Coauthor roundtable: Reflecting on healthcare economics, biomedical research, and medical education

00:00:04.600 --> 00:00:05.904
[MUSIC]    

00:00:05.904 --> 00:00:06.882
[BOOK PASSAGE]

00:00:06.883 --> 00:00:13.840
PETER LEE: “As a society—indeed, as a species—we 
have a choice to make. Do we constrain or even  

00:00:13.840 --> 00:00:20.080
kill artificial intelligence out of fear of its 
risks and obvious ability to create new harms? Do  

00:00:20.080 --> 00:00:26.080
we submit ourselves to Al and allow it to freely 
replace us, make us less useful and less needed?  

00:00:26.080 --> 00:00:31.520
Or do we start, today, shaping our Al future 
together, with the aspiration to accomplish things  

00:00:31.520 --> 00:00:42.189
that humans alone, and Al alone, can't do but that 
humans+Al can? The choice is in our hands … .”

00:00:42.189 --> 00:00:42.240
[END OF BOOK PASSAGE]
[THEME MUSIC]

00:00:42.240 --> 00:00:55.440
This is The AI Revolution in Medicine, 
Revisited. I’m your host, Peter Lee.    

00:00:55.440 --> 00:01:01.200
Shortly after OpenAI's GPT-4 was publicly 
released, Carey Goldberg, Dr. Zak Kohane,  

00:01:01.200 --> 00:01:06.640
and I published The AI Revolution in Medicine 
to help educate the world of healthcare and  

00:01:06.640 --> 00:01:11.120
medical research about the transformative 
impact this new generative AI technology  

00:01:11.120 --> 00:01:16.240
could have. But because we wrote the 
book when GPT-4 was still a secret,  

00:01:16.240 --> 00:01:23.680
we had to speculate. Now, two years later, what 
did we get right, and what did we get wrong?     

00:01:23.680 --> 00:01:28.320
In this series, we’ll talk to clinicians, 
patients, hospital administrators,  

00:01:28.320 --> 00:01:41.020
and others to understand the reality of AI 
in the field and where we go from here.      

00:01:41.020 --> 00:01:41.040
[THEME MUSIC FADES]

00:01:41.040 --> 00:01:44.080
The book passage I read at 
the top is from the epilogue,  

00:01:44.080 --> 00:01:47.760
and I think it’s a truly fitting 
closing sentiment for the conclusion  

00:01:47.760 --> 00:01:52.400
of this podcast series—because it 
calls back to the very beginning.

00:01:52.400 --> 00:01:59.200
As I’ve mentioned before, Carey, Zak, and I wrote 
The AI Revolution in Medicine as a guide to help  

00:01:59.200 --> 00:02:05.760
answer these big questions, particularly as they 
pertain to medicine. You know, we wrote the book  

00:02:05.760 --> 00:02:13.760
to empower people to make a choice about AI’s 
development and use. Well, have they? Have we?

00:02:13.760 --> 00:02:20.240
Perhaps we’ll need more time to tell. But over the 
course of this podcast series, I’ve had the honor  

00:02:20.240 --> 00:02:25.920
of speaking with folks from across the healthcare 
ecosystem. And my takeaway? They’re all committed  

00:02:25.920 --> 00:02:33.200
to shaping AI into a tool that can improve the 
industry for practitioners and patients alike.

00:02:33.200 --> 00:02:37.280
In this final episode, I’m thrilled 
to welcome back my coauthors,  

00:02:37.280 --> 00:02:44.480
Carey Goldberg and Dr. Zak Kohane. We’ll examine 
the insights from the second half of the season.

00:02:44.480 --> 00:02:50.080
[TRANSITION MUSIC]

00:02:50.080 --> 00:02:53.984
Carey, Zak—it’s really great 
to have you here again!

00:02:53.984 --> 00:02:54.868
CAREY GOLDBERG: Hey, Peter!

00:02:54.868 --> 00:02:55.760
ZAK KOHANE: Hi, Peter.

00:02:55.760 --> 00:03:03.600
LEE: So this is the second roundtable. And just to 
recap, you know, we had several early episodes of  

00:03:03.600 --> 00:03:11.360
the podcast where we talked to some doctors, 
some technology developers, some people who  

00:03:11.360 --> 00:03:19.760
think about regulation and public policy, patient 
advocates, a venture capitalist who invests in,  

00:03:19.760 --> 00:03:27.280
kind of, consumer and patient-facing 
medical ventures, and some bioethicists.

00:03:27.280 --> 00:03:31.920
And I think we had a great conversation 
there. I think, you know, it felt mostly  

00:03:31.920 --> 00:03:35.360
validating. A lot of the things that 
we predicted might happen happened,  

00:03:35.360 --> 00:03:43.280
and then we learned a lot of new things. 
But now we have five more episodes, and  

00:03:43.280 --> 00:03:49.920
the mix of kinds of people that we talk 
to here is different than the original.

00:03:49.920 --> 00:03:56.480
And so I thought it would be great for us 
to have a conversation and recap what we  

00:03:56.480 --> 00:04:02.880
think we heard from all of them. 
So let's just start at the top.

00:04:02.880 --> 00:04:07.120
So in this first episode in the 
second half of this podcast series,  

00:04:07.120 --> 00:04:13.840
we talked to economists Azeem Azhar and Ethan 
Mollick. And I thought those conversations were  

00:04:13.840 --> 00:04:20.000
really interesting. Maybe there were, kind of, 
two things, two main topics. One was just the  

00:04:20.000 --> 00:04:31.120
broader impact on the economy, on the cost 
of healthcare, on overall workforce issues.

00:04:31.840 --> 00:04:36.960
One of the things that I thought was really 
interesting was something that Ethan Mollick  

00:04:36.960 --> 00:04:43.520
brought up. And maybe just to refresh our 
memories, let's play this little clip from Ethan.

00:04:43.520 --> 00:04:45.760
ETHAN MOLLICK: So we’re in this really 
interesting period where there’s incredible  

00:04:45.760 --> 00:04:49.760
amounts of individual innovation in productivity 
and performance improvements in this field,  

00:04:49.760 --> 00:04:53.120
like very high levels of it. … We’re 
seeing that in nonmedical problems,  

00:04:53.120 --> 00:04:55.360
the same kind of thing, which is, you 
know, we’ve got research showing 20  

00:04:55.360 --> 00:04:58.720
and 40% performance improvements. … But 
then the organization doesn’t capture it;  

00:04:58.720 --> 00:05:01.680
the system doesn’t capture it. Because 
the individuals are doing their own work,  

00:05:01.680 --> 00:05:05.560
and the systems don’t have the ability 
to, kind of, learn or adapt as a result.

00:05:05.560 --> 00:05:11.200
LEE: So let me start with you, Zak. Does that make 
sense to you? Are you seeing something similar?

00:05:11.200 --> 00:05:17.600
KOHANE: I thought it was incredibly insightful 
because we discussed on our earlier podcast how  

00:05:18.240 --> 00:05:22.560
a chief AI officer in one 
of the healthcare hospitals,  

00:05:22.560 --> 00:05:28.080
in one of the healthcare systems, 
was highly regulating the use of AI,  

00:05:28.080 --> 00:05:34.400
but yet in her own practice on her smartphone 
was using all these AI technologies.

00:05:34.400 --> 00:05:42.296
And so it's insightful that on the one hand, 
she is increasing her personal productivity, …

00:05:42.296 --> 00:05:42.308
LEE: Right.

00:05:42.308 --> 00:05:50.720
KOHANE: … and perhaps she's increasing her 
quality of her care. But it's very hard for  

00:05:50.720 --> 00:05:56.800
the healthcare system to actually realize 
any gains. It's unlikely … let's put it this  

00:05:56.800 --> 00:06:03.480
way. It would be for her a defeat if they 
said, “Now you should see more patients.”

00:06:03.480 --> 00:06:04.072
LEE: Yes. [LAUGHS]

00:06:04.072 --> 00:06:10.400
KOHANE: Now, I'm not saying that won't 
happen. It could happen. But, you know,  

00:06:10.400 --> 00:06:17.360
gains of productivity are really at the individual 
level of the doctors. And that's why they're  

00:06:17.360 --> 00:06:25.440
adopting it. That's why the ambient dictation 
tools are so successful. But really turning it  

00:06:25.440 --> 00:06:32.640
into things that matter in terms of productivity 
for healthcare, namely making sure that patients  

00:06:32.640 --> 00:06:37.440
are getting healthy, requires that every piece 
of the puzzle works well together. You know,  

00:06:37.440 --> 00:06:42.320
it's well-tread ground to talk about how 
patients get very expensive procedures,  

00:06:42.320 --> 00:06:46.936
like a cardiac transplant, and then go home, 
and they’re not put on blood thinners …

00:06:46.936 --> 00:06:46.948
LEE: Right.

00:06:46.948 --> 00:06:52.560
KOHANE: … and then they get a stroke. You know, 
the chain is as strong as the weakest link. And  

00:06:52.560 --> 00:07:00.880
just having AI in one part of it is not going 
to do it. And so hospitals, I think, are doubly  

00:07:01.440 --> 00:07:10.560
burdened by the fact that, (A) they tend to not 
like innovation because they are high-revenue,  

00:07:10.560 --> 00:07:17.520
low-margin companies. But if they want it 
implemented effectively, they have to do it  

00:07:17.520 --> 00:07:25.120
across the entire processes of healthcare, which 
are vast and not completely under their control.

00:07:25.120 --> 00:07:28.320
LEE: Yeah. Yep. You know, that was Sara Murray,  

00:07:28.320 --> 00:07:32.160
who's the chief health AI 
officer at UC San Francisco.

00:07:32.160 --> 00:07:36.320
And then, you know, Carey, remember, 
we were puzzled by Chris Longhurst’s  

00:07:36.320 --> 00:07:44.880
finding in a controlled study that the, you 
know, having an AI respond to patient emails  

00:07:44.880 --> 00:07:52.000
didn't seem to lead to any, I guess you would 
call it, productivity benefits. I remember we  

00:07:52.000 --> 00:07:57.360
were both kind of puzzled by that. I wonder if 
that's related to what Ethan is saying here.

00:07:57.360 --> 00:08:02.160
GOLDBERG: I mean, possibly, but I 
think we've seen since then that  

00:08:02.160 --> 00:08:06.560
there have been multiple studies 
showing that in fact using AI can  

00:08:06.560 --> 00:08:11.920
be extremely effective or helpful, 
even, for example, for diagnosis.

00:08:11.920 --> 00:08:16.320
And so I find just from the patient point of view,  

00:08:16.320 --> 00:08:22.240
it kind of drives me crazy that you have 
individual physicians using AI because  

00:08:22.240 --> 00:08:26.560
they know that it will improve the care 
that they're offering. And yet you don't  

00:08:26.560 --> 00:08:31.520
have their institutions kind of stepping up 
and saying, “OK, these are the new norms.”

00:08:31.520 --> 00:08:34.720
By the way, Ethan Mollick is a 
national treasure, right. Like,  

00:08:34.720 --> 00:08:39.335
he is the classic example of someone 
who just stepped up at this moment …

00:08:39.335 --> 00:08:39.346
LEE: Yeah.

00:08:39.346 --> 00:08:44.080
GOLDBERG: … when we saw this extraordinary 
technological advance. And he's not only stepping  

00:08:44.080 --> 00:08:50.320
up for himself. He's spreading the word to the 
masses that this is what these things can do.

00:08:50.320 --> 00:08:54.800
And so it's frustrating to see the 
institutions not stepping up and  

00:08:54.800 --> 00:08:57.360
instead the individual doctors having to do it.

00:08:57.360 --> 00:09:00.000
KOHANE: But he made another 
very interesting point,  

00:09:00.000 --> 00:09:07.360
which was that the reason that he could 
be so informative to not only the public  

00:09:07.360 --> 00:09:13.360
but practitioners of AI is these things would 
emerge out of the shop, and they would not be  

00:09:15.200 --> 00:09:20.240
aged too long, like a fine wine, before 
they were just released to the public.

00:09:20.240 --> 00:09:28.480
And so he was getting exposure to these models 
just weeks after some of the progenitors had first  

00:09:28.480 --> 00:09:34.160
seen it. And therefore, because he's actually 
a really creative person in terms of how he  

00:09:34.160 --> 00:09:44.400
exercises models, he sees uses and problems 
very early on. But the point is institutions,  

00:09:44.400 --> 00:09:50.080
think about how much they are disadvantaged. 
They're not Ethan Mollick. They're not the  

00:09:50.080 --> 00:09:55.920
progenitors. So they're even further behind. 
So it's very hard. If you talk to most of the  

00:09:55.920 --> 00:10:01.720
C-suite of hospitals, they'd be delighted to 
know as much about the impact as Ethan Mollick.

