: I think they would say,
[14:33] SPEAKER_00: oh,
[14:33] SPEAKER_00: I can't work there.
[14:34] I mean,
[14:34] that's amazing because this is a technology that didn't really
[14:37] exist three years ago.
[14:38] SPEAKER_00: I mean, we had scribes at UCSF three, four, five years ago.
[14:42] SPEAKER_00: They were pre-medical students who we hired a company to bring them in.
[14:46] SPEAKER_00: And only in healthcare could we computerize the doctor-patient encounter.
[14:50] SPEAKER_00: And rather than laying people off,
[14:51] SPEAKER_00: we would add an extra person in the exam room to scribe the note.
[14:56] SPEAKER_00: But that's a task that could be, and it has been largely replaced.
[15:00] So we now offer an AI scribe to every single one of our
[15:03] outpatient doctors.
[15:04] SPEAKER_00: Now, actually inpatient doctors, I'm using it as well.
[15:07] SPEAKER_00: Not everybody has to use it.
[15:08] SPEAKER_00: About half people do.
[15:09] SPEAKER_00: It's been a massive satisfier, both to doctors and patients.
[15:12] The patients actually notice that the doctors are making eye
[15:16] contact and doesn't
[15:17] SPEAKER_00: seem as grumpy as they used to seem.
[15:19] SPEAKER_00: And so, to me, and I feel sorry for the companies because
[15:25] SPEAKER_00: It went from two years ago being this,
[15:26] SPEAKER_00: oh,
[15:26] SPEAKER_00: my God,
[15:27] SPEAKER_00: I can't believe this thing works,
[15:28] to now like a ho-hum commodity where everybody's bidding on the
[15:32] lowest price.
[15:33] SPEAKER_00: But I think it was a very important test case because when AI,
[15:36] SPEAKER_00: the new AI,
[15:37] SPEAKER_00: entered the world of healthcare,
[15:38] it was very important that we don't start with diagnosis or
[15:42] clinical decision
[15:42] SPEAKER_00: support.
[15:43] SPEAKER_00: Start with low-hanging fruit, get the buy-in from the docs.
[15:46] SPEAKER_00: And I think that's what's happened.
[15:47] SPEAKER_00: The docs are using it.
[15:49] SPEAKER_00: and saying it's great.
[15:50] SPEAKER_00: And now I'm open to other tools that we're already using.
[15:53] SPEAKER_00: There's a chart summarization tool.
[15:55] SPEAKER_00: One out of five patients has a chart longer than Moby Dick.
[16:00] SPEAKER_00: So now there's a tool.
[16:01] SPEAKER_00: I can press a button.
[16:02] SPEAKER_00: It will summarize a patient's whole chart.
[16:03] We're just rolling out a draft discharge summary tool that will
[16:06] take a hospital
[16:07] SPEAKER_00: course and at least draft the discharge summary.
[16:10] SPEAKER_00: So
[16:11] SPEAKER_00: I think we're building trust in these tools.
[16:14] SPEAKER_00: These are singles.
[16:14] SPEAKER_00: They're not home runs.
[16:15] SPEAKER_00: The home run is clinical decision support.
[16:17] SPEAKER_00: The home run is the tool is embedded in the electronic health record.
[16:21] SPEAKER_00: It is reading the notes.
[16:23] It is maybe suggesting diagnoses or I've made a diagnosis and
[16:27] it says to me,
[16:27] SPEAKER_00: ah,
[16:28] SPEAKER_00: it's now day three and your patient still has a fever and a white count.
[16:31] SPEAKER_00: Maybe, you know, maybe that's not right.
[16:33] SPEAKER_00: It is teeing up orders, those sort of things.
[16:35] SPEAKER_00: That's coming.
[16:36] But I'm actually glad we didn't start with that because I think
[16:38] we needed to build trust first.
[16:40] SPEAKER_01: Yeah,
[16:41] SPEAKER_01: and we're also going to talk about the patient side of this too,
[16:43] SPEAKER_01: not just the physician or clinician side.
[16:46] SPEAKER_01: Absolutely.
[16:47] SPEAKER_01: Okay, so these are gaining steam, the things you just reviewed.
[16:54] SPEAKER_01: There's still people objecting to them, by the way.
[16:56] SPEAKER_01: There's still articles.
[16:56] SPEAKER_01: There was one in this.
[16:57] SPEAKER_01: Of course.
[16:58] SPEAKER_01: Of course.
[16:59] SPEAKER_01: Of course.
[17:01] SPEAKER_01: So we don't have a total buy-in,
[17:03] but I am in touch with that many physicians feel this is a big
[17:07] morale booster.
[17:08] SPEAKER_01: They're getting more time in their life for their family.
[17:12] SPEAKER_01: No question.
[17:14] SPEAKER_01: So that's going to keep building, too.
[17:16] SPEAKER_01: There'll be more hours.
[17:17] SPEAKER_01: And the gift of time,
[17:18] SPEAKER_01: that's something that I,
[17:20] SPEAKER_01: of course,
[17:21] SPEAKER_01: have been very enthusiastic about humanizing medical care.
[17:25] SPEAKER_00: Absolutely.
[17:25] SPEAKER_00: And if you think about it,
[17:27] SPEAKER_00: in some ways,
[17:27] the gift of time is being given back because it was actually
[17:30] taken away by the
[17:31] SPEAKER_00: electronic health record.
[17:32] In some ways,
[17:32] this is mitigating a flaw in our rollout of EHRs over the last
[17:37] decade.
[17:38] SPEAKER_01: I totally agree.
[17:39] SPEAKER_01: But beyond that, it wasn't just EHR.
[17:42] SPEAKER_01: It was the big business of medicine.
[17:44] SPEAKER_01: Yeah, of course.
[17:47] SPEAKER_01: appointments to do a lot and too much for patients.
[17:51] SPEAKER_01: And of course,
[17:52] SPEAKER_01: as you say,
[17:53] when they weren't even looking at patients and they were just
[17:55] on a keyboard,
[17:56] SPEAKER_01: that just took it to another low level.
[17:58] SPEAKER_01: So you introduced the book with the avatar at Mayo Clinic, Dr. Mark.
