podcast Peter Attia 2024-07-15 topics

#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.

Audio

Show notes

Isaac “Zak” Kohane, a pioneering physician-scientist and chair of the Department of Biomedical Informatics at Harvard Medical School, has authored numerous papers and influential books on artificial intelligence (AI), including The AI Revolution in Medicine: GPT-4 and Beyond. In this episode, Zak explores the evolution of AI, from its early iterations to the current third generation, illuminating how it is transforming medicine today and unlocking astonishing possibilities for the future. He shares insights from his unconventional journey and early interactions with GPT-4, highlighting significant AI advancements in image-based medical specialties, early disease diagnosis, and the potential for autonomous robotic surgery. He also delves into the ethical concerns and regulatory challenges of AI, its potential to augment clinicians, and the broader implications of AI achieving human-like creativity and expertise.

Subscribe on: APPLE PODCASTS | RSS | GOOGLE | OVERCAST | STITCHER

We discuss:

Zak’s unconventional journey to becoming a pioneering physician-scientist, and his early interactions with GPT-4 [2:15];
The evolution of AI from the earliest versions to today’s neural networks, and the shifting definitions of intelligence over time [8:00];
How vast data sets, advanced neural networks, and powerful GPU technology have driven AI from its early limitations to achieving remarkable successes in medicine and other fields [19:00];
An AI breakthrough in medicine: the ability to accurately recognize retinopathy [29:00];
Third generation AI: how improvements in natural language processing significantly advanced AI capabilities [32:00];
AI concerns and regulation: misuse by individuals, military applications, displacement of jobs, and potential existential concerns [37:30];
How AI is enhancing image-based medical specialties like radiology [49:15];
The use of AI by patients and doctors [55:45];
The potential for AI to augment clinicians and address physician shortages [1:02:45];
The potential for AI to revolutionize early diagnosis and prediction of diseases: Alzheimer’s disease, CVD, autism, and more [1:08:00];
The future of AI in healthcare: integration of patient data, improved diagnostics, and the challenges of data accessibility and regulatory compliance [1:17:00];
The future of autonomous robotic surgery [1:25:00];
AI and the future of mental health care [1:31:30];
How AI may transform and disrupt the medical industry: new business models and the potential resistance from established medical institutions [1:34:45];
Potential positive and negative impacts of AI outside of medicine over the next decade [1:38:30];
The implications of AI achieving a level of creativity and expertise comparable to exceptional human talents [1:42:00];
Digital immortality and legacy: the potential to emulate an individual’s personality and responses and the ethical questions surrounding it [1:45:45];
Parting thoughts [1:50:15]; and
More.

Show Notes

Notes from intro :
Isaac Kohane goes by Zak and is a physician scientist and chair of the department of Biomedical Informatics at Harvard Medical School He’s an Associate Professor of Medicine at Brigham and Women’s Hospital
Zak has published several hundred papers in the medical literature and authored the widely-used books Microarrays for Integrative Genomics and The AI Revolution in Medicine: GPT-4 and Beyond
He is the Editor-in-Chief of the newly launched New England Journal of Medicine AI
In this episode we talk about the evolution of AI It wasn’t clear to Peter until this interview that we’re really in the 3rd generation of AI Zak has been part of both the 2nd and current generation
We talk about AI’s ability to impact medicine today (where is it having an impact), and where will it have an impact in the near term? What seems likely
We talk about what the future can hold The difference between science fiction and the potential for where we hope it could go
This is a very interesting podcast for Peter, and a topic that he knows so little about
He’s an Associate Professor of Medicine at Brigham and Women’s Hospital
It wasn’t clear to Peter until this interview that we’re really in the 3rd generation of AI
Zak has been part of both the 2nd and current generation
What seems likely
The difference between science fiction and the potential for where we hope it could go

Zak’s unconventional journey to becoming a pioneering physician-scientist, and his early interactions with GPT-4 [2:15]

Give folks a sense of your background. Your path through medical school and training was not very typical

Zak grew up in Switzerland
Nobody in his family was a doctor
He came to the US, decided to major in biology, and then got nerd sniped by computing in the late 70s, so he minored in computer science
He goes to medical school, and in the middle of the first year he realized it was not what he expected It’s a noble profession, but it’s not a science (it’s an art) He thought he was going into science
He bails out for a while to do a PhD in computer science This is in the early 80s, and it’s a heyday of AI (actually a 2nd heyday) We’re currently in the 3rd heyday It was a time of great promise With retrospective scope, it’s very clear that it was not going to be successful But unlike today, we had not released it to the public It was not actually working in the way that we thought it was going to work, and it certainly didn’t scale
His thesis advisor at MIT ( Peter Szolovits ) said, “ Zak, you should finish your clinical training because I’m not getting a lot of respect from clinicians. And so, to bring rational decision making to the clinic, you really want to finish your clinical training. ”
Zak finished medical school, did a residency in pediatrics and then pediatric endocrinology, which was actually extremely enjoyable
When he was done, he restarted his research in computing
He started a lab at Children’s Hospital in Boston and then a center of biomedical informatics at the medical school
Zak was getting a lot of grants, and like almost every other endeavor, getting money gets attention from the powers that be
They asked him to start a center and then eventually a new department of biomedical informatics that he’s the chair of They now have 16 professors or assistant professors of biomedical informatics
Zak has been involved in a lot of machine learning projects
It’s a noble profession, but it’s not a science (it’s an art)
He thought he was going into science
This is in the early 80s, and it’s a heyday of AI (actually a 2nd heyday) We’re currently in the 3rd heyday
It was a time of great promise
With retrospective scope, it’s very clear that it was not going to be successful But unlike today, we had not released it to the public It was not actually working in the way that we thought it was going to work, and it certainly didn’t scale
We’re currently in the 3rd heyday
But unlike today, we had not released it to the public
It was not actually working in the way that we thought it was going to work, and it certainly didn’t scale
They now have 16 professors or assistant professors of biomedical informatics

Like everybody else, Zak was taken by surprise by large language models

He got an email from Peter Lee in October ‘22 (it was right out of a Michael Crichton novel), that said, “ Zak, if you’ll answer the phone, I can’t tell you what it’s about, but it’ll be well worth your while .” Peter was a Professor of Computer Science at CMU and also department chair there Then, he went to DARPA and then he went to Microsoft He told Zak about GPT-4, and this was before any of us had heard about ChatGPT (which is initially GPT-3.5) He gets Zak early access to it when no one else knows that it exists, and Zak starts trying it against hard cases
Zak remembers from his training, being called down to the nursery There’s a child with a small phallus and a hole at the base of the phallus, and they can’t palpate testicles They want to know what to do because Zak is a pediatric endocrinologist
Peter was a Professor of Computer Science at CMU and also department chair there
Then, he went to DARPA and then he went to Microsoft
He told Zak about GPT-4, and this was before any of us had heard about ChatGPT (which is initially GPT-3.5)
He gets Zak early access to it when no one else knows that it exists, and Zak starts trying it against hard cases
There’s a child with a small phallus and a hole at the base of the phallus, and they can’t palpate testicles
They want to know what to do because Zak is a pediatric endocrinologist

So Zak asked GPT-4 about this case, “ What would you do? What are you thinking about? ”

It runs him through the whole workup of these very rare cases of ambiguous genitalia
In this case, it was congenital adrenal hyperplasia where the making of excess androgens during pregnancy and then subsequently birth causes the clitoris to swell, form the glans of the penis, of the phallus, and the labia minora to fuse to form the shaft of what looks like a penis But there’s no testicles, there’s ovaries
There’s a whole endocrine workup with genetic tests, hormonal tests, ultrasound, and it does it all
Zak explains, “ It really blows my mind, because very few of us in computer science really thought that these large language models would scale up the way they do. It was just not expected .”
Talking to Bill Gates about this, he told me that his line engineers in Microsoft Research, a lot of his fanciest computer scientists did not expect this But the line engineers at Microsoft were just watching the scale-up, GPT-0, 1, 2, and they just saw it was going to keep on scaling up with the size of the data and with the size of the model And they said, “ Yeah, of course this is going to achieve this kind of expertise .”
But there’s no testicles, there’s ovaries
But the line engineers at Microsoft were just watching the scale-up, GPT-0, 1, 2, and they just saw it was going to keep on scaling up with the size of the data and with the size of the model
And they said, “ Yeah, of course this is going to achieve this kind of expertise .”

“ But the rest of us, I think because we value our own intellects so much, we couldn’t imagine how we would get that kind of conversational expertise just by scaling up the model and the data set .”‒ Zak Kohane

The evolution of AI from the earliest versions to today’s neural networks, and the shifting definitions of intelligence over time [8:00]

Zak alluded to the fact that when he was doing his PhD in the early ‘80s, he was in the 2nd generation of AI
This leads Peter to assume that the 1st generation was shortly following World War II Alan Turing and the Turing test
Alan Turing and the Turing test

Talk us through what Alan Turing posited, what the Turing test was and proposed to be, and really what the 1st generation of AI was

After World War II, we had computing machines, and anybody who was a serious computer scientist could see that you could have these processes that could generate other processes You could see how these processes could take inputs and become more sophisticated
As a result, shortly after World War II, we actually had artificial neural networks The perceptron , which was modeled on the ideas of a neuron that could take inputs from the environment and then have certain expectations, and if you updated the neuron as to what was going on, it would update the weights going into that artificial neuron
Going back to Turing, he came up with a test that said essentially if a computational entity could maintain essentially its side of the conversation without revealing that it was a computer and that others would mistake it for a human Then for all intents and purposes that would be intelligent behavior
You could see how these processes could take inputs and become more sophisticated
The perceptron , which was modeled on the ideas of a neuron that could take inputs from the environment and then have certain expectations, and if you updated the neuron as to what was going on, it would update the weights going into that artificial neuron
Then for all intents and purposes that would be intelligent behavior

There’s been all sorts of additional constraints put on it, and one of the hallmarks of AI is that it keeps on moving the goalposts of what we consider to be intelligent behavior

If you had told someone in the ’60s that the world chess masters were going to be beaten by a computer program they’d say, “ Well, that’s AI. ” And then, when Kasparov was beaten by the Deep Blue (by the IBM machine), people said, “ Well, it’s just doing search very well. It’s searching through all the possible moves in the future. It also knows all the grandmaster moves. It has a huge encyclopedic store of all the different grandmaster moves. This is not really intelligent behavior .”
If you had told people it could recognize human faces and find your grandmother in a picture, on any picture in the internet, they’d say, “ Well, that’s intelligence .” And of course, when we did it, “ No, that was not intelligent. ”
And then, when we said it could write a rap poem about Peter Attia based on your web page, and it did that, well, that would be intelligent That would be creative
But then, if you said it’s doing it based on having created a computational model based on all the text ever generated by human beings, as much as we can gather, which is 1-6 terabytes of data, and this computational model basically is predicting what is the next word that’s going to be said, not just the next word, but of the millions of words that could be, what are the probabilities of that next word, that is what’s generating that rap, there’s people who are arguing that’s not intelligence
And then, when Kasparov was beaten by the Deep Blue (by the IBM machine), people said, “ Well, it’s just doing search very well. It’s searching through all the possible moves in the future. It also knows all the grandmaster moves. It has a huge encyclopedic store of all the different grandmaster moves. This is not really intelligent behavior .”
And of course, when we did it, “ No, that was not intelligent. ”
That would be creative

