Welcome to Real Talk on AI and I care with your host, Dr. Scott Morris. And I'm co host, Dr. Rehan Ahmed.
In each episode, we discuss current news, add in a little AI education, and debate innovative AI tools and topics that will change the industry. Welcome to another episode of Real Talk. My name is Dr. Rehan Ahmed and we're Dr.
Scott Morris. Scott, how have you been? Good. It's been a really crazy week, Rehan.
There's just so many cool things going on. You know, maybe it's the new year and you know, I have a new power driving me in terms of getting stuff done, you know, the GSD gene, but it's been a busy week. I think it's been a lot, you know, last week we talked a little bit about chat GPT health and if it's cool with you, I think we're going to dig in a little more to that this week. Yeah, I think that's a huge topic, a lot to discuss there.
So yeah, go for it. Scott, tell us about it. Well, let's do our knowledge by Rehan. And I think it come to me that people often they, we use GPT as this thing and do people really understand what GPT means.
And so I want to cover a knowledge bite of what does that really mean? So GPT stands for generative, pre-trained transformer. So, you know, I was kind of hoping it sounded cooler, right? But generative, pre-trained transformer, you know, and all this stuff, it's like one of those, you know, Michael Bay movies, like this is transformer that's, but it's already trained.
No, so, so you'll just kind of think about what it means. So first of all, let's talk about pre-trained. So you kind of think of pre-trained like as our models childhood. You know, we ran you and I often talk about this is like childhood, teenage years, adulthood, that kind of stuff.
Pre-trained is the adult is the every model's childhood. So before it, we, for these models are ever being released, they're sent these digital libraries containing, you know, pretty much everything humans have ever written, you know, Wikipedia, coding forums, classical novels, music, movies, recipe blogs. You know, and it, it, it, so it really doesn't know things like a human does. It learns to recognize patterns.
You know, it's spent, these models have spent billions of hours seeing the word, let's say salt, usually followed by pepper. So when they see salt, they think, okay, well, the next word's probably highly likely it's going to be pepper. Or the E equals MC is almost followed by squared. Right?
So it's learned, let's call it the shape or the pattern of human knowledge. So it's pre-trained on all of this information. And we, and we'll get into more of the detail of that piece when we get into our fireside. But part two is transformer, you know, and you can think about this as, oh, it's the technology.
And so older AI used to read one word at a time, like a person pointing their finger, reading a book, one word at a time. Like when we did, when we were little kids, and the sentence was too long, many times it just kind of forgot how it started. But transformers are kind of different. It looks at an entire page at once, and it uses something called self-attention to realize that when you say it at the end of the paragraph, you're referring to the umbrella mentioned in the same sentence, or maybe many sentence ago, it starts to understand context, not just words.
It's not reading one word at a time. It's reading sentences or paragraphs at a time to process this. And then part three is that word generative, right? And so generative is that, you know, it's, we've talked about, it's not like a search engine that points you to some website, it takes all of those different pieces that it learned.
And all the different stuff in context, and says, I'm going to put all this together, and I'm going to build you something new. I'm going to generate a new answer from everything that I currently know, whether it's writing like a legal brief or debugging your code or explaining quantum physics. You know, it's generating some unique response based on the probability of what you asked your prompt to what it thinks you really want to learn. And so pre-trained is it's pre-trained in all the world's information, and then it uses a transformer, that technology, to understand context.
And then it's generative because it takes that information, create something new. In short, I guess it's kind of thinking about like the world's most well-read assistant, ever. But as we talked, that comes with challenges, like was the information correct? Did the information it learned from?
Was it something it created itself? You know, and so there's lots of problems with it, but we just wanted to, hopefully, for the knowledge by everybody's a little bit understanding what GPT really means. Yeah, and we've talked a ton about how it's like auto-correct on steroids, and that's ultimately a very simplistic way of looking at it. Scott's always great to hear about what GPT is, what it stands for, because sometimes you get lost in all of this stuff.