[18:06] SPEAKER_01: And that's another potentially early win,
[18:10] SPEAKER_01: which is the front door,
[18:12] SPEAKER_01: if you will,
[18:12] SPEAKER_01: for patients,
[18:13] SPEAKER_01: where you have Dr.
[18:14] SPEAKER_01: Morris's avatar,
[18:16] SPEAKER_01: looks just like him,
[18:17] SPEAKER_01: talks just like him,
[18:19] SPEAKER_01: interacting with patients.
[18:21] SPEAKER_01: You want to talk about that?
[18:23] SPEAKER_01: Yeah, it was a dramatic story.
[18:24] SPEAKER_01: Is that something that is going to be ready soon?
[18:27] SPEAKER_00: Well, they're trying it at Mayo.
[18:30] I was meeting with the CEO of the Mayo Clinic and he was
[18:32] reluctant to show this to
[18:33] SPEAKER_00: me,
[18:33] SPEAKER_00: but he just said,
[18:34] SPEAKER_00: the look on his face was,
[18:35] SPEAKER_00: this is so cool,
[18:36] SPEAKER_00: I better,
[18:37] SPEAKER_00: and showed it to me.
[18:38] SPEAKER_00: So yeah, they are developing digital twins of trusted Mayo physicians.
[18:42] SPEAKER_00: And importantly, it's not trained on the entire internet.
[18:47] It's not trained on The Onion and it's not trained on Reddit
[18:50] conversations.
[18:50] SPEAKER_00: It's trained on Mayo
[18:52] SPEAKER_00: specific data.
[18:53] SPEAKER_00: And I think,
[18:54] SPEAKER_00: therefore,
[18:55] because they recognize it better be trustworthy if it's going
[18:57] to play the part of a
[18:59] SPEAKER_00: Mayo clinician.
[19:01] As we think about how do we scale the expertise of Dr.
[19:05] Eric Topol,
[19:06] there's only so much you.
[19:08] SPEAKER_00: At some point,
[19:09] SPEAKER_00: if we said,
[19:10] SPEAKER_00: if we could build an avatar that looks like you,
[19:12] SPEAKER_00: sounds like you,
[19:13] SPEAKER_00: and says the things you would have said in a given circumstance,
[19:16] you now have scaled the ability to provide that kind of
[19:19] expertise to more patients
[19:21] SPEAKER_00: in more places.
[19:22] SPEAKER_00: think that's immensely exciting.
[19:24] SPEAKER_00: It's a little creepy.
[19:26] SPEAKER_00: And of the things I'm very enthusiastic about in AI,
[19:31] SPEAKER_00: in healthcare,
[19:32] SPEAKER_00: the one that worries me the most is the dark side of that,
[19:35] SPEAKER_00: which,
[19:35] SPEAKER_00: of course,
[19:35] SPEAKER_00: is deep fakes.
[19:37] SPEAKER_00: If we can create a facsimile of you,
[19:40] SPEAKER_00: we can also create a facsimile of you saying,
[19:42] SPEAKER_00: you know,
[19:42] SPEAKER_00: you shouldn't get vaccines because they can tell people.
[19:45] SPEAKER_00: So we're going to have to figure that one out.
[19:48] SPEAKER_00: And this is one that has a very bright side, but also a very dark side.
[19:51] SPEAKER_01: Yeah,
[19:52] SPEAKER_01: no,
[19:52] SPEAKER_01: I think we're already seeing avatars all over the place,
[19:56] SPEAKER_01: and this is going to come to medicine.
[19:59] SPEAKER_01: It's a front tour glorified graphically,
[20:02] and it's potentially even more alluring to patients because it
[20:11] gives them the sense
[20:11] of actually connecting more than just through a text or through
[20:15] a… But again,
[20:16] SPEAKER_00: the enemy,
[20:16] SPEAKER_00: the perfect versus the good…
[20:20] you know,
[20:21] try to find a primary care doctor in San Francisco or in San
[20:23] Diego.
[20:24] SPEAKER_00: It ain't that easy.
[20:26] And one can imagine a world where this kind of technology is
[20:30] basically doing
[20:31] primary care and managing your blood pressure and managing your
[20:34] lipids and managing
[20:35] SPEAKER_00: your weight and your GLPs.
[20:38] SPEAKER_00: And I think that'd probably be okay.
[20:39] SPEAKER_00: I'm guessing it would do a very,
[20:41] very good job as long as there's a triage protocol that gets
[20:44] you to a doctor.
[20:45] SPEAKER_00: And if you have really complex needs, you can still see a
[20:48] human,
[20:49] but primary care in its current incarnation is not,
[20:52] it's just not doable.
[20:53] And so I think we're going to need stuff like this in order for
[20:55] patients to get the
[20:56] SPEAKER_00: care they need.
[20:57] SPEAKER_01: Well, you're bringing up these massive unmet needs.
[21:00] SPEAKER_01: You already touched on mental health.
[21:02] SPEAKER_01: You're talking about primary care.
[21:03] SPEAKER_01: I mean, we don't have any sense of the supply-demand relationship.
[21:09] SPEAKER_01: Impossible.
[21:09] SPEAKER_01: The question is, can we build this, that it really works well?
[21:12] SPEAKER_01: We'll get to that in just a little bit.
[21:16] SPEAKER_01: The concern about AI in several studies looking for performance,
[21:23] SPEAKER_01: whether it was radiology,
[21:25] SPEAKER_01: diagnosis,
[21:26] SPEAKER_01: various things,
[21:27] SPEAKER_01: is
[21:29] is that the AI did better than the doctor with the AI in seven
[21:33] different studies,
[21:34] SPEAKER_01: different groups,
[21:37] SPEAKER_01: different tasks.
[21:39] SPEAKER_01: And you say, wait a minute, it's not supposed to be that way.
[21:42] SPEAKER_01: The doctor and AI was supposed to be the best of all worlds.
[21:47] SPEAKER_01: And we hadn't seen that.
[21:48] SPEAKER_01: And of course,
[21:49] SPEAKER_01: there are some studies,
[21:50] SPEAKER_01: admittedly,
[21:51] SPEAKER_01: that do show the fusion of human and AI being superior.
[21:57] SPEAKER_01: But this brings up the concern about automation bias
[22:02] SPEAKER_01: lack of grounding of physicians and how to use the AI.