So, the goalposts around the Turing test keep getting moved

Zak no longer finds that an interesting topic, because what it’s actually doing and whether you want to call it intelligent or not, that’s up to you It’s like discussing whether a dog is intelligent Is a baby intelligent before it can recognize the consistency of objects? Initially, babies, if you hide something from it, it’s gone; and it comes back, it’s a surprise
It’s like discussing whether a dog is intelligent
Is a baby intelligent before it can recognize the consistency of objects? Initially, babies, if you hide something from it, it’s gone; and it comes back, it’s a surprise
Initially, babies, if you hide something from it, it’s gone; and it comes back, it’s a surprise

There’s this spectrum of intelligent behavior

There’s a very simple computational model of predicting the next word called a Markov model
Several years ago, people were studying songbirds, and they were able to predict the full song, the next note, and the next note of a song just using a very simple Markov model
So from that perspective, we think that we’re all very smart, but Zak explains, “ The fact that you and I without thinking too hard about it can come up with fluid speech… ”

The model now

The model is now a trillion parameters
It’s not a simple Markov model but still a model
Unfortunately, the late Kahneman’s notions of Thinking Fast and Slow , and this notion of system 1, which is this pattern in recognition, which is very much similar to what we’re seeing here, and system 2, which is the more deliberate and much more conscious kind of thinking that we pride ourselves on
But a lot of what we do is this reflexive, very fast pattern recognition

Advances seem to be happening at a nonlinear pace

Peter thinks back to World War II, where we saw rule-based computing come of age Anybody who’s gone back and watched movies about the Manhattan Project or the decoding of all the sorts of things that took place, Enigma, for example, that’s straight rules-based computational power, and we’re obviously at the limits That can only go so far
It seems that there was a long hiatus before we went from there to maybe what some have called context-based computation What your Siri does or Alexa, which is a step quite beyond that
Then, from there we go to what Zak has already talked about: Blue or Watson Where you have computers that are probably going even 1 step further
Then, where we are now is GPT-4
When you look at the output, it’s staggeringly different
Anybody who’s gone back and watched movies about the Manhattan Project or the decoding of all the sorts of things that took place, Enigma, for example, that’s straight rules-based computational power, and we’re obviously at the limits
That can only go so far
What your Siri does or Alexa, which is a step quite beyond that
Where you have computers that are probably going even 1 step further

Peter asks, “ What was it that was taking place during the period of your PhD? What you’re calling wave 2 of AI, what was the objective, and where was the failure? ”

The objective in the 1st era was, you wrote computer programs in assembler language (or in languages like Fortran ) and there was a limit of what you could do You had to be a real computational programmer to do something in that mode
In wave 2 in the 1970s, we came up with these rule-based systems where we set rules in what looked like English If there is a patient who has a fever and you get an isolate from the lab, and that bacteria in the isolate is gram positive, then you might have a streptococcal infection with a probability of so-and- so
You had to be a real computational programmer to do something in that mode
If there is a patient who has a fever and you get an isolate from the lab, and that bacteria in the isolate is gram positive, then you might have a streptococcal infection with a probability of so-and- so

The problem with these rule-based systems was several fold

With rule-based systems, you’re now programming in the level of human knowledge, not in computer code
A – You’re going to generate tens of thousands of these rules, and these rules would interact in ways that you could not anticipate
B – And we did not know enough and we could not pull out of human beings the right probabilities What is the right probability of: you have a fever and you don’t see anything on the blood test? What else is going on?
There’s a large set of possibilities, and getting all those rules out of human beings ended up being extremely expensive, and the results were not stable
For that reason, because we didn’t have much data online, we could not go to the next step, which is to have data to actually drive these models
What is the right probability of: you have a fever and you don’t see anything on the blood test? What else is going on?

What were the data sources then?

Books, textbooks, and journals as interpreted by human experts
That’s why some of these were called expert systems , because they were derived from introspection by experts who would then come up with the rules, with the probabilities
For example, there was a program called MYCIN run by Ted Shortliffe out of Stanford, who developed an antibiotic advisor that was a set of rules based on what he and his colleagues sussed out from the different infectious disease textbooks and infectious disease experts It stayed only up to date as long as they kept on looking at the literature, adding rules, fine-tuning it as an interaction between two rules that was not desirable, then you had to adjust that Very labor-intensive: if there’s a new thing, you’d have to add some new rules There’s no way you could keep it up to date It was very hard to keep up, and people didn’t The language it was programmed in was called e-MYCIN (these looked like English) There were no electronic medical records, so it was not informed by what was going on in the clinic
It stayed only up to date as long as they kept on looking at the literature, adding rules, fine-tuning it as an interaction between two rules that was not desirable, then you had to adjust that Very labor-intensive: if there’s a new thing, you’d have to add some new rules There’s no way you could keep it up to date
It was very hard to keep up, and people didn’t
The language it was programmed in was called e-MYCIN (these looked like English)
There were no electronic medical records, so it was not informed by what was going on in the clinic
Very labor-intensive: if there’s a new thing, you’d have to add some new rules
There’s no way you could keep it up to date

3 revolutions had to happen in order for us to have what we have today, and that’s why Zak thinks we’ve had such a quantum jump recently

How vast data sets, advanced neural networks, and powerful GPU technology have driven AI from its early limitations to achieving remarkable successes in medicine and other fields [19:00]

In generation 2, were there other industries that were having more success than medicine?

Were their applications in the military? Elsewhere in government?

Back in the 1970s, there were a whole bunch of computer companies around what we called 128 in Boston These were companies that were famous back then, like Wang computer , like Digital Equipment Corporation It’s a very sad story for Boston because that was before Silicon Valley got its pearl of computer companies around it
Digital Equipment Corporation, built a program called R1, and R1 was an expert in configuring the mini computers that you ordered If you wanted some capabilities, it would actually configure all the industrial components, the processors, the disk, and it would know about all the exceptions and what you needed to know What cabling, what memory, configuration, all that was done It basically replaced several individuals who had that very, very rare knowledge to configure their systems
It was also used in several government logistics effort
Although they were successful and used commercially, those efforts were limited
Because it turns out human beings, once you got to about 3, 4, 5, 6-thousand rules, no single human being could keep track of all the ways these rules couldn’t work
We used to call this the complexity barrier , that these rules would interact in unexpected ways, and you’d get incorrect answers, things that were not commonsensical because you had actually not captured everything about the real world
And so, it was very narrowly focused
If the expertise was a little bit outside the area of focus, if let’s say it was an infectious disease program and there was a little bit of influence from the cardiac status of the patient and you had not accurately modeled that, its performance would degrade rapidly
Similarly, if there was in Digital Equipment a new model that had a completely different part not included and that there were some dependencies that were not modeled, it would degrade in performance
These were companies that were famous back then, like Wang computer , like Digital Equipment Corporation
It’s a very sad story for Boston because that was before Silicon Valley got its pearl of computer companies around it
If you wanted some capabilities, it would actually configure all the industrial components, the processors, the disk, and it would know about all the exceptions and what you needed to know What cabling, what memory, configuration, all that was done
It basically replaced several individuals who had that very, very rare knowledge to configure their systems
What cabling, what memory, configuration, all that was done

These systems were very brittle, did not show common sense. They had expert behavior, but it was very narrowly done.

There were applications in medicine back then that survived till today

For example, already back then we’ve had the systems doing interpretation of ECGs pretty competently, at least a first pass until they would be reviewed by an expert cardiologist
There’s also a program that interpreted what’s called serum protein electrophoresis , where you look at a protein separated out by an electric gradient to make a diagnosis, let’s say, of myeloma or other protein disorders
Those also were deployed clinically, but they only worked very much in narrow areas
They were by no stretch of an imagination general purpose reasoning machines

Walk us through the 3 things that have taken the relative failures of the 1st and 2nd attempts at AI to get us where we are today

1 – Lots of data We needed to have a lot of online data to be able to develop models of interesting performance and quality
ImageNet was one of the first such data sets, collections of millions of images, with annotations (importantly: this has a cat in it, this has a dog in it, this is a blueberry muffin, this has a human in it) Annotations were absolutely essential to allow us to train the first very successful neural network models
We did not have a lot of textual information about medicine until PubMed went online All the medical literature, at least an abstract of it is online in PubMed A subset of that is open-access, which has the full text) ( PubMed Central ) All of the sudden, that has opened up over the last 10 years
After Obama signed the HITECH Act (electronic health records), this generated a lot of text for the use in these systems Which also ruined the lives of many doctors
2 – The neural network models themselves
The perceptron (mentioned earlier) was developed not too long after World War II, and was shown by one of the pioneers of AI ( Marvin Minsky ) to have fundamental limitations in that it could not do certain mathematical functions like what’s called an exclusive-or gate Because of that, people said these neural networks are not going to scale
But there were a few true believers who kept on pushing and making more and more advanced architectures and those multi-level, deep neural networks
So, instead of having one neural network, we layer on top of one neural network another one and another one and another one, so that the output of the first layer gets propagated up to the second layer of neurons to the third layer and fourth layer, and so on
We needed to have a lot of online data to be able to develop models of interesting performance and quality
Annotations were absolutely essential to allow us to train the first very successful neural network models
All the medical literature, at least an abstract of it is online in PubMed A subset of that is open-access, which has the full text) ( PubMed Central ) All of the sudden, that has opened up over the last 10 years
A subset of that is open-access, which has the full text) ( PubMed Central )
All of the sudden, that has opened up over the last 10 years
Which also ruined the lives of many doctors
Because of that, people said these neural networks are not going to scale

Peter asks, “ Was this a theoretical mathematical breakthrough or a technological breakthrough? ”

It was both, because having insight that we could actually come up with all the mathematical functions that we needed to, we could simulate them with these multi-level networks
Whereas, it was a theoretical insight, but we’d have never made anything out of it, if not for the fact of mostly teenage boys playing video games In order to have first-person shooters capable of running high-resolution pictures of aliens or monsters in high-resolution 24-bit color, 60 frames per second, we needed to have very parallel processors that would allow you to do the linear algebra that allowed you to calculate what was going to be the intensity of color on every dot of the screen at 60 frames per second You need huge matrices, and it turns out that’s something that can be run in parallel You want to have multiple parallel processors capable of rendering those images, again, at 60 frames per second, so basically millions of bits on your screen being rendered at 24 or 32-bit color In order to do that, you need to have that linear algebra that you just referred to being run in parallel
And so, these parallel processors called graphical processing units (GPUs) were developed They were developed by several companies, and some of them stayed in business They were absolutely essential to the success of video game
It then occurred to many smart mathematicians and computer scientists that the same linear algebra that was used to drive that computation for images could also be used to calculate the weights of the edges between the neurons in a neural network
So, the mathematics of updating the weights in response to stimuli, let’s say, of a neural network, updating of those weights can be done all in linear algebra
A typical computer has a central processing unit (that’s one processing unit)
A GPU has tens of thousands of processors that do this one very simple thing, linear algebra
And so, by having this parallelism that only supercomputers would have typically on your simple PC (because you needed to show the graphics at 60 frames per second), gave us all of a sudden these commodity chips that allowed us to calculate the performance of these multi-level neural networks
In order to have first-person shooters capable of running high-resolution pictures of aliens or monsters in high-resolution 24-bit color, 60 frames per second, we needed to have very parallel processors that would allow you to do the linear algebra that allowed you to calculate what was going to be the intensity of color on every dot of the screen at 60 frames per second
You need huge matrices, and it turns out that’s something that can be run in parallel You want to have multiple parallel processors capable of rendering those images, again, at 60 frames per second, so basically millions of bits on your screen being rendered at 24 or 32-bit color In order to do that, you need to have that linear algebra that you just referred to being run in parallel
You want to have multiple parallel processors capable of rendering those images, again, at 60 frames per second, so basically millions of bits on your screen being rendered at 24 or 32-bit color
In order to do that, you need to have that linear algebra that you just referred to being run in parallel
They were developed by several companies, and some of them stayed in business
They were absolutely essential to the success of video game

That theoretical breakthrough was a 2nd part but would not have happened without the actual implementation capability that we had with the GPUs

Nvidia was the most successful example, and it created an ecosystem of implementers who built their neural network, deep learning systems on top of the Nvidia architecture

An AI breakthrough in medicine: the ability to accurately recognize retinopathy [29:00]

Would you go back and look at the calendar and say this was the year or quarter when there was escape velocity achieved there?