And understanding the core architecture is helpful in sort of seeing about where it's best deployed and sort of what makes sense. And in talking about last, we talked about chat GPT health. And I know you wanted to dive in, and let's get into a little bit more about the details for our fireside chat and how we think something, because this is, you know, we both said this is going to be a big deal. And it was coming, and sort of your thoughts on this.
Yeah, you know, I think I come up with this last name. I was like, you know, is it GPT health? Is this really healing, or is it just hype? The more I thought about it, Rayhan, the more I'm worried that it might be more hype than it is actual healing.
But so let's talk about, for the audience, you know, let's talk about what are the goods, what are the bads of this program, at least in its current form, and obviously like everything else, it will get better. But let's talk about the goods, let's talk about the bad. So when we have our patients ask us this, or when we're using it ourselves, we have maybe a little clear context. Remember a couple of months ago, we talked about black box, gray box, clear box.
Let's give some transparency to the goods and bads, the advantages and the pitfalls of what health GPT might just mean. Well, let me, I don't know if I'm playing the pro side here or not, but when we talked about GPT health last week, I think the stat was 20% or more of all conversations that I currently had with chat GPT and probably other LLMs like Gemini and Anthropic, our health related questions. I certainly do it where I'll ask about health things about me or my family, I'll upload labs, and I'll ask, hey, you know, what should I be worried about? What should I ask my doctor?
You know, I think does this look normal? And you know, I upload like dozens of medical reports, MRI reports. Is it 100? It is pretty good.
I mean, well, I'm just kind of like I, it's already happening. It would be, so the pro argument is, wouldn't it be great if companies make sure it's guardrailed by good evidence? Well, they're grounded in some, you know, there's some guardrails there with physician review or something like that. Well, I think that's a, that's a, that's a, that's a, that's a, that's a, that's a, that's a, that's a, that's a, that's a, that's a, that's a, I'm going to play devil's advocate.
You're absolutely right. I think it definitely is going to be helpful and advantageous for people in terms of preparing and understanding, you know, a lot of our health talk is very complicated. Lingo, for people who didn't go to medical school or didn't go to what op school or didn't go to whatever, you know, PA school. You know, I think that, I think that's great.
I think it's going to definitely help with preparation and education and learning some more of the things. My concern, though, is this not a substitute, you know, we always talk about AI is not going to replace doctors with the doctors who use AI will. I still think you can't diagnose, prescribe or manage, right? And definitely can't manage.
You can give suggestions. And I think that it's going to be a valuable tool for clinicians and patients to use and talk about and think and see all those things. But kind of like we talked about in the knowledge bite with, you know, in terms of the pre-trained is is it actually pre-trained and right information and will it start learning from bad information and thinking it's good information and then creating bad results. You know, is it going to hallucinate bad information?
You know, the challenges the data were feeding and you yourself said you know you've feeding all these things into it. But did the radiologist, the red DMRI read it correctly? You know, so is it taking that information and making a conclusion from that? Is it generating?
Because remember, it's chat GPT is generative. Is it generating an answer you want to hear or is it really generating a correct answer? I don't know. That's a really good point.
I mean, garbage in, garbage out. One would hope that it's been sort of validated or reassessed that whatever outputs it comes up with, it sort of makes sense and people have sort of looked at that before. I will say when I've used it, I've been surprised at how accurate I think it's been. It hasn't been way off.
So I think the proof is in the pudding. We'll see one nice thing about chat GPT is that it connects with EMR systems, which is really, you know, I think that's that's kind of like the missing link that a company of that size can easily do. They can integrate with Epic in a way that's very quick. And the news is.
It definitely has to have a get there too, right? Because we talk about this when we talk about EHR. EHR is a clickbox, right? It's not capturing all the information.