[22:07] SPEAKER_01: What are your thoughts about that?
[22:08] When you saw these studies popping up,
[22:10] we wrote a New York Times op-ed about that too.
[22:13] SPEAKER_01: Did you think,
[22:14] SPEAKER_01: oh,
[22:14] SPEAKER_01: this is just a passing problem,
[22:16] SPEAKER_01: or do you think this is going to be a persistent?
[22:19] Yeah,
[22:19] as I said,
[22:19] you know,
[22:20] I was looking,
[22:20] what are the big picture issues?
[22:21] I think one of the biggest picture issues here is you have AI
[22:26] tools that are good
[22:28] SPEAKER_00: enough to be useful,
[22:29] SPEAKER_00: but not perfect.
[22:30] SPEAKER_00: And so the system that we will create is the human in the loop,
[22:33] is the doctor who is the final arbiter of the diagnosis or the
[22:37] reading of the
[22:38] SPEAKER_00: mammogram or you name it.
[22:41] SPEAKER_00: And at least that sounds great.
[22:45] SPEAKER_00: And yet there are some real problems with it.
[22:46] And one of them,
[22:48] of course,
[22:48] is humans are not very good at remaining vigilant.
[22:52] SPEAKER_00: if they're overseeing a technology that they've learned to trust.
[22:54] SPEAKER_00: It was right the last 49 times.
[22:56] SPEAKER_00: Are you going to catch it when it's wrong on number 50?
[22:59] SPEAKER_00: The another problem is that humans will de-skill.
[23:02] SPEAKER_00: But I'd say that in some ways,
[23:04] SPEAKER_00: the new problem is the one that you're getting at here,
[23:07] SPEAKER_00: which is it may get to be,
[23:09] and we're seeing examples of where the AI is so good that then
[23:12] having the human
[23:13] SPEAKER_00: intervene actually degrades the performance of the overall dyad.
[23:19] SPEAKER_00: You know,
[23:19] SPEAKER_00: the answer to that,
[23:20] SPEAKER_00: I think,
[23:20] SPEAKER_00: is to be aware of that as a,
[23:22] SPEAKER_00: you know,
[23:23] SPEAKER_00: almost to have that,
[23:24] SPEAKER_00: that is a possibility as we test these tools,
[23:26] SPEAKER_00: that in fact,
[23:27] one of the possible outcomes is the human intervention actually
[23:30] makes things worse.
[23:32] SPEAKER_00: And if that's true,
[23:33] SPEAKER_00: and the humans probably are not going to get any better over time,
[23:36] then that may be a task that we actually say is ready for
[23:38] autonomous AI to do its
[23:40] SPEAKER_00: thing.
[23:41] SPEAKER_00: And I think that that's probably the way this is going to go,
[23:43] SPEAKER_00: that we're going to be looking,
[23:44] SPEAKER_00: we're going to be deconstructing what a doctor does.
[23:47] And there's certain things we're going to say,
[23:48] we really need a human to do that.
[23:50] SPEAKER_00: This is a judgment call.
[23:52] SPEAKER_00: The stakes are high.
[23:53] It involves a whole lot of coordination across getting the
[23:56] patient's input or
[23:57] SPEAKER_00: family's input,
[23:58] SPEAKER_00: et cetera.
[23:58] SPEAKER_00: But there may be certain tasks where we say,
[24:01] actually having human intervention here performs worse as a
[24:06] dyad than if the AI
[24:07] SPEAKER_00: does its own thing.
[24:08] And I think we're just going to have to have that be one of the
[24:10] testable hypotheses
[24:11] SPEAKER_00: as we go forward.
[24:13] SPEAKER_01: Yeah,
[24:13] SPEAKER_01: I mean,
[24:13] and one of the things just to emphasize that when these studies
[24:16] get published,
[24:17] SPEAKER_01: they don't really hit the real world of medicine.
[24:22] SPEAKER_01: They're not testing it in the messy world of medical practice,
[24:26] SPEAKER_01: but rather,
[24:27] SPEAKER_01: you know,
[24:27] SPEAKER_01: giving vignettes to doctors or the combination and,
[24:31] SPEAKER_01: you know,
[24:32] SPEAKER_01: patient actors.
[24:33] SPEAKER_01: And, you know, these are simulations of they're not the real world.
[24:36] SPEAKER_01: And we'll see.
[24:36] SPEAKER_00: They're getting,
[24:37] SPEAKER_00: you know,
[24:37] Google's just rolled out a study yesterday of where they are
[24:40] going to try it in a
[24:41] SPEAKER_00: real world.
[24:42] SPEAKER_00: setting.
[24:43] SPEAKER_00: And,
[24:43] SPEAKER_00: you know,
[24:43] SPEAKER_00: one of the points I made in the book is,
[24:45] SPEAKER_00: you know,
[24:46] SPEAKER_00: you look at these tools that can do AI physical therapy and it's like,
[24:49] SPEAKER_00: look how good this is.
[24:50] SPEAKER_00: It can replace physical therapy as well.
[24:52] Try it on a 90 year old patient who lives in a third floor
[24:56] Chinatown walk up and
[24:57] SPEAKER_00: doesn't speak English.
[24:58] SPEAKER_00: I mean,
[24:59] SPEAKER_00: that it's,
[25:00] SPEAKER_00: you know,
[25:00] the tech companies are not used to building tools that have to
[25:03] work on everybody.
[25:05] SPEAKER_00: And in the health care world, we take care of everybody.
[25:08] And that adds a level of complexity that I think most companies
[25:11] are not used to dealing with.
[25:12] SPEAKER_01: Totally agree.
[25:13] SPEAKER_01: All right.
[25:14] SPEAKER_01: Now,
[25:15] SPEAKER_01: there are some issues besides hallucinations,
[25:20] SPEAKER_01: besides the fact that it can be confabulation and bullshit.
[25:24] There's this problem of sycophancy and basically just playing
[25:31] up to the prompter.
[25:34] SPEAKER_01: And there's also a major problem of having this assured confidence.
[25:46] SPEAKER_01: What do you think about those two concerns?
[25:49] SPEAKER_00: The sycophancy has gotten a little bit better.