00:10:01.720 --> 00:10:08.560
LEE: Yeah. By the way, you know, I picked 
out this quote because within Microsoft,  

00:10:08.560 --> 00:10:12.400
and I suspect every other software company, 
we're seeing something very similar,  

00:10:12.400 --> 00:10:19.360
where individual programmers are 20 to 30% more 
productive just in the number of lines of code  

00:10:19.360 --> 00:10:23.840
they write per day or the number of pull 
requests per week. Any way you measure it,  

00:10:23.840 --> 00:10:31.680
it's very consistent. And yet by the time you get 
to, say, a 25-person software engineering team,  

00:10:31.680 --> 00:10:37.360
the productivity of that whole 
team isn't 25% more productive.

00:10:37.360 --> 00:10:40.640
Now, that is starting to change because 
we're starting to figure out that, well,  

00:10:40.640 --> 00:10:48.400
maybe we should reshape how the team operates. 
And there's more of an orientation towards having,  

00:10:48.400 --> 00:10:53.040
you know, smaller teams of full-stack 
developers. And then you start to see  

00:10:53.040 --> 00:11:00.000
the gains. But if you just keep the team 
organized in the usual way, there seems to  

00:11:00.000 --> 00:11:04.320
be a loss. So there's something about what Ethan 
was saying that resonated very strongly with me.

00:11:04.320 --> 00:11:08.240
GOLDBERG: But I would argue that it's not 
just productivity we're talking about. There's  

00:11:08.240 --> 00:11:14.160
a moral imperative to improve the care. 
And if you have tools that will do that,  

00:11:14.160 --> 00:11:17.040
you should be using them or trying harder to.

00:11:17.040 --> 00:11:18.160
LEE: Right. Yep.

00:11:18.160 --> 00:11:25.520
KOHANE: I think, yes, first of all, 
absolutely you would. Unfortunately,  

00:11:25.520 --> 00:11:29.920
most of the short-term productivity 
measures will not measure improvements  

00:11:29.920 --> 00:11:35.520
in the quality of care because it takes 
a long time to die even with bad care.

00:11:35.520 --> 00:11:40.400
And so that doesn't show up right away. 
But I think what Peter just said actually  

00:11:40.400 --> 00:11:49.120
came across in several of the podcasts, 
which is that it's very tricky trying to  

00:11:49.120 --> 00:11:53.040
shoehorn these things into making what 
we're already doing more productive.

00:11:53.040 --> 00:11:54.720
GOLDBERG: Yeah. Existing structures.

00:11:54.720 --> 00:11:59.120
KOHANE: Yeah. And I know, Carey, 
that you've raised this issue  

00:11:59.680 --> 00:12:08.160
many times. But it really calls into question, 
what should we be doing with our time with  

00:12:08.160 --> 00:12:16.640
doctors? And they are a scarce resource. And 
what is the most efficient way to use them?

00:12:16.640 --> 00:12:18.640
You know, I remember we [The New England Journal 
of Medicine AI] published a paper of someone who  

00:12:18.640 --> 00:12:25.680
was able to use AI to increase the throughput 
of their emergency room by actually more  

00:12:25.680 --> 00:12:34.320
appropriately having the truly sick people in the 
sick queue, in the triage queue, for urgent care.

00:12:34.320 --> 00:12:38.240
And so I think we're going to have 
to think that way more broadly,  

00:12:38.240 --> 00:12:43.680
about we don't have to now look at 
every patient as an unknown with  

00:12:44.240 --> 00:12:48.880
maybe a few pointers on diagnosis. We 
can have a fairly extensive profiling.

00:12:48.880 --> 00:12:53.520
And I know that colleagues in Clalit 
[Health Services] in Israel, for example,  

00:12:53.520 --> 00:12:58.720
are using the overall trajectory of the 
patient and some considerations about  

00:12:58.720 --> 00:13:01.840
utilities to actually figure 
out who to see next week.

00:13:01.840 --> 00:13:08.000
LEE: Yeah, you know, what you said brings up 
another maybe connection to one thing that we  

00:13:08.000 --> 00:13:15.040
see also in software development. And it relates 
to also what we were discussing earlier: about the  

00:13:15.040 --> 00:13:21.120
last thing a doctor wants is to have a tool that 
allows them to see even yet more patients per day.

00:13:21.680 --> 00:13:25.760
So in software development, 
there's always this tension. Like,  

00:13:25.760 --> 00:13:29.200
how many lines of code can you write per 
day? That's one productivity measure.

00:13:29.920 --> 00:13:35.840
But sometimes we're taught, well, 
don't write more lines of code per day,  

00:13:35.840 --> 00:13:42.240
but make sure that your code is well structured. 
Take the time to document it. Make sure it's fully  

00:13:42.240 --> 00:13:46.080
commented. Take the time to talk to your 
fellow software engineering team members  

00:13:46.080 --> 00:13:50.560
to make sure that it's well coordinated. 
And in the long run, even if you're writing  

00:13:50.560 --> 00:13:55.920
half the number of lines of code per day, the 
software process will be far more efficient.

00:13:55.920 --> 00:14:00.160
And so I've wondered whether there's a 
similar thing where doctors could see  

00:14:00.160 --> 00:14:06.560
20% fewer patients in a day, but if they take 
the time and also had AI help to coordinate,  

00:14:06.560 --> 00:14:12.240
maybe a patient's journey might be half 
as long. And therefore, the health system  

00:14:12.240 --> 00:14:16.840
would be able to see twice as many patients 
in a year's period or something like that.

00:14:16.840 --> 00:14:21.840
KOHANE: So I think you've 
“nerd sniped” me because you  

00:14:21.840 --> 00:14:28.240
[LAUGHTER]—which is all too easy—but I 
think there's a central issue here. And  

00:14:28.240 --> 00:14:32.640
I think this is the stumbling block between 
what Ethan's telling us about between the  

00:14:32.640 --> 00:14:37.600
individual productivity and the larger 
productivity, is the team's productivity.

00:14:37.600 --> 00:14:43.600
And there is actually a good analogy 
in computer science and that's,  

00:14:43.600 --> 00:14:45.600
uh, Brooks’s “mythical man-month,” ...

00:14:45.600 --> 00:14:46.720
LEE: Yes, exactly.

00:14:46.720 --> 00:14:50.400
KOHANE: … where he shows how you 
can have more and more resources,  

00:14:50.400 --> 00:14:56.400
but when the coordination starts failing, because 
you have so many, uh, individuals on the team,  

00:14:56.400 --> 00:15:03.360
you start falling apart. And so even if the, 
uh, individual doctors get that much better,  

00:15:03.360 --> 00:15:06.400
yeah, they take better care of 
patients, make less stupid things.

00:15:06.400 --> 00:15:11.680
But in terms of giving the “I get you into 
the emergency room, and I get you out of  

00:15:11.680 --> 00:15:15.600
a hospital as fast as possible, as safely 
as possible, as effectively as possible,”  

00:15:15.600 --> 00:15:21.920
that's teamwork. And we don't do it. And we're 
not really optimizing our tools for that.

00:15:21.920 --> 00:15:26.560
GOLDBERG: And just to throw in a little reality 
check, I'm not aware of any indication yet that  

00:15:26.560 --> 00:15:32.635
AI is in any way shortening medical journeys 
or making physicians more efficient. Yet …

00:15:32.635 --> 00:15:32.946
LEE: Right.

00:15:32.946 --> 00:15:33.840
GOLDBERG: … at least. Yeah.

00:15:33.840 --> 00:15:38.400
LEE: Yes. So I think, you know, with respect 
to our book, critiquing our book, you know,  

00:15:38.400 --> 00:15:42.880
I think it's fair to say we were fairly 
focused or maybe even fixated on the  

00:15:42.880 --> 00:15:51.360
individual doctor or nurse or patient, and we 
didn't really, at least I never had a time where  

00:15:51.360 --> 00:15:57.480
I stepped back to think about the whole care 
coordination team or the whole health system.

00:15:57.480 --> 00:16:02.800
KOHANE: And I think that's right. It's 
because, first of all, you weren’t thinking  

00:16:02.800 --> 00:16:06.560
about it? It's not what we're taught in 
medical school. We're not taught to talk  

00:16:06.560 --> 00:16:11.840
about team communication excellence. 
And I think it's absolutely essential.

00:16:12.800 --> 00:16:17.200
There’s a … what’s the … there was 
an early … [Terry] Winograd. And  

00:16:17.200 --> 00:16:22.720
he was trying to capture what are the 
different kinds of actions related to  

00:16:22.720 --> 00:16:27.040
pronouncements that you could expect and how could 
AI use that. And that was beginning to get at it.

00:16:27.040 --> 00:16:33.200
But I actually think this is dark matter of 
human organizational technology that is not  

00:16:33.200 --> 00:16:38.480
well understood. And our products don't do 
well. You know, we can talk about all the  

00:16:38.480 --> 00:16:44.859
groupware things that are out there. But 
they all don't quite get to that thing.

00:16:44.859 --> 00:16:44.872
LEE: Right.

00:16:44.872 --> 00:16:48.960
KOHANE: And I can imagine an 
AI serving as a team leader,  

00:16:48.960 --> 00:16:54.400
a really active team leader, a real 
quarterback of, let's say, a care team.

00:16:54.400 --> 00:16:58.800
LEE: Well, in fact, you know, we have been 
trying to experiment with this. My colleague,  

00:16:58.800 --> 00:17:03.920
Matt Lungren, who was also one of the interviewees 
early on, has been working with Stanford  

00:17:04.960 --> 00:17:12.480
Medicine on a tumor board AI agent—something 
that would facilitate tumor board meetings.

00:17:12.480 --> 00:17:17.760
And the early experiences are pretty 
interesting. Whether it relates to  

00:17:17.760 --> 00:17:22.480
efficiency or productivity I think remains to 
be seen, but it does seem pretty interesting.

00:17:22.480 --> 00:17:23.440
But let's move on.

00:17:23.440 --> 00:17:24.706
GOLDBERG: Well, actually, Peter, …

LEE: Oh, go ahead.

00:17:24.706 --> 00:17:26.942
GOLDBERG: … if you're willing 
to not quite move on yet …

00:17:26.942 --> 00:17:27.600
LEE: [LAUGHS] All right.

00:17:27.600 --> 00:17:33.200
GOLDBERG: … this kind of segues into one of, 
I think, the most provocative questions that  

00:17:33.200 --> 00:17:40.160
arose in the course of these episodes and 
that I'd love to have you answer, which was,  

00:17:40.160 --> 00:17:46.160
remember, it was a question at a gathering 
that you were at, and you were asked, “Well,  

00:17:46.160 --> 00:17:53.040
you're focusing a lot on potential AI effects 
on individual patient and physician experiences.  

00:17:53.040 --> 00:17:59.360
But what about the revolution, right? What 
about, like, can you be more big-picture and  

00:17:59.360 --> 00:18:07.360
envision how generative AI could actually, kind 
of, overturn or fix the broken system, right?”

00:18:07.360 --> 00:18:10.840
I'm sure you've thought about that 
a lot. Like, what's your answer?

00:18:10.840 --> 00:18:17.360
LEE: You know, I think ultimately, it will 
have to. For it to really make a difference,  

00:18:17.360 --> 00:18:23.600
I think that the normal processes, our normal 
concept of how healthcare is delivered—how  

00:18:23.600 --> 00:18:27.200
new medical discoveries are made and brought  

00:18:27.200 --> 00:18:32.000
into practice—I think those things 
are going to have to change a lot.

00:18:32.800 --> 00:18:38.160
You know, one of the things I think about a 
lot right at the moment is, you know, we tend  

00:18:38.160 --> 00:18:45.920
to think about, let's say, medical diagnosis 
as a problem-solving exercise. And I think,  

00:18:46.720 --> 00:18:51.520
at least at the Kaiser Permanente 
School of Medicine, the instruction  

00:18:51.520 --> 00:18:58.000
really treats it as a kind of detective 
thing based on a lot of knowledge about  

00:18:58.000 --> 00:19:01.680
biology and biomedicine and 
human condition, and so on.

