00:00There are two megatrends that are going to define medicine in the next five years, and
00:07then going onwards.
00:08One is that AI's capabilities keep on improving literally month to month.
00:15The large language models that I used a year ago are nowhere near as good as the ones that
00:21I'm using today.
00:22At the same time, primary care is falling apart.
00:25I cannot find primary care doctors to refer my friends or faculty to.
00:30And that provides an opportunity on a patient-centered focus, where a lot of the missing functions
00:38that the primary care doctors used to provide, we can provide and augment, maybe with nurse
00:44practitioners, maybe with physician assistants, or maybe just with the social network of patients
00:50in a consumer-focused way.
00:56AI's been around a long time.
00:58It has become the buzz because I think the end user is now seeing a lot more of it directly.
01:04But there's so much opportunity for AI and technology, machine learning in general, in
01:09healthcare.
01:10And I think about it, especially in three particular areas.
01:13I think about the ways in which we can take costs out of the system.
01:16There's still a lot of phone calls and faxes and the humans involved in work that doesn't
01:21really add that much value that machines can do and should do, because then we can
01:25invest that savings in the experience that actually helps patients get healthier, making
01:32more providers available, improving that experience.
01:35The second area is the provider experience.
01:38So providers often have to go through pages and pages, not just dozens, sometimes hundreds
01:42of pages of history to try to find insights so that they can focus with a patient on the
01:49topics that are going to matter most.
01:51AI can help extract a lot of insights from that history.
01:54AI can also help a provider take notes without having to spend their time facing a computer,
02:01so they can be facing the patient and spending energy with the patient.
02:05And then, of course, I think there is an opportunity for customers and patients to find out more
02:10about their choices so that they can be more informed in their own health and actually
02:16perhaps be navigated using technology without necessarily taking up the time of a provider
02:22to do that navigation for them.
02:29When GPT-4 came out, that was a two-year sprint, basically, from GPT-3.
02:39But it keeps on accelerating.
02:41A year later, many of the large language models started having visual and other multimodal
02:47capabilities.
02:48And so I took a picture that had never been seen by any doctor.
02:52It was just out of the New England Journal of Medicine.
02:53It was a picture of the back of a man, an older man, who had developed a lot of itchiness
03:00over the past day and felt really bad.
03:03And there was a photograph of his back, and there were these, looked like scratch marks,
03:08lots of lines.
03:10I took everything from the puzzler, all the text and the picture, but I removed the most
03:15important clue.
03:16And I gave it to GPT-4 and said, what is this?
03:20And this was with a clue.
03:22Most doctors didn't get it.
03:23And it said, it could be something called bleomycin toxicity, reaction to a chemotherapy
03:28treatment.
03:29The other is shiitake mushroom toxicity.
03:32I'd never heard of it.
03:34And what I removed from the history of the patient is that they're eating mushrooms the
03:39day before.
03:40And so we're having this phenomenal, fast-moving capability that is behaving with all the hallucinations,
03:52with all the problems, better on average than most doctors in terms of depth of knowledge.
03:58How do we address the problem I just asserted?
04:01Well, it turns out that right out of the box, when you train an AI on all the information
04:07you have in the internet, it'll say a lot of weird things, some of which are politically
04:13incorrect.
04:14So OpenAI and Google put a lot of effort into something called alignment.
04:18They give examples of questions and how they want it answered.
04:22And you give that tens of thousands of such examples, and it starts behaving in a certain
04:27way.
04:28But it's not aligned to a certain behavior.
04:31So you can actually align AIs to maximize certain outcomes.
04:38If you're making the AI, you could align it to the patient, you can align it to the hospital.
04:44No one's talking about this, but it's actually a well-known procedure how to align.
04:49It's something called fine-tuning and reinforcement learning with human feedback.
04:53These are all technical terms, which basically say, I'm taking this whole big machine and
04:59I'm telling it how I want it to behave.
05:01What are the concerns about AI?
05:04There are concerns around bias, around hallucinations, but in my opinion, the biggest concern is
05:12who is the AI serving?
05:15Is it maximizing the interests of the hospital, of the insurer, of public health, or the patient?
05:24And answering that question, I think, is perhaps the most important concern.
05:29And right now, we're not talking a lot about it.
Comments