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, adding a little AI education, and debate innovative AI tools and topics that will change the industry. All right, hey, so let's get started again with the news of the week. And today I want to unpack and talk a little bit more about the Nvidia open AI deal that I'm sure you've heard about. I think it's this massive deal that sort of raised a lot of eyebrows in the industry.
So basically here, the facts are that open AI is planning to deploy 10, get this gigawatts of Nvidia systems. I mean, this is like what sort of like small utilities are doing. And in return, Nvidia says it will invest up to 100 billion with a B into open AI as each sort of gigawatt comes online. So it's kind of an interesting dynamic.
So Nvidia is investing in open AI. And then open AI is then using that money to then buy Nvidia GPUs, the Nvidia evaluation climbs and then Nvidia then reinvest kind of circular investments, self reinforcing kind of in some ways, Ponzi scheme ish a little bit definitely not there. Not fraudulent, but definitely self reinforcing. And this is where people and I wrote my column about this just just this week, or we're talking about the AI hype cycle.
I mean, this is where we're talking about sort of inflated expectations because we're I mean, these are huge mind boggling numbers. And we just believe that the next AI model is going to transform every industry overnight. And that drives more investment, which drives GPU orders and drives more belief. And it's just it's concentrated within just a few companies.
But I want to turn some and we're talking about this and I think it's giving a lot of people pause, right. But in healthcare, especially in I care, we're kind of like in a different part of that curve. We're kind of when I'm talking to clinicians, we're in a slow but disillusionment in my opinion. I think everyone's really excited about AI.
But when I talk to my friends, the adoption is slow, slower than maybe they would have expected this time as we approach sort of the end of 2025. And so if we're expecting this type of adoption, we compute the actual compute could be a type of gatekeeper, right. And I'm worried about that because this next generation of retinal diagnostics and OCT interpretations and call them predictions and ocomics or body surgery. Everything we've been talking about is going to depend on massive, massive compute.
And so all this is centered on these mega companies, not smaller I care startups. And I advise several small companies and you know, in talking about the unit economics of making money and making revenue, which has to be demonstrated, we definitely are thinking about compute cost. And how much is each sort of AI generation going to cost, et cetera, et cetera. And if it starts to be concentrated within a few, just a few handful of companies as these models get increasingly powerful, that does put the center of gravity squarely in their hands.
And so the takeaway really in my mind is that this NVIDIA Open AI loop is sort of symptom of the type cycle, but it's really telling us who's going to control these the next wave of tools that we hope will actually shape clinical care. And so if I care wants to participate in the next wave of AI, we definitely need to align with these ecosystems that have the compute to build it. No answers here, a lot of concerns. And the AI bubble everyone saying it's it's we're in it and it's just a matter of timing and reality sort of setting in.
But I love the Bill Gates quote that we often overestimate what's going to happen in two years, but in 10 years, directionally, I think we are both very bullish on AI's impact on health care. Oh, for sure, for sure. You know, it's funny, Rayon, because I today in the knowledge bite those guys who listen last week, Ruponser from TopCon was on and I told you guys in the knowledge bite last week, that was going to talk about explainable AI or X AI. And you said it right.
So why is our industry? Why are our peers so slow? And I to adopt and I think, well, first of all, there's some tools are just starting to come out. But I think we have a trust issue, right?
I mean, we're in a relatively conservative profession as a whole and medicine, even more so maybe an eye care. And so I think there's a huge trust issue and maybe for today's knowledge bite, I want to kind of unpack what is X AI, otherwise known as explainable AI. And that trust issue, like you said, maybe that's what our big, maybe that's what our big issue kind of really is. And so, you know, for, I guess let's talk about this.
And we first of all, we go from talking about last week, the black box problem to understanding what, how do we get around that? Like what is explainable AI and what's transparency and what's opaqueness? And you know, first we have to look at the problem that explainable AI solving, which is how do we get out of the black box? And so, you know, for the last decade, AI and medicine has been, I mean, it's powerful and it's getting more powerful.
And you talked about, I mean, the compute and the data and the logistic, the, you know, just the data volume that's going to be turned out in the next few years is just massive. If you think of, if you feed a retinal scan into the computer and then it spits out some diagnosis, which I think we're not that far away and says it's a 98% chance of glaucoma. Well, the accuracy is amazing, but do we actually know it's accurate? I mean, and then if you ask the computer, you know, if you ask the system, well, why do you think it's glaucoma?