00:19:01.680 --> 00:19:08.480
But there's another way to think about it, 
given AI, which is when you see a patient and  

00:19:08.480 --> 00:19:14.160
you develop some data, maybe through a physical 
exam, labs, and so on, you can just simply ask,  

00:19:14.160 --> 00:19:21.040
“You know, what did the 500 other people who are 
most similar to this experience, how were they  

00:19:21.040 --> 00:19:25.040
diagnosed? How were they treated? What were 
their outcomes? What were their experiences?”

00:19:25.040 --> 00:19:28.960
And that's really a fundamentally 
different paradigm. And it just  

00:19:28.960 --> 00:19:30.880
seems like at least the technical means  

00:19:31.520 --> 00:19:36.720
will be there. And by the way, that also then 
relates to [the questions]: “And what was most  

00:19:36.720 --> 00:19:43.600
efficacious cost-wise? What was most efficient in 
terms of the total length of the patient journey?  

00:19:43.600 --> 00:19:48.800
How does this relate to my quality scores so I 
can get more money from Medicare and Medicaid?”

00:19:48.800 --> 00:19:54.400
All of those things, I think, you 
know, we're starting to confront.

00:19:54.400 --> 00:19:56.960
One of the other episodes that 
we're going to talk about,  

00:19:56.960 --> 00:20:03.760
was my interview with two medical students. 
Actually, thinking of a Morgan Cheatham as  

00:20:03.760 --> 00:20:09.440
just a medical student or medical resident 
[LAUGHTER] is a little strange. But he is.

00:20:10.160 --> 00:20:17.760
One of the things he talks about is the 
importance that he placed in his medical  

00:20:17.760 --> 00:20:25.520
training about adopting AI. So, Zak, 
I assume you see this also with some  

00:20:25.520 --> 00:20:29.440
students at Harvard Medical School. And 
the other medical student we interviewed,  

00:20:29.440 --> 00:20:33.760
Daniel Chen, seemed to indicate this, too, 
where it seems like it's the students who are  

00:20:33.760 --> 00:20:38.880
bringing AI into the medical education ahead 
of the faculty. Does that resonate with you?

00:20:38.880 --> 00:20:46.080
KOHANE: It absolutely resonates with me. 
There are students I run into who, honestly,  

00:20:46.080 --> 00:20:50.960
my first thought when I'm talking to them is, why 
am I teaching you [LAUGHTER], and why are you not  

00:20:50.960 --> 00:20:56.000
starting a big AI company, AI medicine company, 
now and really change healthcare instead of going  

00:20:56.000 --> 00:21:04.160
through the rest of the rigmarole? And I think 
broadly, higher education has a problem there,  

00:21:04.160 --> 00:21:11.760
which is we have not embraced, again, going 
back to Ethan, a lot of the tools that  

00:21:12.800 --> 00:21:17.280
can be used. And it's because we don't 
know necessarily the right way to teach  

00:21:17.280 --> 00:21:23.280
them. And so far, the only lasting heuristic 
seems to be: use them and use them often.

00:21:23.280 --> 00:21:29.920
And so it's an awkward thing, where the 
person who knows how to use the AI tools  

00:21:29.920 --> 00:21:34.800
now in the first-year medical school can 
teach themselves better and faster than  

00:21:34.800 --> 00:21:39.720
anybody else in their class who is just 
relying on the medical school curriculum.

00:21:39.720 --> 00:21:44.560
LEE: Now, the reason I brought up Morgan now 
after our discussion with Ethan Mollick is  

00:21:46.160 --> 00:21:49.820
Morgan also talked about AI 
collapsing medical specialties.

00:21:49.820 --> 00:21:53.360
GOLDBERG: Yes.
LEE: And so let's hear this snippet from him.

00:21:53.360 --> 00:21:58.160
MORGAN CHEATHAM: AI collapses medical specialties 
onto themselves, right. You have the canonical  

00:21:58.160 --> 00:22:03.200
example of the cardiologist, you know, arguing 
that we should diuresis and maybe the nephrologist  

00:22:03.200 --> 00:22:09.360
arguing that we should, you know, protect the 
kidneys. And how do two disciplines disagree  

00:22:09.360 --> 00:22:14.160
on what is right for the patient when in 
theory, there is an objective best answer  

00:22:14.160 --> 00:22:19.680
given that patient’s clinical status? … So I’m 
interested in this question of whether medical  

00:22:19.680 --> 00:22:24.720
specialties themselves need to evolve. And if we 
look back in the history of medical technology,  

00:22:24.720 --> 00:22:29.760
there are many times where a new technology 
forced a medical specialty to evolve.

00:22:29.760 --> 00:22:33.760
LEE: So on the specific 
question about specialties,  

00:22:33.760 --> 00:22:38.000
Zak, do you have a point of view? And 
let me admit, first of all, for us,  

00:22:38.000 --> 00:22:42.880
all three of us, we didn't have any clue 
about this in our book. I don't think.

00:22:42.880 --> 00:22:44.960
KOHANE: Not much. Not much of a clue.

00:22:46.400 --> 00:22:52.160
So I'm reminded of a New Yorker cartoon where 
you see a bunch of surgeons around the patient,  

00:22:52.160 --> 00:22:55.360
and someone says, “Is that a spleen?” And 
it says, “I don't know. I slept during  

00:22:55.360 --> 00:23:01.040
the spleen lecture,” [LAUGHTER] and … 
or “I didn't take the spleen course.”

00:23:01.040 --> 00:23:05.680
And yet when we measure things, we 
measure things much more than we  

00:23:05.680 --> 00:23:09.040
think we are doing. So for example, we 
[NEJM AI] just published a paper where  

00:23:09.680 --> 00:23:13.520
echocardiograms were being done. And it 
turns out those ultrasound waves just  

00:23:13.520 --> 00:23:19.360
happen to also permeate the liver. And you can 
actually diagnose on the way with AI all the  

00:23:19.360 --> 00:23:24.960
liver disease that is in—and treatable 
liver disease—that's in those patients.

00:23:24.960 --> 00:23:31.920
But if you're a cardiologist, “Liver? You know, 
I slept through liver lecture.” [LAUGHTER] And  

00:23:31.920 --> 00:23:44.800
so I do think that, (A) the natural, often 
guild/dollar-driven silos in medicine are  

00:23:44.800 --> 00:23:52.800
less obvious to AI, despite the fact that they 
do exist in departments and often in chapters.

00:23:52.800 --> 00:23:56.720
But Morgan's absolutely right. I 
can tell you as an endocrinologist,  

00:23:57.520 --> 00:23:59.760
if I have a child in the ICU, the endocrinologist,  

00:23:59.760 --> 00:24:04.160
the nephrologist, and the neurosurgeon 
will argue about the right thing to do.

00:24:04.160 --> 00:24:13.120
And so in my mind, the truly revolutionary thing 
to do is to go back to 1994 with Pete Szolovits,  

00:24:13.680 --> 00:24:19.760
the Guardian Angel Project. What I think 
you need is a process. And the process is  

00:24:19.760 --> 00:24:25.200
the quarterback. And the quarterback has 
only one job: take care of the patient.

00:24:25.200 --> 00:24:30.400
And it should be thinking all the time about 
the patient. What's the right thing? And can  

00:24:30.400 --> 00:24:34.480
be as school-marmish or not about, 
“Zak, you're eating this or that or  

00:24:34.480 --> 00:24:42.720
exercise or sleep,” but also, “Hey, surgeons and 
endocrinologists, you're talking about my host,  

00:24:42.720 --> 00:24:48.080
Zak. This is the right way because this 
problem and this problem and our best  

00:24:48.080 --> 00:24:52.240
evidence is this is the right way to get rid 
of the fluid. The other ways will kill him.”


00:24:52.240 --> 00:24:55.440
And I think you need an authoritative quarterback  

00:24:55.440 --> 00:24:59.040
that has the view of the others 
but then makes the calls.

00:24:59.040 --> 00:25:02.880
LEE: Is that quarterback going to be AI or human?

00:25:02.880 --> 00:25:09.120
KOHANE: Well, for the very lucky people, it'll 
be a human augmented by AI, super concierge.

00:25:09.120 --> 00:25:14.800
But I think we're running out of doctors. And so 
realistically, it's going to be an AI that will  

00:25:14.800 --> 00:25:20.480
have to be certified in very different ways, along 
the ways Dave Blumenthal says, essentially, trial  

00:25:20.480 --> 00:25:25.920
by fire. Like putting residents into clinics, 
we're going to be putting AIs into clinics.

00:25:26.720 --> 00:25:32.400
But what's worse, by the way, than the three 
doctors arguing about care in front of the patient  

00:25:32.400 --> 00:25:37.840
is, what happens so frequently, is then you see 
them outpatient, and each one of them gives you  

00:25:37.840 --> 00:25:46.320
a different set of decisions to make. Sometimes 
that actually interact pathologically, unhealthily  

00:25:46.320 --> 00:25:52.000
with each other. And only the very smart nurses 
or primary care physicians will actually notice  

00:25:52.000 --> 00:25:55.680
that and call, quote, a “family meeting,” or 
bring everybody in the same room to align them.

00:25:55.680 --> 00:26:03.200
LEE: Yeah, I think this idea of quarterback 
is really very, very topical right now because  

00:26:03.200 --> 00:26:07.520
there's so much intensity in the AI 
space around agents. And in fact,  

00:26:07.520 --> 00:26:16.000
you know, the Microsoft AI team under Mustafa 
Suleyman and Dominic King, Harsha Nori,  

00:26:16.000 --> 00:26:22.560
and team just recently posted a paper on something 
called sequential diagnosis, which is basically an  

00:26:22.560 --> 00:26:29.120
AI quarterback that is supposed to smartly consult 
with other AI specialties. And interestingly, one  

00:26:29.120 --> 00:26:35.323
of the AI agents is sort of the devil's advocate 
that's always criticizing and questioning things.

00:26:35.323 --> 00:26:38.320
GOLDBERG: That’s interesting.
LEE: And at least on very, very hard, rare cases,

00:26:38.320 --> 00:26:44.800
it can develop some impressive results. There's 
something to this that I think is emerging.

00:26:44.800 --> 00:26:50.080
GOLDBERG: And, Peter, Morgan said something 
that blew me away even more, which was,  

00:26:50.080 --> 00:26:55.360
well, why do we even need specialists if 
the reason for a specialist is because  

00:26:55.360 --> 00:27:00.320
there's so much medical knowledge that 
no single physician can know all of it,  

00:27:00.320 --> 00:27:06.240
and therefore we create specialists, but 
that limitation does not exist for AI.

00:27:06.240 --> 00:27:06.290
LEE: Yeah. Yeah.

00:27:06.290 --> 00:27:11.920
GOLDBERG: And so there he was kind of undermining 
this whole elaborate structure that has grown  

00:27:11.920 --> 00:27:16.240
up because of human limitations that 
may not ultimately need to be there.

00:27:16.240 --> 00:27:20.880
LEE: Right. So now that gives me a good 
segue to get back to our economist and  

00:27:20.880 --> 00:27:28.640
get to something that Azeem Azhar said. 
And so there's a clip here from Azeem.

00:27:28.640 --> 00:27:35.200
AZEEM AZHAR: We didn’t talk about, you know, AI 
in its ability to potentially do this, which is  

00:27:35.200 --> 00:27:44.960
to extend the clinician’s presence throughout the 
week. You know, the idea that maybe some part of  

00:27:44.960 --> 00:27:49.440
what the clinician would do if you could talk to 
them on Wednesday, Thursday, and Friday could be  

00:27:49.440 --> 00:27:56.800
delivered through an app or a chatbot just as a 
way of encouraging the compliance, which is often,  

00:27:56.800 --> 00:28:02.080
especially with older patients, one reason why 
conditions, you know, linger on for longer.

00:28:02.080 --> 00:28:09.200
LEE: And, you know, in the same conversation, he 
also talked about his own management of asthma and  

00:28:09.200 --> 00:28:17.200
the fact that he's been managing this for several 
decades and knows more than any other human being,  

00:28:17.200 --> 00:28:22.400
no matter how well medically trained, could 
possibly know. And it's also very highly  

00:28:22.400 --> 00:28:29.120
personalized. And it's not a big leap to imagine 
AI having that sort of lifelong understanding.