[25:51] I mean,
[25:51] there was a version of GPT a year ago that was like,
[25:53] oh,
[25:53] that was so smart of you.
[25:55] SPEAKER_00: It was like, stop that already.
[25:58] SPEAKER_00: My ego is fine.
[25:58] SPEAKER_00: I don't need you to tell me this.
[26:00] SPEAKER_00: You know,
[26:00] it gets us into trouble when it basically tells you you're so
[26:04] brilliant when you're
[26:05] SPEAKER_00: wrong.
[26:06] And I think that that's part of the problem in the mental
[26:09] health chatbots where,
[26:11] SPEAKER_00: you know,
[26:12] SPEAKER_00: some of these,
[26:12] SPEAKER_00: when you read the transcripts,
[26:14] SPEAKER_00: you know,
[26:14] SPEAKER_00: it basically agrees with the kid that the kid should kill himself.
[26:17] SPEAKER_00: I mean, it's heinous.
[26:19] SPEAKER_00: So, yeah, I mean, these tools have to be trustworthy.
[26:23] SPEAKER_00: And to the extent that it's basically BSing you,
[26:26] SPEAKER_00: whether it's BSing you with the wrong answer or BSing you telling you how
[26:30] SPEAKER_00: unbelievably wonderful you are,
[26:32] SPEAKER_00: That has to be engineered out.
[26:33] SPEAKER_00: I think that's very important.
[26:36] SPEAKER_00: I think one of my worries,
[26:38] SPEAKER_00: of course,
[26:38] SPEAKER_00: is that we will,
[26:40] both because of sick and fancy and because of the incredible
[26:42] humanness of the tools
[26:43] SPEAKER_00: and the way they respond,
[26:45] SPEAKER_00: that we will bestow on it more trust than it deserves.
[26:49] SPEAKER_00: Yeah.
[26:49] SPEAKER_00: Because we trust humans more than we trust machines.
[26:52] But when it talks to you in a human voice and appears to be
[26:55] laughing at your jokes and all that.
[26:57] SPEAKER_00: So I just think we've got to be super hard nosed about that.
[26:59] And like,
[27:00] you know,
[27:00] when we test it,
[27:01] it's got to be in real world circumstances.
[27:04] It has to be,
[27:04] you know,
[27:05] the messiness of medicine has to be part of what we're testing.
[27:08] SPEAKER_00: And we need to be testing also sort of the machine human thing.
[27:14] SPEAKER_00: which if it's sycophantic and the human unduly trusts it because of that,
[27:18] that needs to be something that we're weighing as we put these
[27:21] tools through their
[27:21] SPEAKER_00: paces.
[27:22] SPEAKER_00: Because you're absolutely right.
[27:23] SPEAKER_00: If you're just testing,
[27:24] SPEAKER_00: you know,
[27:24] SPEAKER_00: that it can do well on a test or do well on a scenario,
[27:29] SPEAKER_00: but the scenario is fed to them.
[27:30] SPEAKER_00: I mean,
[27:31] SPEAKER_00: I guess one tiny departure,
[27:33] SPEAKER_00: I kind of tended to look at how I use GPT or Gemini and say,
[27:38] SPEAKER_00: oh,
[27:38] SPEAKER_00: you know,
[27:39] SPEAKER_00: it works pretty well for me,
[27:40] SPEAKER_00: and therefore it's pretty good for a patient.
[27:43] I think experts don't recognize the degree to which their
[27:46] brains work differently
[27:47] SPEAKER_00: than they used to work when they were non-experts.
[27:50] For the patient to take the 15 different symptoms or signs they
[27:54] have and know which
[27:55] SPEAKER_00: ones to feed in—
[27:57] is a pretty high-level cognitive act that for physicians,
[28:00] we just learn to do.
[28:01] SPEAKER_00: It's like breathing.
[28:02] SPEAKER_00: But for a patient, it's not natural.
[28:03] And therefore,
[28:04] are they going to get it exactly right which symptoms to put
[28:06] in?
[28:07] SPEAKER_00: And then on the output side,
[28:09] SPEAKER_00: I can look at the results or you can look at the results and say,
[28:11] SPEAKER_00: oh,
[28:11] SPEAKER_00: that's a smart answer,
[28:12] SPEAKER_00: and that's a smart answer,
[28:13] SPEAKER_00: and that one is really stupid.
[28:14] SPEAKER_00: I'm discarding that.
[28:15] SPEAKER_00: And patients, by and large, have no ability to do that.
[28:18] SPEAKER_00: So the tools that we need to build for patients,
[28:21] SPEAKER_00: I think,
[28:21] SPEAKER_00: are going to be more sophisticated and have to be more doctor-like,
[28:25] SPEAKER_00: more iterative,
[28:26] SPEAKER_00: more you tell me your symptoms,
[28:27] SPEAKER_00: I ask you a set of questions,
[28:29] SPEAKER_00: you answer them back and forth,
[28:31] SPEAKER_00: than just a generic GPT or Gemini or Claude,
[28:34] SPEAKER_00: these generic chatbots tools.
[28:35] SPEAKER_00: I don't think they're quite ready for prime time for patients.
[28:38] SPEAKER_01: Yeah,
[28:38] SPEAKER_01: and that's actually the area I want to delve in now,
[28:41] SPEAKER_01: which is the patient side of this,
[28:43] SPEAKER_01: which is
[28:44] SPEAKER_01: really big.
[28:46] SPEAKER_01: And as you already touched on, we have ChatGPT Health recently released.
[28:50] SPEAKER_01: The day after was Claude Health.
[28:52] SPEAKER_01: Today there's Lotus Health.
[28:54] Tomorrow there'll be Perflexity Health,
[28:56] and they'll just keep going,
[28:57] right?
[28:58] SPEAKER_01: Now, they're not customized models for AI.
[29:03] SPEAKER_01: They're not doing what you say.
[29:05] They don't have context of all the person's data,
[29:08] no less other parts of their stack of data.
[29:13] SPEAKER_01: that seems to be a really big unmet need now.
[29:16] SPEAKER_01: What do you think?
[29:17] SPEAKER_00: Yeah, I mean, I think you could see the end game here.
