The Peter Attia Drive - #309 ‒ AI in medicine： its potential to revolutionize disease prediction, diagnosis, and outcomes, causes for concern in medicine and beyond, and more ｜ Isaac Kohane, M.D., Ph.D.

00:00:00.000 Hey, everyone. Welcome to the Drive podcast. I'm your host, Peter Atiyah. This podcast,

00:00:16.540 my website, and my weekly newsletter all focus on the goal of translating the science of longevity

00:00:21.520 into something accessible for everyone. Our goal is to provide the best content in health and

00:00:26.720 wellness, and we've established a great team of analysts to make this happen. It is extremely

00:00:31.660 important to me to provide all of this content without relying on paid ads. To do this, our work

00:00:36.960 is made entirely possible by our members, and in return, we offer exclusive member-only content

00:00:42.700 and benefits above and beyond what is available for free. If you want to take your knowledge of

00:00:47.940 this space to the next level, it's our goal to ensure members get back much more than the price

00:00:53.200 of the subscription. If you want to learn more about the benefits of our premium membership,

00:00:57.980 head over to peteratiyahmd.com forward slash subscribe.

00:01:04.180 My guest this week is Isaac Cohen, who goes by Zach. Zach is a physician scientist and chair

00:01:10.180 of the Department of Biomedical Informatics at Harvard Medical School, and he's an associate

00:01:14.460 professor of medicine at the Brigham and Women's Hospital. Zach has published several hundred papers

00:01:18.940 in the medical literature and authored the widely used books, Microarrays for Integrative Genomics

00:01:24.040 and the AI Revolution in Medicine, GPT-4 and beyond. He is also the editor-in-chief of the newly launched

00:01:30.660 New England Journal of Medicine AI. In this episode, we talk about the evolution of AI. It wasn't really

00:01:37.280 clear to me until we did this interview that we're really in the third generation of AI, and Zach has been

00:01:41.780 a part of both the second and obviously the current generation. We talk about AI's abilities to impact

00:01:47.560 medicine today. In other words, where is it having an impact? And where will it have an impact in the

00:01:52.820 near term? What seems very likely? And of course, we talk about what the future can hold. And obviously

00:01:58.280 here, you're starting to think a little bit about the difference between science fiction and potentially

00:02:02.300 where we hope it could go. Very interesting podcast for me, really a topic I know so little about,

00:02:08.120 which tend to be some of my favorite episodes. So without further delay, please enjoy my conversation

00:02:13.100 with Zach Cohen. Well, Zach, thank you so much for joining me today. This is a topic that's highly

00:02:23.480 relevant and one that I've wanted to talk about for some time, but wasn't sure who to speak with. And

00:02:29.280 we eventually kind of found our way to you. So again, thanks for making the time and sharing your

00:02:33.580 expertise. Give folks a little bit of a sense of your background. What was your path through medical

00:02:38.780 school and training? It was not a very typical path. No. So what happened was I grew up in

00:02:44.760 Switzerland. Nobody in my family was a doctor, come to United States, decide to major in biology.

00:02:52.340 And then I get nerd sniped by computing back in the seventies and the late seventies. And so I minor in

00:02:59.900 computer science, but I still complete my degree in biology and I go to medical school. And then in the

00:03:06.820 middle of medical school's first year, I realized, holy smokes, this is not what I expected. It's a

00:03:12.380 noble profession, but it's not a science. It's an art. It's not a science. And I thought I was going

00:03:17.580 into science. And so I bail out for a while to do a PhD in computer science. And this is during the

00:03:24.960 1980s now, early 1980s. And it's a heyday of AI. It's actually a second heyday. We're going through the

00:03:32.660 third heyday. And it was a time of great promise. And with a retrospective scope, very clear that it

00:03:41.600 was not going to be successful. There was a lot of over-promising. There is today, but unlike today,

00:03:47.780 we had not released it to the public. It was not actually working in the way that we thought it

00:03:53.940 would go on at work. And it certainly didn't scale. It was a very interesting period. And

00:04:00.420 my thesis advisor, Peter Solovich, a professor at MIT, said, Zach, you should finish your clinical

00:04:06.620 training because I'm not getting a lot of respect from clinicians. And so to bring rational decision

00:04:13.280 making to the clinic, you really want to finish your clinical training. And so I finished medical

00:04:19.040 school, did a residency in pediatrics and pediatric endocrinology, which was actually extremely

00:04:25.180 enjoyable. But when I was done, I restarted my research in computing, started a lab at Children's

00:04:32.580 Hospital in Boston, and then a center of biomedical informatics at the medical school. Like in almost

00:04:39.840 every other endeavor, getting money gets attention from the powers that be. And so I was getting a lot

00:04:46.560 of grants. And so they asked me to start the center and then eventually a new department of

00:04:52.840 biomedical informatics that I'm the chair of. We have now 16 professors or assistant professors

00:04:58.900 of biomedical informatics. Then I had been involved in a lot of machine learning projects,

00:05:06.120 but like everybody else, I was taken by surprise, except perhaps a little bit earlier than most,

00:05:10.520 by large language models. I got a call from Peter Lee in October of 22. And actually I didn't get a

00:05:17.140 call. It was an email right out of a Michael Crichton novel. It said, Zach, if you'll answer

00:05:22.780 the phone, I can't tell you what it's about, but it'd be well worth your while. And so I get a call

00:05:29.440 from Peter Lee and I knew him from before. He was a professor of computer science at CMU and also

00:05:36.000 department chair there. And then he went to ARPA and then he went to Microsoft and he tells me about

00:05:40.800 GPT-4. And this was before any of us had heard about chat GPT, which is initially GPT-3.5. He tells

00:05:48.040 me about GPT-4 and he gets me early access to it when no one else knows that exists. Only a few

00:05:55.400 people do. And I start trying it against hard cases. I get called down. I just remember from my

00:06:03.200 training. I get called down to the nursery. It's a child with a small phallus and a hole at the base

00:06:10.540 of the phallus and they can't palpate testicles and they want to know what to do because I'm a

00:06:16.560 pediatric endocrinologist. And so I asked GPT-4, what would you do? What are you thinking about?

00:06:22.940 And it runs me through the whole workup of these very rare cases of ambiguous genitalia. In this case,

00:06:30.860 it was congenital adrenal hyperplasia where the making of excess androgens during pregnancy and 0.92

00:06:38.420 then subsequently in birth causes the clitoris to swell from the glands of the penis of the phallus 1.00

00:06:46.360 and the labia minora to fuse to form the shaft of what looks like a penis. But there's no testicles, 0.99

00:06:55.920 there's ovaries. And so there's a whole endocrine workup with genetic tests, hormonal tests,

00:07:04.620 ultrasound, and it does it all. And it blows my mind. It really blows my mind because very few of us in

00:07:12.720 computer science really thought that these large language models would scale up the way they do.

00:07:18.420 It was just not expected. And talking to Bill Gates about this after Peter Leon introduced me to

00:07:26.500 problem, and he told me that his line engineers in Microsoft research, a lot of his fanciest computer

00:07:32.160 scientists did not expect this. But the line engineers at Microsoft were just watching the

00:07:37.660 scale up, you know, GPT 0, 1, 2, and they just saw it was going to keep on scaling up with the size of

00:07:44.600 the data and with the size of the model. And they said, yeah, of course, it's going to achieve this

00:07:49.520 kind of expertise. But the rest of us, I think because we value our own intellects so much, we

00:07:56.160 couldn't imagine how we'd get that kind of conversational expertise just by scaling up the

00:08:03.340 model and the data set. Well, Zach, that's actually kind of a perfect introduction to how I want to

00:08:09.460 think about this today, which is to say, look, there's nobody listening to us who hasn't heard

00:08:13.700 the term AI, and yet virtually no one really understands what is going on. So if we want to

00:08:19.280 talk about how AI can change medicine, I think we have to first invest some serious bandwidth in

00:08:26.200 understanding AI. Now, you alluded to the fact that when you were doing your PhD in the early 80s,

00:08:30.240 you were in the second generation of AI, which leads me to assume that the first generation was

00:08:36.460 shortly following World War II. And that's probably why someone by the name of Alan Turing

00:08:41.820 has his name on something called the Turing test. So maybe you can talk us through what Alan Turing

00:08:48.300 posited, what the Turing test was and proposed to be, and really what Gen 1 AI was. We don't have to

00:08:54.740 spend too much time on it, but clearly it didn't work. But let's maybe talk a little bit about

00:08:58.740 the postulates around it and what it was. After World War II, we had computing machines.

00:09:07.000 And anybody who was a serious computer scientist could see that you could have these processes

00:09:13.920 that could generate other processes. And you could see how these processes could take inputs

00:09:19.800 and become more sophisticated. And as a result, shortly after World War II, we actually had artificial

00:09:30.960 neural networks, the perceptron, which was modeled, roughly speaking, on the ideas of a neuron that could

00:09:40.740 take inputs from the environment and then have certain expectations.

00:09:47.120 And if you updated the neurons as to what was going on, it would update the weights going into

00:09:55.640 that artificial neuron. And so going back to Turing, he just came up with a test that said,

00:10:03.420 essentially, if a computational entity could maintain, essentially, its side of the conversation

00:10:10.820 without revealing that it was a computer and that others would mistake it for a human, then for all

00:10:19.780 intents and purposes, that would be intelligent behavior. And there's been all sorts of additional

00:10:26.900 constraints put on it. And one of the hallmarks of AI, frankly, is that it keeps on moving the goalposts

00:10:34.840 of what we consider to be intelligent behavior. If you had told someone in the 60s that the world chess

00:10:43.360 masters were going to be beaten by a computer program, they'd say, well, that's AI, really, that's AI.

00:10:49.820 And then when Kasparov was beaten by Deep Blue, by the IBM machine, people said, well, it's just doing

00:10:57.260 search very well. It's searching through all the possible moves in the future. It also knows all the

00:11:02.540 grandmaster moves. It has a huge encyclopedia store of all the different grandmaster moves.

00:11:07.900 And this is not really intelligent behavior. If you told people it could recognize human faces

00:11:14.640 and find your grandmother in a picture on any picture in the internet, they'd say, well, that's

00:11:20.920 intelligence. And of course, when we did it, no, that was not intelligent. And then when we said it

00:11:28.280 could write a rap poem about Peter Atia based on your webpage, and it did that, well, that would be

00:11:37.700 intelligent, that would be creative. But then if you said it's doing it based on having created a

00:11:44.220 computational model based on all the text ever generated by human beings, as much as we can gather,

00:11:50.940 which is one to six terabytes of data. And this computational model basically is predicting what is

00:11:57.860 the next word that's going to say, not just the next word, but of the millions of words that could be,

00:12:02.880 what are the probabilities of that next word? That is what's generating that rap. There's

00:12:08.720 people who are arguing that's not intelligence. So the goalposts around the Turing test keep getting

00:12:14.980 moved. So I just have to say that I no longer find that an interesting topic because it's what it's

00:12:22.000 actually doing. And whether you want to call it intelligent or not, that's up to you. It's like

00:12:26.480 discussing whether is a dog intelligent, is a baby intelligent before it can recognize

00:12:33.620 constancy of objects. Initially, babies, if you hide something from it, it's gone and it comes back.

00:12:39.880 It's a surprise. But at some point early on, they learn there's constancy of objects, even when they

00:12:44.720 don't see them. There's this spectrum of intelligent behavior. And I'd just like to remind myself that

00:12:53.780 there's a very simple computational model predicting the next word called a Markov model.

00:13:00.260 And several years ago, people were studying songbirds, and they were able to predict the full

00:13:07.000 song, the next note, and the next note of the songbird, just using a very simple Markov model.

00:13:14.240 So from that perspective, I know we think that we're all very smart, but the fact that you and I,

00:13:20.680 without thinking too hard about it, can come up with fluid speech. Okay, so the model's now

00:13:25.000 a trillion parameters. It's not a simple Markov model, but it's still a model. And perhaps later,

00:13:31.140 we'll talk about how this plays into, unfortunately, the late Kahneman's notions of thinking fast and

00:13:37.820 thinking slow. And his notion of system one, which is this sort of pattern recognition,

00:13:42.100 which is very much similar to what I think we're seeing here. And system two, which is the more

00:13:46.920 deliberate and much more conscious kind of thinking that we pride ourselves on. But a lot of what we

00:13:53.200 do is this sort of reflexive, very fast pattern recognition.