Yeah
It was probably around 2012 when there was an ongoing contest every year saying who has the best image recognition software
These deep neural networks running off GPUs were able to outperform significantly all their other competitors in image recognition in 2012
That’s very clearly when everybody just woke up and said, “ Whoa, we knew about neural networks. We didn’t realize that these convolutional neural networks were going to be this effective. And seems that the only thing that’s going to stop us is computational speed and the size of our data sets. ”
That moved things very fast along in the imaging space with, very soon, consequences in medicine

AI recognizes retinopathy

It was only 6 years later that we saw journal articles about recognition of retinopathy (diseases affecting the retina, the back of your eye in diabetes), and a paper coming out, of all places, from Google saying, “ We can recognize different stages of retinopathy based on the images of the back of the eye .”
That also was a wake-up call because yes, part of the goal-post moving is great that we could recognize cats and dogs in web pages But now all of a sudden, this thing that we thought was a specialized human expertise could be done by that same stack of software Just if you gave it enough cases of these retinopathy’s, it would actually work well
Furthermore, what was wild was that there’s something called transfer learning , where you tune up these networks, get them to recognize cats and dogs, and in the process of recognizing cats and dogs, it learns how to recognize the little circles and lines and fuzziness and so on You did a lot better in training up the neural network first on the entire set of images and then on the retinas And if you just went straight to: I’m just going to train on the retinas That transfer alerting was impressive
Zak was asked to write an editorial for the Journal of the American Medical Association in 2018 when that Google article was written What was impressive to doctors was that what was the main role of doctors in that publication was just twofold 1 – To just label the images that were used for training: this retinopathy is not retinopathy 2 -To serve as judges of its performance The rest of it was computer scientists working with GPUs and images tuning it
But now all of a sudden, this thing that we thought was a specialized human expertise could be done by that same stack of software Just if you gave it enough cases of these retinopathy’s, it would actually work well
Just if you gave it enough cases of these retinopathy’s, it would actually work well
You did a lot better in training up the neural network first on the entire set of images and then on the retinas
And if you just went straight to: I’m just going to train on the retinas That transfer alerting was impressive
That transfer alerting was impressive
What was impressive to doctors was that what was the main role of doctors in that publication was just twofold
1 – To just label the images that were used for training: this retinopathy is not retinopathy
2 -To serve as judges of its performance
The rest of it was computer scientists working with GPUs and images tuning it

“ It didn’t look anything like medical school and you were having expert level recognition of retinopathy. That was a wake-up call .”‒ Zak Kohane

Third generation AI: how improvements in natural language processing significantly advanced AI capabilities [32:00]

Peter brings up the 2017 paper by Google, Attention is All You Need

This is a great paper that was about the invention of the transformer , which is a specific type of neural network architecture
Zak was talking about these were fairly vanilla convolutional, neural networks, the same one that can detect dogs and cats
Retinopathy was a big medical application
Except for computer scientists, no one noticed the Attention Is All You Need paper
And Google had this wonderful paper that said, “ If we recognize not just text that colocates together ”
So we’re going to get back away from images for a second
Because previously there was this notion that I can recognize a lot of similarities in text If I see which words occur together, I can implicate the meaning of a word by the company it keeps And so if I see this word and it has around it kingdom, crown, throne, it’s about a king and similarly for queen and so on
That kind of association in which we created what was called embedding vectors , which in plain English: it’s a string of numbers that says for any given word, what’s the probability? How often do these other words co- occur with it?
Just using those embeddings, those vectors, those lists of numbers that describe the co-occurrence of other words, we were able to do a lot of what’s called natural language processing, which is looking at text and saying, “ This is what it means. This is what’s going on .”
But then in the 2017 paper, they actually took a next step, which was the insight that where exactly the thing that we were focusing on was in a sentence, what was before and after the actual ordering of it mattered, not just the simple co-occurrence: that knowing what position that word was in the sentence actually made a difference
If I see which words occur together, I can implicate the meaning of a word by the company it keeps
And so if I see this word and it has around it kingdom, crown, throne, it’s about a king and similarly for queen and so on
How often do these other words co- occur with it?

That paper showed the performance went way up in terms of recognition, and that transformer architecture came from that paper, made it clear for number of researchers that if you scaled that transformer architecture up to a larger model so that the position dependence and this vector was learned across many, many more texts, the whole internet, you could train it to do various tasks

This transformer model, which is called the pre-trained model : think of it as equivalent of an equation with multiple variables
In the case of GPT-4 , we think it’s about a trillion variables It’s like an equation where you have a number in front of each variable, a coefficient that’s about a trillion long
It’s like an equation where you have a number in front of each variable, a coefficient that’s about a trillion long

This model can be used for various purposes

One is the chatbot purpose, which is given this sequence of words, what is the next word that’s going to be said?
Now, that’s not the only thing you could use this model for, but that turns out to have been the breakthrough application of the transformer model for text

Peter asks, “ Would you say that is the third thing that enabled this third wave of AI, the transformer? ”

That’s not what Zak was thinking about
The real breakthrough in data-driven AI was around the 2012 era If you talked to him in 2018, he would’ve already told you we’re in a new heyday and everybody would agree with you, there was a lot of excitement about AI just because of the image recognition capabilities
This was an additional capability that’s beyond what many of us were expecting just from the scale up of the neural network
If you talked to him in 2018, he would’ve already told you we’re in a new heyday and everybody would agree with you, there was a lot of excitement about AI just because of the image recognition capabilities

Zak’s summary : “ The 3, just to make sure I’m consistent, was: large data sets, multi-level neural networks (aka deep neural networks), and the GPU infrastructure. That brought us well through the 2012 to 2018. ”

The 2017 blip that became what we now know to be this whole large language model transformer architecture , that development was unanticipated for many of us That was already on the heels of an ascendant AI era There was already billions of dollars of frothy investment in frothy companies, some of which did well and many of which did not do so well
The transformer architecture has revolutionized many parts of the human condition, but it was already part of the third wave
That was already on the heels of an ascendant AI era
There was already billions of dollars of frothy investment in frothy companies, some of which did well and many of which did not do so well

AI concerns and regulation: misuse by individuals, military applications, displacement of jobs, and potential existential concerns [37:30]

By the time GPT-3.5 came out, it was now becoming a much a verb as Google was in the early 2000s

There were clearly people who knew what Google was in ‘96 and ‘97, but by 2000 everybody knew what Google was
Something about GPT-3.504 was the tipping point where you cannot know what it is at this point

Does this speak to the trajectory we’re on?

On the topic of regulation: is there a chance for this thing to be harmful to us in some way that we do not yet perceive?

There was no public discussion of this in the 80s, maybe because AI just wasn’t powerful enough to pose a threat
Peter assumes that one side says: pedal to the metal, let’s go forth on development and don’t regulate this The other side is: no, we need to have some brakes and barriers
Zak thinks this is not quite right
He agrees that chatbots have now become a commonly used noun That probably happened with the emergence of GPT 3.5, and that appeared around December of 2022
The out the box, that pre-trained model Zac mentioned earlier could tell you things like: how do I kill myself? How do I manufacture a toxin? It could allow you to do a lot of harmful things So there was that level of concern We can talk about what’s been done about those 1st order efforts
Then there’s been a group of scientists who interestingly went from saying, “ We’ll never actually get general intelligence from this particular architecture ,” to saying, “ Oh my gosh, this technology is able to inference in a way that I had not anticipated. And now I’m so worried that either because it’s malevolent or just because it’s trying to do something that has bad side effects for humanity, it presents an existential threat .”
On the other side, Zak doesn’t believe anybody is saying, “ Let’s just go heads down and let’s see how fast we can get to artificial general intelligence .” Or if they do think that they’re not saying it openly
pedal to the metal, let’s go forth on development and don’t regulate this
The other side is: no, we need to have some brakes and barriers
That probably happened with the emergence of GPT 3.5, and that appeared around December of 2022
It could allow you to do a lot of harmful things
So there was that level of concern
We can talk about what’s been done about those 1st order efforts
Or if they do think that they’re not saying it openly

Define AGI

Zak explains, “ That was an unfortunate slip because a rtificial g eneral i ntelligence means a lot of things to a lot of people… it’s a moving target, and it’s very much in the beholder .”
There’s a guy called Eliezer Yudkowsky , one of the so-called doomers, and he comes up with great scenarios of how a sufficiently intelligent system could figure out how to persuade human beings to do bad things Or control of our infrastructure to bring down our communications infrastructure or airplanes out of the sky And we can talk about whether that’s relevant or not
On the other side, we have OpenAI and Google
What was fascinating to Zak is that OpenAI (which working with Microsoft) generated GPT-4, and they were not saying publicly at all, “ Let’s not regulate it. ” In fact, they were saying, “ Please regulate me. ” Sam Altman went on a world tour where he said, “ We should be very concerned about this. We should regulate AI. ” Said this before Congress
Sam was kind enough to write foreword to the book Zak wrote with Peter Lee and Carey Goldberg on GPT-4 and the revolution in medicine, and Zak feels a bit churlish in wondering why they were insisting so much on regulation
Or control of our infrastructure to bring down our communications infrastructure or airplanes out of the sky
And we can talk about whether that’s relevant or not
In fact, they were saying, “ Please regulate me. ”
Sam Altman went on a world tour where he said, “ We should be very concerned about this. We should regulate AI. ” Said this before Congress
Said this before Congress

There’s 2 interpretations to why they were insisting on regulation

1 – A sincere wish that it’d be regulated So we check these machines, these programs, to make sure they don’t actually do anything harmful
2 – Something called regulatory lock-in which means: I’m a very well-funded company, and I’m going to create regulations with Congress about what is required, which boxes do you have to check in order to be allowed to run If you’re a small company, you’re not going to have a bevy of lawyers with big checks to comply with all the regulatory requirements
Zak doesn’t know Sam personally, but he imagines he’s a very well-motivated individual
So we check these machines, these programs, to make sure they don’t actually do anything harmful
If you’re a small company, you’re not going to have a bevy of lawyers with big checks to comply with all the regulatory requirements

There is someone just as potentially evil that we have to worry about, another intelligence, and that’s human beings

How do human beings use these great tools?
We know for a fact that one of the earliest users of GPT-4 were high schoolers trying to do their homework and solve hard puzzles given to them
We also know that various parties have been using the amazing text generation and interactive capabilities of these programs to spread misinformation to chatbots
And there’s a variety of maligned things that could be done by third parties using these engines

For Zak, the clear and present danger today is the question: how do individuals decide to use these general purpose programs?