00:28:29.120 --> 00:28:36.480
KOHANE: So in fact, I want to give credit back 
to our book since you insulted us. [LAUGHTER] You  

00:28:36.480 --> 00:28:46.560
challenged us. You doubted us. We do have at the 
end of the book a AI which is helping this woman  

00:28:46.560 --> 00:28:53.022
manage her way through life. It's quarterbacking 
for the woman all these different services.

00:28:53.022 --> 00:28:53.880
LEE: Yes.
KOHANE: So there.

00:28:53.880 --> 00:28:58.480
LEE: Ah, you're right. Yes. In 
fact, it's very much, I think,  

00:28:58.480 --> 00:29:02.760
along the lines of the vision that 
Azeem laid out in our conversation.

00:29:02.760 --> 00:29:06.560
GOLDBERG: Yeah. It also reminded me of 
the piece Zak wrote about his mother at  

00:29:06.560 --> 00:29:11.120
one point when she was managing congestive 
heart failure and she needed to watch her  

00:29:11.120 --> 00:29:16.640
weight very carefully to see her fluid status. 
And absolutely, there's no … I see no reason  

00:29:16.640 --> 00:29:22.160
whatsoever why that couldn't be done with AI 
right now. Actually, although back then, Zak,  

00:29:22.160 --> 00:29:27.600
you were writing that it takes much more than 
an AI [LAUGHS] to manage such a thing, right?

00:29:27.600 --> 00:29:32.880
KOHANE: You need an AI that you can 
trust. Now, my mother was born in 1927,  

00:29:32.880 --> 00:29:36.160
and she'd learned through the school of hard 
knocks that you can't trust too many people,  

00:29:36.160 --> 00:29:40.040
maybe even not your son, MD, PhD [LAUGHTER].

00:29:40.040 --> 00:29:43.920
But what I've been surprised [by] is how, 
for example, how many people are willing  

00:29:43.920 --> 00:29:49.640
to trust and actually see effective use of 
AI as mental health counselors, for example.

00:29:49.640 --> 00:29:50.160
GOLDBERG: Yeah

00:29:50.160 --> 00:29:54.080
KOHANE: So it may in fact be that 
there's a generational thing going on,  

00:29:54.080 --> 00:29:58.160
and at least there'll be some very 
large subset of patients which will  

00:29:58.160 --> 00:30:02.000
be completely comfortable in ways that 
my mother would have never tolerated.

00:30:02.000 --> 00:30:05.920
LEE: Yeah. Now, I think we're starting 
to veer into some of the core AI.

00:30:05.920 --> 00:30:11.200
And so I think maybe one of the 
most fun conversations I had  

00:30:11.200 --> 00:30:18.640
was in the episode with both Sébastien Bubeck, 
my former colleague at Microsoft Research,  

00:30:18.640 --> 00:30:25.120
and now he's at OpenAI, and Bill Gates. 
And there was so much that was, I thought,  

00:30:25.120 --> 00:30:30.720
interesting there. And there was one point, 
I think that sort of touches tangentially  

00:30:30.720 --> 00:30:37.581
on what we were just conversing about, that 
Sébastien said. So let's hear this snippet.

00:30:37.581 --> 00:30:41.360
SÉBASTIEN BUBECK: And one example that I 
really like, a study that recently appeared  

00:30:41.360 --> 00:30:47.520
where … they were comparing doctors without and 
with ChatGPT. … So this was a set of cases where  

00:30:47.520 --> 00:30:55.600
the accuracy of the doctors alone was around 
75%. ChatGPT alone was 90%. … But then the  

00:30:55.600 --> 00:31:02.800
kicker is that doctors with ChatGPT was 80%. 
Intelligence alone is not enough. It’s also  

00:31:02.800 --> 00:31:08.000
how it’s presented, how you interact with it. 
And ChatGPT, it’s an amazing tool. Obviously,  

00:31:08.000 --> 00:31:12.800
I absolutely love it. But it’s not … you 
don’t want a doctor to have to type in,  

00:31:12.800 --> 00:31:16.800
you know, prompts and use it that 
way. It should be, as Bill was saying,  

00:31:16.800 --> 00:31:21.160
kind of running continuously in the 
background, sending you notifications.

00:31:21.160 --> 00:31:28.480
LEE: So I thought Sébastien was saying something 
really profound, but I haven't been able to quite  

00:31:28.480 --> 00:31:34.080
decide or settle in my mind what it is. 
What do you make of what Seb just said?

00:31:34.080 --> 00:31:41.120
KOHANE: I think it's context. I think that 
it requires an enormous amount of energy,  

00:31:41.120 --> 00:31:46.080
brain energy, to actually correctly 
provide the context that you want  

00:31:46.080 --> 00:31:51.920
this thing to work on. And it's only going 
to really feel like we're in a different  

00:31:51.920 --> 00:31:57.600
playing field when it's listening all 
the time, and it just steps right in.

00:31:57.600 --> 00:32:05.120
There is an advantage that, for example, a good 
programmer can have in prompting Cursor or any  

00:32:05.120 --> 00:32:12.400
of these tools to do so. But it takes effort. 
And I think being in the conversation all the  

00:32:12.400 --> 00:32:18.880
time so that you understand the context in the 
widest possible way is incredibly important. And  

00:32:18.880 --> 00:32:26.320
I think that's what Seb is getting at, which 
is if we spoon feed these machines, yes, 90%.

00:32:27.120 --> 00:32:31.760
But then, talking to a human being who then 
has to interact and gets distracted from  

00:32:31.760 --> 00:32:38.400
whatever flow they're in and maybe even makes 
them feel like an early bicycle rider who all  

00:32:38.400 --> 00:32:43.200
of a sudden realizes, “I'm balancing on two 
wheels—oh no!” And they fall over. You know,  

00:32:43.200 --> 00:32:46.960
there's that interaction which 
is negatively synergistic.

00:32:46.960 --> 00:32:54.800
And so I do think it's a very hard human-computer 
engineering problem. How do we make these two  

00:32:54.800 --> 00:33:00.720
agents, human and computational, work in 
an ongoing way in the flow? I don't think  

00:33:00.720 --> 00:33:05.600
I'm seeing anything that's particularly new. And 
the things that you're beginning to hint about,  

00:33:05.600 --> 00:33:12.475
Peter, in terms of agentic coordination, 
I think we'll get to some of that. 33:25

00:33:12.475 --> 00:33:18.960
LEE: Yeah. Carey, does this give you any 
pause? The kind of results that … they're  

00:33:18.960 --> 00:33:24.480
puzzling results. I mean, the idea of 
doctors with AI seeming at least in  

00:33:24.480 --> 00:33:30.720
this one test—it's just one test—but it's 
odd that it does worse than the AI alone.

00:33:30.720 --> 00:33:36.560
GOLDBERG: Yes. I would want to understand more 
about the actual conditions of that study.

00:33:37.200 --> 00:33:43.440
From what Bill Gates said, I was most 
struck by the question of resource-poor  

00:33:43.440 --> 00:33:48.080
environments. That even though this was 
absolutely one of the most promising,  

00:33:48.080 --> 00:33:54.160
brightest perspectives that we highlighted in 
the book, we still don't seem to be seeing a  

00:33:54.160 --> 00:34:00.560
lot of use among the one half of humanity 
that lacks decent access to healthcare.

00:34:00.560 --> 00:34:04.720
I mean, there are access problems everywhere, 
including here in the United States. And it  

00:34:04.720 --> 00:34:10.720
is one of the most potentially promising uses of 
AI. And I thought if anyone would know about it,  

00:34:10.720 --> 00:34:14.280
he would with the work that 
the Gates Foundation does.

00:34:14.280 --> 00:34:20.240
LEE: You know, I think both you and Bill, 
I felt, are really simpatico. You know,  

00:34:20.240 --> 00:34:25.760
Bill expressed genuine surprise that more 
isn't happening yet. And it really echoed,  

00:34:25.760 --> 00:34:33.600
in fact, maybe even using some of the exact same 
words that you've used. And so two years on,  

00:34:33.600 --> 00:34:38.560
you've expressed repeatedly expecting 
to have seen more out in the field by  

00:34:38.560 --> 00:34:44.560
now. And then I thought Bill was saying 
something in our conversation very similar.

00:34:44.560 --> 00:34:45.280
GOLDBERG: Yeah.

00:34:45.280 --> 00:34:51.120
LEE: You know, for me, I see it both ways. 
I see the world of medicine really moving  

00:34:51.120 --> 00:34:56.640
fast in confronting the reality of AI in 
such a serious way. But at the same time,  

00:34:56.640 --> 00:35:02.480
it's also hard to escape the feeling that 
somehow, we should be seeing even more.

00:35:02.480 --> 00:35:07.920
So it's an odd thing, a little bit paradoxical.

00:35:07.920 --> 00:35:11.760
GOLDBERG: Yeah. I think one thing that we 
didn't focus on hardly at all in the book  

00:35:11.760 --> 00:35:17.120
but that we are seeing is these companies 
rising up, stepping up to the challenge,  

00:35:17.120 --> 00:35:22.640
Abridge and OpenEvidence, and what 
Morgan describes as a new stack, right.

00:35:23.360 --> 00:35:26.240
So there is that on the flip side.

00:35:26.240 --> 00:35:32.160
LEE: Now, I want to get back to this 
thing that Seb was saying. And, you know,  

00:35:32.160 --> 00:35:39.120
I had to bring up the issue of sycophancy, which 
we discussed at our last roundtable also. But it  

00:35:39.120 --> 00:35:44.400
was particularly … at the time that Seb, Bill, 
and I had our conversation, OpenAI had just gone  

00:35:44.400 --> 00:35:52.720
through having to retract a fresh update of 
GPT-4o because it had become too sycophantic.

00:35:53.680 --> 00:36:02.160
So I can't escape the feeling that some of 
these human-computer interaction issues are  

00:36:02.160 --> 00:36:08.720
related to this tension between you want AI to 
follow your directions and be faithful to you,  

00:36:08.720 --> 00:36:16.640
but at the same time not agree with 
you so often that it becomes a fault.

00:36:16.640 --> 00:36:23.360
KOHANE: I think it’s asking the AI to 
enter into a fundamental human conundrum,  

00:36:23.360 --> 00:36:28.240
which is there are extreme versions of 
doublethink, and there's everyday things,  

00:36:28.240 --> 00:36:33.200
everyday asks of doublethink, which 
is how to be an effective citizen.

00:36:33.200 --> 00:36:39.440
And even if you're thinking, “Hmm. I'm thinking 
this. I'm just not going to say it because that  

00:36:39.440 --> 00:36:44.160
would be rude or counterproductive.” 
Or some of the official doublethinks,  

00:36:44.160 --> 00:36:50.640
where you're actually told you must say this, 
even if you think something else. And I think  

00:36:50.640 --> 00:36:59.120
we're giving a very tough mission for these 
things: be nice to the user and be useful.

00:36:59.120 --> 00:37:03.760
And, in education, where the 
thing is not always one in  

00:37:03.760 --> 00:37:09.040
the same. Sometimes you have to give a 
little tough love to educate someone,  

00:37:09.040 --> 00:37:15.680
and doing that well is both an art 
and it's also very difficult. And so,  

00:37:15.680 --> 00:37:21.360
you know, I'm willing to believe that the 
latest frontier models that have made the  

00:37:21.360 --> 00:37:27.755
news in the last month are very high-performing, 
but they're also all highlighting that tension …

00:37:27.755 --> 00:37:28.788
LEE: Yes.

00:37:28.788 --> 00:37:39.920
KOHANE: … that tension between behaving 
like a good citizen and being helpful.  

00:37:39.920 --> 00:37:45.440
And this gets back to what are the fundamental 
values that we hope these things are following.

00:37:45.440 --> 00:37:50.240
It's not, you know, “Are these things going 
to develop us into the paperclip factory?”  

00:37:50.240 --> 00:37:56.480
It's more of, “Which of our values are going to 
be elevated, and which one will be suppressed?”

00:37:56.480 --> 00:38:01.680
LEE: Well, since I criticized our book before, let 
me pat ourselves on the back this time because,  

00:38:01.680 --> 00:38:07.520
I think, pervasive throughout our book, 
we were touching on some of these issues.

00:38:07.520 --> 00:38:13.440
In fact, we started the book, you know, 
with GPT-4 scolding me for wanting it  

00:38:13.440 --> 00:38:20.960
to impersonate Zak. And there was the 
whole example of asking it to rewrite  

00:38:20.960 --> 00:38:27.360
a poem in a certain way, and it kind of 
silently just tried to slide, you know,  

00:38:27.360 --> 00:38:33.200
without me knowing, slide by without 
following through on the whole thing.