And the system goes, well, because it is, well, okay, that's not explainable, right? That's just an opinion. And is that any different than our gray box where if I ask a glaucoma specialist or anybody else said, well, what is this? And they go to glaucoma and I said, well, why is it glaucoma?
They said, well, because it is because my experience I'm like, but that's not explainable either, right? So I don't know if there's a difference between gray box and black box other than the black box people don't trust. So we're going to need to make it transparent so people can see the thinking and that may never happen in a gray box situation. You know, so I mean, this is a make process is hidden in a black box and we have to make it open source or somewhere where everybody can see.
Now, if you're a patient, I always think about this way, you know, if you're a patient ready to undergo eye surgery based on some machine saying you have glaucoma. Well, that could be really concerning because what if the black box is wrong? You don't want to guess. I mean, you want rationale like, is there a genotype, Venus that reason?
Have we seen time? Have we seen, you know, what's happening over time? There's a lot of questions we have to know or questions we have to answer and the demands on a black box are much higher than they've ever been for the gray box of humanity that we've called medicine for the last. Who knows how many thousands of years, right?
And so now, you know, I think about this is where explainable eye AI comes in. It's like turning and I said this last week, you know, it's like turning a black box into a glass box. Because like all these boxes, that's just that's really good. You're talking about gray boxes and black box and now glass box.
A glass box, you know, and I care it explainable AI doesn't just give a diagnosis. It needs to show its work. And this is I was thinking about like, when I'm teaching my students, I'm like, well, talk to me about why you got from point A to point D. Well, I just did no tell me what B and C are.
You know, so it's it's it's almost like we're talking to our resident and saying, you're presenting to me and say, okay, explain to me how you got there. I just want to hear you think so I know that you're thinking on the right track. And once you teach me you're thinking the right track. I'm probably not going to question about this particular subject again because I know your thinking is on the right path.
You know, and tell me, well, I believe it's, I don't know, let's say diabetic retinopathy. Well, I think it's diabetic retinopathy because I see micro hemorrhages, micro heems here, here and here. And the lens they've had a myopic shift. And you know, I mean, they start thinking through the process.
And I think that it's, you know, I always think I don't know if I ever told you this, right? I used to be a co owner of a marketing company. And these were in the days when, you know, Google was really just coming out. We had these things called heat maps.
And I was think about AI kind of like a heat map, right? Is if we look through the world as this heat map and said, like, let's take for example, I am trying to like how I put this in an explainable way, I would talk about explainable AI. Imagine like a digital photo of the back of the eye, you know, the red, a standard AI just gives you score. It says, well, here's what it is.
Well, we don't know how they come up with that. In an explainable AI system, it overlays a heat map over that image and highlights the pixels. We're seeing this in some of the OCTs saying, oh, look here, this is different than it was from the last time. And it's given even red or orange or whatever color that particular, you know, it's a glowing hot spot, a heat map that says these blood vessels, these nerve fibers, this area is different.
And it's changed. And this is why we're making this diagnosis. And that allows the I care provider to look at the image and say, oh, so I get what the AI is looking at. Now, this is indeed something we need to worry about.
Or maybe as looking at shadow on the lens and not the retina. So it's a false alarm. And I think that when we start having it, and you and I've talked to us before, I know you're a GPT guy and I'm a Gemini guy and I love turning on and Gemini says show thinking, right. If you have one of the more advanced pro models, it'll actually show you step by step, I step by step, I step what it's thinking.
I'm like, oh, so it's on the right track. Sometimes you get your answer and you look at it and you're like, I don't know where it came from. You look at it and you go, oh, here's where it went wrong. It started thinking into totally different path.
Well, it's great and Gemini, but we need to have that for healthcare and for I care as well. So if we can shift from prediction, which is kind of where we're at right now with gender V.I. To explanation, I think that changes the game a couple different ways. So first of all, it's going to I wouldn't say it prevents hallucinations, but it makes them transparent that they're there.
So we as human beings in our gray box can look at, oh, you're wrong or you're right or I understand why. So there's an assumption was totally way off base way off and you know, there's still clinical experience is going to come from us going, oh, it was way off. But the models as we learn to retrain them, right. That's we've talked about it before supervised or reinforcement learning.
Humans are going to reinforce learning. Oh, no, you made a mistake here. And the next time the model gets better. So eventually, hopefully it won't make those mistakes, but it's going to take lots of end values to get there.