00:13:58.840 So if we go back to World War II, that's to your point where we saw basically rule-based computing

00:14:05.380 come of age. And anybody who's gone back and watched movies about the Manhattan Project,

00:14:11.140 or the decoding of all the sorts of things took place, Enigma, for example. Again, that's straight

00:14:18.780 rules-based computational power. And we're obviously at the limits of, I can only go so far. But it seems

00:14:25.420 that there was a long hiatus before we went from there to kind of like maybe what some have called

00:14:32.620 context-based computation, what your Siri does or Alexa, which is a step quite beyond that. And then

00:14:42.300 of course, you would go from there to what you've already talked about, Blue or Watson, where you have

00:14:48.640 computers that are probably going even one step further. And then of course, where we are now,

00:14:54.880 which is GPT-4. I want to talk a little bit about the computational side of that. But more what I want

00:15:00.860 to get at is this idea that there seems to be a very non-linear pace at which this is happening.

00:15:07.600 And I hear your point. I'd never thought of it that way. I hear your point about the goalpost moving,

00:15:12.840 but I think your instinct around majoring in the right thing is also relevant, which is let's focus

00:15:20.780 less on the fact that we're never quite hitting the asymptote definitionally. Let's look at the actual

00:15:26.380 output and it is staggeringly different. So what was it that was taking place during the period of your PhD,

00:15:33.200 what you're calling wave two of AI? What was the objective and where was the failure?

00:15:38.800 So the objective was in the first era, you wrote computer programs in assembler language or in languages

00:15:45.960 like Fortran. And there was a limit of what you could do. You had to be a real computational

00:15:52.580 programmer to do something in that mode. In wave two in the 1970s, we came up with these rule-based

00:16:01.700 systems where we said rules in what looked like English. If there is a patient who has a fever

00:16:08.400 and you get an isolate from the lab and that bacteria in the isolate is gram-positive,

00:16:14.700 then you might have a streptococcal infection with a probability of so-and-so. And these rule-based

00:16:21.580 systems, which you're now programming in the level of human knowledge, not in computer code,

00:16:28.600 the problem with that was several fold. A, you're going to generate tens of thousands of these rules

00:16:33.720 rules. And these rules would interact in ways that you could not anticipate. And we did not know

00:16:39.460 enough. And we could not pull out of human beings the right probabilities. And what is the right

00:16:46.460 probability of you have a fever and you don't see anything on the blood test? What else is going on?

00:16:54.920 And there's a large set of possibilities. And getting all those rules out of human beings ended up

00:16:59.580 being extremely expensive and the results were not stable. And for that reason, because we didn't

00:17:06.280 have much data online, we could not go to the next step, which is have data to actually drive these

00:17:13.340 models. What were the data sources then? Books, textbooks, and journals as interpreted by human

00:17:21.460 experts. That's why some of these were called expert systems, because they were derived from

00:17:26.760 introspection by experts who would then come up with the rules, with the probabilities. And some of the

00:17:35.100 early work, like for example, there was a program called Mycin run by Ted Shortliff out of Stanford, who

00:17:41.540 developed a antibiotic advisor that was a set of rules based on what he and his colleagues sussed out from

00:17:52.280 the different infectious disease textbooks and infectious disease experts. And it stayed only up

00:17:58.820 to date as long as they kept on looking at the literature, adding rules, fine-tuning it. There's

00:18:05.060 an interaction between two rules that was not desirable. Then you had to adjust that. Very labor-intensive.

00:18:12.120 And then if there's a new thing, you'd have to add some new rules. If AIDS happened, you'd have to say,

00:18:19.180 oh, there's this new pathogen. I have to make a bunch of rules. The probability is going to be

00:18:25.200 different if you're an IV drug abuser or if you're a male, a homosexual. And so it was very, very hard

00:18:32.040 to keep up. And in fact, people didn't. What was the language that it was programmed in? Was this

00:18:37.000 Fortran? No, no. These were so-called rule-based systems. And so the languages, for example,

00:18:42.720 system mycin was called e-mycin, essential mycin. So these looked like English.

00:18:48.740 Super labor-intensive.

00:18:50.300 Super labor-intensive. And there's no way you could keep it up to date. And at that time,

00:18:56.060 there was no electronic medical records. They were all paper records. So not informed by what was

00:18:59.940 going on in the clinic. Three revolutions had to happen in order for us to have what we have today.

00:19:08.120 And that's why I think we had such a quantum jump recently.

00:19:13.300 Before we get to that, that's the exciting question, but I just want to go back to the

00:19:16.820 Gen 2. Were there other industries that were having more success than medicine? Were there

00:19:21.660 applications in the military? Were there applications elsewhere in government where

00:19:25.620 they got a little closer to utility?

00:19:28.400 Yes. So there was a company which was a remnant of... Back in the 1970s, there were a whole bunch of

00:19:35.900 computer companies around what we called 128 in Boston. And these were companies that were famous

00:19:42.280 back then, like Wang Computer, like Digital Equipment Corporation. And it's a very sad story for Boston

00:19:49.260 because that was before Silicon Valley got its pearl of computer companies around it. And one of the

00:19:57.560 companies, Digital Equipment Corporation, built a program called R1. And R1 was an expert in configuring

00:20:05.380 the mini computers that you ordered. So you wanted some capabilities and it would actually

00:20:09.760 configure all the industrial components, the processors, the disk, and it would know about

00:20:16.520 all the exceptions and what you needed to know, what cabling, what memory configuration, all that was

00:20:22.520 done. And it basically replaced several individuals who had that very, very rare knowledge to configure

00:20:29.720 their systems. It was also used in several government logistics efforts. But even those efforts, although

00:20:37.740 they were successful and used commercially, were limited because it turns out human beings, once you 0.56

00:20:43.160 got to about three, four, five, six thousand rules, no single human being could keep track of all the ways

00:20:50.480 these rules could work. We used to call this the complexity barrier, that these rules would interact in

00:20:57.520 unexpected ways and you'd get incorrect answers, things that were not commonsensical because you'd

00:21:07.220 actually not captured everything about the real world. And so it was very narrowly focused. And if the

00:21:15.380 expertise was a little bit outside the area of focus, if let's say it was an infectious disease program

00:21:20.960 and there was a little bit of influence from the cardiac status of the patient and you had not

00:21:28.020 accurately modeled that, its performance would degrade rapidly. Similarly, if there was in digital

00:21:35.220 equipment a new model that had a complete different part that had not included and that there were some

00:21:43.200 dependencies that were not modeled, it would degrade in performance. So these systems were very brittle,

00:21:48.460 did not show common sense. They had expert behavior, but it was very narrowly done. There were applications

00:21:55.600 of medicine back then that survived till today. For example, already back then we had these systems

00:22:02.040 doing interpretation of ECGs pretty competently, at least a first pass until they would be reviewed by

00:22:10.320 an expert cardiologist. There's also a program that interpreted what's called serum protein

00:22:15.540 electrophoresis, where you look at proteins separated out by an electric gradient to make a diagnosis,

00:22:23.260 let's say of myeloma or other protein disorders. And those also were deployed clinically, but they only

00:22:30.920 worked very much in narrow areas. They were by no stretch of an imagination, general purpose reasoning

00:22:37.340 machines.

00:22:37.900 So let's get back to the three things. There are three things that have taken the relative failures of first and second

00:22:46.000 attempts at AI and got us to where we are today. I can guess what they are, but let's just have you walk us through

00:22:52.880 them.

00:22:53.540 The first one was just lots of data. And we needed to have a lot of online data to be able to develop models of

00:23:03.920 interesting performance and quality. So ImageNet was one of the first such data sets, collections of millions of

00:23:13.080 images with annotations, importantly. This has a cat in it. This has a dog in it. This is a blueberry muffin. This has a

00:23:20.320 human in it. And having that was absolutely essential to allow us to train the first very successful neural network

00:23:30.320 models. And so having those large data sets was extremely important. The other, and there's equivalent in

00:23:39.760 medicine, which is we did not have a lot of textual information about medicine until PubMed went online. So all the

00:23:50.520 literature, medical literature, at least we have an abstract of it in PubMed. Plus we have for a subset of it that's open

00:23:58.380 access because government has paid for it through grants. There's something called PubMed Central, which has the full

00:24:04.100 text. So all of a sudden that has opened up over the last 10 years. And then electronic health records, after Obama

00:24:12.660 signed the HITECH Act, electronic health records, which also ruined the lives of many doctors, also happened to also

00:24:19.540 generate a lot of text for the use in these systems. So that's large amounts of data being generated online. The second was

00:24:28.160 the neural network models themselves. So the perceptron that I mentioned that was developed

00:24:33.240 not too long after World War II was shown by one of the pioneers of AI, Marvin Minsky, to have fundamental

00:24:41.700 limitations in that it could not do certain mathematical functions like what's called an exclusive

00:24:47.760 ore gate. Because of that, people said these neural networks are not going to scale. But there were a few true

00:24:53.360 believers who kept on pushing and making more and more advanced architectures and those multi-level deep

00:25:03.060 neural networks. So instead of having one neural network, we layer on top of one neural network, another

00:25:09.860 one, and another one, and another one, so that the output of the first layer gets propagated up to the

00:25:16.540 second layer of neurons to the third layer and fourth layer and so on.

00:25:20.780 And I'm sorry, was this a theoretical mathematical breakthrough or a technological breakthrough?

00:25:26.420 Both. It was both because having those insight that these, we could actually come up with all the

00:25:33.540 mathematical functions that we needed to, we could simulate them with these multi-level networks,

00:25:38.440 whereas it was a theoretical insight, but we would have never made anything out of it if not for the

00:25:43.880 fact of sweaty teenagers, mostly teenage boys, playing video games. In order to have first-person shooters

00:25:53.120 capable of running high-resolution pictures of aliens or monsters in high-resolution, 24-bit color,

00:26:04.900 60 frames per second, we needed to have processors, very parallel processors,

00:26:10.700 that would allow you to do the linear algebra that allow you to calculate what was going to be the

00:26:17.580 intensity of color on every dot of the screen at 60 frames per second.

00:26:22.320 And that's literally just because of the matrix multiplication

00:26:25.080 math that's required to do this. You have N by M matrices that are so big, and you're crossing and

00:26:33.040 dotting huge matrices.

00:26:34.940 Huge matrices. And it turns out that's something that can be run in parallel. So you want to have

00:26:40.940 multiple parallel processors capable of rendering those images, again, at 60 frames per second. So

00:26:47.840 basically, millions of bits on your screen being rendered at 24 or 32-bit color. And in order to do

00:26:55.180 that, you need to have that linear algebra that you just referred to being run in parallel.

00:26:59.460 And so these parallel processors called graphical processing units, GPUs, were developed. And the GPUs

00:27:09.780 were developed by several companies. And some of them stayed in business, some didn't, but they were

00:27:14.960 aptly essential to the success of video games. Now, it then occurred to many smart mathematicians and

00:27:21.900 computer scientists that the same linear algebra that was used to drive that computation for images

00:27:29.320 could also be used to calculate the weights of the edges between the neurons in a neural network.

00:27:37.860 So the mathematics of updating the weights in response to stimuli, let's say, of a neural network,

00:27:45.520 updating of those weights can be done all in linear algebra. And if you have this processor, so a typical

00:27:55.140 computer has a central processing unit. So that's one processing unit. A GPU has tens of thousands of

00:28:04.840 processors that do this one very simple thing, linear algebra. And so by having this parallelism that

00:28:12.420 only supercomputers would have typically on your simple PC, because you needed to show the graphics

00:28:20.260 at 60 frames per second, gave us all of a sudden these commodity chips that allowed us to calculate

00:28:26.460 the performance of these multi-level neural networks. So that theoretical breakthrough was the second

00:28:31.740 part, but would not have happened without the actual implementation capability that we had with the GPUs.

00:28:40.600 And so NVIDIA would be the most successful example of this, presumably?

00:28:45.680 It was not the first, but it's definitely the most successful example. And there's a variety of

00:28:49.820 reasons why it was successful and created an ecosystem of implementers who built their neural

00:28:56.380 network deep learning systems on top of the NVIDIA architecture.

00:29:02.000 Would you go back and look at the calendar and say this was the year or quarter when there was

00:29:06.500 escape velocity achieved there? Yeah. So it was probably around 2012 when there was an ongoing

00:29:13.340 contest every year saying who has the best image recognition software. And these deep neural networks

00:29:22.060 running off GPUs were able to outperform significantly all their other competitors

00:29:29.400 in image recognition in 2012. That's very clearly when everybody just woke up and said, whoa, we knew about

00:29:36.560 neural networks. We didn't realize that these convolutional neural networks were going to be

00:29:41.980 this effective. And it seems that the only thing that's going to stop us is computational speed and the size

00:29:50.440 of our data sets. That moved things very fast along in the imaging space with very soon consequences in

00:29:59.620 medicine. It was only six years later that we saw journal articles about recognition of retinopathy,

00:30:08.380 diseases affecting the retina, the back of your eye and diabetes. And a paper coming out of all places

00:30:15.360 from Google saying we can recognize different stages of retinopathy based on the images of the back of

00:30:23.020 the eye. And that also was a wake up call because yes, part of the goalpost moving is great that we

00:30:28.660 could recognize cats and dogs in web pages. But now all of a sudden, this thing that we thought was

00:30:35.880 specialized human expertise could be done by that same stack of software. Just if you gave it enough

00:30:43.380 cases of these retinopathies, it would actually work well. And furthermore, what was wild was that

00:30:50.120 there's something called transfer learning, where you tune up these networks, get them to recognize

00:30:55.800 cats and dogs. And in the process of recognizing cats and dogs, it learns how to recognize little

00:31:01.040 circles and lines and fuzziness and so on. You did a lot better in training up the neural network

00:31:08.560 first on the entire set of images and then on the retinas. And if you just went straight to,

00:31:15.180 I'm just going to train on the retinas. And so that transfer learning was impressive.