00:38:33.200 --> 00:38:38.960
And so that early version of GPT-4 was 
definitely not sycophantic at all. In fact,  

00:38:38.960 --> 00:38:43.672
it was just as prone to call you an idiot 
if it thought you were wrong. [LAUGHTER]

00:38:43.672 --> 00:38:47.984
KOHANE: I had some very testy conversations 
around my endocrine diagnosis with it. [LAUGHTER]

00:38:47.984 --> 00:38:51.440
GOLDBERG: Yeah. Well then, Peter, I would ask 
you, I mean last time I asked you about, well,  

00:38:51.440 --> 00:38:54.560
hallucinations, aren't those solvable? 
And this time I would ask you, well,  

00:38:54.560 --> 00:39:01.360
sycophancy, isn't that kind of like a dial 
you can turn? Like, is that not solvable?

00:39:01.360 --> 00:39:08.560
LEE: You know, I think there are several 
interlocking problems. But if we assume  

00:39:08.560 --> 00:39:12.320
superintelligence, even with superintelligence,  

00:39:12.320 --> 00:39:18.960
medicine is such an inexact science that 
there will always be situations that are  

00:39:18.960 --> 00:39:25.840
guesses that take into account other factors 
of a person's life, other value judgments,  

00:39:25.840 --> 00:39:30.800
exactly as Zak had pointed out in 
our previous roundtable conversation.

00:39:31.520 --> 00:39:36.960
And so I think there's always going to 
be an opening for either differences of  

00:39:36.960 --> 00:39:42.960
opinion or agreeing with you too much. 
And there are dangers in both cases.  

00:39:42.960 --> 00:39:49.760
And I think they'll always be present. I 
don't know that, at least in something as  

00:39:49.760 --> 00:39:56.240
inexact as medical science, I don't know 
that it'll ever be completely eliminated.

00:39:56.240 --> 00:39:59.520
KOHANE: And it's interesting because I was 
trying to think what's the right balance,  

00:39:59.520 --> 00:40:04.800
but there are patients who want to 
be told this is what you do. Whereas  

00:40:04.800 --> 00:40:09.840
there's other patients who want to go 
through every detail of the reasoning.

00:40:09.840 --> 00:40:13.520
And it's not a matter of education. 
It's really a temperamental,  

00:40:13.520 --> 00:40:21.995
personality issue. And so we're going to 
have to, I think, develop personalities …

00:40:21.995 --> 00:40:22.240
LEE: Yeah.

00:40:22.240 --> 00:40:30.000
KOHANE: … that are most effective for those 
different kinds of individuals. And so  

00:40:30.000 --> 00:40:34.160
I think that is going to be the real 
frontier. Having human values and  

00:40:34.160 --> 00:40:39.480
behaving in ways that are recognizable and 
yet effective for certain groups of patients.

00:40:39.480 --> 00:40:40.400
LEE: Yeah.

00:40:40.400 --> 00:40:44.080
KOHANE: And lots of deep questions, 
including how paternalistic do we want to be?

00:40:44.080 --> 00:40:47.920
LEE: All right, so we're getting into medical 
science and hallucination. So that gives me  

00:40:47.920 --> 00:40:54.720
a great segue to the conversations in 
the episode on biomedical research. And  

00:40:55.280 --> 00:41:00.720
one of the people that I interviewed was 
Noubar Afeyan from Moderna and Flagship  

00:41:00.720 --> 00:41:03.745
Pioneering. So let's listen to this snippet.

00:41:03.745 --> 00:41:09.440
NOUBAR AFEYAN: We, some hundred or so times a 
year, ask “what if” questions that lead us to  

00:41:09.440 --> 00:41:14.640
totally weird places of thought. We then try 
to iterate, iterate, iterate to come up with  

00:41:14.640 --> 00:41:19.520
something that’s testable. Then we go into a 
lab, and we test it. So in that world, right,  

00:41:19.520 --> 00:41:23.280
sitting there going, like, “How do I know this 
transformer is going to work?” The answer is,  

00:41:23.280 --> 00:41:28.400
“For what?” Like, it’s going to work to 
make something up … well, guess what? We  

00:41:28.400 --> 00:41:34.146
knew early on with LLMs that hallucination was 
a feature, not a bug for what we wanted to do. 

00:41:34.146 --> 00:41:41.600
LEE: [LAUGHS] So I think that really 
touches on just the fact that there's  

00:41:41.600 --> 00:41:46.480
so many unknowns and such lack 
of precision and exactness in  

00:41:46.480 --> 00:41:50.240
our understanding of human biology and 
of medicine. Carey, what do you think?

00:41:50.240 --> 00:41:57.280
GOLDBERG: I mean, I just have this emotional 
reaction, which is that I love the idea of AI  

00:41:57.280 --> 00:42:04.720
marching into biomedical science and everything 
from getting to the virtual cell eventually to,  

00:42:06.000 --> 00:42:09.040
Zak, I think it was a colleague 
of yours who recently published  

00:42:09.040 --> 00:42:16.320
about ... it was a new medication that 
had been sort of discovered by AI,  

00:42:16.320 --> 00:42:20.320
and it was actually testing out up to 
the phase II level or something, right?

00:42:20.320 --> 00:42:21.960
KOHANE: Oh, this is Marinka’s work.

00:42:21.960 --> 00:42:29.840
GOLDBERG: Yeah, Marinka, Marinka Zitnik. And … 
yeah. So, I mean, I think it avoids a lot of the,  

00:42:29.840 --> 00:42:35.200
sort of, dilemmas that are involved with safety 
and so on with AI coming into medicine. And it's  

00:42:35.200 --> 00:42:40.960
just the discovery process, which we all 
want to advance as quickly as possible.  

00:42:40.960 --> 00:42:45.760
And it seems like it actually has a great deal of 
potential that's already starting to be realized.

00:42:45.760 --> 00:42:48.080
LEE: Oh, absolutely.

00:42:48.640 --> 00:42:54.000
KOHANE: I love this topic. First of all, I 
thought, actually, I think Bill and Seb, actually,  

00:42:54.000 --> 00:43:00.080
had interesting things to say on that very topic, 
rationales which I had not really considered why,  

00:43:00.080 --> 00:43:05.520
in fact, things might progress faster in 
the discovery space than in the clinical  

00:43:05.520 --> 00:43:13.680
delivery space, just because we don't know in 
clinical medicine what we're trying to maximize  

00:43:13.680 --> 00:43:18.720
precisely. Whereas for a drug effect, we 
do know what we're trying to maximize.

00:43:18.720 --> 00:43:22.080
LEE: Well, in fact, I happened 
to save that snippet from Bill  

00:43:22.080 --> 00:43:26.468
Gates saying that. So let's cue that up.

00:43:26.468 --> 00:43:33.680
BILL GATES: I think it’s very much within 
the realm of possibility that the AI is  

00:43:33.680 --> 00:43:39.120
not only accelerating healthcare discovery 
but substituting for a lot of the roles of,  

00:43:39.120 --> 00:43:45.120
you know, “I’m an organic chemist,” or “I run 
various types of assays.” I can see those,  

00:43:45.120 --> 00:43:50.880
which are, you know, testable-output-type 
jobs but with still very high value,  

00:43:50.880 --> 00:43:56.120
I can see, you know, some replacement 
in those areas before the doctor.

00:43:56.120 --> 00:44:00.200
LEE: So, Zak, isn't that Bill 
saying exactly what you’re saying?

00:44:00.200 --> 00:44:05.360
KOHANE: That is my point. I have to say that 
this is another great bet; that either we're  

00:44:05.360 --> 00:44:09.920
all going to be surprised or a large group 
of people will be surprised or disappointed.

00:44:09.920 --> 00:44:14.240
There's still a lot of people in 
the sort of medicinal chemist,  

00:44:14.240 --> 00:44:18.000
trialist space who are still extremely 
skeptical that this is going to work.  

00:44:18.000 --> 00:44:23.840
And we haven't quite shown them yet that it is. 
Why have we not shown them? Because we haven't  

00:44:23.840 --> 00:44:28.880
gone all the way to a phase III study, which 
showed that the drug behaves as expected to,  

00:44:28.880 --> 00:44:34.640
is effective, and basically doesn't hurt 
people. That turns out to require a lot of  

00:44:34.640 --> 00:44:39.120
knowledge. I actually think we're getting 
there, but I understand the skepticism.

00:44:39.120 --> 00:44:41.720
LEE: Carey, what are your thoughts?

00:44:41.720 --> 00:44:45.920
GOLDBERG: Yeah. I mean, there will be no 
way around going through full-on clinical  

00:44:45.920 --> 00:44:49.520
trials for anything to ever reach 
the market. But at the same time,  

00:44:50.800 --> 00:44:56.320
you know, it's clearly very promising. And just 
to throw out something for the pure fun of it,  

00:44:56.320 --> 00:45:01.360
Peter, I saw … one of my favorite tweets recently 
was somebody saying, you know, isn't it funny how  

00:45:01.360 --> 00:45:07.760
computer science is actually becoming a lot more 
like biology in that it's just becoming empirical.

00:45:07.760 --> 00:45:11.600
It's like you just throw stuff at the AI and 
see what it does. [LAUGHTER] And I was like,  

00:45:11.600 --> 00:45:15.120
oh, yeah, that's what Peter was 
doing when we wrote the book. I mean,  

00:45:15.120 --> 00:45:18.720
he understood as many innards as 
anybody can. But at the same time,  

00:45:18.720 --> 00:45:25.240
it was a totally empirical exercise in seeing what 
this thing would do when you threw things at it.

00:45:25.240 --> 00:45:25.920
LEE: Right.

00:45:25.920 --> 00:45:28.389
GOLDBERG: So it's the new biology.

00:45:28.389 --> 00:45:34.320
LEE: Well, yeah. So I think we talked in 
our book about accelerating, you know,  

00:45:34.320 --> 00:45:41.520
biomedical knowledge and medical science. 
And that actually seems to be happening. And  

00:45:42.080 --> 00:45:46.320
I really had fun talking to 
Daphne Koller about some of  

00:45:46.320 --> 00:45:51.200
the accomplishments that she's made. And 
so here's a little snippet from Daphne.

00:45:51.200 --> 00:45:56.240
DAPHNE KOLLER: This will impact not only the 
early stages of which hypotheses we interrogate,  

00:45:56.240 --> 00:46:00.400
which molecules we move forward, but also 
hopefully at the end of the day, which molecule  

00:46:00.400 --> 00:46:06.240
we prescribe to which patient. And I think there’s 
been obviously so much narrative over the years  

00:46:06.240 --> 00:46:12.080
about precision medicine, personalized medicine, 
and very little of that has come to fruition, with  

00:46:12.080 --> 00:46:17.520
the exception of, you know, certain islands in 
oncology, primarily on genetically driven cancers.

00:46:17.520 --> 00:46:23.600
LEE: So, Zak, when I was listening to that, 
I was reminded of one of the very first  

00:46:23.600 --> 00:46:29.680
examples that you had where, you know, 
you had a very rare case of a patient,  

00:46:29.680 --> 00:46:38.480
and you're having to narrow down some pretty 
complex and very rare genetic conditions.  

00:46:38.480 --> 00:46:43.920
This thing that Daphne says, that seems to 
be the logical conclusion that everyone who's  

00:46:43.920 --> 00:46:51.160
thinking hard about AI and biology is coming 
to. Does it seem more real now two years on?

00:46:51.160 --> 00:46:57.680
KOHANE: It absolutely seems more real. Here's 
some sad facts. If you are at a cancer center,  

00:46:57.680 --> 00:47:04.640
you will get targeted therapies if you qualify 
for it. Outside cancer centers, you won't. And  

00:47:04.640 --> 00:47:08.960
it's not that the therapies aren't available. 
It's just that you won't have people thinking  

00:47:08.960 --> 00:47:14.240
about it in that way. And especially if you have 
some of the rare and more aggressive cancers,  

00:47:14.240 --> 00:47:20.320
if you're outside one of those cancer centers, 
you're at a significant disadvantage for  

00:47:20.960 --> 00:47:27.840
survival for that reason. And so anything 
that provides just the “simple,” in quotes,  

00:47:28.400 --> 00:47:35.600
dogged investigation of the targeted 
therapies for patients, it's a home run.