[29:20] The end game here is that models that are built on healthcare
[29:25] data and healthcare
[29:26] SPEAKER_00: database,
[29:27] SPEAKER_00: in some ways,
[29:28] SPEAKER_00: if you use open evidence,
[29:29] SPEAKER_00: I mean,
[29:29] SPEAKER_00: the reason it's so good,
[29:31] SPEAKER_00: it was built
[29:32] using not the entire generic internet,
[29:34] but really the medical literature,
[29:36] society guidelines.
[29:38] SPEAKER_00: Tools like that,
[29:39] SPEAKER_00: that are built for a purpose,
[29:41] embedded in the electronic health record and embedded in the
[29:44] patient's portal.
[29:46] SPEAKER_00: So as it is today,
[29:47] why should I take the data that's in my Epic record and have to
[29:50] port it out into
[29:52] SPEAKER_00: GPT,
[29:53] SPEAKER_00: into OpenAI,
[29:54] SPEAKER_00: a company whose,
[29:55] SPEAKER_00: I don't know whether I can trust them with my data.
[29:57] SPEAKER_00: They have no, they're not obliged to follow HIPAA.
[30:02] SPEAKER_00: And then using essentially GPT to query my data.
[30:05] SPEAKER_00: That feels like a very transitional stage.
[30:07] SPEAKER_00: The final stage is it's embedded in my chart, so my patient portal.
[30:13] SPEAKER_00: It has access to all the data in my medical record.
[30:17] And if I am querying it about a symptom or I'm trying to
[30:21] understand what it means
[30:22] because it just said my EKG is abnormal and my PR segment is
[30:26] too long or whatever,
[30:27] SPEAKER_00: that it can answer that question and not just answer it generically,
[30:31] SPEAKER_00: but answer it in the context of me.
[30:33] SPEAKER_00: I think that's where this goes.
[30:34] SPEAKER_00: I don't think at the end of the day,
[30:36] people are going to be using absolutely standalone tools that
[30:39] are using generic AI
[30:40] SPEAKER_00: models to give you the answer.
[30:43] SPEAKER_00: I think these tools that are embedded in the electronic health record,
[30:45] SPEAKER_00: whether they're built by Epic or built by third-party companies,
[30:49] SPEAKER_00: who knows?
[30:49] SPEAKER_00: We'll see.
[30:50] SPEAKER_00: That'll be a battle.
[30:51] But I think it's got to be much more integrated into the way
[30:54] patients get their
[30:54] SPEAKER_00: information and get their care.
[30:57] Yeah,
[30:57] I mean,
[30:57] right now,
[30:59] these are very shallow,
[31:01] very,
[31:01] you know,
[31:02] first pass.
[31:04] SPEAKER_01: And as you say, they're built for this purpose, no less anywhere close.
[31:09] But one of the problems I want to get your comment on,
[31:12] the inputs here are very incomplete.
[31:16] Now,
[31:16] of course,
[31:17] maybe the people that go to UCSF medical system only go there.
[31:23] SPEAKER_01: But rest of the world, they go to eight different providers.
[31:27] SPEAKER_01: They don't have the communication between all these records.
[31:31] SPEAKER_01: So the inputs,
[31:33] SPEAKER_01: even when they start to get together,
[31:35] SPEAKER_01: they're dealing with incomplete ascertainment in this country.
[31:40] SPEAKER_01: No less in this country, as opposed to many others.
[31:44] SPEAKER_01: Yeah.
[31:45] SPEAKER_01: We don't care for the population to try to prevent disease.
[31:49] We're just getting one year at a time with your insurance
[31:52] companies so the employer renews it.
[31:57] SPEAKER_01: We don't do anything to try to...
[32:00] SPEAKER_01: Are you frustrated about the incompleteness of data ascertainment,
[32:06] SPEAKER_01: availability,
[32:08] SPEAKER_01: along with the problem of not harnessing this potential?
[32:16] SPEAKER_00: I am, but I guess it's not in my top two or three problems.
[32:19] SPEAKER_00: And maybe,
[32:20] SPEAKER_00: you know,
[32:20] I'm in a big health system that does everything from primary
[32:23] care to do your
[32:25] SPEAKER_00: transplant.
[32:26] And yeah,
[32:27] I mean,
[32:27] there probably are times where we miss something because the
[32:34] you know,
[32:34] at Kaiser or at Sutter or at a different system,
[32:39] and it's not embedded in our system.
[32:41] SPEAKER_00: But I don't think that comes up as often as the other problems,
[32:47] SPEAKER_00: which are,
[32:47] SPEAKER_00: you know,
[32:48] can we provide good access to care when the patient needs the
[32:51] care that they need?
[32:53] SPEAKER_00: Are we getting the correct answers?
[32:54] SPEAKER_00: I think the interoperability question and the ability to,
[32:57] SPEAKER_00: certainly the world would be,
[32:58] SPEAKER_00: there's no question,
[32:59] SPEAKER_00: the world would be a better place if
[33:01] SPEAKER_00: all of patients' data are in one place,
[33:04] SPEAKER_00: are easily downloadable,
[33:06] follow you around even if you switch health systems and if you
[33:09] switch health
[33:09] SPEAKER_00: insurers.
[33:10] No question that makes for better information for an individual
[33:13] patient and better
[33:15] SPEAKER_00: AI,
[33:15] because ultimately you have a better database to be used and a
[33:18] much more holistic
[33:19] SPEAKER_00: database to use.
[33:20] SPEAKER_00: But in day-to-day work here,
[33:22] SPEAKER_00: it feels like,
[33:23] SPEAKER_00: yeah,
[33:24] SPEAKER_00: it's a real problem,
[33:24] but not the biggest one that we have to solve to markedly
[33:28] improve health and health
[33:29] SPEAKER_00: care.
[33:29] SPEAKER_01: Yeah, I guess I would agree with the way you framed it.
[33:33] SPEAKER_01: The only thing I would say is that,
[33:34] SPEAKER_01: you know,
[33:34] SPEAKER_01: when you have the UK Biobank with 20-year follow-up,
[33:37] SPEAKER_01: when you have,
[33:38] SPEAKER_01: for example,
[33:39] SPEAKER_01: Denmark with 20-year follow-up of millions of people,
[33:42] SPEAKER_01: all their data,
[33:43] SPEAKER_01: nothing missing.