If you look at what’s going on in the Ukraine-Russian War, Zak sees more and more autonomous vehicles flying and carrying weaponry and dropping bombs
And we see in our own military a lot more autonomous drones with greater and greater autonomous capabilities

Those are purpose-built to actually do dangerous things

And a lot of science fiction fans will refer to Skynet from the Terminator series, but we’re literally building it right now

Peter recalls in The Terminator , they refer to a moment, a year (1997 or something) when the Skynet became “self-aware”

And some who when it became self-aware, it decided to destroy humans

Is “self-aware” movie-speak for AGI? Or is it super intelligence? What do you think self-aware means in more technical terms?

Self-awareness means a process by which the intelligent entity can look back, look inwardly at its own processes and recognize itself
Now, that’s very hand-wavy, but Douglas Hofstadter has probably done the most thoughtful and clear writing about what self-awareness means If you really want to read a one full book that spends a whole book trying to explain it, it’s called I Am a Strange Loop In that book, he explains how if you have enough processing power and you can represent the processes that you have, essentially models of the processes that constitute you In other words, if you’re able to look at what you’re thinking, you may have some sense of self-awareness There’s a bit of an act of faith on that ‒ many AI researchers don’t buy that definition
There’s a difference between self-awareness and actual war intelligence
You can imagine a super-powerful computer that would predict everything that was going to happen around you and was not aware of itself as an entity
The fact remains, you do need to have a minimal level of intelligence, be able to be self-aware A fly may not be self-aware; it just goes and finds good smelling poop and does whatever it’s programmed to do on that But dogs have some self-awareness and awareness of their surroundings They don’t have perfect self-awareness; they don’t recognize themselves in the mirror, and they’ll bark at that Birds will recognize themselves in mirrors We recognize ourselves in many, many ways
If you really want to read a one full book that spends a whole book trying to explain it, it’s called I Am a Strange Loop
In that book, he explains how if you have enough processing power and you can represent the processes that you have, essentially models of the processes that constitute you In other words, if you’re able to look at what you’re thinking, you may have some sense of self-awareness There’s a bit of an act of faith on that ‒ many AI researchers don’t buy that definition
In other words, if you’re able to look at what you’re thinking, you may have some sense of self-awareness
There’s a bit of an act of faith on that ‒ many AI researchers don’t buy that definition
A fly may not be self-aware; it just goes and finds good smelling poop and does whatever it’s programmed to do on that
But dogs have some self-awareness and awareness of their surroundings They don’t have perfect self-awareness; they don’t recognize themselves in the mirror, and they’ll bark at that
Birds will recognize themselves in mirrors
We recognize ourselves in many, many ways
They don’t have perfect self-awareness; they don’t recognize themselves in the mirror, and they’ll bark at that

There is some correlation between intelligence and self-awareness, but these are not necessarily dependent functions

Peter’s takeaway : there are clear and present dangers associated with current best AI tools in that humans can use them for nefarious purposes

Is the most scalable example of that still relatively small in that it’s not an existential threat to our species at large?

Yes and no
If one were trying to do gain-of-function research with a virus, you could use these tools very effectively
There’s this disconnect: there’s those real existential threats, and then there’s this more fuzzy thing that we’re worried about (correctly): about bias, incorrect decisions, hallucinations
And there’s concerns about mistakes that might be made
There’s concerns about displacement of workers that just as automation displaced a whole other series of workers, now that we have something that works in the knowledge industry automatically, just as we’re placing a lot of copy editors and illustrators with AI Where’s that going to stop? It’s now much more in the white collar space
There is concern around the harm that could be generated there
In the medical domain, are we getting good advice? Are we getting bad advice? Whose interests are being optimized in these various decision procedures? That’s another level that doesn’t quite rise at all to the level of extinction events, but a lot of policymakers and the public seem to be concerned about it
Where’s that going to stop?
It’s now much more in the white collar space
Are we getting bad advice?
Whose interests are being optimized in these various decision procedures? That’s another level that doesn’t quite rise at all to the level of extinction events, but a lot of policymakers and the public seem to be concerned about it
That’s another level that doesn’t quite rise at all to the level of extinction events, but a lot of policymakers and the public seem to be concerned about it

How AI is enhancing image-based medical specialties like radiology [49:15]

The example Zak gave earlier is almost one we take for granted: you go and get an EKG at the doctor’s office and you get a pretty darn good readout This was true 30 years ago, just as it is today It’s going to tell you if you have an AV block , it’s going to tell you if you have a bundle branch block They read EKGs better than Peter does
This was true 30 years ago, just as it is today
It’s going to tell you if you have an AV block , it’s going to tell you if you have a bundle branch block
They read EKGs better than Peter does

What is the next area where we could see this? Radiology?

It seems to Peter that radiology, image pixel-based medicine would be the most logical next place to see AI do good work

“ In all the visual based medical specialties, it looks like AI can do as well as many experts. ”‒ Zak Kohane

What are the image appreciation subspecialties?

Pathology , when you’re looking at slices of tissue under the microscope
Radiology where you’re looking at X-rays or MRIs
Dermatology where you’re looking at pictures of the skin
In all those visual-based specialties, the computer programs are doing by themselves as well as many experts, but they’re not replacing the doctors because that image recognition process is only part of their job
Now, to be fair to your point in radiology, we already today before AI, many hospitals would send X-rays by satellite to Australia or India where they would be read overnight by a doctor or a specially trained person who had never seen the patient, and then the reports filed back to us because they’re 12 hours away from us, overnight we’d have the results of those reads And that same function can be done automatically by AI So that’s replacing a certain kind of doctor
And that same function can be done automatically by AI
So that’s replacing a certain kind of doctor

Peter asks, “ Will they know why this patient presented? Do they have a previous X-ray to compare it to?… Are we not at the point now where all of that information could be given to the AI to enhance the pre-test probability of whatever diagnosis it comes to? ”

Zak is delighted when Peter says pre-test probability ( Bayes’ theorem )
Zak explains, “ You just said a lot because what you just said actually went beyond what the straight convolutional neural networks would do ’cause they actually could not replace radiologists because they could not do a good job of taking into account the previous history of the patient .” It’s required the emergence of transformers where you can have multimodality You have both the image and the text Now, they’re going to do better than many, many radiologists today
It’s required the emergence of transformers where you can have multimodality You have both the image and the text Now, they’re going to do better than many, many radiologists today
You have both the image and the text
Now, they’re going to do better than many, many radiologists today

Zak doesn’t think there is any threat yet to radiologists as a job

One of the most irritating predictions (to doctors) was made by Geoffrey Hinton , one of the intellectuals leaders of neural network architecture He said (in approximately 2016), “ But in six years we wouldn’t have no need for radiologists. ” And that was just clearly wrong The reason it was wrong is A – They did not have the capabilities that we just talked about, understanding about the clinical context B – It’s also the fact that we don’t have enough radiologists to do the work
If you look at residency programs in American medicine, we’re not getting enough radiologists out We have an overabundance of applicants for interventional radiology They’re making a lot of money It’s high prestige But straight up radiology readers, there’s not enough of them
The same is true for primary care doctors Go around medical schools and ask, “ Who’s becoming a primary care doctor? ” Almost nobody Primary care is disappearing in the United States In fact, Mass General and Brigham announced officially they’re not seeing primary care patients
In Zak’s speciality ( pediatric endocrinology ), half of the slots nationally are not being filled Pediatric developmental disorders like autism, only half of those slots are filled PDID , there’s a huge gap emerging in the available expertise
It’s not what we thought it was going to be that we had a surplus of doctors that had to be replaced
It’s just we have a surplus and a few focused areas, which are very popular
And then for all the work of primary care and primary prevention, you have almost no doctors available
He said (in approximately 2016), “ But in six years we wouldn’t have no need for radiologists. ” And that was just clearly wrong The reason it was wrong is A – They did not have the capabilities that we just talked about, understanding about the clinical context B – It’s also the fact that we don’t have enough radiologists to do the work
The reason it was wrong is A – They did not have the capabilities that we just talked about, understanding about the clinical context
B – It’s also the fact that we don’t have enough radiologists to do the work
We have an overabundance of applicants for interventional radiology They’re making a lot of money It’s high prestige
But straight up radiology readers, there’s not enough of them
They’re making a lot of money
It’s high prestige
Go around medical schools and ask, “ Who’s becoming a primary care doctor? ” Almost nobody
Primary care is disappearing in the United States In fact, Mass General and Brigham announced officially they’re not seeing primary care patients
In fact, Mass General and Brigham announced officially they’re not seeing primary care patients
Pediatric developmental disorders like autism, only half of those slots are filled
PDID , there’s a huge gap emerging in the available expertise

Isn’t radiology a gap that should be fillable by AI?

Peter asks, “ If you’re saying we have a dearth of imaging radiologists who are able to work the emergency rooms, urgent care clinics and hospitals, wouldn’t that be the first place we would want to apply our best of imaging recognition with our super powerful GPUs and now plug them into our transformers with our language models so that I can get clinical history, medical past history, previous images, current images, and they don’t have to send it to a radiologist in Australia to read it, who then has to send it back to a radiologist here to check if we’re just trying to fill a gap, that gap should be fillable, shouldn’t it? ”

That’s exactly where it is being filled

The use of AI by patients and doctors [55:45]

What keep distracting Zak in this conversation is that there’s a whole other group of users of these AIs that we’re not talking about: the patients
Previously, none of these tools were available to patients
With the release of GPT-3.5 and 4 , and now Gemini and Claude 3 , they’re being used by patients all the time in ways that we had not anticipated

An example of how a patient used GPT-4

There was a child who was having trouble walking, having trouble chewing, and then started having intractable headaches Mom brought him to multiple doctors They did multiple imaging studies, no diagnosis; kept on being in intractable pain She just typed into GPT-4 all the reports, and asked GPT-4, what’s the diagnosis? And GPT-4 said tethered cord syndrome She then went with all the imaging studies to a neurosurgeon, said, “ What is this? ” He looked at it, he said, “ Tethered cord syndrome. ”
Mom brought him to multiple doctors
They did multiple imaging studies, no diagnosis; kept on being in intractable pain
She just typed into GPT-4 all the reports, and asked GPT-4, what’s the diagnosis?
And GPT-4 said tethered cord syndrome
She then went with all the imaging studies to a neurosurgeon, said, “ What is this? ”
He looked at it, he said, “ Tethered cord syndrome. ”

We have an epidemic of misdiagnosis and undiagnosed patients

Part of Zak’s background: he’s the principal investigator of the coordinating center of something called the Undiagnosed Network It’s a network with 12 academic hospitals down the west coast from University of Washington, Stanford UCLA to Baylor up the East Coast, Harvard Hospitals, NIH. And we see a few thousand patients every year These are patients who have been undiagnosed and they’re in pain That’s just a small fraction of those who are undiagnosed
It’s a network with 12 academic hospitals down the west coast from University of Washington, Stanford UCLA to Baylor up the East Coast, Harvard Hospitals, NIH. And we see a few thousand patients every year These are patients who have been undiagnosed and they’re in pain That’s just a small fraction of those who are undiagnosed
These are patients who have been undiagnosed and they’re in pain
That’s just a small fraction of those who are undiagnosed

“ We bring to bear a whole bunch of computational techniques and genomic sequencing to actually be able to help these individuals, but it’s very clear that there’s a much larger burden out there of misdiagnosed individuals .”‒ Zak Kohane

Peter asks, “ Does it surprise you that in that example, the mother was the one that went to GPT-4 and inputted that. I mean, she had presumably been to many physicians along the way. Were you surprised that one of the physicians along the way hadn’t been the one to say, “Gee, I don’t know, but let’s see what this GPT-4 thing can do? ”

Most clinicians Zak knows do not have what he used to call the “Google reflex” Zak remembers when he was on the wards and they had a child with dysmorphology (they look different), and he asked one of the fellows, “ What’s the diagnosis? ” He didn’t know Zak asked, “ How would you find out? ” They had no idea Zak said, “ He has this and this and this finding… Let’s take what I just said and type it into Google .” One of the top 3 responses was the diagnosis
Clinicians use Google in their civilian life, but not in the clinic
Doctors being driven very, very hard and they’re being told to use certain technological tools They’re being turned into data entry clerks Who has the time to look up a journal article?
Zak remembers when he was on the wards and they had a child with dysmorphology (they look different), and he asked one of the fellows, “ What’s the diagnosis? ”
He didn’t know
Zak asked, “ How would you find out? ”
They had no idea
Zak said, “ He has this and this and this finding… Let’s take what I just said and type it into Google .”
One of the top 3 responses was the diagnosis
They’re being turned into data entry clerks
Who has the time to look up a journal article?