00:47:36.160 --> 00:47:47.280
So my late graduate student, Atul Butte, died 
recently at UCSF, where he was both a professor  

00:47:47.280 --> 00:47:53.200
and the leader of the Bakar Institute, and he 
was a Zuckerberg Chan Professor of Pediatrics.

00:47:53.200 --> 00:48:00.960
He was diagnosed with a rare tumor two years 
ago. His wife is a PhD biologist, and when he  

00:48:00.960 --> 00:48:07.520
was first diagnosed, she sent me the diagnosis and 
the mutations. And I don't know if you know this,  

00:48:07.520 --> 00:48:11.200
Peter, but this was still when we were writing 
the book and people didn't know about GPT-4.

00:48:11.200 --> 00:48:17.600
I put in those mutations into GPT-4 
and the diagnosis. And I said, “I'd  

00:48:17.600 --> 00:48:24.720
like to help treat my friend. What's the right 
treatment?” And GPT, to paraphrase, GPT-4 said,  

00:48:24.720 --> 00:48:28.560
“Before we start talking about treatment, are 
you sure this is the right diagnosis? Those  

00:48:28.560 --> 00:48:32.640
mutations are not characteristic 
for that tumor.” And he had been  

00:48:32.640 --> 00:48:39.120
misdiagnosed. And then they changed the 
diagnosis therapy and some personnel.

00:48:39.120 --> 00:48:45.680
So I don't have to hallucinate this. It's 
already happened, and we're going to need  

00:48:45.680 --> 00:48:54.080
this. And so I think targeted therapy for 
cancers is the most obvious use. And if God  

00:48:54.080 --> 00:49:00.160
forbid one of you has a family member who 
has cancer, it's moral malpractice not to  

00:49:00.160 --> 00:49:05.100
look at the genetics and run it past GPT-4 
and say, “What are the available therapies?”

00:49:05.100 --> 00:49:07.600
LEE: Yeah.
KOHANE: I really deeply believe that.

00:49:07.600 --> 00:49:11.600
LEE: Carey, I think one thing you've 
always said is that you're surprised  

00:49:11.600 --> 00:49:18.400
that we don't hear more stories along 
these lines. And I think you threw a  

00:49:18.400 --> 00:49:21.600
quote from Mustafa Suleyman back 
at me. Do you want to share that?

00:49:21.600 --> 00:49:26.720
GOLDBERG: Yes. Recently, I believe 
it was a Big Technology interview,  

00:49:26.720 --> 00:49:32.720
and the reporter asked Mustafa Suleyman, 
“So you guys are seeing 50 million queries,  

00:49:32.720 --> 00:49:37.200
medical queries, a day [to Copilot and 
Bing]. You know, how's that going?” And  

00:49:37.840 --> 00:49:45.040
I think I am a bit surprised that we're 
not seeing more stories of all types. Both  

00:49:45.040 --> 00:49:52.000
here's how it helped me and also here was maybe, 
you know, a suggestion that was not optimal.

00:49:52.000 --> 00:49:58.320
LEE: Yeah. I do think in our book, we 
did predict both positive and negative  

00:49:58.320 --> 00:50:05.920
outcomes of this. And it is odd. Atul was 
very open with his story. And of course,  

00:50:05.920 --> 00:50:11.040
he is such … he was such a prominent 
leader in the world of medicine.

00:50:11.040 --> 00:50:19.120
But I think I share your surprise, Carey. I 
expected by now that a lot more public stories  

00:50:19.120 --> 00:50:22.840
would be out. Maybe there is someone writing 
a book collecting these things, I don't know.

00:50:22.840 --> 00:50:26.630
KOHANE: Maybe someone called Carey 
Goldberg should write that book. [LAUGHTER]

00:50:26.630 --> 00:50:29.760
GOLDBERG: Write a book, maybe. 
I mean, we have Patients Use AI,  

00:50:29.760 --> 00:50:34.000
which is a wonderful blog by Dave 
deBronkart, the patient advocate.

00:50:34.000 --> 00:50:38.240
But I wonder if it's also something 
structural, like who would be or what  

00:50:38.240 --> 00:50:42.560
would be the institution that would be 
gathering these stories? I don’t know.

00:50:42.560 --> 00:50:43.108
LEE: Right.

00:50:43.108 --> 00:50:46.160
KOHANE: And that's the problem. You see, this goes 
back to the same problem that [Ethan] Mollick was  

00:50:46.160 --> 00:50:52.560
talking about. Individual doctors are using them. 
The hospital as a whole is not doing that. So it's  

00:50:52.560 --> 00:50:58.960
not judging the quality, as part of its quality 
metrics, of how good the AI is performing and  

00:50:58.960 --> 00:51:05.840
what new has happened. And the other audience, 
namely the patients, have no mechanism. There  

00:51:05.840 --> 00:51:11.680
is no mechanism to go to Better Business Bureau 
and say, “They screwed up,” or “This was great.”

00:51:11.680 --> 00:51:17.840
LEE: So now I want to get a little more 
futuristic. And this gets into whether  

00:51:17.840 --> 00:51:24.240
AI is really going to get almost to the ab 
initio understanding of human biology. And  

00:51:24.240 --> 00:51:30.880
so Eric Topol, who is one of the guests, 
spoke to this a bit. So let's hear this.

00:51:30.880 --> 00:51:36.400
LEE: So you talk about a virtual cell. 
Is that achievable within 10 years,  

00:51:36.400 --> 00:51:38.960
or is that still too far out? 

00:51:38.960 --> 00:51:45.040
ERIC TOPOL: No, I think within 10 years for sure. 
You know, the group that got assembled, that Steve  

00:51:45.040 --> 00:51:50.720
Quake pulled together, I think has 42 authors in 
a paper in Cell. The fact that he could get these  

00:51:50.720 --> 00:51:58.560
42 experts in life science and some in computer 
science to come together and all agree that not  

00:51:58.560 --> 00:52:04.160
only is this a worthy goal, but it’s actually 
going to be realized, that was impressive.

00:52:04.160 --> 00:52:10.080
LEE: You know, I have to say Eric's optimism 
took me aback. Just speaking as a techie,  

00:52:10.080 --> 00:52:16.320
I think I started off being optimistic: as 
soon as we can figure out molecular dynamics,  

00:52:16.320 --> 00:52:22.240
biology can be solved. And then you 
start to learn more about biochemistry,  

00:52:22.240 --> 00:52:30.560
about the human cell, and then you realize, oh, 
my God, this is just so vast and unknowable.  

00:52:30.560 --> 00:52:34.880
And now you have Eric Topol saying, 
“Well, in less than 10 years.”

00:52:34.880 --> 00:52:41.840
KOHANE: So what's delightful about this period 
is that those of us who are cautious were so  

00:52:41.840 --> 00:52:48.000
incredibly wrong about AI two years ago. 
[LAUGHTER] That's a true joy ... I mean,  

00:52:48.000 --> 00:52:55.040
absolute joy. It's great to have your 
futurism made much more positive.

00:52:56.000 --> 00:53:02.640
But I think that we're going 
from, you know, for example,  

00:53:02.640 --> 00:53:08.000
AlphaFold has had tremendous impact. 
But remember, that was built on years  

00:53:08.000 --> 00:53:12.640
of acquisition of crystallography data 
that was annotated. And of course,  

00:53:12.640 --> 00:53:16.960
the annotation process becomes less relevant as 
you go down the pipe, but it started from that.

00:53:16.960 --> 00:53:17.440
LEE: Yes.

00:53:17.440 --> 00:53:21.600
KOHANE: And there's lots of parts of the 
cell. So when people talk about virtual  

00:53:21.600 --> 00:53:26.240
cells—I don't mean to get too technical—mostly 
they're talking about perturbation of gene  

00:53:26.240 --> 00:53:32.880
expression. They're not talking about, 
“Oh, this is how the liposome and the  

00:53:34.320 --> 00:53:37.760
centrosome interact, and notice how 
the Golgi bodies bump into each other.”

00:53:37.760 --> 00:53:41.600
There's a whole bunch of other levels of 
abstraction we know nothing about. This  

00:53:41.600 --> 00:53:48.560
is a complex factory. And right now, we're 
sort of the level from code into loading code  

00:53:48.560 --> 00:53:53.200
into memory. We're not talking about how 
the rest of the robots work in that cell,  

00:53:53.200 --> 00:53:57.760
and how the rest of those robots work in the cell 
turns out to be pretty important to functioning.

00:53:57.760 --> 00:54:06.400
So I'd love to be wrong again. And in 10 years, oh 
yeah, not only, you know, our first in-human study  

00:54:06.400 --> 00:54:12.880
will be you, Dr. Zak. We're going put the drug 
because we fully simulated you. That'd be great.

00:54:12.880 --> 00:54:13.992
LEE: Yes.

00:54:13.992 --> 00:54:17.040
KOHANE: And, by the way, just to give 
people their due, there probably was a  

00:54:17.040 --> 00:54:22.000
lot of animal research that could be done 
in silico and that for various political  

00:54:22.000 --> 00:54:29.520
reasons we're now seeing happen. That's a 
good thing. But I think that sometimes it  

00:54:29.520 --> 00:54:36.080
takes a lot of hubris to get us where we need 
to get, but my horizon is not the same as his.

00:54:36.080 --> 00:54:43.200
LEE: So I guess I have to take this time to brag. 
Just recently out of our AI for Science team did  

00:54:43.200 --> 00:54:55.360
publish in Science a biological emulator that 
does pretty long timespan, very, very precise,  

00:54:55.360 --> 00:55:00.560
and very efficient molecular dynamics, 
biomolecular dynamics emulation. We call it  

00:55:00.560 --> 00:55:05.592
emulation because it's not simulating every single 
time step but giving you the final confirmations.

00:55:05.592 --> 00:55:07.661
KOHANE: That's an amazing result.

00:55:07.661 --> 00:55:07.672
LEE: Yeah.

00:55:07.672 --> 00:55:10.880
KOHANE: But … that is an amazing 
result. And you're doing it in  

00:55:10.880 --> 00:55:14.320
some very important interactions. 
But there's so much more to do.

00:55:14.320 --> 00:55:17.440
LEE: I know, and it's single 
molecules; it’s not even two  

00:55:17.440 --> 00:55:24.320
molecules. There's so much more to 
go for here. But on the other hand,  

00:55:24.880 --> 00:55:34.040
Eric is right, you know, 42 experts writing 
for Cell, you know, that's not a small matter.

00:55:34.040 --> 00:55:37.680
KOHANE: So I think sometimes you 
really need to drink your own  

00:55:38.720 --> 00:55:45.520
hallucinogens to actually succeed. Because 
remember, when the Human Genome Project was  

00:55:46.800 --> 00:55:49.520
launched, we didn't know how to sequence at scale.

00:55:49.520 --> 00:55:53.360
We said maybe we would get there. And 
then in order to get the right funding  

00:55:53.360 --> 00:55:59.360
and excitement and, I think, focus, 
we predicted that by early 2000s we'd  

00:55:59.360 --> 00:56:04.720
be transforming medicine. Has not 
happened yet. Things have happened,  

00:56:04.720 --> 00:56:12.080
but at a much slower pace. And we're 25 years 
out. In fact, we're 35 years out from the launch.

00:56:13.040 --> 00:56:15.440
But again, things are getting faster and faster.  

00:56:15.440 --> 00:56:20.400
Maybe the singularity is going to make a whole 
bunch of things easier. And GPT-6 will just say,  

00:56:20.400 --> 00:56:25.024
“Zak, you are such a pessimist. 
Let me show you how it's done.”

00:56:25.024 --> 00:56:25.040
GOLDBERG: Yeah.

00:56:25.040 --> 00:56:28.240
It really is a pessimism 
versus optimism. Like is it,  

00:56:28.240 --> 00:56:33.280
I mean, biology is such a bitch, right. 
[LAUGHTER] Can we actually get there?

00:56:33.280 --> 00:56:38.400
At the same time, everyone was surprised and 
blown away by the, you know, the quantum leap  

00:56:38.400 --> 00:56:44.080
of GPT-4. Who knows when enough data gets in 
there if we might not have a similar leap.

00:56:44.080 --> 00:56:45.280
LEE: Yeah. All right.