00:31:20.980 And then the other thing as a doctor was impressive to many of us. I was actually asked to write an

00:31:26.200 editorial for the Journal of the American Medical Association in 2018 when a Google article was

00:31:32.560 written. What was impressive to us was that what was the main role of doctors in that publication?

00:31:39.120 It was just twofold. One was to just label the images that were used for training. This is

00:31:45.700 a retinopathy. It's not retinopathy. And then to serve as judges of its performance. And that was it.

00:31:53.580 The rest of it was computer scientists working with GPUs and images, tuning it. And that was it.

00:32:00.480 Didn't look anything like medical school. And you were having expert level recognition of

00:32:05.840 retinopathy. That was a wake-up call. You've alluded to the 2017 paper by Google,

00:32:14.080 Attention is All That is Needed, I think is the title of the paper. Attention is All You Need.

00:32:19.060 That's not what I'm referring to. I'm also referring to a 2018 paper in JAMA.

00:32:23.360 Ah, I'm sorry.

00:32:24.700 You're talking about the great paper, Attention is All You Need. That was about the invention of the

00:32:29.320 transformer, which is a specific type of neural network architecture. I was talking about these

00:32:35.540 were vanilla, fairly vanilla convolutional neural networks, the same one that can detect dogs and

00:32:41.560 cats. It was a big medical application, retinopathy 2018. Except for computer scientists, no one noticed

00:32:47.940 the attention is all you need paper. And Google had this wonderful paper that said, you know,

00:32:56.520 if we recognize not just text that co-locates together, because previously, so we're going to

00:33:03.740 get away from images for a second. There was this notion that I can recognize a lot of similarities

00:33:11.520 in text. If I see which words occur together, I can implicate the meaning of a word by the company

00:33:18.780 it keeps. And so if I see this word and it has around it, kingdom, crown, throne, it's about a king.

00:33:30.080 And similarly for queen and so on. That kind of association in which we created what was called

00:33:37.420 embedding vectors, which just in plain English, it's a string of numbers that says for any given word,

00:33:45.920 what's the probability? How often do these other words co-occur with it? And just using those

00:33:51.960 embeddings, those vectors, those lists of numbers that describe the co-occurrence of other words,

00:33:59.700 we were able to do a lot of what's called natural language processing, which you're looking at text

00:34:04.600 and saying, this is what it means. This is what's going on. But then in the 2017 paper,

00:34:11.220 they actually took a next step, which was the insight that where exactly the thing that we

00:34:19.380 were focusing on was in the sentence, what was before and after the actual ordering of it

00:34:24.980 mattered, not just the simple co-occurrence, that knowing what position that word was in the sentence

00:34:32.480 actually made a difference. That paper showed the performance went way up in terms of recognition.

00:34:40.340 And that transformer architecture that came from that paper made it clear for a number of

00:34:47.780 researchers, not me, that if you scaled that transformer architecture up to a larger model

00:34:54.680 so that the position dependence and this vector was learned across many, many more texts,

00:35:03.120 the whole internet, you could train it to do various tasks. This transformer model, which is called

00:35:08.200 the pre-trained model. So I apologize, I find it very boring to talk about because unless I'm working

00:35:14.080 with fellow nerds, this transformer, this pre-trained model, can think of it as the equivalent of an

00:35:20.180 equation with multiple variables. In the case of GPT-4, we think it's about a trillion variables.

00:35:27.540 It's like an equation where you have a number in front of each variable, a coefficient,

00:35:31.840 that's about a trillion long. And this model can be used for various purposes. One is the chatbot

00:35:41.900 purpose, which is given this sequence of words, what is the next word that's going to be said?

00:35:47.860 Now, that's not the only thing you could use this model for, but that's, turns out to have been

00:35:53.040 the breakthrough application of the transformer model for text.

00:35:57.460 Just to round out what you said earlier, Zach, would you say that is the third thing that enabled

00:36:02.300 this third wave of AI, the transformer? It was not what I was thinking about. For me,

00:36:07.400 I was thinking of the real breakthrough in data-driven AI. I put around the 2012 era. This is

00:36:13.520 yet another, if you talk to me in 2018, I would have already told you we're in a new heyday and

00:36:20.340 everybody would agree with you. There was a lot of excitement about AI just because of the image

00:36:25.300 recognition capabilities. This was an additional capability that's beyond what many of us were

00:36:32.740 expecting just from the scale-up of the neural network. The three, just to make sure I'm consistent,

00:36:39.280 was large data sets, multi-level neural networks, aka deep neural networks, and the GPU infrastructure.

00:36:46.440 That brought us well through the 2012 to 2018. The 2017 blip that became what we now know to be

00:36:58.760 this whole large language model transformer architecture, that development, unanticipated

00:37:04.920 for many of us, but that was already on the heels of a ascendant AI era. There was already billions of

00:37:10.760 dollars of frothy investment in frothy companies, some of which did well and many of which did not

00:37:17.800 do so well. The transformer architecture has revolutionized many parts of the human condition,

00:37:24.520 I think, but it was already part of it. I think the third wave. There's something about GPT where I feel

00:37:34.040 like most people by the time GPT-3 came out or certainly by 3.5, this was now outside of the

00:37:41.880 purview of computer scientists, people in the industry who were investing in it. This was now

00:37:48.920 becoming as much a verb as Google was in probably the early 2000s. There were clearly people who knew

00:37:57.580 what Google was in 96 and 97, but by 2000, everybody knew what Google was, right? Something about GPT 3.5

00:38:05.900 or 4 was kind of the tipping point where I don't think you can not know what it is at this point.

00:38:11.900 I don't know if that's relevant to the story, meaning does that sort of speak to what trajectory

00:38:19.180 we're on now? The other thing that I think, Zach, has become so audible in the past year

00:38:26.140 is the elevation in the discussion of how to regulate this thing, which seems like something

00:38:34.540 you would only argue about if you felt that there were a chance for this thing to be harmful to us

00:38:41.980 in some way that we do not yet perceive. So what can you say about that? Because that's obviously a nod

00:38:48.380 to the technical evolution of AI, that very serious people are having discussions about

00:38:56.620 pausing, moratoriums, regulations. There was no public discussion of that in the 80s,

00:39:02.060 which may have spoke to the fact that in the 80s, it just wasn't powerful enough to pose a threat.

00:39:06.620 So can you maybe give us a sense of what people are debating now? What is the smart, sensible,

00:39:13.420 reasonable argument on both sides of this? And let's just have you decide what the two sides are.

00:39:19.340 I'm assuming one side says, pedal to the metal. Let's go forth on development. Don't regulate this.

00:39:25.260 Let's just go nuts. The other side is, no, we need to have some breaks and barriers.

00:39:30.300 Not quite that. So you're absolutely right that chatbots have now become a commonly used noun. And that

00:39:36.940 probably happened with the emergence of GPT 3.5. And that appeared around, I think, December of 2022.

00:39:44.700 But now, yes, because out of the box, that pre-trained model I told you about could tell

00:39:51.020 you things like, how do I kill myself? How do I manufacture a toxin? It could allow you to do a lot

00:39:58.300 of harmful things. So there was that level of concern. We can talk about what's been done about

00:40:04.860 those first order efforts. Then there's been a group of scientists who interestingly went from

00:40:14.300 saying, we'll never actually get general intelligence from this particular architecture to saying, oh my

00:40:21.420 gosh, this technology is able to inference in a way that I had not anticipated. And now I'm so worried

00:40:30.300 that either because it's malevolent or just because it's trying to do something that has bad side effects

00:40:37.020 for humanity, it presents an existential threat. Now, on the other side, I don't believe is anybody saying,

00:40:44.220 let's just go heads down and let's see how fast we can get to artificial general intelligence.

00:40:51.020 Or if they do think that, they're not saying it openly.

00:40:54.060 Can you just define AGI, Zach? I think we've all heard the term, but is there a quasi-accepted

00:41:00.060 definition? First of all, there's not. And I hate myself for even bringing it up because it starts-

00:41:05.100 I was going to bring it up before you, anyway, it was inevitable.

00:41:08.380 That was an unfortunate slip because artificial general intelligence means a lot of things to a

00:41:13.580 lot of people. And I slipped because I think it's, again, a moving target and it's very much

00:41:19.820 eye on the beholder. There's a guy called Eliezer Yudkowsky, one of the so-called doomers.

00:41:24.620 And he comes up with great scenarios of how a sufficiently intelligent system could figure out

00:41:33.340 how to persuade human beings to do bad things or control of our infrastructure to bring down our

00:41:40.780 communications infrastructure or airplanes out of the sky. And we can talk about whether that's

00:41:45.580 relevant or not. And on the other side, we have, let's say, OpenAI and Google.

00:41:50.780 But what was fascinating to me is that OpenAI, which working with Microsoft generated GPT-4,

00:41:58.060 we're not saying publicly at all, let's not regulate it. In fact, they were saying,

00:42:03.100 please regulate me. Sam Altman went on a world tour where he said, we should be very concerned about

00:42:08.540 this. We should regulate AI. And he was before Congress saying, we should regulate AI. And so,

00:42:15.980 I feel a bit churlish about saying this because Sam was kind enough to write forward to the book I

00:42:21.260 wrote with Peter Lee and Kerry Goldberg on GPT-4 and the revolution in medicine. But I was wondering,

00:42:29.580 why were they insisting so much on regulation? And there's two interpretations. One is just a sincere,

00:42:35.820 and it could very well be that. Sincere wish that it be regulated so we check these machines,

00:42:41.500 these programs to make sure they don't actually do anything harmful. The other possibility,

00:42:46.060 unfortunately, is something called regulatory locking, which means I'm a very well-funded company,

00:42:51.820 and we're going to create regulations with Congress about what is required, which boxes do you have to

00:42:57.660 check in order to be allowed to run. If you're a small company, you're not going to have a

00:43:03.660 bevy of lawyers with big checks to comply with all the regulatory requirements. And so, I think Sam is,

00:43:13.180 I don't know him personally, I imagine he's a very well-motivated individual. But whether it's for

00:43:19.100 the reason of regulatory lock-in or for genuine concern, there has not been any statements of,

00:43:26.460 let's go heads down. They do say, let's be regulated. Now, having said that, before you even

00:43:32.780 go with a doomer scenario, I think there is someone just as potentially evil that we have to worry about,

00:43:38.380 another intelligence, and that's human beings. And how do human beings use these great tools?

00:43:44.540 So, just as we know for a fact that one of the earliest users of GPT-4 were high schoolers trying

00:43:52.940 to do their homework and solve hard puzzles given to them, we also know that various parties have

00:43:59.740 been using the amazing text generation and interactive capabilities of these programs to

00:44:06.380 spread misinformation, to chatbots, and there's a variety of malign things that could be done by

00:44:13.580 third parties using these engines. And I think that's, for me, the clear and present danger today,

00:44:20.140 which is how do individuals decide to use these general purpose programs?

00:44:27.180 If you look at what's going on in the Ukraine-Russian war, I see more and more autonomous

00:44:34.380 vehicles flying and carrying weaponry and dropping bombs. And we see in our own military a lot more

00:44:44.620 autonomous drones with greater and greater autonomous capabilities. Those are purpose-built

00:44:51.980 to actually do dangerous things. And a lot of science fiction fans will refer to Skynet from the

00:45:01.820 Terminator series, but we're literally building it right now.