They don’t do the Google reflex even less do they have the: let’s look at the patient’s history and see what GPT-4 would come up with

Early on, Zak was gratified to see doctors saying, “ Wow, I just took the patient history, plugged into GPT-4 and said, write me a letter of prior authorization .” Because it was saving them 5 minutes to write that letter
He was not pleased because if you use ChatGPT , you’re using a program that is covered by OpenAI as opposed to a version of GPT-4 that is being run on protected Azure cloud by Microsoft, which is HIPAA covered HIPAA is the legal framework under which we protect patient privacy and if you violate it, you can be fined and even go to prison
Because it was saving them 5 minutes to write that letter
HIPAA is the legal framework under which we protect patient privacy and if you violate it, you can be fined and even go to prison

In other words, if a physician wants to put any information into GPT-4, they better not identify it

If they just plugged in a patient note into ChatGPT, that’s a HIPAA violation
If there’s a Microsoft version of it, which is HIPAA-compliant, it’s not

The doctors were using it for improving the administrative part of healthcare, but by and large, only a few doctors use it for diagnostic acumen

What about more involved radiology?

Once we start to look at three-dimensional images such as cross-sectional images, CT scans, MRIs, or even more complicated images like ultrasound and things of that nature, what is the current state of the art with respect to AI in the assistance of reading these types of images?

That’s the very exciting news, 1 of the 3 ingredients for a breakthrough (remember how Zak said it was important to have lots of data?) Having a lot of data around, for example, echocardiogram , the ultrasounds of your heart ‒ normally it takes a lot of training to interpret those images correctly There is a recent study from the EchoCLIP group, and they took a million echocardiograms and a million textual reports and essentially trained the model both to create those embeddings (Zak talked about of the images and of the text) This is very complicated video: you’re putting a multidimensional video because you have timescale, you have Doppler effects It views it from different angles It’s dependent on the radiology tech (they can be good or bad)
Having a lot of data around, for example, echocardiogram , the ultrasounds of your heart ‒ normally it takes a lot of training to interpret those images correctly
There is a recent study from the EchoCLIP group, and they took a million echocardiograms and a million textual reports and essentially trained the model both to create those embeddings (Zak talked about of the images and of the text) This is very complicated video: you’re putting a multidimensional video because you have timescale, you have Doppler effects It views it from different angles It’s dependent on the radiology tech (they can be good or bad)
This is very complicated video: you’re putting a multidimensional video because you have timescale, you have Doppler effects
It views it from different angles
It’s dependent on the radiology tech (they can be good or bad)

We’re going to get rid of the cardiologist before we get rid of the technician

The potential for AI to augment clinicians and address physician shortages [1:02:45]

The potential for AI to replace clinicians

Zak’s target in this conversation is nurse practitioners and physician assistants with these [AI] tools can replace a lot of expert clinicians and there is a big open question: what is the real job for doctors in 10 years from now? Zak doesn’t think we know the answer to that
Peter points out, “ We still haven’t come to proceduralists, so we still have to talk about the interventional radiologist, the interventional cardiologist, and the surgeon. ”
What we’re doing is identifying the pecking order of physicians, and let’s not even think about it through the lens of replacement
Let’s start with the lens of augmentation , which is the radiologist can be the most easily augmented, the pathologist, the dermatologist, the cardiologist who’s looking at Echoes and EKGs and stress tests, people who are interpreting visual data and using visual data will be the most easily augmented
Zak doesn’t think we know the answer to that

The second tranche of that will be people who are interpreting language data plus visual data

Now we’re talking about your internist, your pediatrician, where you have to interpret symptoms and combine them with laboratory values and combine it with a story and an image

In a place where we don’t have primary care physicians

The American Association of Medical Colleges estimates that by 2035 (that’s only 11 years from now), we’ll be missing on the order 50,000 primary care doctors
As Zak mentioned, he can’t get primary care at the Brigham or at MGH today
And in the absence of that, you have to ask yourself: how can we replace these absent primary care practitioners with nurse practitioners, with physician assistants augmented by these AIs because there’s literally no doctor to replace

Peter asks, “ Where are we technologically on that augmentation? ”

If NVIDIA never came out with another chip and this is as good as it’s going to get, do we have good enough GPUs, good enough multilayer neural networks that all you need is more data and training sets that we could now do the augmentation that has been described by us in the last five minutes?
The short answer is yes

Zak gives a concrete example

Most concierge services cost in Boston, somewhere between $5-$20,000 a year
You can get this very low cost of concierge service called One Medical One Medical was acquired by Amazon and they have a lot of nurse practitioners in there and you can make an appointment and you can text with them Zak believes that those individuals could be helped in ordering the right imaging studies, the right EKGs, the right medications, and assess your continuing heart failure and only decide in the very few cases that you need to see a specialist cardiologist or a specialist endocrinologist today
It would just be a matter of just making the current models bette
One Medical was acquired by Amazon and they have a lot of nurse practitioners in there and you can make an appointment and you can text with them
Zak believes that those individuals could be helped in ordering the right imaging studies, the right EKGs, the right medications, and assess your continuing heart failure and only decide in the very few cases that you need to see a specialist cardiologist or a specialist endocrinologist today

A big question for us (this is a regulatory question): which ones do a better job? (and they’re not all equal)

Zak doesn’t think we need technological breakthroughs to just make the current set of paraprofessionals work at a level of entry-level doctors

He adds the old very bad joke: What do you call the medical student who graduates at the bottom of his class? Doctor
And so if you could just merely get the bottom 50% of doctors to be as good as a top 50%, that’ll be transformative for healthcare

There are other superhuman capabilities that we can go towards that require the next generation of algorithms, NVIDIA architectures and datasets

“ If we stop now, we could already transform medicine ”‒ Zak Kohane

It’s just a matter of the sweat equity to create the models, figure out how to include them in the workflow, how to pay for them, how to create a reimbursement system and a business model that works for our society
There’s no technological barrier

Peter’s takeaway : everything we’ve talked about is take the best case example of medicine today and you can augment it with AI such that you can raise everyone’s level of care to that of the best, no gaps and it’s scaled out

The potential for AI to revolutionize early diagnosis and prediction of diseases: Alzheimer’s disease, CVD, autism, and more [1:08:00]

Where do you see the potential for AI in solving problems that we can’t even solve on the best day at the best hospitals with the best doctors?

For example, we can’t really diagnose Alzheimer’s disease until it appears to be at a point that for all intents and purposes is irreversible Maybe on a good day we can halt progression really, really early in a patient with just a whiff of MCI, mild cognitive impairment, maybe with an early amyloid detection and an anti-amyloid drug
Peter asks, “ Is it science fiction to imagine that there will be a day when an AI could listen to a person’s voice, watch the movements of their eyes, study the movements of their gait and predict 20 years in advance when a person is tearing down the barrel of a neurodegenerative disease and act at a time when maybe we could actually reverse it? ”
Zak doesn’t believe it’s science fiction at all Looking at images of retinas today can tell you not just whether you have retinal disease, but if you have hypertension, if you’re a male, if you’re female, how old you are, and some estimate of your longevity This uses straightforward convolutional neural network, not even ones that involve transformers And that’s just looking at the back of your eye and seeing enough data
Zak was a small player in a study that appeared in Nature in 2005 with Bruce Yankner They were looking at frontal lobes of individuals who had died from a variety of reasons, often in accidents of various ages They saw bad news: after age 40, your transcriptome (the genes that are switched on) falls off a cliff 30% of your transcriptome went down That seemed to be a big difference in the expression of genes around age 40 But there was one 90-year-old who looked like the young guy So maybe there’s hope for some of us
Zak thought about it afterwards and there were other things that actually have much smoother functions, like our skin
All our organs age and they age at different rates
Having the right data sets and the ability to see nuances that we don’t notice makes it very clear to Zak that the early detection part can be very straightforward (no problem)
The treatment part, we can talk about it
For example we had early on from the very famous Framingham Heart Study , a predictor of when you were going to have heart disease based on just a handful of variables
Maybe on a good day we can halt progression really, really early in a patient with just a whiff of MCI, mild cognitive impairment, maybe with an early amyloid detection and an anti-amyloid drug
Looking at images of retinas today can tell you not just whether you have retinal disease, but if you have hypertension, if you’re a male, if you’re female, how old you are, and some estimate of your longevity This uses straightforward convolutional neural network, not even ones that involve transformers And that’s just looking at the back of your eye and seeing enough data
This uses straightforward convolutional neural network, not even ones that involve transformers
And that’s just looking at the back of your eye and seeing enough data
They were looking at frontal lobes of individuals who had died from a variety of reasons, often in accidents of various ages
They saw bad news: after age 40, your transcriptome (the genes that are switched on) falls off a cliff 30% of your transcriptome went down That seemed to be a big difference in the expression of genes around age 40
But there was one 90-year-old who looked like the young guy So maybe there’s hope for some of us
30% of your transcriptome went down
That seemed to be a big difference in the expression of genes around age 40
So maybe there’s hope for some of us

Now we have these artificial intelligence models that based on hundreds of variables, can predict various other diseases and Zak believes it will do Alzheimer’s very soon

He thinks you’ll be able to see a combination of gait, speech patterns, picture of your body, picture of skin, and eye movements ‒ that will be a very accurate predictor
Speaking about eyes, Zak just published a very nice study where in a car just by looking at the driver, they can figure out what your blood sugar is because diabetics previously have not been able to get driver licenses sometimes because of the worry about them passing out because of hypoglycemia So there was a very nice study that showed that you could just by looking (have cameras pointed at the eyes) could actually figure out exactly what the blood sugar is That kind of detection is fairly straightforward It’s a different question about what you can do about it
So there was a very nice study that showed that you could just by looking (have cameras pointed at the eyes) could actually figure out exactly what the blood sugar is
That kind of detection is fairly straightforward
It’s a different question about what you can do about it

Could a retinal exam be useful in predicting adverse cardiac events?