It's a checkbox. And you know, so are we going to be creating bias by the fact that we think that everything entered into the EHR is actually the whole picture? I mean, you and I both we've talked about before. It's not even close to the whole picture.
It's just what we document for record keeping and for practice claim adjudication or submission, right? It's not actually the whole picture. So even when it links with the EHRs, is it really going to get the whole picture to get the right information? I think you're definitely right about how narrative clinical assessments have become much more templatized.
You know, I think with with recent EMRs that I know that for a fact, and the EMR we use, it's just so it's just a template now. And whereas before with paper charts, and you'd actually write real information, or what you really think it now is just a lot of Galbady Goop. Now we're feeling more of them. I mean, the connection with labs and radiology reports, I think that's that could be super helpful.
And here's the other thing. So chat GPT recently acquired, I mean, just two days ago acquired a company called Torch for a hundred million dollars. It's like come back to a year old, believe it or not. And they are the secret sauce with integrating a bunch of different types of medical data, including wearables.
So I think this is where an LLM like chat GPT has a distinct advantage of being able to connect to not only an EMR, but then also connect to your or a ring or your smart glasses or your glycemic, you know, the monitor that you're that you're wearing. So it connects both online offline wearable sensors with the MR is all in sort of one nice health. Personal intelligent assistant for you. It's the smartest doctor that you have that knows you better than any of your other doctors.
That's I think that's the the the the the pro case I'm making. All right. Well, let me throw in a double that one too, because I agree. I think that part's great and that it's going to actually build a Git biometric data.
We've never even dreamed of capturing. But the problem with all the GPT programs are it has a memory issue. Can it remember all that data? Or will it eventually start hallucinating what that data means?
You know, all the systems, whether it be Gemini or GPT or Claude or whatever, they have a memory problem. You know, it's like I almost I cannot almost I'm not trying to like poke fun too much of this, but I always think a lot of these large language models, they got a little bit of Alzheimer's going on. You know, they they just don't have great memories. They have great knowledge, but not great memories.
So will all that information we pump in? Will they actually remember it? Or there's there have to be some other piece of some other type of database sitting between the LLMs and the output. Yeah, that's really good.
I've wanted to actually have one of that too. Like why is my chat GPT not remembering conversation from a couple weeks ago? And this is where I grew through. I think that needs to be solved.
But I would also make the point that what are we comparing to? Right? And so as we look at technologies, it's easy to compare it against some mythical gold standard that doesn't exist. But we are compared to the average prime grade box or to the gray box.
And as much as I love my fellow physicians, you know, I think this is the worst chat GPT health there is going to ever be. That's good point next week. And next month, it's going to be a lot better. So yeah, I'm sort of making a I'm I'm I'm I'm I'm manning this case a little bit.
But yeah, I'm really excited about where it's going and all the different applications, especially with with I care. And we talked about a last time too. Yeah, I mean, I hope it does implement with I care at some point, but we know how our EHRs compared to like Epic, which is one big beast, our EHRs are a little slower to adapt to any kind of change. But I mean, I look forward to I'm not trying to be all Debbie down around this one.
I mean, I think it has tremendous possibility. I just think for the audience, we need we and our patients need to go in this with eyes open. Thinking using our gray matter to interpret does this make sense? Is it right?
Other problems? Let's go into this with an open mind. Yeah, 100%. I think that's a that's great parting wisdom there Scott.
Yeah, everybody stay tuned. Next week we have a really good show coming up. We're looking forward to having you and listen, always, always, please do not hesitate to let Rayon know. If you're like it, if you don't like something, if you'd rather you want to hear something, we haven't covered yet.
It's either Scott S. E O T or Rayon our E H A N at AI in I care dot AI. Those are email addresses. So Scott it AI and I care dot A or Rayon at AI in I care dot AI.
We look forward to hearing from you. Know what we can make better, what you like, and how we can change your life. You've been listening to Real Talk an AI and I care your weekly podcast to keep you informed about AI technologies revolutionizing I care.