[33:44] SPEAKER_00: You can get a lot more out of that than you can.
[33:48] SPEAKER_00: Including all the people in Greenland.
[33:49] SPEAKER_00: Yeah, exactly.
[33:52] SPEAKER_00: So, yeah, I mean, look at COVID.
[33:54] I mean,
[33:55] look at,
[33:56] you know,
[33:56] you were doing this incredible job annotating the literature.
[33:59] SPEAKER_00: How much of the literature came from the UK?
[34:01] SPEAKER_00: Oh, yeah.
[34:01] Because they had a complete data set of all their patients in
[34:04] ways that we couldn't
[34:05] SPEAKER_00: possibly do in the U.S.
[34:06] SPEAKER_00: So there are major advantages to that.
[34:08] SPEAKER_00: There's no question.
[34:09] SPEAKER_01: Yeah.
[34:10] SPEAKER_01: The reason I'm getting at it is, like, for example...
[34:13] SPEAKER_01: In China,
[34:14] they've already instituted using abdominal and chest CT to
[34:18] detect pancreatic cancer
[34:20] and override the radiologist's interpretation because they
[34:27] don't want to miss any
[34:27] SPEAKER_01: pancreatic cancer.
[34:29] SPEAKER_01: We don't do that.
[34:31] SPEAKER_01: I don't know when we're going to do that.
[34:33] There are a lot of things on scans that don't rely on large
[34:36] language models that
[34:37] SPEAKER_01: are opportunistic,
[34:38] SPEAKER_01: that we could pick up all sorts of things.
[34:40] There are companies now that are giving retina photos and
[34:44] telling patients whether
[34:46] SPEAKER_01: they have Alzheimer's disease or cardiovascular disease.
[34:49] SPEAKER_01: We're not doing any of that stuff here.
[34:51] SPEAKER_01: Yeah.
[34:52] SPEAKER_01: One of the problems is,
[34:54] SPEAKER_01: as we talk about the use of AI and for helping patients,
[34:59] SPEAKER_01: a lot of things that we could do that we're not.
[35:04] There's different thresholds around the world for how and when
[35:09] to use these things.
[35:13] SPEAKER_01: When you put it all together,
[35:14] SPEAKER_01: I guess towards the very last part of the book,
[35:19] SPEAKER_01: I think you kind of said it all.
[35:23] I want to just get you back onto that statement because I think
[35:26] it's just a good way to sum up.
[35:33] But I am equally convinced that particularly for patients with
[35:36] acute and complex
[35:38] SPEAKER_01: chronic conditions,
[35:39] SPEAKER_01: there will always be a need for a human guide.
[35:44] Someone with not only deep medical knowledge and refined
[35:48] clinical judgment,
[35:50] SPEAKER_01: but also the emotional intelligence.
[35:52] SPEAKER_01: to recognize and address the unspoken fears and on.
[35:57] SPEAKER_01: So that brings it back to the human side here.
[36:01] And I wanted to give you a chance to do that because some
[36:04] people might say,
[36:04] SPEAKER_01: oh,
[36:05] SPEAKER_01: well,
[36:06] SPEAKER_01: Dr.
[36:06] SPEAKER_01: Wachter's all onto this AI and he's really becoming super.
[36:10] SPEAKER_01: And he went from the anti-digital to the pro AI here.
[36:16] SPEAKER_01: Tell us about this
[36:18] SPEAKER_01: the human part of medicine,
[36:20] SPEAKER_01: the essentiality of medicine in your view that gets the right,
[36:23] SPEAKER_01: strikes ultimately the right balance in the future.
[36:26] SPEAKER_00: Yeah, I, thank you.
[36:27] SPEAKER_00: I,
[36:28] I landed in a place and not purposefully because I am fully
[36:32] capable of writing a
[36:33] pretty grumpy book because I wrote one 10 years ago about the
[36:35] electronic health
[36:36] SPEAKER_00: records.
[36:37] SPEAKER_00: But in this time,
[36:40] I landed in a pretty optimistic place because I think we need
[36:43] it and I think these
[36:43] SPEAKER_00: tools are amazingly good.
[36:45] SPEAKER_00: But one of the things I wanted to check myself on is,
[36:49] SPEAKER_00: you know,
[36:50] SPEAKER_00: is it good enough to replace physicians or nurses,
[36:53] SPEAKER_00: et cetera?
[36:55] SPEAKER_00: And to do that with an open mind.
[36:58] SPEAKER_00: Obviously, I'm rooting for the humans.
[37:00] I've got a daughter and son-in-law who are pulmonary fellows
[37:02] here and I'd like them
[37:03] SPEAKER_00: to have jobs,
[37:04] SPEAKER_00: et cetera,
[37:04] SPEAKER_00: et cetera.
[37:05] And I do not want a bot telling me that I have cancer or I have
[37:08] diabetes.
[37:09] SPEAKER_00: There's certain parts of it that seem clear.
[37:12] SPEAKER_00: And so as I began using these tools over the last couple of years,
[37:15] SPEAKER_00: I would use them when I'm ward attending.
[37:17] SPEAKER_00: And it struck me that these tools are going to make me a better doctor.
[37:21] It's going to make me a happier doctor because it's going to
[37:24] take a bunch of the
[37:24] SPEAKER_00: silly paperwork off my plate.
[37:26] It's going to give me access to knowledge that really does help
[37:29] me do a better job
[37:30] SPEAKER_00: taking your patients.
[37:31] SPEAKER_00: But it does not replace it.
[37:33] the human doctor on all sorts of areas,
[37:36] empathy,
[37:37] communication,
[37:38] coordination.
[37:40] SPEAKER_00: The medicine's messy and complex.
[37:42] SPEAKER_00: There are things that I think the AI can do by itself.
[37:45] SPEAKER_00: And anybody who says,
[37:46] we can't have autonomous AI doing something in a high-stakes
[37:49] thing like medicine.
[37:51] SPEAKER_00: I take a Waymo once a week, and, you know, you talk about high stakes.
[37:55] SPEAKER_00: Try making a left turn across to Visadero Street during rush hour.