00:56:45.280 --> 00:56:51.920
So let's get back to healthcare delivery. 
Besides Morgan Cheatham, we talked to [a]  

00:56:51.920 --> 00:56:57.920
more junior medical student who's at the Kaiser 
Permanente School of Medicine, Daniel Chen. And,  

00:56:57.920 --> 00:57:03.040
you know, I asked him about this question 
of patients who come in armed [LAUGHS] with  

00:57:03.040 --> 00:57:08.240
a lot of their own information. 
Let's hear what he said about this.

00:57:08.240 --> 00:57:12.320
DANIEL CHEN: But for those that come in 
with a list, I sometimes sit down with them,  

00:57:12.320 --> 00:57:17.200
and we’ll have a discussion, honestly. … 
“I don’t think you have meningitis because,  

00:57:17.200 --> 00:57:21.520
you know, you’re not having a fever. Some of 
the physical exam maneuvers we did were also  

00:57:21.520 --> 00:57:26.480
negative. So I don’t think you have anything 
to worry about that,” you know. So I think  

00:57:26.480 --> 00:57:32.240
it’s having that very candid conversation with 
the patient that helps build that initial trust.

00:57:32.240 --> 00:57:37.760
LEE: So, Zak, as far as I can tell, Daniel 
and Morgan are figuring this out on their own  

00:57:37.760 --> 00:57:42.920
as medical students. I don't think this is 
part of the curriculum. Does it need to be?

00:57:42.920 --> 00:57:48.720
KOHANE: It's missing the bigger point. The 
incentives and economic forces are such that  

00:57:48.720 --> 00:57:58.240
even if you were Daniel, and things have not 
changed in terms of incentives, and it's 2030,  

00:57:58.240 --> 00:58:02.000
he still has to see this many patients in an hour.

00:58:02.000 --> 00:58:09.280
And sitting down, going over that with a patient, 
let's say some might need more ... in fact,  

00:58:09.280 --> 00:58:13.520
I think computer scientists are enriched for 
these sort of neurotic “explain [to] me why  

00:58:13.520 --> 00:58:18.320
this works,” when often the answer is, 
“I have no idea; empirically it does.”

00:58:19.520 --> 00:58:26.880
And patients in some sense deserve that 
conversation, and we're taught about joint  

00:58:26.880 --> 00:58:32.800
decision making, but in practice, there's a 
lot of skills that are deployed to actually  

00:58:32.800 --> 00:58:39.840
deflect so that you can get through the 
appointment and see enough patients per hour.

00:58:40.800 --> 00:58:46.560
And that's why I think that one of the 
central … another task for AI is how to  

00:58:46.560 --> 00:58:51.120
engage with patients to actually 
explain to them why their doctor  

00:58:51.120 --> 00:58:55.680
is doing what he's doing and perhaps ask 
the one or two questions that you should  

00:58:55.680 --> 00:59:00.061
be asking the doctor in order to reassure 
you that they're doing the right thing.

00:59:00.061 --> 00:59:00.072
LEE: Yeah.

00:59:00.072 --> 00:59:05.920
KOHANE: I just … right now, we are going to 
have less doctor time, not more doctor time.

00:59:05.920 --> 00:59:14.480
And so I've always been struck by the 
divide between medicine that we're taught  

00:59:14.480 --> 00:59:24.000
as it should be practiced as a gentle person's 
vocation or sport as opposed to assembly line,  

00:59:24.000 --> 00:59:27.920
heads down “you've got to see those 
patients by the end of the day” because,  

00:59:27.920 --> 00:59:30.560
otherwise, you haven't seen all 
the patients at the end of the day.

00:59:30.560 --> 00:59:35.520
LEE: Yeah. Carey, I've been dying to ask 
you this, and I have not asked you this  

00:59:35.520 --> 00:59:42.720
before. When you go see a doctor, are you 
coming in armed with ChatGPT information?

00:59:42.720 --> 00:59:48.080
GOLDBERG: I haven't needed to yet, but 
I certainly would. And also my reaction  

00:59:49.680 --> 00:59:57.200
to the medical student description was, I think 
we need to distinguish between the last 20 years,  

00:59:57.200 --> 01:00:03.200
when patients would come in armed with Google, and 
what they're coming in with now because at least  

01:00:03.200 --> 01:00:12.320
the experiences that I've witnessed, it is miles 
better to have gone back and forth with GPT-4 than  

01:00:12.320 --> 01:00:20.880
with, you know, dredging what you can from Google. 
And so I think we should make that distinction.

01:00:20.880 --> 01:00:25.520
And also, the other thing that most 
interested me was this question for  

01:00:25.520 --> 01:00:32.696
medical students of whether they should not 
use AI for a while so that they can learn …

01:00:32.696 --> 01:00:32.706
LEE: Yes.

01:00:32.706 --> 01:00:37.120
GOLDBERG: … how to think and similarly maybe 
don't use the automated scribes for a while  

01:00:37.120 --> 01:00:42.000
so they can learn how to do a note. And at 
what point should they then start being able  

01:00:42.000 --> 01:00:47.920
to use AI? And I suspect it's fairly early 
on that, in fact, they're going be using it  

01:00:47.920 --> 01:00:53.400
so consistently that there's not that much they 
need to learn before they start using the tools.

01:00:53.400 --> 01:00:59.520
LEE: These two students were incredibly 
impressive. And so I have wondered, you know,  

01:00:59.520 --> 01:01:04.800
if we got a skewed view of things. I mean, Morgan 
is, of course, a very, very impressive person.  

01:01:04.800 --> 01:01:12.880
And Daniel was handpicked by the dean of the 
medical school to be a subject of this interview.

01:01:12.880 --> 01:01:18.000
KOHANE: You know, we filter our students, by and 
large, I mean, there's exceptions, but students in  

01:01:18.000 --> 01:01:23.840
medical school are so starry eyed. And they are 
really ... they got into medical school—I mean,  

01:01:23.840 --> 01:01:27.040
some of them may have faked it—but a lot of 
them because they really wanted to do good.

01:01:27.040 --> 01:01:27.480
LEE: Right.

01:01:27.480 --> 01:01:30.640
KOHANE: And they really wanted to 
help. And so this is very constant  

01:01:30.640 --> 01:01:35.440
with them. And it's only when they're 
in the machine, past medical school,  

01:01:35.440 --> 01:01:39.360
that they realize, oh my God, this 
is a very, very different story.

01:01:39.360 --> 01:01:43.520
And I can tell you, because I teach a 
course in computational-enabled medicine,  

01:01:43.520 --> 01:01:49.280
so I get a lot of these nerd medical students, 
and I'm telling them, “You're going to experience  

01:01:49.280 --> 01:01:53.200
this. And you're going to say, ‘I'm not going 
to able to change medicine until I get enough  

01:01:53.200 --> 01:01:58.960
cred 10, 15 years from now, whereas I could start 
my own company and immediately change medicine.’”

01:01:58.960 --> 01:02:03.120
And increasingly I'm getting calls 
in like residency and saying, “Zak,  

01:02:03.120 --> 01:02:05.025
help me. How do I get out of this?”

01:02:05.025 --> 01:02:05.040
GOLDBERG: Wow.

01:02:05.040 --> 01:02:10.240
KOHANE: And so I think there's 
a real disillusionment of, like,  

01:02:10.240 --> 01:02:13.920
between what we're asking for people 
coming to medical school—we're looking  

01:02:13.920 --> 01:02:21.280
for a phenotype—and then we're disappointing 
them massively, not everywhere, but massively.

01:02:21.280 --> 01:02:29.760
And for me, it's very sad because among 
our best and brightest, and then because  

01:02:29.760 --> 01:02:38.080
of economics and expectations and the nature 
of the beast, they're not getting to enjoy the  

01:02:38.080 --> 01:02:44.000
most precious part of being a doctor, which is 
that real human connection, and longitudinality,  

01:02:44.000 --> 01:02:49.200
you know, the connection between the same doctor 
visit after visit, is more and more of a luxury.

01:02:49.200 --> 01:02:55.760
LEE: Well, maybe this gets us to the last 
episode, you know, where I talk to a former,  

01:02:55.760 --> 01:03:01.600
you know, state director of public health, 
Umair Shah, and with Gianrico Farrugia,  

01:03:01.600 --> 01:03:08.320
who's the CEO of Mayo Clinic. And I think 
if there's one theme that I took away from  

01:03:08.320 --> 01:03:14.160
those conversations is that we're not 
thinking broadly enough nor big enough.

01:03:14.720 --> 01:03:23.200
And so here's a little quote of exchange that 
Umair Shah, who was the former head of public  

01:03:23.200 --> 01:03:29.360
health in the State of Washington and prior 
to that in Harris County, Texas, and we had  

01:03:29.360 --> 01:03:35.520
a conversation about what techies tend to focus 
on when they're thinking about AI and medicine.

01:03:35.520 --> 01:03:41.120
UMAIR SHAH: I think one of the real 
challenges is that when even tech companies,  

01:03:41.120 --> 01:03:46.560
and you can name all of them, when they 
look at what they're doing in the AI space,  

01:03:46.560 --> 01:03:49.800
they gravitate towards healthcare delivery.

01:03:49.800 --> 01:03:54.960
LEE: Yes. And in fact, it's not even 
delivery. I think techies—I did this,  

01:03:54.960 --> 01:04:00.000
too—tend to gravitate specifically to diagnosis.


01:04:00.000 --> 01:04:03.520
LEE: I have been definitely guilty. I think Umair,  

01:04:03.520 --> 01:04:08.480
of course, was speaking as a former 
frustrated public health official  

01:04:08.480 --> 01:04:15.040
in just thinking about all the other things that 
are important to maintain a healthy population.

01:04:15.040 --> 01:04:18.080
Is there some lesson that we should take away? I  

01:04:18.080 --> 01:04:22.480
think our book also focused a 
lot on things like diagnosis.

01:04:22.480 --> 01:04:27.440
KOHANE: Yeah. Well, first of all, I 
think we just have to have humility. And  

01:04:27.440 --> 01:04:36.960
I think it's a really important ingredient. I 
found myself staring at the increase in lifespan  

01:04:36.960 --> 01:04:42.640
in human beings over the last two centuries 
and looking for bumps that were attributable.

01:04:42.640 --> 01:04:45.600
I'm in medical school. I've already made this 
major commitment. What are the bumps that  

01:04:45.600 --> 01:04:49.200
are attributable to medicine? And there 
was one bump that was due to vaccines,  

01:04:49.200 --> 01:04:57.440
a small bump. Another small bump that was due 
to antibiotics. And the rest of it is nutrition,  

01:04:57.440 --> 01:05:01.040
sanitation, yeah, nutrition and sanitation.

01:05:01.040 --> 01:05:09.440
And so I think doctors can be incredibly 
valuable, but not all the time. And we're  

01:05:09.440 --> 01:05:16.800
spending now one-sixth of our GDP on it. The 
majority of it is not effectively prolonging  

01:05:16.800 --> 01:05:22.160
life. And so the humility has to be 
the right medicine at the right time.

01:05:22.160 --> 01:05:26.320
But that runs, (A) against a bunch of 
business models. It runs against the  

01:05:26.320 --> 01:05:32.640
primacy of doctors in healthcare. It was 
one thing when there were no textbooks;  

01:05:33.440 --> 01:05:38.160
there was no PubMed. You know, the doctor 
was the repository of all the probably  

01:05:38.160 --> 01:05:45.280
knowledge that we have. But I think your 
guests were right. We have to think more  

01:05:45.280 --> 01:05:51.440
broadly in the public health way. How do we 
make knowledge pervasive like sanitation?

01:05:51.440 --> 01:05:56.960
GOLDBERG: Although I would add that 
since what we're talking about is AI,  

01:05:57.520 --> 01:06:01.280
it's harder to see if ... and if what 
you're talking about is public health,  

01:06:01.280 --> 01:06:07.120
I mean, it was certainly very important to have 
good data during the pandemic, for example.

01:06:07.120 --> 01:06:11.600
But most of the ways to improve public health, 
like getting people to stop smoking and eat  

01:06:11.600 --> 01:06:17.280
better and sleep better and exercise more, 
are not things that AI can help with that  

01:06:17.280 --> 01:06:25.840
much. Whereas diagnosis or trying to improve 
treatment are places that it could tackle.

01:06:25.840 --> 01:06:27.440
And in fact, Peter, I wanted to put you—oh,  

01:06:27.440 --> 01:06:30.640
wait, Zak's going to say something—but, 
Peter, I wanted to put you on the spot.