00:45:05.260 In the Terminator, Zach, they kind of refer to a moment, I don't remember the year, like 1997 or

00:45:11.900 something. And I think they talk about how Skynet became, quote, self-aware. And somehow when it became

00:45:17.340 self-aware, it just decided to destroy humans. Is self-aware movie speak for AGI? Like, what do you

00:45:25.100 think self-aware means in more technical terms? Or is it super intelligence? There's so many terms here,

00:45:32.300 and I don't know what they mean. Okay. So self-awareness means a process by which the

00:45:38.620 intelligent entity can look back, look inwardly at its own processes and recognize itself. Now, that's

00:45:46.060 very hand-wavy, but Douglas Hofstra has probably done the most thoughtful and clear writing about what

00:45:56.140 self-awareness means. I will not do it justice, but if you really want to read a wonderful book that

00:46:02.460 spends a whole book trying to explain it, it's called I Am A Strange Loop. And in I Am A Strange

00:46:08.540 Loop, he explains how if you have enough processing power and you can represent the processes that you

00:46:17.660 have essentially models of the processes that constitute you. In other words, you're able to look at what

00:46:23.340 you're thinking. You may have some sense of self-awareness. There's a bit of an act of faith

00:46:28.300 on that. Many AI researchers don't buy that definition. There's a difference between self-awareness

00:46:35.500 and actual raw intelligence. You can imagine a super powerful computer that would predict everything

00:46:43.740 that was going to happen around you and was not aware of itself as an entity. The fact remains,

00:46:49.180 you do need to have a minimal level of intelligence to be able to be self-aware. So a fly may not be

00:46:56.540 self-aware. It just goes and finds good-smelling poop and does whatever it's programmed to do on that.

00:47:04.220 But dogs have some self-awareness and awareness of their surroundings. They don't have perfect

00:47:11.900 self-awareness, like they don't recognize themselves in the mirror and they'll bark at that. Birds will

00:47:17.180 recognize themselves in mirrors. We recognize ourselves in many, many ways. So there is some

00:47:23.580 correlation between intelligence and self-awareness, but these are not necessarily dependent functions.

00:47:28.700 So what I'm hearing you say is, look, there are clear and present dangers associated with current

00:47:34.780 best AI tools in that humans can use them for nefarious purposes. It seems to me that the most scalable

00:47:43.100 example of that is still relatively small in that it's not existential threat to our species large,

00:47:50.780 correct? Well, yes and no. If I was trying to do gain of function research with a virus,

00:47:58.140 good point, I could use these tools very effectively. Yeah. That's a great example.

00:48:03.980 There's this disconnect and perhaps you understand the disconnect better than I do.

00:48:07.580 There's those real existential threats. And then there's this more fuzzy thing that we're worried

00:48:14.780 about correctly about bias, incorrect decisions, hallucinations. We can get into what that might be

00:48:22.140 and our use in the everyday of human condition. And there's concerns about mistakes that might be

00:48:28.220 made. There's concerns about displacement of workers that just as automation displaced a whole other series

00:48:37.420 of workers. Now that we have something that works in the knowledge industry automatically, just as

00:48:43.660 we're placing a lot of copy editors and illustrators with AI, where's that going to stop? It's now much

00:48:50.700 more in the white collar space. And so there is concern around the harm that could be generated there.

00:48:57.260 In the medical domain, are we getting good advice? Are we getting bad advice? Whose interests are being

00:49:02.940 optimized in these various decision procedures? That's another level that doesn't quite rise at

00:49:08.540 all to the level of extinction events. But a lot of policymakers and the public seem to be concerned

00:49:14.860 about it. Those are fair points. Let's now talk about that state of play within medicine. So I liked

00:49:20.700 your first example, almost one we take for granted, but you go and get an EKG at the doctor's office. This was

00:49:25.820 true 30 years ago, just as it is today. You get a pretty darn good readout. It's going to tell you if

00:49:31.500 you have an AV block. It's going to tell you if you have a bundle branch block. Put it this way,

00:49:36.540 they read EKGs better than I do. That's not saying much anymore, but they do. What was the next area

00:49:42.300 where we could see this? It seems to me that radiology is a field of medicine, which is of course,

00:49:48.940 image pixel based medicine that would be the most logical next place to see AI do good work. What

00:50:00.140 is the current state of AI in radiology? In all the visual based medical specialties,

00:50:07.740 it looks like AI can do as well as many experts. So what are the image appreciation subspecialties?

00:50:19.020 Pathology, when you're looking at slices of tissue under the microscope. Radiology,

00:50:23.020 where you're looking at x-rays or MRIs. Dermatology, where you're looking at pictures of the skin.

00:50:31.580 So in all those visual based specialties, the computer programs are doing by themselves as well

00:50:42.300 as many experts, but they're not replacing the doctors because that image recognition process

00:50:50.460 is only part of their job. Now, to be fair to your point, in radiology, we already today,

00:50:58.220 before AI in many hospitals would send x-rays by satellite to Australia or India where they would

00:51:06.220 be read overnight by a doctor or a specially trained person who had never seen the patient.

00:51:12.060 And then the reports filed back to us because they're 12 hours away from us overnight, we'd have

00:51:17.660 the results of those reads. And that same kind of function can be done automatically by AI. So that's

00:51:23.260 replacing a certain kind of doctor.

00:51:24.140 Let me dig into that a little bit more. So let's start with a relatively simple

00:51:29.740 type of image, such as a mammogram or a chest x-ray. So it's a single image. I mean, I guess

00:51:35.980 with a chest x-ray, you'll get an AP and a lateral, but let's just say you're looking at an AP

00:51:40.780 or a single mammogram. A radiologist will look at that. A radiologist will have clinical information

00:51:47.500 as well. So they will know why this patient presented in the case of the chest x-ray,

00:51:52.940 for example, in the ER in the middle of the night. Were they short of breath?

00:51:56.300 Do they have a fever? Do they have a previous x-ray? I can compare it to all sorts of information.

00:52:02.620 Are we not at the point now where all of that information could be given to the AI to enhance

00:52:09.420 the pre-test probability of whatever diagnosis it comes to?

00:52:12.860 I am delighted when you say pre-test probability. Don't talk dirty around me.

00:52:16.940 Love my Bayes theorem over here.

00:52:18.780 Yep. So you just said a lot, because what you just said actually went beyond what the straight

00:52:26.060 convolutional neural networks would do, because they actually could not replace radiologists,

00:52:30.380 because they could not do a good job of taking into account the previous history of the patient.

00:52:36.540 And it's required the emergence of transformers, where you can have multimodality. You have both

00:52:43.980 the image and the text. Now, they're going to do better than many, many radiologists today.

00:52:52.460 There is, I don't think, any threat yet to radiologists as a job. One of the most irritating

00:52:58.220 to doctors predictions was by Jeffrey Hinton, one of the intellectuals leaders of neural network

00:53:03.900 architecture. He said, I think it was in 2016, I have this approximately wrong, but in six years,

00:53:09.740 we wouldn't have no need for radiologists. And that was just clearly wrong. And the reason it was

00:53:16.220 wrong is A, they did not have these capabilities that we just talked about, about understanding

00:53:21.180 about the clinical context. But it's also the fact that we just don't have enough radiologists.

00:53:25.260 Meaning to do the training?

00:53:27.180 To actually do the work. So if you look at American medicine, I'll let you shut me down.

00:53:33.980 But if you look at the residency programs, we're not getting enough radiologists out. We have an

00:53:42.060 overabundance of applicants for interventional radiology. They're making a lot of money. It's

00:53:47.100 high prestige. But straight up radiology readers, not enough of them. Primary care doctors, I go around

00:53:54.620 medical schools and ask who's becoming a primary care doctor. Almost nobody. So primary care is

00:53:59.660 disappearing in the United States. In fact, Mass General and Brigham announced officially they're

00:54:05.020 not seeing primary care patients. People are still going to dermatology and they're still going to

00:54:10.060 plastic surgery. What I did, pediatric endocrinology, half of the slots nationally are not being filled.

00:54:17.500 Pediatric developmental disorders like autism, those slots, half of them filled.

00:54:23.740 PDID. There's a huge gap emerging in the available expertise. So it's not what we thought it was

00:54:33.180 going to be that we had a surplus of doctors that had to be replaced. It's just we have a surplus in

00:54:39.980 a few focused areas which are very popular. And then for all the work of primary care and primary

00:54:46.140 prevention, kind of stuff that you're interested in, we have almost no doctors available.

00:54:50.540 Yeah, let's go back to the radiologist for a second because, again, I'm fixated on this one

00:54:55.660 because it seems like the most, well, the closest one to address. And again, if you're saying, look,

00:55:01.500 we have a dearth of imaging radiologists who are able to work the emergency rooms, urgent care clinics,

00:55:08.140 and hospitals, wouldn't that be the first place we would want to apply our best of imaging recognition

00:55:15.820 with our super powerful GPUs and now plug them into our transformers with our language models

00:55:23.260 so that I can get clinical history, medical past history, previous images, current images,

00:55:30.380 and you don't have to send it to a radiologist in Australia to read it, who then has to send it back

00:55:35.900 to a radiologist here to check. Like, if we're just trying to fill a gap, that gap should be fillable,

00:55:40.780 shouldn't it? And that's exactly where it is being filled. And what keeps distracting me in this

00:55:46.380 conversation is that there's a whole other group of users of these AIs that we're not talking about,

00:55:52.620 which is the patients. And previously, none of these tools were available to patients. With the release

00:55:58.700 of GPT 3.5 and 4, and now Gemini and Claude III, they're being used by patients all the time in ways

00:56:06.540 that we had not anticipated. Let me give you an example. So there's a child who was having trouble

00:56:14.780 walking, having trouble chewing, and then started having intractable headaches. Mom brought him to

00:56:21.980 multiple doctors, they did multiple imaging studies, no diagnosis, kept on being in intractable

00:56:28.540 pain. She just typed into GPT-4 all the reports and asked GPT-4, what's the diagnosis? And GPT-4

00:56:36.220 said, tethered cord syndrome. She then went with all the imaging studies to a neurosurgeon and said,

00:56:42.220 what is this? He looked at it and said, tethered cord syndrome. And we have such an epidemic of

00:56:48.460 misdiagnosis and undiagnosed patients. Part of my background that I'll just mention briefly,

00:56:55.580 I'm the principal investigator of the coordinating center of something called the Undiagnosed Network.

00:56:59.740 It's a network with 12 academic hospitals down the West Coast from University of Washington,

00:57:05.580 Stanford, UCLA, to Baylor, up the East Coast, Harvard hospitals, NIH. And we see a few thousand

00:57:12.300 patients every year. And these are patients who have been undiagnosed and they're in pain. That's just a

00:57:17.980 small fraction of those who are undiagnosed. And yes, we bring to bear a whole bunch of computational

00:57:23.020 techniques and genomic sequencing to actually be able to help these individuals. But it's very clear

00:57:29.100 that there's a much larger burden out there of misdiagnosed individuals.

00:57:33.340 But the question for you, Zach, which is, does it surprise you that in that example, the mother

00:57:37.980 was the one that went to GPT-4 and inputted that? I mean, she had presumably been to many physicians

00:57:45.100 along the way. Were you surprised that one of the physicians along the way hadn't been the one to say,

00:57:51.260 gee, I don't know, but let's see what this GPT-4 thing can do?

00:57:54.460 Most clinicians I know do not have what I used to call the Google reflex. I remember when I was

00:58:02.620 on the wards and we had a child with dysmorphology, they look different. And I said to the fellows,

00:58:10.700 this is after residency, what is the diagnosis? And they said, I don't know, I don't know. I said,

00:58:16.140 he has this and this and this finding. What's the diagnosis? And I said, how would you find out?

00:58:20.780 They had no idea. I just said, let's take what I just said and type it into Google.

00:58:24.620 In the top three responses, there was the diagnosis. And that reflex, which they do

00:58:31.100 use in a civilian life, they did not have in the clinic. And doctors are in a very unhappy position

00:58:38.220 these days. They're really being driven very, very hard. And they're being told to use certain

00:58:44.140 technological tools. They're being turned into data entry clerks. They don't have the Google reflex.

00:58:49.740 They don't have the reflex, who has the time to look up a journal article? They don't do the Google

00:58:55.500 reflex. Even less, do they have the, let's look at the patient's history and see what GPT-4 would

00:59:02.380 come up with. I was gratified to see early on doctors saying, wow, look, I just took the patient

00:59:09.260 history, plugged into GPT-4 and said, write me a letter of prior authorization. And they were

00:59:14.380 actually tweeting about doing this, which on the one hand, I was very, very pleased for them

00:59:19.180 because it was saving them five minutes to write that letter to the insurance company saying,

00:59:24.300 please authorize my patient for this procedure. I was not pleased for them because if you use chat GPT,

00:59:30.460 you're using a program that is covered by open AI, as opposed to a version of GPT-4 that is being run

00:59:38.860 on protected Azure cloud by Microsoft, which is HIPAA covered. For those of you, the audience

00:59:45.260 doesn't know, HIPAA is the legal framework under which we protect patient privacy. And if you violate

00:59:50.940 it, you can be fined and even go to prison. So in other words, if a physician wants to put any

00:59:57.260 information into GPT-4, they better not identify it. That's right. So they just plunked in a patient

01:00:04.300 note into chat GPT. That's a HIPAA violation. If there's a Microsoft version of it, which is HIPAA

01:00:10.620 compliant, it's not. So they were using it to improve their lives. The doctors were using it

01:00:15.580 for improving the business, the administrative part of healthcare, which is incredibly important.

01:00:20.140 But by and large, only a few doctors use it for diagnostic acumen.