[37:59] You know,
[37:59] you're putting your life in your hand,
[38:00] and the thing is wildly safe.
[38:03] SPEAKER_00: So I think there is a role.
[38:04] SPEAKER_00: I think there are things that we'll do right now that we count on humans.
[38:06] SPEAKER_00: But I think there is a plenty of territory left for the human clinicians.
[38:11] SPEAKER_00: I think patients will insist on it in certain circumstances.
[38:16] SPEAKER_00: And I think that lands us in a really good place.
[38:17] SPEAKER_00: I think it lands us in a place where there are still doctors,
[38:21] SPEAKER_00: but the doctors are practicing at the top of their license.
[38:23] SPEAKER_00: They're not wasting their time doing paperwork.
[38:25] SPEAKER_00: They have access to better knowledge than what they have now.
[38:28] SPEAKER_00: And that's partly why I landed in a very optimistic place,
[38:31] SPEAKER_00: because I think this could be net good for clinicians and,
[38:34] SPEAKER_00: more importantly,
[38:34] SPEAKER_00: net good for patients.
[38:36] SPEAKER_01: Yeah.
[38:36] SPEAKER_01: No, I think that's a nice message.
[38:40] SPEAKER_01: Did I capture some key parts of your philosophy?
[38:44] SPEAKER_00: You captured it beautifully.
[38:45] Not beautifully enough that someone shouldn't go out and buy
[38:47] the book,
[38:48] but absolutely perfectly.
[38:49] That's true.
[38:52] SPEAKER_01: Is there anything you want to add to what we discussed?
[38:55] I mean,
[38:55] obviously,
[38:55] we do on for hours,
[38:56] and I hope someday we'll have that chance.
[38:59] SPEAKER_00: No,
[38:59] SPEAKER_00: I think,
[38:59] SPEAKER_00: I mean,
[39:00] SPEAKER_00: I guess one question that comes up a lot,
[39:02] SPEAKER_00: Eric,
[39:02] SPEAKER_00: is,
[39:02] SPEAKER_00: you know,
[39:03] SPEAKER_00: the training of future physicians.
[39:04] And I think that is a really interesting question that we're
[39:07] really struggling with.
[39:09] It's very clear that they need to learn a lot about AI and how
[39:12] it works and when to
[39:13] SPEAKER_00: trust it and when not to.
[39:14] SPEAKER_00: I think the central question, which we're really grappling with,
[39:18] Are there things that you can take off the curriculum in terms
[39:20] of knowledge and
[39:22] SPEAKER_00: clinical reasoning?
[39:23] SPEAKER_00: My own bias is it's premature to do that,
[39:26] SPEAKER_00: that for the foreseeable future,
[39:28] SPEAKER_00: the AI will be doing its thing,
[39:30] SPEAKER_00: but there will be a doctor who's looking over its shoulder.
[39:33] And you can easily see us entering a death spiral that if
[39:36] people aren't good at
[39:38] SPEAKER_00: that anymore,
[39:38] SPEAKER_00: then yes,
[39:39] SPEAKER_00: the AI will be better than they are.
[39:40] SPEAKER_00: But I think we need them to still be good at it.
[39:42] SPEAKER_00: So the only area of consensus,
[39:45] the only area where everybody agrees maybe we can take this off
[39:48] the curriculum is
[39:49] SPEAKER_00: the Krebs cycle.
[39:52] SPEAKER_00: For the non-physicians who don't have PTSD,
[39:55] SPEAKER_00: when I say that,
[39:56] it's this incredibly obscure biochemical pathway that we all
[40:00] learned in med school,
[40:01] SPEAKER_00: never,
[40:01] SPEAKER_00: ever used again.
[40:02] SPEAKER_00: I think there's a consensus that we can get by without the Krebs cycle.
[40:05] SPEAKER_01: Yeah, well, and in the book, Along Those Lines, you had a quote
[40:09] SPEAKER_01: I can't remember from whom,
[40:10] SPEAKER_01: but about doctors need to learn to be the orchestra,
[40:16] SPEAKER_01: the leader,
[40:16] SPEAKER_01: not to learn.
[40:17] SPEAKER_00: Yeah.
[40:17] SPEAKER_00: So this is Mark Triola,
[40:18] who's the dean for students and a real leader in AI meets
[40:21] education at NYU.
[40:23] SPEAKER_00: And as I was Lee,
[40:24] SPEAKER_00: I spent a couple of days at NYU and I said,
[40:25] SPEAKER_00: and Mark said,
[40:26] SPEAKER_00: you know,
[40:26] we teach the students now to play the viola and play the
[40:30] bassoon and play the
[40:31] SPEAKER_00: drums.
[40:31] SPEAKER_00: We need to teach them to be orchestra conductors.
[40:34] SPEAKER_00: And I said,
[40:35] SPEAKER_00: That's interesting, Mark.
[40:36] SPEAKER_00: What exactly does that mean?
[40:37] SPEAKER_00: And he said, I have no idea.
[40:38] SPEAKER_00: I just thought it sounded good.
[40:42] SPEAKER_01: Okay.
[40:42] SPEAKER_01: Well,
[40:42] SPEAKER_01: one last thing I got to ask you about,
[40:45] SPEAKER_01: because when I was doing deep medicine,
[40:49] SPEAKER_01: I had a chapter on deep empathy.
[40:51] SPEAKER_01: And I said, that's the one thing AI will never do.
[40:56] SPEAKER_01: And I just completely missed it,
[40:59] SPEAKER_01: that when there would be these large language models,
[41:02] SPEAKER_01: it would channel empathy.
[41:04] SPEAKER_01: AI has no clue what it is.
[41:06] SPEAKER_01: Of course.
[41:08] SPEAKER_01: The machine wouldn't know what the hell it is.
[41:10] SPEAKER_01: But it is now,
[41:12] SPEAKER_01: in 13 studies,
[41:14] SPEAKER_01: 12 have found,
[41:17] SPEAKER_01: scored by doctors,
[41:18] SPEAKER_01: that the AI was more empathetic than the doctors.
[41:21] SPEAKER_01: So I want to just get your views, and this will be the last thing.
[41:25] SPEAKER_01: And do you foresee, when you have these notes that are automated,
[41:31] And will they be reviewed by AI to coach doctors to be more
[41:35] empathetic and better
[41:36] SPEAKER_01: communicators?