The 2 most popular models for looking at major adverse cardiac event risk prediction are the Framingham model and the Multi-Ethnic Study on Atherosclerosis (MESA model)
But you needed something else to build those models, which was enough time to see the outcome
In the Framingham cohort, which was the late 70s and early 80s, you then had the Framingham offspring cohort and then you had to be able to follow these people with their LDL-C and HDL-C and triglycerides and later eventually they incorporated calcium scores
If today we said, look, we want to be able to predict 30-year mortality, which is something no model can do today
This is a big pet peeve of Peter’s: we generally talk about cardiovascular disease through the lens of 10-year risk (which he thinks is ridiculous) We should talk about lifetime risk He would settle for 30-year risk
If we had a 30-year model where we could take many more inputs, Peter would love to be looking at the retina
We should talk about lifetime risk
He would settle for 30-year risk

Peter believes a retinal examination should be a part of medicine today for everybody

Peter would take a retinal exam over a hemoglobin A1c all day every day
The point is, we could define the data set, and we want to see these 50 things in everybody to predict every disease Let’s overdo it and we can prune things later
Let’s overdo it and we can prune things later

Peter asks, “ Is there any way to get around the fact that we’re going to need 30 years to see this come to fruition in terms of watching how the story plays out? Or are we basically going to say, no, we’re going to do this over five years. It won’t be that useful because a five-year predictor basically means you’re already catching people in the throes of the disease. ”

“ I’ll say three words, electronic health records. So that turns out not to be the answer in the United States. ”‒ Zak Kohane

In the US we move around and don’t stay in any given healthcare system that long
Very rarely will Zak have all the measurements on a patient (glycohemoglobin, all your blood pressures, all your clinic visits, all the imaging studies that you’ve had)
That’s not the case in Israel for example In Israel, they have these HMOs (health maintenance organizations) and one of them ( Clalit ), Zak has a good relationship with because they published all the big Covid studies looking at the efficacy of the vaccine and why could they do that? Because they had the whole population available and they have about 20, 25 years worth of data on all their patients in detail and family relationships
Kaiser Permanente also has that kind of data, and with that data Zak thinks you can actually come close
Now these are noisier measurements, and so those of us who are data junkies (like Zak) always keep mumbling to ourselves, “perfect is the enemy of good” Waiting 30 years to have the perfect data set is not the right answer to help patients now
In Israel, they have these HMOs (health maintenance organizations) and one of them ( Clalit ), Zak has a good relationship with because they published all the big Covid studies looking at the efficacy of the vaccine and why could they do that?
Because they had the whole population available and they have about 20, 25 years worth of data on all their patients in detail and family relationships
Waiting 30 years to have the perfect data set is not the right answer to help patients now

There are things that we could know now that are knowable today that we just don’t know because we haven’t bothered to look

A quick example

Zak did a study of autism using electronic health records maybe 15 years ago, and he saw there was a lot of GI problems
When he talked to a pediatric expert, they were a little bit dismissive They said, “ Brain bad, tummy hurt. ” They’ve seen a lot of inflammatory bowel disease, and it doesn’t make sense to that this is somehow an effect of brain function
To make a long story short, Zak did a massive study looking forward to tens of thousands of individuals, and sure enough, he found subgroups of patients who had immunological problems associated with their autism They had type 1 diabetes, inflammatory bowel disease, lots of infections
They said, “ Brain bad, tummy hurt. ”
They’ve seen a lot of inflammatory bowel disease, and it doesn’t make sense to that this is somehow an effect of brain function
They had type 1 diabetes, inflammatory bowel disease, lots of infections

Those were knowable, but they were not known

Zak had parents coming to him more thankful that for anything else he had ever done for them clinically because he was telling these parents they weren’t hallucinating, that these kids had these problems They just weren’t being recognized by medicine because no one had the big wide angle to see these trends
So without knowing the field of Alzheimer’s the way Zak does other fields, he bets you there are trends in Alzheimer’s that you can pick up today by looking at enough patients that you’ll find some that have more frontotemporal components, some that have more effective components, some that have more of an infectious and immunological component Those are knowable today
They just weren’t being recognized by medicine because no one had the big wide angle to see these trends
Those are knowable today

The future of AI in healthcare: integration of patient data, improved diagnostics, and the challenges of data accessibility and regulatory compliance [1:17:00]

If the physician is the customer (who is not necessarily the most tech forward customer), and truthfully, like many customers of AI runs the risk of being marginalized by the technology if the technology gets good enough
Yet you need the customer to access the patient to make the data system better, to make the training set better

How do you see the interplay over the next decade of that dynamic?

Zak agrees, that’s the right question
In order for these AI models to work, you need a lot of data, a lot of patients

Where is that data going to come from?

There are some healthcare systems (like the Mayo Institute) who think they can get enough data in that fashion
There are some data companies that are trying to get relationships with healthcare systems where they can get de-identified data

Zak is betting on something else

There is a trend where consumers are going to have increased access to their own data
The 21st Century Cures Act was passed by Congress and it said that patients should be given access to their own data programmatically
Now, they’re not expecting your grandmother to write a program to access the data programmatically, but by having a right to it, it enables others to do so
For example, Apple has something called Apple Health It has this big heart icon on it If you’re one of the 800 hospitals that they’ve already hooked up with Mass General or Brigham Women’s and you’re a patient there, if you authenticate yourself to it, if you give it your username and password, it’ll download into your iPhone your labs, your meds, your diagnoses, your procedures, as well as all the wearable stuff, your blood pressure that you get as an outpatient and various other forms of data That’s already happening now
There’s not a lot of companies that are taking advantage of that, but right now that data is available on tens of millions of Americans
It has this big heart icon on it
If you’re one of the 800 hospitals that they’ve already hooked up with Mass General or Brigham Women’s and you’re a patient there, if you authenticate yourself to it, if you give it your username and password, it’ll download into your iPhone your labs, your meds, your diagnoses, your procedures, as well as all the wearable stuff, your blood pressure that you get as an outpatient and various other forms of data That’s already happening now
That’s already happening now

Isn’t it interesting how unfriendly that data is in its current form?

An example from Peter’s practice

If we send a patient to Labcorp or Boston Heart (or pick your favorite lab), and we want to generate our own internal reports based on those where we want to do some analysis on that, lay out trend sheets, we have to use our own internal software
It’s almost impossible to scrape those data out of the labs because they’re sending you PDF reports
Their APIs are garbage
Nothing about this is user-friendly

Even if you have the My Health thing come on your phone, it’s not navigable, it’s not searchable, it doesn’t show you trends over time

Peter asks, “ Is there a more user hostile industry from a data perspective than the health industry right now? ”

No
There’s a good reason why: because they’re keeping you captive
The good news is you’re speaking to a real nerd
Zak know 2 ways where we could solve this problem
1 – If it’s in the Apple Health thing, someone can actually write a program, an app on the iPhone, which will take those data as numbers (and not have to scrape it) and it can run it through your own trending programs You could actually use it directly
2 – With Gemini and GPT-4, you can actually give it those PDFs and with the right prompting, it will actually take those data and turn them into tabular spreadsheets
You could actually use it directly

Peter asks, “ We can’t do that because of HIPAA, correct? ”

If the patient gets it from the patient portal, and the patient gives it to you, you absolutely can do that It doesn’t matter that it’s not de-identified It’s not doable through ChatGPT because your lawyers would say, Peter, you’re going to get a million dollars in fines from HIPAA But if you do GPT on the Azure cloud , that’s HIPAA protected and you absolutely can use it with patient consent (100% you can do it) Zak is not a shill for Microsoft (he doesn’t own any stock)
GPT is being used with patient data out of Stanford right now
Epic’s using GPT-4 , and it’s absolutely legitimately usable by you People don’t understand that
Peter’s practice has now totally bypassed OCRs (optical character recognition) , which is what they were using 15 years ago to scrape together this data
It doesn’t matter that it’s not de-identified
It’s not doable through ChatGPT because your lawyers would say, Peter, you’re going to get a million dollars in fines from HIPAA
But if you do GPT on the Azure cloud , that’s HIPAA protected and you absolutely can use it with patient consent (100% you can do it) Zak is not a shill for Microsoft (he doesn’t own any stock)
Zak is not a shill for Microsoft (he doesn’t own any stock)
People don’t understand that

An example where Zak used GPT-4 for diagnosis

3 months ago, the NEJM published a picture of the week [shown below]: the back of a 72-year-old that looks like someone scratched themselves

Figure 1. NEJM Image challenge from October 5, 2023 . Image credit: NEJM

Zak put this image into GPT-4 and the text (minus 1 fact) and it came up with 2 things it could be: bleomycin toxicity or shiitake mushroom toxicity What Zak removed is that they guy had eaten mushrooms the day before
Peter doesn’t think most doctors know this
Further, when Peter tried to send a picture of his kid’s rash to the pediatrician and they don’t know what it is It’s it’s like we’re rubbing 2 sticks together and Zak telling him about the Zippo lighter
What Zak removed is that they guy had eaten mushrooms the day before
It’s it’s like we’re rubbing 2 sticks together and Zak telling him about the Zippo lighter

This is Zak’s message to patients without a primary care doctor: AI is better than no doctor, maybe better

Zak pulls down the adverse events reporting files from the FDA and uses GTP-4 to analyze it

He doesn’t know squat about the FDA
It’s a big zip file (compressed file), and he went and said to GPT-4: please analyze this data It says: unzipping, based on this table, I think this is about the adverse events and this is the locations, what do you want to know? Zak says, “ Tell me what adverse events for disease modifying drugs for arthritis. ” It says: to do that, I’ll have to join these 2 tables And it just does it; it creates its own Python code, and it gives me a report
It says: unzipping, based on this table, I think this is about the adverse events and this is the locations, what do you want to know?
Zak says, “ Tell me what adverse events for disease modifying drugs for arthritis. ”
It says: to do that, I’ll have to join these 2 tables And it just does it; it creates its own Python code, and it gives me a report
And it just does it; it creates its own Python code, and it gives me a report

Peter asks, “ Is this a part of medical education now? You’re at Harvard, right? (…one of the 3 best medical schools in the US) ”

Peter wonders if they spend as much time on this as they do histology, where he spent 1000 hours looking at slides under a microscope that he never once tried to understand He doesn’t want to say there wasn’t a value in doing that, but he wants to understand the relative balance of education
Zak explains that it’s like the stethoscope Arguably, we should be using things other than the stethoscope “ Let me make sure I don’t get fired, or at least beaten severely by telling you that George Daley , our dean of the medical school, has said explicitly he wants to change all of medical education so these learnings are infused throughout the 4 years, but it’s going to take some doin g.”
He doesn’t want to say there wasn’t a value in doing that, but he wants to understand the relative balance of education
Arguably, we should be using things other than the stethoscope
“ Let me make sure I don’t get fired, or at least beaten severely by telling you that George Daley , our dean of the medical school, has said explicitly he wants to change all of medical education so these learnings are infused throughout the 4 years, but it’s going to take some doin g.”

The future of autonomous robotic surgery [1:25:00]

We’ve gone from purely the recognition/ image-based to how do I combine image with voice, story, text?
Zak has made a very compelling case that we don’t need any more technological breakthroughs to augment those It’s purely a data set problem at this point and a willingness
Let’s now move to the procedural
It’s purely a data set problem at this point and a willingness

Today, the surgeon needs to move the robot. Are we getting to the point where that could change?