Kind of like it takes to teach the resident or next turn now. Same idea. You know, part part two is, you know, I think discovery because sometimes A.I. is going to find patterns that we just simply don't you know, I talk as many times patterns that we simply didn't go exist.
It's going to highlight some section of the retiner, the corroid or the RP and say that we always thought was normal and healthy. Look at it. It's going to go, no, look, look right here. That's a change.
And we're going to we never even saw that never, never even notice day because we're looking at this in the context of where five patients behind and two tech short. We just take a quick scan and go, nope, looks the same. The eye doesn't look at it that way instantly goes, nope, look right here. There's a difference.
And then trust, you know, we kind of started this conversation with, you know, I think this is explainable. I was going to keep the human. Dr. In the driver's seat and it's going to take this from AI being this mysterious replacement to this augmentative tool that helps our decision making.
And, you know, I mean, I guess the conclusion in a, you know, explainable eyes, just it's not just going to make what we do smarter. But it's also going to help make it safer. Once again, as you and I talked about many times on these podcasts, right on is we're going to work together with these AI tools to reinforce each other and make better decisions. And who wins on that?
The patient wins the provider wins the system wins the AI wins. I, you know, we just we have to make things transparent and get out of these silos that that we're in. That whether it be gray box or black box is it said before we have to get into a glass box or everybody can kind of see what's happening. So hopefully for all you guys listen, that took explainable AI and made it explainable.
I love it Scott. And I think that's a one, it's a great, I think preview of what we should be talking about in future knowledge, but it's on what you know why are certain techniques within AI. Maybe are things that are not explainable and we should definitely go go deep into that. And it's a great segue into our fireside chat and we're going to be talking about interoperability.
And so you ended by talking about, you know, not being in silos. And this is kind of a bug that we've been talking about a little bit. And interoperability is something, you know, we've everyone supports it. It's like one of those things like, you know, I don't know healthcare.
I mean, it's like everyone says they support about. I'm not sure they actually do right now. Yeah, you know, we've, it seems like every, for a decade, I've been hearing about, oh, they're standards and mandates and laws and incentives penalties. But like Scott, I still get facts referrals from an adaptation other day.
They're like, oh, they fact you something over. I'm like, and I just kind of like thought for a second, like, why are my still getting a facts from someone. I'm still making facts here. That's right.
And why can't it be a tab. It can't be a tab in my EHR. Just like easily look at it. Why is communication between providers.
Like we're from the 1980s, right? And or earlier. Or earlier, right? Like, are we using carrier bridges next thing?
Like what's going on? Why can't you know? And so we've talked about those maybe graphs remember the blue maybe graphs. And you know, in I care, we do we're a great stress test for interoperability.
I mean, we even within our clinic, like why can't the visual field machine talk to my HR, why can't my OS easily like why isn't why is it so complicated. Why is it not plug and play. Why can't I run an AI model natively within a system? Why don't we have multi modality where it doesn't matter what system you use.
And this is that interoperability piece, right? So this is just for the listeners. This is one array on eyes big pet peeves. And so we're lying down the challenge to to EHR companies that we need interoperability.
We want data flows of information between systems and devices. We want things to be integrated. We want to be able to have non siloed intelligence software architecture easily deployed. I mean, it can be done.
It's these it's not. You know, not rocket science. And I think the issue is and I'd love to hear your thoughts that there is actually it's not a technical problem. It's an incentive.
It's an incentive problem. Well said, right? Yeah, it is an incentive problem. And this is just one of my, I mean, I listed my top 10 pet peeves and innovation.
This might be this might be one or two is that we got as you said, we got equipped being siloed. I mean, the only way we're going to make better medicine have better care delivery is to have the data. And that comes down to, you know, I think about all of the data we create that's video and audio and text and imagery and we need to have open architecture with so AI can read. I mean, the ability for multimodal AI is there.
I mean, this is not new. And this is not, as you said, this is not a technical problem. This is a, I don't want to share my information problem. And for all the audience, I am crossing my arms right now going, this is the problem.
I'm just going to have a problem. Every is going, nope, you can't have my data. And I know, Rand, you know, when you say I and my, you're not talking about the patient. You're talking about the patient.
I'm talking about a lot of companies, digital diagnostic companies, EHR companies, everybody's, I mean, you could say frame companies, lens companies, private echo. I mean, no, date is mine. Can't have it. And I agree with that.