[41:37] SPEAKER_01: Will that be a license requirement or an annual?
[41:41] SPEAKER_01: Is it true?
[41:42] Is it true that the AI is,
[41:45] without knowing what empathy is,
[41:47] is a real channel for empathy?
[41:50] SPEAKER_00: Well, first of all, you got that right.
[41:51] I mean,
[41:51] the AI doesn't know the difference between empathy and a recipe
[41:55] for chicken marsala.
[41:56] SPEAKER_00: Right.
[41:56] It's just a thing that it has learned to mimic,
[41:58] but it's an incredibly good actor.
[42:00] SPEAKER_00: And as you say,
[42:02] the studies show that blinded observers hearing the response of
[42:06] AI and the response
[42:07] SPEAKER_00: of a human often find the AI's response to be more empathetic.
[42:10] SPEAKER_00: Now,
[42:10] you might say that's unfair because often it's because the AI
[42:12] has more time and
[42:13] SPEAKER_00: gives you more expansive responses,
[42:15] SPEAKER_00: but actually...
[42:16] SPEAKER_00: True.
[42:17] SPEAKER_00: It has more time, and therefore it may do that.
[42:19] So I think the use of AI in coaching is going to be really
[42:23] quite profound.
[42:24] SPEAKER_00: We're building an AI tool here for giving our trainees,
[42:28] SPEAKER_00: our house staff,
[42:29] SPEAKER_00: and our students feedback,
[42:32] SPEAKER_00: where it will listen to a presentation,
[42:34] SPEAKER_00: in some ways sometimes draft the feedback,
[42:36] SPEAKER_00: or if I have
[42:37] SPEAKER_00: feedback on that presentation, I may dictate it in.
[42:39] And then at the end,
[42:40] it'll consolidate all that and give it to the to the trainee.
[42:42] SPEAKER_00: So I think there's a huge amount that can be done with coaching.
[42:45] And,
[42:46] you know,
[42:46] it doesn't know anything about empathy,
[42:49] but it knows what empathy sounds like.
[42:52] SPEAKER_00: And therefore,
[42:54] SPEAKER_00: and,
[42:54] SPEAKER_00: you know,
[42:54] SPEAKER_00: when I talk to my palliative care colleagues about this,
[42:56] SPEAKER_00: in some ways they say,
[42:57] SPEAKER_00: you know,
[42:58] SPEAKER_00: we teach our trainees a formula.
[43:01] SPEAKER_00: for how they talk to a patient about terrible news.
[43:04] SPEAKER_00: And that may seem artificial, but of course it is.
[43:07] SPEAKER_00: We've learned that this is the best way to do this thing.
[43:10] SPEAKER_00: And the AI knows that.
[43:12] And so therefore it's capable of coaching and therefore capable
[43:14] of scaling coaching
[43:16] SPEAKER_00: in a way that it's hard to figure out with humans.
[43:18] SPEAKER_00: I guess the final point is,
[43:20] how many times do I go into a patient's room and watch my
[43:23] medical student interview
[43:24] SPEAKER_00: the patient?
[43:25] Most of the time I'm waiting till the student comes out and
[43:27] they present the case to me.
[43:29] SPEAKER_00: It'd be like they went into a concert hall,
[43:31] SPEAKER_00: played the piano,
[43:32] SPEAKER_00: and came out and told me how good it sounded.
[43:34] SPEAKER_00: Whereas the AI could be listening to them have that conversation and say,
[43:37] SPEAKER_00: you know,
[43:37] SPEAKER_00: when you said this,
[43:38] SPEAKER_00: it didn't work very well,
[43:39] SPEAKER_00: and I might try saying it a different way next time.
[43:41] SPEAKER_00: So I think it's pretty exciting.
[43:44] SPEAKER_01: I'm with you.
[43:46] SPEAKER_01: I'm totally with you on that.
[43:47] I brought that up in part because of connecting with the
[43:50] training of the future
[43:51] SPEAKER_01: physicians and how this could actually be a surprisingly helpful tool.
[43:57] SPEAKER_01: So on your back cover of your book, I wrote,
[44:01] SPEAKER_01: an engaging, hype-free case for informed optimism.
[44:05] SPEAKER_01: And indeed, that's what the book is.
[44:07] SPEAKER_01: And I think people who are into AI and medicine are going to learn a lot.
[44:11] SPEAKER_01: We've covered just a little smidgen of it.
[44:14] SPEAKER_01: But I think it's really been fun to talk with you and
[44:18] SPEAKER_01: It's really a nice contribution you've made to your thoughtfulness.
[44:22] SPEAKER_01: You're talking to,
[44:24] SPEAKER_01: you know,
[44:25] SPEAKER_01: 100 plus people in the field and some great anecdotes,
[44:29] SPEAKER_01: good storytelling.
[44:30] SPEAKER_01: You know, it really comes together well.
[44:32] SPEAKER_01: So thank you, Bob.
[44:34] SPEAKER_00: Thank you.
[44:36] SPEAKER_00: Coming from you, that means a ton.
[44:37] SPEAKER_00: And we remain super proud of everything you've done.
[44:39] SPEAKER_00: And I really appreciate the endorsement.
[44:42] SPEAKER_00: It was very helpful.
[44:43] SPEAKER_01: Yeah, well, you got a lot of them, that's for sure.
[44:45] SPEAKER_01: Now,
[44:46] I want to just think we have just about 700 people who took the
[44:49] time to join us in
[44:50] SPEAKER_01: the middle of the day.
[44:51] SPEAKER_01: Really appreciate that.
[44:53] SPEAKER_01: I'm sure if you have lingering questions, Bob has an avatar.
[45:01] SPEAKER_01: No, I'm sure you can find him if there's something.
[45:04] SPEAKER_01: I'm easy to find.
[45:06] SPEAKER_01: Yeah, very easy to find.
[45:08] So Bob,
[45:09] really enjoyed it and look forward to further discussions with
[45:12] you.
[45:12] SPEAKER_02: Thank you, Eric.
[45:13] SPEAKER_02: It was a real pleasure.
[45:15] SPEAKER_02: Thanks, everybody.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.