For example, consider the radical prostatectomy

Currently, this is never done open
This is a procedure that the Da Vinci (a robot) has revolutionized There’s no blood loss anymore
When Peter was a resident, this was one of the bloodiest operations they did It was the only operation by the way, for which they had the patients donate their own blood 2 months ahead of time (that’s how guaranteed it was that they were going to need blood transfusions)
Today, it’s insane how successful this operation is, on a large part of the robot, but the surgeon needs to move the robot
Zak explains, “ That’s already happening. Based on what we’re seeing with robotics in the general world, I think the Da Vinci controlled by a robot, 10 years is a very safe bet. ”
There’s no blood loss anymore
It was the only operation by the way, for which they had the patients donate their own blood 2 months ahead of time (that’s how guaranteed it was that they were going to need blood transfusions)

Surgery is a data problem because you can turn it into a pixel and movement problem

Comparing autonomous vehicles to robotic surgery

Remember, there’s a lot of degrees of freedom in moving a car around traffic
Medicine is not the only job where lives are at stake
Driving a ton of metal at 60 miles per hour in traffic is also putting lives at stake
Last time Zak looked, there’s several manufacturers who are saying that some appreciable fraction of that effort, they’re controlling multiple degrees of freedom with a robot
Peter also spoke with somebody recently (one of the companies that’s deep in the space of autonomous vehicles), and they very boldly made a compelling case for autonomous vehicles, saying, “ If every vehicle on the road was at their level of technology and autonomous driving, you wouldn’t have fatalities anymore. ” But the key was that every vehicle had to be at that level
But the key was that every vehicle had to be at that level

Peter asks, “ Does that sense check to you? ”

Zak begins by explaining that he is a terrible driver, but he’s a better driver now in a Tesla
There is a good message for medicine in this
When Zak is driving his Tesla , he knows he needs to jiggle the steering when because otherwise it will assume he’s zoning out
What he didn’t realize was is that he will pick up his phone and look at it (very bad), and the car was looking at him telling him to put down the phone 3 minutes later, he picks up his phone again and switches off the autopilot When he gets home, it says, “ All right, that was bad. You do that 4 more times, I’m switching off autopilot until the next software update. ” Zak mentions this because it takes a certain amount of confidence to do that to your customer base
3 minutes later, he picks up his phone again and switches off the autopilot
When he gets home, it says, “ All right, that was bad. You do that 4 more times, I’m switching off autopilot until the next software update. ” Zak mentions this because it takes a certain amount of confidence to do that to your customer base
Zak mentions this because it takes a certain amount of confidence to do that to your customer base

Zak asks, “ In medicine, how likely is it that we’re going to fall asleep at the wheel if we have an AI thinking for us? ”

It’s a real issue
For example, we know for a fact back in the ‘90s that doses for a drug like Ondansetron where people would talk endlessly about how frequently should be given it, with what dose The moment you put it in the order entry system, 95% of doctors would just use the default there
The moment you put it in the order entry system, 95% of doctors would just use the default there

How in medicine are we going to keep doctors awake at the wheel and will we dare to do the kind of challenges that Zak just described the car doing?

Zak believes because of what he’s seen with autonomy and robots that controlling a dementia robot will probably have less bad outcomes Every once in a while, someone nicks something and you have to go into full surgery or they go home and they die on the way home ’cause they exsanguinate
Every once in a while, someone nicks something and you have to go into full surgery or they go home and they die on the way home ’cause they exsanguinate

Zak thinks robotic surgery is going to be safer

It’s unbelievable for Peter to wrap his head around that, but he reserves the right to be startled
Peter thinks certain things seem much easier than others
He has an easier time believing we’re going to be able to replace interventional cardiologists where the number of degrees of freedom, the complexity and the relationship between what the image shows, what the cath shows and what the input is, the stent, that gap is much narrower
But when you talk about doing a Whipple procedure , when you talk about what it means to cell by cell, take a tumor off the superior mesenteric vessels, he’s thinking, “ Oh my God. ”

Zak predicts that in 10 years there will be prostatectomy by robot

AI and the future of mental health care [1:31:30]

This is a field of medicine today that is grossly underserved
Everything Zak has said resonates with Peter from his own experiences in pediatrics and primary care
Zak points out that at Harvard, 60% of undergraduates are getting some sort of mental health support and it’s completely outdoing all the resources available to the university health services They have to outsource some of their mental health, and this is a very richly endowed university (yet they don’t have the resources)
We live in a world where when a person is depressed, when a person is anxious, when a person has any sort of mental or emotional illness, pharmacotherapy plays a role, but it can’t displace psychotherapy You have to be able to put these 2 things together and the data would suggest that the knowledge of your psychotherapist is important, but it’s less important than the rapport you can generate with that individual
They have to outsource some of their mental health, and this is a very richly endowed university (yet they don’t have the resources)
You have to be able to put these 2 things together and the data would suggest that the knowledge of your psychotherapist is important, but it’s less important than the rapport you can generate with that individual

Peter asks, “ Based on that, do you believe that the most sacred, protected (if you want to use that term) profession within all of medicine, will then be psychiatry? ”

Zak would like to think that
If he had a psychiatric GPT speaking to him, he wouldn’t think that it understood him
On the other hand, back in the 1960s or ‘70s, there was a program called ELIZA , and it was a simple pattern matching program It would just emulate what’s called a Rogerian therapist “ Where I really hate my mother .” “ Why do you say you hate your mother? ” “ Oh, it’s because I don’t like the way she fed me. ” “ What is it about the way she fed you? ” This very, very simple pattern matching and this ELIZA program, which was developed by Joe Weitzbaum at MIT His own secretary would lock herself in her office to have sessions with this thing because it’s not judgmental
It turns out that there’s a large group of patients who actually would rather have a non-human, non-judgmental person who remembers what they’ve said from last time and shows empathy verbally
Zak wrote this book with Peter Lee and Peter made a big deal in the book about how GPT-4 was showing empathy In the book, Zak argued with him that this is not that big a deal, and said, “I remember from medical school being told that some of the most popular doctors are popular, because they’re very deep empaths, not necessarily the best doctors. ”
For Zak, cognitive behavioral therapy with AI would not be acceptable to him
But if it’s giving you insight into yourself and it’s based on the wisdom culled from millions of patients, who’s to say that it’s worse? And it’s certainly not judgmental and maybe it’ll bill less
It would just emulate what’s called a Rogerian therapist “ Where I really hate my mother .” “ Why do you say you hate your mother? ” “ Oh, it’s because I don’t like the way she fed me. ” “ What is it about the way she fed you? ”
This very, very simple pattern matching and this ELIZA program, which was developed by Joe Weitzbaum at MIT His own secretary would lock herself in her office to have sessions with this thing because it’s not judgmental
“ Where I really hate my mother .”
“ Why do you say you hate your mother? ”
“ Oh, it’s because I don’t like the way she fed me. ”
“ What is it about the way she fed you? ”
His own secretary would lock herself in her office to have sessions with this thing because it’s not judgmental
In the book, Zak argued with him that this is not that big a deal, and said, “I remember from medical school being told that some of the most popular doctors are popular, because they’re very deep empaths, not necessarily the best doctors. ”
And it’s certainly not judgmental and maybe it’ll bill less

How AI may transform and disrupt the medical industry: new business models and the potential resistance from established medical institutions [1:34:45]

As you look out over the next decade, and we’ll start with medicine, what are you most excited about and what are you most afraid of with respect to AI?

With regard to medicine, Zak is most concerned about how AI could be used by the medical establishment to keep things the way they are, to pour concrete over practices
What he’s most excited about is alternative business models: young doctors who create businesses outside the mold of hospitals
Hospitals are these very, very complex entities Some of the bigger ones make billions of dollars, but with very small margins, 1-2%
Some of the bigger ones make billions of dollars, but with very small margins, 1-2%

When you have huge revenue but very small margins, you’re going to be very risk averse and you’re not going to want to change, and so what Zak is excited about is the opportunity for new businesses and new ways of delivering to patients insights that are data-driven

What he’s worried about is hospitals doing a bunch of information blocking and regulations that will make it harder for these new businesses to get created Understandably, they don’t want to be disrupted That’s the danger
Understandably, they don’t want to be disrupted
That’s the danger

Peter asks, “ In that latter case or that case that you’re afraid of, Zak, can patients themselves work around the hospitals with these new companies, these disruptive companies and say, ‘Look, we have the legal framework that says, I own my data. As a patient, I own my data. ’”

He knows from his practice that just because patients own the data, doesn’t make it easy to get There is no aspect of his practice that is more miserable and more inefficient than data acquisition from hospitals It’s actually comical
Zak agrees, and he pays hundreds of dollars to get his data from his patients with rare and unknown diseases in this network extracted from the hospitals because it’s worth it to pay someone to do that extraction, and it’s doable
There is no aspect of his practice that is more miserable and more inefficient than data acquisition from hospitals
It’s actually comical

Do you think that patients who have their data coupled with AI and companies will be a sufficient hedge against your biggest fear?

Unlike his 10-year prostatectomy by robot prediction, he’s not as certain
But he would give better than 50% odds that in the next 10 years there’ll be at least one company that figures out how to use that patient’s right to access through dirty APIs Using AI to clean it up, and provide decision support with human doctors or health professionals to create alternative businesses
Zak is convinced because the demand is there, and he thinks you’ll see companies that are even willing to put themselves at risk To take the medical risk on that If they do better than a certain level of performance, they get paid more, and if they do worse, they don’t get paid
Using AI to clean it up, and provide decision support with human doctors or health professionals to create alternative businesses
To take the medical risk on that
If they do better than a certain level of performance, they get paid more, and if they do worse, they don’t get paid

Zak believes companies are going to want to be in that space, but he doesn’t want to underestimate the medical establishment’s ability to squish threats

Potential positive and negative impacts of AI outside of medicine over the next decade [1:38:30]

Let’s pivot to AI outside of medicine. What are you most afraid of over the next decade, and what are you most excited about?

Zak is most concerned about…

What Zak is most afraid of is a lot of the ills of social networks being magnified by use of these AI’s to further accelerate cognitive chaos and vitriol that fills our social experiences on the net It could be used to accelerate them
Peter saw an article 2 weeks ago where an individual (maybe currently or formerly part of the FBI) stated they believed that somewhere between 75-90% of individuals on social media were not in fact individuals
Zak believes that “It’s going to be harder to actually distinguish reality from human beings, harder and harder and that’s the real problem ”
It could be used to accelerate them
We are fundamentally social animals and if we cannot understand our social context in most of our interactions, it’s going to make us crazier

Zak’s most positive aspect would be…

He thinks that these tools can be used to expand the creative expression of all people
If you’re a poor driver like he is, he’s going to be a better driver
If you’re a lousy musician, but have a great ear, you’re going to be able to express yourself musically in ways that you could not do before
You’re going to see filmmakers who were never meant to be filmmakers before express themselves

“ I think human expression is going to be expanded because just like [the] printing press… allowed also expression of all literature in the ways that would’ve not been possible without the printing press. I’m looking forward to human expression and creativity. ”‒ Zak Kohane

Zak strongly recommends playing with some of the picture generation or music generation capabilities of AI
Peter is ashamed to admit that his interactions with AI are limited to problem solving with ChatGPT-4 He hasn’t done anything creatively with it
He hasn’t done anything creatively with it

The implications of AI achieving a level of creativity and expertise comparable to exceptional human talents [1:42:00]

One of Peter’s favorites writers when it comes to science and medicine is Sid Mukhergee [Sid was the guest in episodes #32 and #244 ]
[Sid was the guest in episodes #32 and #244 ]
When Peter reads something Sid has written he thinks, “ [Sid] has a special gift that I can appreciate. ” Peter has written a book, but he’ll never write like Sid
Peter has written a book, but he’ll never write like Sid

Just as we can appreciate watching an exceptional athlete, artist, or musician, does it mean anything if that line becomes blurred?