01:00:26.620 And then what about more involved radiology? So obviously a plain film is one of the more

01:00:32.460 straightforward things to do, although it's far from straightforward, as anybody knows who's

01:00:36.460 stared at a chest x-ray. But once we start to look at three-dimensional images, such as

01:00:41.420 cross-sectional images, CT scans, MRIs, or even more complicated images like ultrasound and things of

01:00:48.220 that nature, what is the current state of the art with respect to AI in the assistance of reading

01:00:55.900 these types of images?

01:00:57.580 So that's the very exciting news, which is, remember how I said it was important to have

01:01:03.340 a lot of data, one of the three ingredients in the breakthrough. So all of a sudden having

01:01:07.740 a lot of data around, for example, echocardiograms, the ultrasounds of your heart. Normally it takes

01:01:14.300 a lot of training to interpret those images correctly. So there is a recent study from the

01:01:20.700 Echo Clip Group led, I think, out of UCLA. And they took a million echocardiograms and a million

01:01:30.380 textual reports and essentially trained the model, both to create those embeddings I talked about

01:01:36.780 of the images and of the text.

01:01:39.020 Just to make sure people understand what we're talking about, this is not, here's a picture of

01:01:44.700 a cat, here's a description, cat. When you put the image in, you're putting a video in. Now you're

01:01:51.420 putting a multi-dimensional video because you have time scale, you have Doppler effects. This is a very

01:01:59.180 complicated video that is going in.

01:02:01.180 It's a very complicated video and it's three-dimensional and it's weird views from different angles.

01:02:08.300 And it's dependent on the user. In other words, the tech, the radiology tech can be good or bad.

01:02:16.860 If I was the one doing it, it would be awful.

01:02:20.780 The echo tech does not have medical school debt. They don't have to go to medical school. They don't

01:02:25.660 have to learn calculus. They don't have to learn physical chemistry, all the hoops that you have to

01:02:29.740 go through in medical school. You don't have the attitudinal debt of doctors. So in two years,

01:02:34.460 they get all those skills and they actually do a pretty good job.

01:02:37.020 They do a fantastic job. But my point is, their skill is very much an important determinant of

01:02:42.700 the quality of the image.

01:02:44.300 Yes. But what we still require these days is a cardiologist to then read it and interpret it.

01:02:50.460 Right. That's sort of where I'm going, by the way, is we're going to get rid of the

01:02:52.940 cardiologist before we get rid of the technician.

01:02:55.500 We're on the same page. My target in this conversation is nurse practitioners

01:03:00.220 and physician assistants with these tools can replace a lot of expert clinicians.

01:03:06.140 And there is a big open question. What is the real job for doctors in 10 years from now?

01:03:13.100 And I don't think we know the answer to that because you fast forward to the conversation just now.

01:03:18.780 Excellent. Well, let's think about it. We still haven't come to proceduralists.

01:03:22.780 So we still have to talk about the interventional radiologist, the interventional cardiologist,

01:03:26.460 and the surgeon. We can talk about the role of the surgeon and the da Vinci robot in a moment.

01:03:31.500 But I think what we're doing is we're kind of identifying the pecking order of physicians.

01:03:36.700 And let's not even think about it through the lens of replacement. Let's start with the lens of

01:03:40.780 augmentation, which is the radiologist can be the most easily augmented, the pathologist,

01:03:47.900 the dermatologist, the cardiologist who's looking at echoes and EKGs and stress tests. People who are

01:03:56.460 interpreting visual data and using visual data will be the most easily augmented. The second tranche of

01:04:03.500 that will be people who are interpreting language data plus visual data. So now we're talking about

01:04:09.020 your internist, your pediatrician, where you have to interpret symptoms and combine them with laboratory

01:04:15.260 values and combine it with a story and an image. Is that a fair assessment in terms of tier?

01:04:21.260 Absolutely a fair assessment. My only quibble, it's not a quibble, I'm going to keep on going back to

01:04:25.820 this, is in a place where we don't have primary care. The American Association of Medical Colleges

01:04:31.580 estimates that by 35, that's only 11 years from now, we'll be missing on the order of 50,000 primary

01:04:36.940 care doctors. As I told you, I can't get primary care at the Brigham or at MGH today. And in the absence of

01:04:43.420 that, you have to ask yourself, how can we replace these absent primary care practitioners with

01:04:50.220 nurse practitioners with physician assistants augmented by these AIs? Because there's literally 0.83

01:04:57.980 no doctor to replace. So tell me, Zach, where are we technologically on that augmentation? If NVIDIA

01:05:06.060 never came out with another chip, if they literally said, you know what, we are only interested in building

01:05:12.860 golf simulators, and we're done with the progress of this, and this is as good as it's going to get. Do we have

01:05:19.980 good enough GPUs, good enough multi-layer neural networks, that all you need is more data and training sets, that we

01:05:28.780 could now do the augmentation that has been described by us in the last five minutes?

01:05:33.740 The short answer is yes. Let me make it very concrete. Most concierge services cost in Boston

01:05:39.900 somewhere between $5,000 and $20,000 a year. You can get this very low cost concierge service that I'm

01:05:46.060 just amazed that have not done the following, called One Medical. One Medical was acquired by Amazon.

01:05:51.500 And they have a lot of nurse practitioners in there. And you can make an appointment,

01:05:55.100 and you can text with them. I believe that those individuals could be helped in ordering the right

01:06:02.380 imaging studies, the right EKGs, the right medications, and assess your continuing heart failure,

01:06:11.980 and only decide in the very few cases that you need to see a specialist cardiologist or a specialist

01:06:19.660 endocrinologist today. Just be a matter of just making the current models better, evaluating them,

01:06:26.540 because not all models are equal. A big question for us, this is the regulatory question, which is,

01:06:32.140 which ones do a better job? And they're not all equal. I don't think we need technological breakthroughs

01:06:38.860 to just make the current set of paraprofessionals work at the level of entry-level doctors. Let me quickly

01:06:47.180 say the old very bad joke. What do you call the medical student who graduates at the bottom of

01:06:52.300 his class? Doctor. And so if you could just merely get the bottom 50% of doctors to be as good as the

01:07:01.580 top 50%, that would be transformative for healthcare. Now, there are other superhuman capabilities that we

01:07:10.220 can go towards, and we can talk about if we want, that do require the next generation of algorithms,

01:07:18.060 NVIDIA architectures, and data sets. Everything stopped now, we could already transform medicine.

01:07:24.300 It's just a matter of the sweat equity to create the models, figure out how to include them in the

01:07:30.780 workflow, how to pay for them, how to create a reimbursement system, and a business model that

01:07:37.420 works for our society. But there's no technological barrier.

01:07:42.940 In my mind, everything we've talked about is take the best case example of medicine today

01:07:49.900 and augment it with AI such that you can raise everyone's level of care to that of the best,

01:07:56.300 best, no gaps, and it's scaled out. Okay, now let's talk about another problem, which is where do you

01:08:04.780 see the potential for AI in solving problems that we can't even solve on the best day at the best

01:08:13.740 hospitals with the best doctors? So let me give you an example. We can't really diagnose Alzheimer's

01:08:21.180 disease until it appears to be at a point that for all intents and purposes is irreversible.

01:08:29.580 Maybe on a good day, we can halt progression really, really early in a patient with just a whiff of MCI,

01:08:36.300 mild cognitive impairment, maybe with an early amyloid detection and an anti-amyloid drug.

01:08:43.100 But is it science fiction to imagine that there will be a day when an AI could listen to a person's

01:08:48.620 voice, watch the movements of their eyes, study the movements of their gait, and predict 20 years

01:08:57.020 in advance when a person is staring down the barrel of a neurodegenerative disease and act at a time

01:09:03.260 when maybe we could actually reverse it? How science fiction-y is that?

01:09:07.580 I don't believe it's science fiction at all. Do you know that looking at retinas today, images of retina,

01:09:13.660 straightforward convolutional neural network, not even ones that involve transformers,

01:09:17.820 can already tell you by looking at your retina, not just whether you have retinal disease,

01:09:23.020 but if you have hypertension, if you're a male, if you're female, how old you are,

01:09:28.700 and some estimate of your longevity. And that's just looking at the back of your eye

01:09:33.580 and seeing enough data. I was a small player in a study that appeared in Nature in 2005 with Bruce

01:09:40.540 Yankner. We were looking at the frontal lobes of individuals who had died for a variety of reasons,

01:09:47.500 often in accidents of various ages. And we saw, bad news for people like me, that after age 40,

01:09:54.380 your transcriptome, the genes that are switched on, fell off a cliff. Thirty percent of your transcriptome

01:10:00.860 went down. And so there seemed to be a big difference in the expression of genes around age 40,

01:10:07.900 but there was one 90-year-old who looked like the young guy. So maybe there's hope for some of us.

01:10:12.140 But then I thought about it afterwards, and there were other things that actually have much smoother

01:10:16.380 functions, which don't have quite a fall off, like our skin. So our skin ages. In fact, all our organs

01:10:23.980 age and they age at different rates. You're saying that the transcriptome of the skin,

01:10:28.940 you did not see this cliff-like effect at a given age, the way you saw it in the frontal cortex.

01:10:34.540 So different organs age at different rates, but having the right data sets and the ability to see

01:10:42.540 nuances that we don't notice makes it very clear to me that the early detection part, no problem.

01:10:49.580 It can be very straightforward. The treatment part, we can talk about it as well. But again,

01:10:54.940 we had early on from the very famous Framium Heart Study, a predictor of when you had going to have

01:11:01.500 heart disease based on just a handful of variables. Now we have these artificial intelligence models

01:11:07.500 that, based on hundreds of variables, can predict various other diseases. And it will do Alzheimer's,

01:11:16.220 I believe, very soon. I think you'll be able to see a combination of

01:11:21.500 gait, speech patterns, picture of your body, picture of your skin, and eye movements. Like you said,

01:11:29.740 will be a very accurate predictor. We just published, by the way recently, speaking about eyes,

01:11:34.300 a very nice study where in a car, just by looking at the driver, it can figure out what your blood sugar is.

01:11:42.460 Because diabetics previously have not been able to get driver licenses sometimes because of the worry about

01:11:49.260 them passing out because of hypogycemia. So there was a very nice study that showed that you could

01:11:52.940 just, by looking, have cameras pointed at the eyes, could actually figure out exactly what the blood

01:11:57.500 sugar is. So that kind of detection is, I think, fairly straightforward. It's a different question

01:12:03.740 about what you can do about it. Before we go to the what you can do about it,

01:12:07.100 I just want to go a little deeper on the predictive side. You brought up the Framingham model or the

01:12:12.220 multi-ethnic study on atherosclerosis, the MESA model. These are the two most popular models by far

01:12:17.100 looking at a major adverse cardiac event risk prediction. But you needed something else to

01:12:21.820 build those models, which was enough time to see the outcome. In the Framingham cohort,

01:12:27.100 which was the late 70s and early 80s, you then had the Framingham offspring cohort.

01:12:31.740 And then you had to be able to follow these people with their LDL-C and HDL-C and triglycerides.

01:12:36.700 And later, eventually, they incorporated calcium scores. So if today we said, look,

01:12:43.100 we want to be able to predict 30-year mortality, which is something no model can do today,

01:12:50.700 this is a big pet peeve of mine, is we generally talk about cardiovascular disease through the lens

01:12:55.580 of 10-year risk, which I think is ridiculous. We should talk about lifetime risk. But I would

01:13:00.220 settle for 30-year risk, frankly. And if we had a 30-year risk model where we could take

01:13:07.100 many more inputs, I would absolutely love to be looking at the retina. I believe, by the way,

01:13:12.540 Zach, that retinal examination should be a part of medicine today for everybody.

01:13:17.180 I would take a retinal exam over a hemoglobin A1C all day, every day. I'd never look at another A1C

01:13:24.620 again if I could see the retina of every one of my patients. But my point is, even if effective today,

01:13:30.540 we could define the data set, and let's overdo it, and we can prune things later. But we want to see

01:13:35.820 these 50 things in everybody to predict every disease. Is there any way to get around the fact

01:13:41.340 that we're going to need 30 years to see this come to fruition in terms of watching how the story plays

01:13:46.300 out? Or are we basically going to say, no, we're going to do this over five years? It won't be that

01:13:51.340 useful because a five-year predictor basically means you're already catching people in the throes of

01:13:55.100 the disease. I'll say three words, electronic health records. So that turns out not to be the

01:14:02.060 answer in the United States. Why? Because in the United States, we move around. We don't stay in any

01:14:09.020 given healthcare system that long. So very rarely will I have all the measurements made on you,

01:14:15.820 Peter, all your glycohemoglobins, all your blood pressures, all your clinic visits, all the imaging

01:14:21.180 studies that you've had. However, that's not the case in Israel, for example. In Israel, they have

01:14:28.460 these HMOs, health maintenance organizations. And one of them, Clarit, I have a good relationship with

01:14:35.020 because they published all the big COVID studies looking at the efficacy of the vaccine. And why

01:14:42.540 could they do that? Because they had the whole population available. And they have about 20,

01:14:48.220 25 years worth of data on all their patients in detail and family relationships. So if you have

01:14:55.980 that kind of data, and Kaiser Permanente also has that kind of data, I think you can actually come

01:15:02.060 close. But you're not going to be able to get retina, gait, voice, because we still have to get those

01:15:07.900 prospectively. I'm going to claim that there are proxies, rough proxies, but for gait, false.