01:06:30.640 --> 01:06:30.790
LEE: Yeah.

01:06:30.790 --> 01:06:36.880
GOLDBERG: I mean, if you had a medical 
issue now, and you went to a physician,  

01:06:36.880 --> 01:06:40.800
would you be OK with them not using generative AI?

01:06:40.800 --> 01:06:45.840
LEE: I think if it's a complex or a mysterious 
case, I would want them to use generative  

01:06:45.840 --> 01:06:53.760
AI. I would want that second opinion on 
things. And I would personally be using it.  

01:06:56.240 --> 01:07:02.240
If for no other reason than just to 
understand what the chart is saying.

01:07:02.240 --> 01:07:06.800
I don't see, you know, how or 
why one wouldn't do that now.

01:07:06.800 --> 01:07:13.440
KOHANE: It's such a cheap second opinion, 
and people are making mistakes. And even if  

01:07:13.440 --> 01:07:18.640
there are mistakes on the part of AI, 
if there's a collision, discrepancy,  

01:07:18.640 --> 01:07:22.080
that's worth having a discussion. And 
again, this is something that we used  

01:07:22.080 --> 01:07:26.789
to do more of when we had more time with 
the patients; we'd have clinic conferences.

01:07:26.789 --> 01:07:26.800
LEE: Yeah.

01:07:26.800 --> 01:07:35.280
KOHANE: And we don't have that now. So I do think 
that there is a role for AI. But I think again,  

01:07:35.280 --> 01:07:42.640
it's much more of a continual presence, being part 
of a continued conversation rather than an oracle.

01:07:43.200 --> 01:07:47.840
And I think that's when you'll start 
seeing, when the AI is truly a colleague,  

01:07:47.840 --> 01:07:51.280
and saying, “You know, Zak, that's 
the second time you made that mistake.  

01:07:52.160 --> 01:07:55.840
You know, that's not obesity. That's 
the effect of your drugs that you're  

01:07:55.840 --> 01:08:01.440
giving her. You better back off of it.” 
And that's what we need to see happen.

01:08:01.440 --> 01:08:06.480
LEE: Well, and for the business of healthcare, 
that also relates directly to quality scores,  

01:08:06.480 --> 01:08:11.200
which translates into money 
for healthcare providers.

01:08:11.920 --> 01:08:16.880
So the last person that we interviewed 
was Gianrico Farrugia. And, you know,  

01:08:16.880 --> 01:08:21.280
I was sort of wondering, I was expecting 
to get a story from a CEO saying, “Oh,  

01:08:21.280 --> 01:08:27.280
my God, this has been so disruptive, incredibly 
important, meaningful, but wow, what a headache.”

01:08:28.320 --> 01:08:34.301
At least Gianrico didn't expose any of that. 
Here's one of the snippets to give you a sense.

01:08:34.301 --> 01:08:38.640
GIANRICO FARRUGIA: When generative AI came, for 
us, it's like, I wouldn't say we told you so,  

01:08:38.640 --> 01:08:43.280
but it's like, ah, there you go. Here's another 
tool. This is what we've been talking about. Now  

01:08:43.280 --> 01:08:47.760
we can do it even better. Now we can move 
even faster. Now we can do more for our  

01:08:47.760 --> 01:08:54.560
patients. It truly never was disruptive. It truly 
immediately became enabling, which is strange,  

01:08:54.560 --> 01:09:00.280
right, because something as disruptive as 
that instantly became enabling at Mayo Clinic.

01:09:00.280 --> 01:09:08.560
LEE: So I tried pretty hard in that interview to 
get Gianrico to admit that there was a period of  

01:09:08.560 --> 01:09:16.480
headache and disruption here. And he never, ever 
gave me that. And so I take him at his word.

01:09:16.480 --> 01:09:21.480
Zak, maybe I should ask you, what about Harvard 
and the whole Harvard medical ecosystem?

01:09:21.480 --> 01:09:30.000
KOHANE: I would be surprised if there are 
system-wide measurable gains in health quality  

01:09:30.000 --> 01:09:38.000
right now from AI. And I do have to say that 
Mayo is one of the most marvelous organizations  

01:09:38.000 --> 01:09:45.440
in terms of team behavior. So if there's someone 
who's gotten the team part of it right, they've  

01:09:45.440 --> 01:09:50.058
come the closest, which relates to our prior 
conversation. They have the quarterback idea …

01:09:50.058 --> 01:09:52.880
LEE: Yes.
KOHANE: … pretty well down compared to others.

01:09:52.880 --> 01:10:01.280
Nonetheless, I take him at his word, that 
it hasn't disrupted them. But I'm also,  

01:10:01.280 --> 01:10:07.440
I have yet to see the evidence that there's 
been a quantum leap in quality or efficacy.  

01:10:07.440 --> 01:10:16.080
And I do believe that it's possible to have a 
quantum leap in efficacy in the right system.

01:10:16.080 --> 01:10:22.800
So if they haven't been disrupted, I 
would venture that they've absorbed it,  

01:10:22.800 --> 01:10:30.800
but they haven't used it to its fullest potential. 
And the way I could be proven wrong is next year,  

01:10:31.520 --> 01:10:35.040
also the metrics showing that over 
the last year, they've had, you know,  

01:10:35.040 --> 01:10:38.080
decreased readmissions, decreased complications,  

01:10:38.080 --> 01:10:43.040
decreased errors and all that. And if so, God 
bless them. And we should all be more like Mayo.

01:10:43.040 --> 01:10:50.240
LEE: So I thought a little bit about 
two other quotes from the interviews  

01:10:50.240 --> 01:10:57.600
that sort of maybe would send us 
off with some more inspirational  

01:10:58.320 --> 01:11:04.560
kind of view of the future. And so there's one 
from Bill Gates and one from Gianrico Farrugia.  

01:11:04.560 --> 01:11:12.320
So what I'd like to do is to play both of those 
and then maybe we can have our last comments.

01:11:12.320 --> 01:11:18.240
BILL GATES: You know, I’ve gone so far as to 
tell politicians with national health systems  

01:11:18.240 --> 01:11:27.440
that if they deploy AI appropriately, that the 
quality of care, the overload of the doctors,  

01:11:27.440 --> 01:11:33.120
the improvement in the economics will be enough 
that their voters will be stunned because they  

01:11:33.120 --> 01:11:39.760
just don’t expect this, and, you know, they could 
be reelected just on this one thing of fixing  

01:11:39.760 --> 01:11:46.720
what is a very overloaded and economically 
challenged health system in these rich countries.

01:11:46.720 --> 01:11:48.701
And now Gianrico.

01:11:48.701 --> 01:11:52.800
GIANRICO FARRUGIA: And we seemed to be on a 
linear path, which is, let's try and reduce  

01:11:52.800 --> 01:12:00.160
administrative burden. Let's try and truly be a 
companion to a physician or other provider. … And  

01:12:00.160 --> 01:12:04.720
then in the next step, we keep going until 
we get to, now we can call it agentic AI,  

01:12:04.720 --> 01:12:11.840
whatever we want to talk about. And my view 
was, no, is that let's start with that aim,  

01:12:11.840 --> 01:12:17.440
the last aim … because the others will come 
automatically if you're working on that harder  

01:12:17.440 --> 01:12:23.280
problem. Because one, to get to that harder 
problem, you'll find all the other solutions.

01:12:23.280 --> 01:12:33.440
All right. I think these are both kind of calls 
to be more assertive about this and more forward  

01:12:33.440 --> 01:12:43.120
leaning. I think two years into the GPT-4 era, 
those are pretty significant and pretty optimistic  

01:12:43.120 --> 01:12:49.280
calls to action. So maybe just to give 
you both one last word. What would be  

01:12:49.280 --> 01:12:54.080
one hope that you would have for the world of 
healthcare and medicine two years from now?

01:12:54.080 --> 01:13:03.440
KOHANE: I would hope for businesses that whoever 
actually owns them at some holding company level,  

01:13:03.440 --> 01:13:09.840
regardless of who owns them, are 
truly patient-focused companies,  

01:13:09.840 --> 01:13:14.240
companies where the whole AI 
is about improving your care,  

01:13:14.240 --> 01:13:20.000
and it's only trying to maximize your care and 
it doesn't care about resource limitations.

01:13:20.000 --> 01:13:26.800
And as I was listening to Bill, and the problem 
with what he was saying about saving dollars for  

01:13:27.440 --> 01:13:36.400
governments is for many things, we have some very 
expensive things that work. And if the AI says,  

01:13:36.400 --> 01:13:42.400
“This is the best thing,” it’s going 
to break your bank. And instead,  

01:13:42.400 --> 01:13:48.400
because of research limitations, we play a 
human-based fancy footwork to get out of it.

01:13:48.400 --> 01:13:52.480
That's a hard game to play, and I leave it 
to the politicians and the public health  

01:13:52.480 --> 01:13:55.360
officials who have to do 
those trades of utilities.

01:13:55.360 --> 01:14:03.840
In my role as doctor and patient, I'd like 
to see very informed, authoritative agents  

01:14:03.840 --> 01:14:10.880
acting only on our behalf so that when we go 
and we seek to have our maladies addressed,  

01:14:10.880 --> 01:14:18.080
the only issue is, what's the best and 
right thing for me now? And I think that  

01:14:18.080 --> 01:14:24.080
is both technically realizable. And even in 
our weird system, there are business plans  

01:14:24.080 --> 01:14:28.280
that will work that can achieve that. 
That's my hope for two years from now.

01:14:28.280 --> 01:14:29.590
LEE: Yeah, fantastic. Carey.

01:14:29.590 --> 01:14:35.440
GOLDBERG: Yeah. I second that so 
enthusiastically. And I think, you know,  

01:14:35.440 --> 01:14:42.800
we have this very glass half full/glass half empty 
phenomenon two years after the book came out.

01:14:42.800 --> 01:14:48.960
And it's certainly very nice to see, you know, 
new approaches to administrative complexity and  

01:14:48.960 --> 01:14:57.440
to prior authorization and all kinds of ways 
to make physicians' lives easier. But really  

01:14:57.440 --> 01:15:04.000
what we all care about is our own health and 
that we would like to be able to optimize the  

01:15:04.000 --> 01:15:10.240
use of this truly glorious technological 
achievement to be able to live longer  

01:15:10.240 --> 01:15:17.736
and better lives. And I think what Zak just 
described is the most logical way to do that.

01:15:17.736 --> 01:15:17.755
[TRANSITION MUSIC]

01:15:17.755 --> 01:15:23.040
LEE: Yeah, I think for me, two years 
from now, I would like to see all of  

01:15:23.040 --> 01:15:30.720
this digital data that's been so painful, such 
a burden on every doctor and nurse to record,  

01:15:30.720 --> 01:15:38.525
actually amount to something meaningful in the 
care of patients. And I think it's possible.

01:15:38.525 --> 01:15:38.555
KOHANE: Amen.
GOLDBERG: Yeah.

01:15:38.555 --> 01:15:41.760
LEE: All right, so it's been quite a journey.  

01:15:42.320 --> 01:15:46.560
We were joking before we're still on speaking 
terms after having written a book. [LAUGHS]

01:15:46.560 --> 01:15:51.920
And then, um, I think listeners might 
enjoy knowing that we debated amongst  

01:15:51.920 --> 01:15:55.520
ourselves what to do about a second edition,  

01:15:55.520 --> 01:15:59.680
which seemed too painful to me, and so I 
suggested the podcast, which seemed too  

01:15:59.680 --> 01:16:04.640
painful to the two of you [LAUGHTER]. And in the 
end, I don't know what would have been easier,  

01:16:04.640 --> 01:16:10.000
writing a book or doing this podcast series, 
but I do think that we learned a lot.

01:16:10.000 --> 01:16:14.880
Now, last bit of business here. To avoid having  

01:16:14.880 --> 01:16:21.200
the three of us try to write a book again and 
do this podcast, I leaned on the production  

01:16:21.200 --> 01:16:25.680
team in Microsoft Research and the Microsoft 
Research Podcast. And I thought it would be  

01:16:25.680 --> 01:16:30.400
good to give an explicit acknowledgment to 
all the people who've contributed to this.

01:16:30.400 --> 01:16:34.160
So it's a long list of names. I'm going 
to read through them all. And then  

01:16:34.160 --> 01:16:40.320
I suggest that we all give an applaud 
[LAUGHTER] to them. And so here we go.