Zak agrees, that’s the right question
He has heard an answer many times that he doesn’t like, “ Deep blue beat Kasparov in chess, but chess is more popular than it ever was, even though we know that the best chess players in the world are computers. ”
Zak doesn’t like that answer because if we create Sid GPT and it wrote Alzheimer’s, the Second Greatest Malady , and it wrote it in full Sid style, but it was not Sid But it was just as emphatic (i.e., family references, etc.) And it was just a computer How would you feel about it? This is the jugular question
Peter thinks he would enjoy it, probably just as much But he doesn’t know who he would praise
Peter is not a religious person, but he loves to see greatness He loves to look at someone who wrote something amazing and say, “ That amazes me ” He loves to be able to look at the best driver in the history of Formula One and study everything about what they did to make them so great
He’s not sure what it means in terms of that He doesn’t know how it would change that
Zak grew up in Switzerland, in Geneva, both of his parents were from Poland
The reason he has an American accent is he went to an international school with a lot of Americans
All he read was whatever his dad would get him from England and science fiction He’s a big science fiction fan
But it was just as emphatic (i.e., family references, etc.)
And it was just a computer
How would you feel about it? This is the jugular question
But he doesn’t know who he would praise
He loves to look at someone who wrote something amazing and say, “ That amazes me ”
He loves to be able to look at the best driver in the history of Formula One and study everything about what they did to make them so great
He doesn’t know how it would change that
He’s a big science fiction fan

Zak goes to science fiction to answer this question

It’s not going to be in 10 years, but it could be in 50 years
You’ll have idols, and the idols will be, “ Yes, Greg Orovich wrote a great novel, but AI-521, their understanding of the human condition is wonderful. I cry when I read their novels .”
[AI] They’ll be a part of the ecosystem
They’ll be entities within us
Whether they are self-aware or not will become a philosophical question
Let’s not go that narrow path, that disgusting rabbit hole where I wonder, “ Does Peter actually have consciousness or not? Does he have the same processes as I do? ”
We won’t know that about these or maybe we will, but will it matter?
They’re just among us and they’ll have brands, they’ll have companies around them
They (AI brands) will be superstars: For example, they’ll be Dr. Fubar from Kansas trained on Ayurvedic medicine , the key person for our alternative medicine Not a human, but we love what they do
For example, they’ll be Dr. Fubar from Kansas trained on Ayurvedic medicine , the key person for our alternative medicine
Not a human, but we love what they do

Digital immortality and legacy: the potential to emulate an individual’s personality and responses and the ethical questions surrounding it [1:45:45]

How long until, from at least an intellectual perspective, we are immortal?

Peter points out, “ If I died today, my children will not have access to my thoughts and musings any longer .”
He wonders if there will be a point during his lifetime when an AI can be trained to be identical to him (at least from a goalpost perspective) to the point where after his death, his children could say, “ Dad, what should I do about this situation? ” And it can answer them in a way that he would have?
Zak explains that was an early business plan that was generated shortly after GPT-4 came out
He talked briefly to Mark Cuban because he saw GPT-4 and he got trademarks or copyrights on his voice, all his work and likeness so that someone could not create a Mark who responded in all the ways he does
It sounds crazy, but there’s a company called rewind.ai , and Zak has it running right now It’s recording everything that appears on his screen, every sound that it hears, and if characters appear on the screen, it’ll OCR them
Then if Zak has a question, he says, “ When did I speak with Peter Attia? ” It’ll find it for him If he asks, “ Who was I talking about AI and Alzheimer’s with? ” It’ll find this video on a timeline
It’s recording everything that appears on his screen, every sound that it hears, and if characters appear on the screen, it’ll OCR them
It’ll find it for him
If he asks, “ Who was I talking about AI and Alzheimer’s with? ”
It’ll find this video on a timeline

Peter asks, “ How many terabytes of data is this Zak? ”

Amazingly small; it’s just gigabytes
It’s possible because compresses it down in real time using Apple Silicon Secondly, you’re old and you don’t realize that gigabytes are not big on a standard Mac that has a terabyte (that’s a thousand gigabytes) Also you can compress audio immensely It’s actually not taking video, it’s just taking multiple snapshots every time the screen changes by a certain amount
It creates a timeline that has enough data, with enough conversations, that someone could create a pretty good approximation of “public Zak”
Secondly, you’re old and you don’t realize that gigabytes are not big on a standard Mac that has a terabyte (that’s a thousand gigabytes)
Also you can compress audio immensely
It’s actually not taking video, it’s just taking multiple snapshots every time the screen changes by a certain amount

Peter asks if Zak is willing to have rewind.ai on his phone with him 24/7 to record the entire range of experiences (from the good, the bad, the ugly) that are necessary if we want to formulate the essence of ourselves

Such as intimate moments, when he is arguing with his wife, when he’s upset at his kids, when he is having the most amazing experience with his postdoc
Is he willing to take those risks in order to have this data set to be turned into a legacy?
Zak thinks it’s pretty creepy to come back from the dead to talk to your children He thinks it’s messing with your kid’s head to have you come back from the dead and give advice, even though they might be tempted Technically, he thinks it won’t be that difficult
He has other goals
He thinks it’s messing with your kid’s head to have you come back from the dead and give advice, even though they might be tempted
Technically, he thinks it won’t be that difficult

We are being monitored all the time

We have iPhones, we have Alexa devices
Zak doesn’t know what is actually being stored by whom and what
People are going to use this data in ways that we do or don’t know

Zak feels that if the little guy, if we have our own copy then we can say, “Well, actually look, this is what I said then… That was taken out of context.” And that’s good

Zak has no stake in rewind.ai
He may have paid them for a license to run on his computer, the microphone is always on So when he’s talking to students in his office, it’s taking that down
Zak adds, “ There are some moments in my life where I don’t want to be on record. There are big chunks of my life that are actually being stored this way. ”
So when he’s talking to students in his office, it’s taking that down

Parting thoughts [1:50:15]

Peter has learned a lot in this conversation
Nothing surprises him more than the timescale that Zak has painted for the evolution of AI within medicine He had no clue we were getting this close to that level of intelligence
Zak advises Peter to take advantage of AI in his clinic now He would get those videos and sounds and get consent from all his his patients to follow their progress Not just the way they report it, but by their gait and how they look
He had no clue we were getting this close to that level of intelligence
He would get those videos and sounds and get consent from all his his patients to follow their progress Not just the way they report it, but by their gait and how they look
Not just the way they report it, but by their gait and how they look

“ You can be pushing the envelope using these technologies as just another very smart, comprehensive assistant. ”‒ Zak Kohane

Selected Links / Related Material

Books by Zak : [1:15]

Microarrays for Integrative Genomics by I Kohane, A Kho, & A Butte (2002)
The AI Revolution in Medicine: GPT-4 and Beyond by P Lee, C Goldberg, & I Kohane (2023) | [1:15, 40:02, 1:30:41]

New journal on AI in medicine : NEJM AI | [1:15]

Zak’s lab website : ZAK LAB | [4:30]

Center for biomedical informatics at Harvard Medical School : About DBMI | Harvard Medical School: Blavatnik Institute Biomedical Informatics (2024) | [4:30]

With AI Google can recognize different stages of retinopathy : Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs | JAMA (V Gulshan et al. 2016) | [30:00]

Zak’s editorial for JAMA on AI in health care : Big Data and Machine Learning in Health Care | JAMA (A Beam, I Kohane 2018) | [31:15]

Google research paper on transformer : Attention is All You Need | arXiv (A Vaswani et al 2017) | [32:00]

Book about what self-awareness means : I Am a Strange Loop by Douglas Hofstadter (2007) [46:00]

Microsoft GPT on the Azure cloud (HIPAA covered) : Azure. Limitless Innovation | Microsoft.com (2024) | [59:30, 1:21:30]

Using AI to interpret echocardiograms : Vision–language foundation model for echocardiogram interpretation | Nature Medicine (M Christensen et al 2024) | [1:01:15]

After age 40, 30% of your transcriptome goes down : Gene regulation and DNA damage in the ageing human brain | Nature (T Lu et al 2004) | [1:09:30]

Detecting hypoglycemia in drivers using cameras focused on their eyes : [1:11:30]

Leveraging Large Language Models to Analyze Continuous Glucose Monitoring Data: A Case Study | medRxiv (E Healey et al 2024)
Machine Learning to Infer a Health State Using Biomedical Signals — Detection of Hypoglycemia in People with Diabetes while Driving Real Cars | NEJM AI (V Lehmann et al 2024)

Study of electronic health records of autism patients : The co-morbidity burden of children and young adults with autism spectrum disorders | PLOS ONE (I Kohane et al 2012) | [1:15:45]

NEJM picture of the week : Image Challenge October 5, 2023 | [1:22:15]

App that records everything on your screen : Your AI assistant that has all the context | rewind.ai (2024) | [1:46:45]

More about GPT-4 and AI in medicine : Resources for introduction to ASI, post 2022 | zaklab.org (Z Kohane 2024)

People Mentioned

Peter Szolovits (Professor of Computer Science and Engineering and head of the Clinical Decision-Making Group within CSAIL at MIT) [4:00]
Peter Lee (President of Microsoft Research, expert in machine learning and data science) [5:15, 42:15]
Bill Gates (Founder of microsoft) [7:15]
Alan Turing (1912-1954, English mathematician and computer scientist, considered to be the father of AI) [8:30]
Garry Kasparov (Russian chess grandmaster and former World Chess Champion) [10:45, 1:43:00]
Edward Shortliffe (Chair Emeritus & Adjunct Professor of Biomedical Informatics at Columbia University, pioneer of AI use in medicine) [17:30]
Marvin Minsky (1927-2016, Pioneer of AI who co-founded MIT’s AI laboratory) [24:30]
Eliezer Yudkowsky (AI researcher who has written about decision theory and ethics) [41:15]
Sam Altman (CEO of OpenAI) [42:00]
Carey Goldberg (Science/health/medicine reporter) [42:15]
Douglas Hofstadter (cognitive and computer scientist) [45:45]
Geoffrey Hinton (one of the intellectuals leaders of neural network architecture) [53:00]
Bruce Yankner (Co-Director of the Paul F Glenn Center for the tBiology of Aging and Professor of Genetics and Neurology at Harvard Medical School) [1:09:30]
George Daley (Dean of Harvard Medical School) [1:24:45]
Joe Weitzbaum (1923-2008, professor emeritus of computer science who created an empathic psychologist program) [1:33:30]
Sid Mukhergee (Assistant Professor of Medicine at Columbia University and Pulitzer Prize-winning author) [1:42:15]
Mark Cuban (businessman, film producer, investor, and television personality) [1:46:30]

Isaac (Zak) Kohane earned his MD/PhD from Boston University. He completed his postdoctoral work at Boston Children’s Hospital, where he has since worked as a pediatric endocrinologist. He joined the faculty at Harvard Medical School in 1992, serving as Director of Countway Library from 2005 to 2015 and as Co-Director of the Center for Biomedical Informatics during the same period, before it became the Department of Biomedical Informatics in July 2015. Dr. Kohane is the inaugural Chair of the Department of Biomedical Informatics and the Marion V. Nelson Professor of Biomedical Informatics at Harvard Medical School.

Dr. Kohane is a member of the Institute of Medicine and the American Society for Clinical Investigation. Kohane has published several hundred papers in the medical literature and authored the widely-used books Microarrays for an Integrative Genomics (2003) and The AI Revolution in Medicine: GPT-4 and Beyond (2023). He is also Editor-in-Chief of NEJM AI . He served as co-author of the Institute of Medicine Report on Precision Medicine that has been the template for national efforts. He develops and applies computational techniques to address disease at multiple scales: from whole healthcare systems as “living laboratories” to the functional genomics of neurodevelopment with a focus on autism. [ Harvard ]

X: @zakkohane

Zak’s lab website: ZAKLAB

Transcript

Show transcript