01:15:15.900 And for hearing problems, visits to the audiologist. Now, these are noisier measurements. And so,

01:15:23.820 those of us who are data junkies, like I am, always keep mumbling to ourselves, perfect is the

01:15:30.060 enemy of good. Waiting 30 years to have the perfect data set is not the right answer to help patients

01:15:36.300 now. And there are things that we could know now that are knowable today that we just don't know

01:15:42.860 because we haven't bothered to look. I'll give you a quick example. I did a study of autism using

01:15:49.180 electronic health records, maybe 15 years ago. And I saw there was a lot of GI problems. And I

01:15:55.660 talked to a pediatric expert, and they said, it was a little bit dismissive. They said, brain bad,

01:16:00.940 tummy hurt. I've seen a lot of inflammatory bowel disease. It just doesn't make sense to me that this

01:16:06.220 is somehow effect of brain function. To make a long story short, we did a massive study. We're

01:16:11.660 looking forward to tens of thousands of individuals. And sure enough, we found subgroups of patients who

01:16:15.820 had immunological problems associated with their autism, and they had type 1 diabetes,

01:16:21.500 inflammatory bowel disease, lots of infections. Those were knowable, but they were not known. And I had,

01:16:27.020 frankly, parents coming to me more thankful than for anything else I had ever done for them

01:16:30.860 clinically, because I was telling these parents, they weren't hallucinating that these kids have

01:16:35.420 these problems. They just weren't being recognized by medicine because no one had the big wide angle

01:16:41.340 to see these trends. So, without knowing the field of Alzheimer's the way I do other fields,

01:16:48.140 I bet you there are trends in Alzheimer's that you can pick up today by looking at enough patients

01:16:53.660 that you'll find some that have more frontotemporal components, some that have more effective

01:16:58.780 components, some that have more of an infectious and immunological component. Those are knowable

01:17:04.220 today. Zach, you've already alluded to the fact that we're dealing with a customer, if the physician

01:17:11.580 is the customer, who is not necessarily the most tech-forward customer. And truthfully, like many customers

01:17:20.060 of AI, runs the risk of being marginalized by the technology if the technology gets good enough.

01:17:25.900 And yet, you need the customer to access the patient to make the data system better,

01:17:33.900 to make the training set better. So, how do you see the interplay over the next decade of that dynamic?

01:17:42.620 That's the right question. Because in order for these AI models to work, you need a lot of data,

01:17:47.900 a lot of patients. Where is that data going to come from? So, there are some healthcare systems,

01:17:52.860 like the Mayo Institute, who think they can get enough data in that fashion. There are some data

01:18:00.940 companies that are trying to get relationships with healthcare systems where they can get de-identified

01:18:06.540 data. I'm betting on something else. There is a trend where consumers are going to have increased

01:18:13.580 access to their own data. The 21st Century Cures Act was passed by Congress, and it said that patients

01:18:20.220 should be given access to their own data programmatically. Now, they're not expecting

01:18:24.860 your grandmother to write a program to access the data programmatically, but by having a right to it, 0.97

01:18:30.540 it enables others to do so. So, for example, Apple has something called Apple Health. It has this big

01:18:36.060 heart icon on it. If you're one of the 800 hospitals that they've already hooked up with,

01:18:40.780 Pass General or Brigham Women's, and you're a patient there, if you authenticate yourself to it,

01:18:45.260 if you give it your username and password, it will download into your iPhone, your labs, your meds,

01:18:51.900 your diagnoses, your procedures, as well as all the wearable stuff, your blood pressure that you get

01:18:58.220 as an outpatient, and various other forms of data. That's already happening now. There's not a lot of

01:19:04.140 companies that are taking advantage of that, but right now that data is available on tens of millions

01:19:08.780 of Americans. Isn't it interesting, Zach, how unfriendly that data is in its current form? I'll

01:19:15.420 give you just a silly example in our practice. So, if we send a patient to LabCorp or Boston Heart or

01:19:21.340 Pick Your Favorite Lab, and we want to generate our own internal reports based on those where we want

01:19:28.940 to do some analysis on that, lay out trend sheets, we have to use our own internal software. It's almost

01:19:37.420 impossible to scrape those data out of the labs because they're sending you PDF reports. Their

01:19:44.860 APIs are garbage. Nothing about this is user-friendly. So, even if you have the My Heart thing or whatever, 0.92

01:19:53.020 the My Health thing come on your phone, it's not navigable. It's not searchable. It doesn't show you

01:19:58.540 trends over time. Like, is there a more user-hostile industry from a data perspective than the health

01:20:04.140 industry right now? No, no. And there's a good reason why, because they're keeping you captive.

01:20:10.860 But Peter, the good news is you're speaking to a real nerd. Let me tell you two ways where we could

01:20:17.020 solve your problem. One, if it's in that Apple Health thing, someone can actually write a program,

01:20:22.140 an app on the iPhone, which will take those data as numbers and not have to scrape it. And it could run

01:20:28.620 it through your own trending programs. You could actually use it directly. Also, Gemini and GPT-4,

01:20:34.540 you can actually give it those PDFs. And actually, with the right prompting, it will actually take

01:20:41.100 those data and turn them into tabular spreadsheets. We can't do that because of HIPAA, correct?

01:20:47.340 If the patient gets it from the patient portal, absolutely, you can do that.

01:20:50.700 The patient can do that, but I can't use a patient's data that way.

01:20:54.380 If the patient gives it to you, absolutely. Really? Oh, yes. But it's not de-identified.

01:21:00.940 It doesn't matter. If a patient says, Peter, you can take my 50 LabCorp reports for the last 10 years,

01:21:09.260 and you can run them through ChatGPT to scrape it out and give me an Excel spreadsheet that will

01:21:15.900 perfectly tabularize everything that we can then run into our model to build trends and look for

01:21:21.100 things. I didn't think that was doable, actually. So it's not doable through ChatGPT because your

01:21:26.220 lawyers would say, Peter, you're going to get a million dollars in fines from HIPAA. I'm not a

01:21:31.100 shill for Microsoft. I don't own any stock. But if you do GPT on the Azure cloud that's HIPAA protected,

01:21:37.740 you absolutely can use it with patient consent. 100% you can do it. GPT is being used

01:21:43.660 with patient data out of Stanford right now. EPIC's using GPT-4, and it's absolutely legitimately

01:21:51.980 usable by you. People don't understand that. We've now just totally bypassed OCRs.

01:21:57.420 We do not need to waste our time with, for people not in the acronyms, optical character recognition,

01:22:02.460 which is 15 years ago what we were trying to do to scrape this data. Peter, let me tell you,

01:22:08.940 there's New England Journal of Medicine. I'm on the editorial board there. And we just published

01:22:12.700 three months ago, a picture of the week back of this 72-year-old. And it looks like a bunch of red

01:22:18.620 marks. To me, it looks like something to scratch themselves. And it says, blah, blah, blah. They

01:22:23.100 had trouble sleeping. This is the image of the week. Image of the week. And I took that whole thing,

01:22:27.820 and I took out one important fact, and then gave it to GPT-4, the image and the text. And it came up with

01:22:35.500 the two things I thought it would be, either bleomycin toxicity, which I don't know what

01:22:39.660 that looks like, and shiitake mushroom toxicity. What I'd removed is the fact that the guy had eaten

01:22:47.100 mushrooms the day before. So this thing just like looking at the picture.

01:22:52.860 GPT-4 spit this out? Yes.

01:22:55.900 I don't think most doctors know this, Zach. I don't think most doctors understand. First of all,

01:23:01.820 I can't tell you how many times I get a rash. Well, I try to send a picture to my doctor,

01:23:06.700 or my kid gets a rash, and I'm trying to send a picture to their pediatrician,

01:23:10.700 and they don't know what it is. And it's like, we're rubbing two sticks together,

01:23:14.380 and you're telling me about the Zippo lighter. Yes. And that's what I'm saying is patients without

01:23:19.260 primary care doctors. I know I keep repeating myself. They understand that they have a Zippo

01:23:22.860 lighter waiting three months because of a rash or the symptoms. They say, I'll use this Zippo lighter.

01:23:28.540 It's better than no doctor for sure, and maybe better. That's now. Quickly illustrate it. I don't

01:23:34.460 know squat about the FDA. And so I pulled down from the FDA the adverse event reporting files.

01:23:41.180 It's a big zip file, compressed file. And I said to GPT-4, please analyze this data. And it says,

01:23:47.260 unzipping. Based on this table, I think this is about the adverse events, and this is the

01:23:51.900 locations. What do you want to know? I say, tell me what adverse events for disease-modifying

01:23:57.740 drugs for arthritis. It says, oh, to do that, I'll have to join these two tables.

01:24:02.220 And it just does it. It creates its own Python code. It does it, and it gives me a report.

01:24:08.140 Is this a part of medical education now? You're at Harvard, right? You're at one of the three best

01:24:12.620 medical schools in the United States, arguably in the world. Is this an integral part of the

01:24:17.660 education of medical students today? Do they spend as much time on this as they do histology,

01:24:23.500 where I spent a thousand hours looking at slides under a microscope that I've never once tried to

01:24:29.740 understand? Again, I don't want to say there wasn't a value in doing that. There was,

01:24:34.620 and I'm grateful for having done it. But I want to understand the relative balance of education.

01:24:40.060 It's like the stethoscope. Arguably, we should be using things other than the stethoscope.

01:24:44.540 Let me make sure I don't get fired, or at least beaten severely, by telling you that George Daly,

01:24:50.220 our dean of the medical school, has said explicitly he wants to change all of medical education.

01:24:54.940 So these learnings are infused throughout the four years, but it's going to take some doing.

01:25:01.180 Let's now move on to the next piece of medicine. So we've gone from purely the recognition image-based

01:25:09.980 to how do I combine image with voice, story, text. You've made a very compelling case that we don't

01:25:17.820 need any more technological breakthroughs to augment those. It's purely a data set problem at this point,

01:25:22.780 and a willingness. Let's now move to the procedural. Is there, in our lifetimes,

01:25:28.540 say, Zach, the probability that if you need to have a radical prostatectomy, which currently,

01:25:35.980 by the way, is never done open. This is a procedure that the da Vinci, a robot, has revolutionized.

01:25:41.580 There's no blood loss anymore. When I was a resident, this was one of the bloodiest operations

01:25:46.380 we did. It was the only operation, by the way, for which we had the patients donate their own blood

01:25:52.300 two months ahead of time. That's how guaranteed it was that they were going to need blood transfusions,

01:25:57.580 so we just said, to hell with it. Come in a couple of months before, give your own blood,

01:26:01.180 because you're going to need at least two units following this procedure.

01:26:05.260 Today, it's insane how successful this operation is on a large part of the robot.

01:26:11.180 But the surgeon needs to move the robot. Are we getting to the point where that could change?

01:26:17.340 So let me tell you where we are today. Today, there's been studies where it's collected a bunch

01:26:21.660 of YouTube videos of surgery and traded up one of these general models. So it says, oh, they're

01:26:29.020 putting on the scalpel to cut this ligament. And by the way, that's too close to the blood vessel.

01:26:35.580 They should move it a little bit to the side. That's already happening. Based on what we're seeing with

01:26:40.780 robotics in the general world, I think the da Vinci controlled by a robot 10 years is a very safe bet.

01:26:50.940 It's a very safe bet. In some ways, 10 years is nothing.

01:26:54.620 It's nothing. But it's a very safe bet. The fact is, right now, I can do a better job,

01:26:59.500 by the way, just to go back to our previous discussion, giving you a genetic diagnosis

01:27:04.780 based on your findings than any primary care provider interpreting a genomic test.

01:27:10.860 So are you using that example, Zach, because it's a huge data problem? In other words, that's obvious

01:27:17.900 that you would be able to do that because the amount of data, I mean, there's 3 billion base pairs

01:27:22.780 to be analyzed. So of course, you're going to do a better job.

01:27:25.180 But linking it to symptoms.

01:27:27.100 Yeah, yeah. But you're saying surgery is a data problem because if you turn it into a pixel problem?

01:27:33.420 Pixel and movement.

01:27:34.780 And degrees of freedom. Yeah.