You know, we talk about date. You bring it up. I quote you all the time on this. So now, Ray, data is the new oil, right.
And I know everybody wants to go, well, I own the oil well. I don't want to share the oil well. I want to make all the money off my oil well. And then I get that, right.
That's just basic capitalism. But somewhere there has to be, there has to be a way. And I don't know if I know the way. I think I have some ideas.
But there has to be a way where we can share this data and open architecture format where we can all learn and all get better and all make money and all saved money to the system. There's so much money out there to be saved that we can all profit from. Yeah. I mean, you, the oil is of no value in the ground.
You, it's only valuable after it gets manufactured in place in a car so you can use it to go somewhere. In the same way that the data that they have is only captures any value when, when it's used in a way that's helpful to the patient consumer doctor, whoever to make a better decision, better outcome. So yeah, I think I think you're right. I think there is a sense of they're holding on to the data for the data's sake, but they're not, you know, what do they do?
But they're not doing anything with. Yeah. There's the frustrating problem is if you were holding on to your data and actually creating something that was going to move medicine forward. Wonderful.
All good. But that's not happened in either. Right. It's just.
It's just data being wasted and, you know, for all the audience wake up call it's being wasted when it goes into a digital filing cabinet and nobody ever looks at it again. That's an end of one that goes nowhere and we've got to fix that. And, you know, I mean, I know their systems out there. I've seen them that they're going to fix this right.
It's common. And, you know, I just, I wish maybe this is the altruistic side of me that's sometimes a little naive rayon. I just think that there's got to be a better way. There's got to be a way that we can share data and have interoperability where it doesn't matter what device you have.
There's a whole other subject. You know, I'm going to go off on a tangent from in around. There's a whole other subject just bugs the crud on me. I am so sick of being subscription to death lately about everything's an ad on it's this ad on doing each are.
And this ad on to need are and this ad on to need are and this. Pretty soon I look at my subscriptions every month and I'm like. Before I even get to my fun stuff like my serious exam in the office, I got $3,000 worth of subscription services like. It just doesn't have to be this way.
I mean, why can't we just have an open architecture integrated model. Where we pay for what we use and we get away from the part of my language is the damn subscription models. But once again, that's an in drop or ability problem, right? Yeah.
Well, you have to get a tack on or an add on or extra piece to make a bad system try to work better. Yeah, on an extent, it's kind of like a middleman fees. And so I feel the pain of smaller companies trying to do something cool and unique and valuable to their customer, but then having to have as a prescription fee because they have to cover the implementation and integration costs. And so there's real pain to go around.
I'm not blaming all about you know, the the HRs have their own issues. I think this is a very. There's a this is a systemic problem. I think there's there are regulations that make things difficult.
There's our maybe our cake laws privacy issues that may or may not be true issues. And I think that's full head which have a good reason to be there, but may have inadvertently held us back from from more substantial innovation because I know in speaking when I speak to patients, they are really frustrated that it seems hard for two doctors to communicate. In a way that's helpful to them. Just get on a 30 second teleconference call within a universal system.
Like let's just talk. I know the mean we're at the spot where if you and I you know you're in Dallas right. Yeah, Dallas and I'm in Colorado. You know outside of Denver and why couldn't I call you and say, hey, Ray, this is the case I got.
This is what I think we're to do. I'm a first person down to you for this surgery for this one. And they I take it and transcribes it all puts in the system and blah blah blah. And get rid of the damn facts.
Right. I mean, let's just give you the give you the summary of exactly what you need to know. Do it right instead of using this. What do you call it?
1980s technology, right? I mean, just I just I struggle so much audience with why we have to keep working in the dark ages. We're phrasing it from a far side chat to a far side brand here. I know, I know.
But I think we are identifying the problem and asking our audience to talk about solutions until we love comments that I was always at a and I care. And everyone's valuable emails and contributions and thoughts and grants themselves. It's been super helpful for me to learn as well. So everybody, thank you so much for joining us.
Hope you learned a little bit and like Ray on set is, you know, feel free to email us. We would love to chat with you and get ideas. I think one of the next things one of these days, Ray on maybe after the first of the year, we need to sit down and start going, OK, here's our ideas for solutions. Yeah, who's listening?
Exactly right. Sounds good. Have a good night everybody. You've been listening to Reel Talk and AI and I care your weekly podcast to keep you informed about AI technologies revolutionizing I care.