01:27:36.620 That's it. Remember, there's a lot of degrees of freedom in moving a car around traffic. And by the

01:27:42.300 way, lives are on the line there too. Now, medicine is not the only job where lives are at stake.

01:27:49.740 Driving a ton of metal at 60 miles per hour in traffic is also putting lives at stake. And last

01:27:57.100 time I looked, there's several manufacturers who are saying that, or some appreciable fraction

01:28:05.020 of that effort, they're controlling multiple degrees of freedom with a robot.

01:28:08.620 Yeah. I very recently spoke with somebody, I won't name the company, I suppose, but it's one

01:28:15.260 of the companies that's deep in the space of autonomous vehicles. And they very boldly stated,

01:28:21.820 they made a pretty compelling case for it, that if every vehicle on the road was at their level of

01:28:27.100 technology and autonomous driving, you wouldn't have fatalities anymore. But the key was that every

01:28:32.780 vehicle had to be at that level. I don't know if you know enough about that field, but does that

01:28:36.620 sense check to you? Well, first of all, I'm a terrible driver. I am a better driver. It's not

01:28:42.140 for an ad, but the fact is I'm a better driver because I'm not on a Tesla, because I'm a terrible

01:28:46.300 driver. And there's actually a very good message for medicine, because I will paraphrase this.

01:28:51.900 I knew enough to know that I need to jiggle the steering wheel when I'm driving with a Tesla,

01:28:56.060 because otherwise it will assume that I'm just zoning out. But what I didn't realize is this,

01:29:01.260 I'm very bad. I'll pick up my phone and I'll look at it. I didn't realize it was looking at me and

01:29:06.060 it says, because Zach put down the phone. So I, okay, I put down. Three minutes later,

01:29:10.460 I pick it up again and it says, okay, that's it. I'm switching off autopilot. So it switches off

01:29:16.140 autopilot and now I have to pay attention, full attention. Then I get home and it says, all right,

01:29:21.740 that was bad. You do that four more times. I'm switching off autopilot until the next software

01:29:27.820 update. And the reason I mentioned that is it takes a certain amount of confidence to do that

01:29:33.180 to your customer base saying, I'm switching off the thing that they bought me for. In medicine,

01:29:38.780 how likely is it that we're going to fall asleep at the wheel if we have an AI thinking for us?

01:29:43.980 It's a real issue. We know for a fact, for example, back in the nineties,

01:29:47.740 that doses for a drug like a dancetron, where people would talk endlessly about how frequently

01:29:52.860 you should be given it with what dose. The moment you put it in the order entry system,

01:29:56.540 95% of doctors would just use the default there. And so how in medicine are we going to keep doctors

01:30:03.100 awake at the wheel? And will we dare to do that kind of challenges that I just described the car

01:30:09.340 doing? So just to get back to it, I do believe because of what I've seen with autonomy and robots

01:30:16.940 that as fancy as we think that is, controlling a dementia robot will probably have less

01:30:23.980 bad outcomes. Every once in a while, someone nicks something and you have to go into full

01:30:29.340 surgery or they go home and they die on the way home because they exsanguinate. I think it's just

01:30:34.620 going to be safer. It's just unbelievable for me to wrap my head around that. But truthfully,

01:30:41.500 it's impossible for me to wrap my head around what's already happened. So I guess I'll try to

01:30:45.740 retain the humility that says I reserve the right to be startled. Again, there are certain things that

01:30:51.980 seem much easier than others. Like I have an easier time believing we're going to be able to replace

01:30:56.140 interventional cardiologists where the number of degrees of freedom, the complexity and the

01:31:02.060 relationship between what the image shows, what the cath shows and what the input is, the stent,

01:31:09.180 that gap is much narrower. Yeah, I can see a bridge to that. But when you talk about doing a Whipple

01:31:14.460 procedure, when you talk about what it means to cell by cell take a tumor off the superior mesenteric

01:31:22.380 vessels, I'm thinking, oh my God. Since we're on record, I'm going to say, I'm talking about your

01:31:28.380 routine prostate removal. Yeah. First 10 years, I would take that bet today. Wow. Let's go one layer further.

01:31:36.940 Sure. Let's talk about mental health. This is a field of medicine today that I would also argue

01:31:42.940 is grossly underserved. Everything you've said to date resonates. I completely agree from my own

01:31:49.980 experience that the resources in pediatrics and primary care, I mean, these things are unfortunate

01:31:55.980 at the moment. Harvard has, I think, 60% of undergraduates are getting some sort of mental

01:32:01.180 health support and it's completely outdoing all the resources available to the university health

01:32:06.780 services. And so we have to outsource some of our mental health. And this is a very richly endowed

01:32:11.660 university. In general, we don't have the resources. So here we live in a world where I think the evidence

01:32:18.300 is very clear that when a person is depressed, when a person is anxious, when a person has any sort of

01:32:24.580 mental or emotional illness, pharmacotherapy plays a role, but it can't display psychotherapy.

01:32:30.100 You have to be able to put these two things together. And the data would suggest that the

01:32:35.660 knowledge of your psychotherapist is important, but it's less important than the rapport you can

01:32:41.700 generate with that individual. Now, based on that, do you believe that the most sacred, protected,

01:32:49.620 if you want to use that term, profession within all of medicine will then be psychiatry?

01:32:54.980 I'd like to think that if I had a psychiatric GPT speaking to me,

01:33:00.100 I wouldn't think that it understood me. On the other hand, back in the 1960s or 70s,

01:33:08.900 there was a program called Eliza and it was a simple pattern matching program. It would just emulate

01:33:14.740 what's called a Rogerian therapist, where I really hate my mother. Why do you say you hate your mother?

01:33:21.140 Oh, it's because I don't like the way she fed me. What is it about the way she fed you? Just very,

01:33:26.980 very simple pattern matching. And this Eliza program, which was developed by Joe Weizenbaum at MIT,

01:33:34.660 his own secretary would lock herself in her office to have sessions with this thing because it's

01:33:39.700 non-judgmental. This was in the 80s?

01:33:42.260 70s or 60s.

01:33:43.700 Wow.

01:33:44.420 Yeah. And it turns out that there's a large group of patients who actually would rather have a non-human,

01:33:51.380 non-judgmental person who remembers what they've said from last time, shows empathy verbally. Again,

01:33:58.500 I wrote this book with Peter Lee and Peter Lee made a big deal in the book about how GPT-4 is showing

01:34:05.220 empathy. In the book, I argued with him that this is not that big a deal. And I said, I remember from

01:34:11.140 medical school being told that some of the most popular doctors are popular because they're very

01:34:16.980 deep empaths, not necessarily the best doctors. And so I said, you know, for certain things,

01:34:22.100 that's just me. I could imagine a lot of, for example, cognitive behavioral therapy being done

01:34:29.380 and be found acceptable by a subset of human beings. It's not, wouldn't be for me. I'd say,

01:34:34.100 I'm just speaking to some stupid program. But if it's giving you insight into yourself and it is 1.00

01:34:39.220 based on the wisdom called for millions of patients, who's to say that it's worse? And it's certainly not

01:34:46.420 judgmental. And maybe it'll be less.

01:34:48.420 So Zach, you're born probably just after the first AI boom. You come of age, intellectually,

01:34:58.740 academically in the second. And now in the mature part of your career, when you're at the height of

01:35:05.780 your esteem, you're riding the wave of this third version, which I don't think anybody would argue

01:35:13.220 is going anywhere. As you look out over the next decade, and we'll start with medicine,

01:35:20.180 what are you most excited about? And what are you most afraid of with respect to AI?

01:35:25.780 Specifically with regard to medicine, what I'm most concerned about is how it could be used by

01:35:33.300 the medical establishment to keep things the way they are, to pour concrete over practices.

01:35:39.700 What I'm most excited about is alternative business models, young doctors who create businesses

01:35:48.260 outside the mold of hospitals. Hospitals are these very, very complex entities.

01:35:55.300 They make billions of dollars, some of the bigger ones, but with very small margins, one to 2%.

01:36:01.220 When you make, have huge revenue, but very small margins, you're going to be very risk averse.

01:36:06.180 And you're not going to want to change. And so what I'm excited about is the opportunity for new

01:36:13.380 businesses and new ways of delivering to patients insights that are data-driven. What I'm worried

01:36:19.940 about is hospitals doing a bunch of information blocking and regulations that will make it harder for

01:36:28.820 these new businesses to get created. Understandably, they don't want to be disrupted. That's the danger.

01:36:34.260 In that latter case or that case that you're afraid of, Zach, can patients themselves work around the

01:36:42.180 hospitals with these new companies, these disruptive companies and say, look, we have the legal framework

01:36:48.980 that says, I own my data as a patient. I own my data. Believe me, we know this in our practice.

01:36:54.340 Just because our patients own the data doesn't make it easy to get. There is no aspect of my practice

01:37:00.180 that is more miserable and more inefficient than data acquisition from hospitals. It's actually

01:37:06.740 comical.

01:37:07.140 Absolutely comical. And I do pay hundreds of dollars to get my data from my patients with rare and unknown

01:37:13.940 diseases in this network extracted from the hospitals because it's worth it to pay someone to do that

01:37:19.780 extraction. Yeah. But now I'm telling you it is doable.

01:37:23.860 So you're saying because of that, are you confident that the legal framework for patients to have their

01:37:29.540 data coupled with AI and companies, do you think that that will be a sufficient hedge against your

01:37:36.420 biggest fear?

01:37:37.060 I think that unlike my 10-year prostatectomy by robot prediction, I'm not as certain, but I would

01:37:44.020 give better than 50% odds that in the next 10 years, there'll be a company, at least one company,

01:37:50.100 that figures out how to use that patient's right to access through dirty APIs, using AI to clean it up,

01:37:59.300 provide decision support with human doctors or health professionals to create alternative

01:38:05.460 businesses. I am convinced because the demand is there. And I think that you'll see companies that

01:38:12.820 are even willing to put themselves at risk. What I mean by that, are willing to take the medical risk

01:38:18.100 on that if they do better than a certain level of performance, they get paid more. And if they do

01:38:24.340 worse- They don't get paid. Yeah.

01:38:26.660 I believe there were companies that are going to be in that space, but that is because I don't want to

01:38:32.420 underestimate the medical establishment's ability to squish threats. So we'll see.

01:38:38.100 Okay. Now let's just pivot to AI outside of medicine. Same question. What are you most afraid

01:38:44.340 of over the next decade? So maybe we're not talking about self-awareness and Skynet, but next decade,

01:38:51.700 what are you most afraid of? And what are you most excited about?

01:38:55.140 What I'm most afraid of is a lot of the ills of social networks being magnified by use of these

01:39:06.500 AIs to further accelerate cognitive chaos and vitriol that fills our social experiences on the net. 0.96

01:39:17.620 It could be used to accelerate them. So that's my biggest fear.

01:39:21.300 I saw an article two weeks ago that was an individual. I can't remember if they were

01:39:26.580 currently in or formerly part of the FBI. And they stated that they believed, I think it was

01:39:32.100 somewhere between 75 and 90% of quote unquote individuals on social media were not in fact

01:39:38.500 individuals. I don't know if you spend enough time on social media to have a point of view on that.

01:39:43.060 Unfortunately, I have to admit to the fact that my daughter, who's now 20 years old,

01:39:47.620 been four years ago, she bought me a mug that says on it, Twitter addict. I spent enough time.

01:39:53.380 I would not be surprised if some large fraction, our bots could get worse. And it's going to be

01:39:59.380 harder to actually distinguish reality from human beings, harder and harder and harder.

01:40:05.780 That's the real problem. We are fundamentally social animals. And if we cannot understand our social

01:40:12.980 context in most of our interactions, it's going to make us crazy, or I should say crazier. And my most

01:40:21.300 positive aspect is, I think that these tools can be used to expand the creative expression of all

01:40:28.740 people. If you're a poor driver like me, I'm going to be a better driver. If you're a lousy musician,

01:40:36.580 but have a great ear, you're going to be able to express yourself musically in ways that you could not do

01:40:41.300 before. I think you're going to see filmmakers who were never meant to be filmmakers before

01:40:46.820 express themselves. I think human expression is going to be expanded because just like printing

01:40:53.700 press allowed all sorts of... In fact, it's a good analogy because the printing press also created a

01:40:58.180 bunch of wars because it allowed people to make clear their opposition to the church and so on,

01:41:03.300 enabled a number of bad things to happen. But it allowed also expression of all literature in ways

01:41:08.420 that would have not been possible without the printing press. I'm looking forward to human

01:41:12.740 expression and creativity. I can't imagine you haven't played with some of the picture generation

01:41:17.460 or music generation capabilities of AI, or if you haven't, I strongly recommend. You're going to be

01:41:22.260 amazed. I have not. I am ashamed maybe to admit my interactions with AI are limited to really chat GPT

01:41:29.860 for and basically problem solving. Solve this problem for me. And by the way, I think I'm doing it at a very

01:41:35.860 JV level. I could really up my game there. Just before we started this podcast, I thought of a 0.82

01:41:40.900 problem. I've been asking my assistant to solve because A, I don't have the time to solve it and

01:41:45.860 I'm not even sure how I would solve it. It would take me a long time. I've been asking her to solve

01:41:49.700 it and it's actually pretty hard. And then I realized, oh my God, why am I not asking chat GPT for

01:41:54.340 to do it? So I just started typing in the question. It's a bit of an elaborate question. As soon as we're

01:41:59.620 done with this podcast, I'll probably go right back to it, but I haven't done anything creatively with it.

01:42:03.540 What I will say is what does this mean for human greatness? So right now, if you look at a book

01:42:12.560 that's been written and someone who's won a Pulitzer Prize, you sort of recognize like, I don't know if

01:42:17.800 you read Sid Mukherjee, right? He's one of my favorite writers when it comes to writing about science and

01:42:22.960 medicine. When I read something that Sid has written, I think to myself, there's a reason that he is so

01:42:30.180 special. He and he almost alone can do something we can't do. I've written a book. It doesn't matter.

01:42:38.500 I could write a hundred books. I'll never write like Sid and that's okay. I'm no worse a person.

01:42:43.860 I'm no worse a person than Sid, but he has a special gift that I can appreciate just as we could all

01:42:50.120 appreciate watching an exceptional athlete or an exceptional artist or musician. Does it mean anything

01:42:56.080 if that line becomes blurred? That's the right question. And yes, Sid writes like poetry.

01:43:04.440 Here's an answer which I don't like. I've heard many times people said, oh, you know that Deep Blue

01:43:10.620 beat Kasparov in chess, but chess is more popular than it ever was, even though we know that the best

01:43:17.580 chess players in the world are computers. So that's one answer. I don't like that answer at all.

01:43:22.720 Yeah. Because if we create Sid GPT and Sid wrote Alzheimer's, the second greatest malady,

01:43:31.180 and he wrote it in full Sid style, but it was not Sid, but it was just as empathic family references.

01:43:39.060 Right. The weaving of history with story with science. Yeah.

01:43:42.060 If it did that and it was just a computer, how would you feel about it, Peter?

01:43:45.120 I mean, Zach, you are asking the jugular question. I would enjoy it, I think, just as much,

01:43:51.840 but I don't know who I would praise. Maybe I have in me a weakness slash tendency to want to idolize.

01:44:00.180 You know, I'm not a religious person, so my idols aren't religious, but I do tend to love to see

01:44:06.340 greatness. I love to look at someone who wrote something who's amazing and say, that amazes me.

01:44:11.720 I love to be able to look at the best driver in the history of Formula One and study everything

01:44:17.900 about what they did to make them so great. So I'm not sure what it means in terms of that.

01:44:22.940 I don't know how it would change that.

01:44:24.560 I grew up in Switzerland, in Geneva. And even though I have this American accent,

01:44:28.860 both my parents were from Poland. And so the reason I have an American accent is I went to

01:44:33.180 international school with a lot of Americans. All I read was whatever my dad would get me from

01:44:37.900 England in science fiction. So I'm a big science fiction fan. So let me go science fiction on you

01:44:43.080 to answer this question. It's not going to be in 10 years, but it could be in 50 years.

01:44:48.240 You'll have idols and idols will be, yes, Greg Orovich wrote a great novel, but you know,

01:44:53.780 AI 521, their understanding of the human condition is wonderful. I cry when I read their novels.

01:45:01.160 They'll be a part of the ecosystem. They'll be entities within us, whether they are self-aware or not,

01:45:06.440 will become a philosophical question. Let's not go that narrow path, that disgusting rabbit hole

01:45:11.540 where I wonder, does Peter actually have consciousness or not? Does he have the same

01:45:15.820 processes as I do? We won't know that about these, or maybe we will, but will it matter

01:45:21.060 if they're just among us? And they'll have brands, they'll have companies around them.

01:45:26.420 They'll be superstars. And they'll be Dr. Fubar from Kansas, trained on iridivic medicine,

01:45:37.580 the key person for alternative medicine, not a human, but we love what they do.

01:45:43.260 Okay. Last question. How long until, from at least an intellectual perspective, we are immortal?

01:45:51.400 So if I died today, my children will not have access to my thoughts and musings any longer.

01:46:00.940 Will there be a point at which during my lifetime, an AI can be trained to be identical to me,

01:46:10.320 at least from a goalpost perspective, to the point where after my death, my children could say,

01:46:17.980 dad, what should I do about this situation? And it can answer them in a way that I would have.

01:46:25.940 It's a great question because that was an early business plan that was

01:46:29.940 generated shortly after GPT-4 came out. In fact, I was talking very briefly to Mark Cuban

01:46:36.240 because he saw GPT-4. I think he got trademarks or copyrights on his voice,

01:46:42.180 all his work and likeness so that someone could not create a mark who responded in all the ways he

01:46:50.440 does. And I'll tell you that it sounds crazy, but there's a company called rewind.ai. And I have

01:46:58.580 it running right now. And everything that appears on my screen, it's recording. Every sound that it

01:47:06.720 hears, it's recording. And if characters appear on the screen, it'll OCR them. If a voice appears,

01:47:13.460 and then if I have a question, I say, when did I speak with Peter Atiyah? They'll find it for me.

01:47:18.540 I'll say, who was I talking about AI and Alzheimer's? And they'll find this video on a timeline.

01:47:28.180 How many terabytes of data is this, Zach?

01:47:30.740 Amazingly small. It's just gigabytes.

01:47:33.780 How is that possible?

01:47:35.320 Because A, it compresses it down in real time with using Apple Silicon. And second of all,

01:47:40.360 you and I, you're old and you don't realize that gigabytes are not big on a standard Mac that has

01:47:46.880 a terabyte. That's a thousand gigabytes. And so you can compress audio immensely. It's actually not

01:47:54.120 taking video. It's just taking multiple snapshots every time the screen changes by a certain amount.

01:47:59.020 Yeah. It's not trying to get video resolution per se.

01:48:02.020 No. And it's doing it. And I can see a timeline. It's quite remarkable. And so that is enough,

01:48:09.240 in my opinion, data so that with enough conversations like this, someone could create

01:48:14.840 a pretty good approximation of at least public Zach.

01:48:18.780 So then the next question is, is Zach willing to have rewind AI on a recording device, his phone

01:48:26.660 with him 24 seven in his private moments, in his intimate moments, when he is arguing with his

01:48:33.400 wife, when he's upset at his kids, when he's having the most amazing experience with his postdoc. Like

01:48:40.120 if you think about the entire range of experiences we have from the good, the bad, the ugly, those are

01:48:45.800 probably necessary. If we want to formulate the essence of ourselves, you envision a day in which

01:48:51.420 people can say, look, I'm willing to take the risks associated with that. And there are clear risks

01:48:55.900 associated with doing that, but I'm willing to take those risks in order to have this legacy,

01:49:02.800 this data set to be turned into a legacy.

01:49:05.300 I think it's actually pretty creepy to come back from the dead to talk to your children.

01:49:09.600 So I actually have other goals. Here's where I take it. We are being monitored all the time.

01:49:14.740 We have iPhones, we have Alexa devices. I don't know what is actually being stored by whom and what.

01:49:20.600 And people are going to use this data in ways that we do or don't know. I feel it's us,

01:49:26.140 the little guy, if we have our own copy and we can say, well, actually, look, this is what I said then.

01:49:32.160 Yeah. That was taken out of context.

01:49:33.480 Yeah. That was taken out of context. And I can do it. I have an assistant that can

01:49:37.980 just find it and find exactly and find all the times I said it. I think that's good. I think it's

01:49:45.240 messing with your kid's head to have you come back from the dead and give advice, even though

01:49:49.820 they might be tempted. Technically, I think it's going to be not that difficult. And again,

01:49:55.080 speaking about Rewind AI, again, I have no stake in them. I think I might have paid them for a license

01:50:01.260 to run on my computer, but the microphone is always on. So when I'm talking to students in

01:50:07.100 my office, it's taking that down. So there are some moments in my life where I don't want to be

01:50:12.860 on record. There are big chunks of my life that are actually being stored this way.

01:50:16.780 Well, Zach, this has been a very interesting discussion. I've learned a lot.

01:50:20.100 I probably came into this discussion with about the same level of knowledge, maybe slightly more

01:50:25.540 than the average person, but clearly not much more on just the general principles of AI, the

01:50:30.760 evolution of AI. I guess if anything surprises me, and a lot does, but nothing surprises me more

01:50:36.780 than the timescale that you've painted for the evolution within my particular field and your

01:50:43.120 particular field, which is medicine. I had no clue that we were getting this close to that level of

01:50:53.260 intelligence. Peter, if I were you, this is not an offer because I'm too busy, but you're a capable guy

01:50:59.260 and you have a great network. If I was running the clinic that you're running, I would take advantage

01:51:03.360 of now. I would get those videos and those sounds and get all my patients with, of course, their

01:51:11.300 consent to be part of this and to actually follow their progress, not just the way to report it,

01:51:17.560 but by their gait, by the way they look. You can do great things in what you're doing and advance the

01:51:24.200 state of art. You're asking who's going to do it. You're doing some interesting things. You could be

01:51:29.180 pushing the envelope using these technologies as just another very smart, comprehensive assistant.

01:51:36.200 Zach, you've given me a lot to think about. I'm grateful for your time and obviously for your

01:51:40.640 insight and years of dedication that have allowed us to be sitting here having this discussion.

01:51:45.320 Thank you very much. It was a great pleasure. Thank you for your time.

01:51:48.900 Thank you for listening to this week's episode of The Drive. It's extremely important to me to

01:51:53.880 provide all of this content without relying on paid ads. To do this, our work is made entirely

01:51:58.860 possible by our members. And in return, we offer exclusive member-only content and benefits above and

01:52:05.500 beyond what is available for free. So if you want to take your knowledge of this space to the next

01:52:09.820 level, it's our goal to ensure members get back much more than the price of the subscription.

01:52:14.980 Premium membership includes several benefits. First, comprehensive podcast show notes that detail

01:52:21.100 every topic, paper, person, and thing that we discuss in each episode. And the word on the street is

01:52:26.940 nobody's show notes rival ours. Second, monthly Ask Me Anything or AMA episodes. These episodes are

01:52:34.820 comprised of detailed responses to subscriber questions typically focused on a single topic

01:52:39.680 and are designed to offer a great deal of clarity and detail on topics of special interest to our

01:52:45.040 members. You'll also get access to the show notes for these episodes, of course. Third, delivery of our

01:52:50.940 premium newsletter, which is put together by our dedicated team of research analysts. This newsletter

01:52:56.320 covers a wide range of topics related to longevity and provides much more detail than our free weekly

01:53:02.560 newsletter. Fourth, access to our private podcast feed that provides you with access to every episode,

01:53:09.440 including AMA's sans the spiel you're listening to now and in your regular podcast feed. Fifth,

01:53:16.360 the Qualies, an additional member-only podcast we put together that serves as a highlight reel featuring

01:53:22.560 the best excerpts from previous episodes of The Drive. This is a great way to catch up on previous episodes

01:53:28.280 without having to go back and listen to each one of them. And finally, other benefits that are added

01:53:33.100 along the way. If you want to learn more and access these member-only benefits, you can head over to

01:53:38.660 peteratiamd.com forward slash subscribe. You can also find me on YouTube, Instagram, and Twitter,

01:53:45.600 all with the handle peteratiamd. You can also leave us a review on Apple Podcasts or whatever podcast

01:53:52.160 player you use. This podcast is for general informational purposes only and does not

01:53:57.380 constitute the practice of medicine, nursing, or other professional healthcare services,

01:54:01.320 including the giving of medical advice. No doctor-patient relationship is formed. The use

01:54:07.140 of this information and the materials linked to this podcast is at the user's own risk. The content

01:54:12.920 on this podcast is not intended to be a substitute for professional medical advice, diagnosis, or treatment.

01:54:18.360 Users should not disregard or delay in obtaining medical advice from any medical condition they

01:54:23.560 have, and they should seek the assistance of their healthcare professionals for any such conditions.

01:54:28.960 Finally, I take all conflicts of interest very seriously. For all of my disclosures and the

01:54:34.080 companies I invest in or advise, please visit peteratiamd.com forward slash about where I keep an up-to-date

01:54:42.040 and active list of all disclosures.

01:54:48.360 Thank you.

The Peter Attia Drive - July 15, 2024

#309 ‒ AI in medicine： its potential to revolutionize disease prediction, diagnosis, and outcomes, causes for concern in medicine and beyond, and more ｜ Isaac Kohane, M.D., Ph.D.

Episode Stats

Summary

Transcript