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. Well, everybody, welcome to another edition of Real Talk brought to you by AI and I care of jobs and publication. Thanks for joining me, Dr. Scott Morris and my co host.
Dr. Rehan Ahmed Scott, I know you're about to leave for vacation. So this is a, you're seeing like Aaron, a good, good mood tonight, huh? Well, it's one of those weeks that, you know, I'm sure everybody in the audience can attest to that, right?
You know you're going and you got to, you're going to try to do everything for this week and everything for next week and everything for the week after. And you're just completely buried, but there's at least a light at the end of the tunnel, right? And you know when you get back, it's going to be the same way, but at least I'm going to go into the glimpse of I can't wait for vacation. Yeah.
Yeah, you know, exactly. So tonight for our knowledge, we are going to be talking about the word token. If you spend, you know, just like 10 minutes reading about AI, you'll probably come across this word. And I love doing this podcast.
Scott is because I get to learn. I read about tokens all the time. I'm like, I think I know what that is. But what is it exactly?
So let's just sort of demystify it for us and for the audience. So here's what AI sort of doesn't tell you. And we think we've told you, but we want to re, to rehit this point again. AI actually doesn't know English.
Scott or Jeff. Which is what you're going or Spanish or any other language, right? Or any other language. Well, that's not true.
That's not true. They does know a language. It's called zeros and ones. Yes.
Yes. That's right. It has, in terms of tokens, it sort of has a fixed dictionary of fragments of word pieces, punctuation. So every piece of text that you have that you give the model gets chopped into fragments before it could read any single character.
And it's those fragments that are tokens. And the chopping of that text is called tokenization. So think of it like scrapbook and AI is as a bag of tiles. So.
So if you say something like the patient, I, you know, the is a tile, patients, a tile, I has a tile one token rare, you know, those those are like small pieces in a bag. But then it gets a little bit more complicated. And let's just walk down our specialty. Right.
And let's talk about more tokens. So cat erect one, one two tokens, glaucoma pie two tokens, refraction two tokens. Those are cheap. But you talk about little longer words, astigmatism, you know, that's us.
Some are like syllables astigmatism is them. Probably three or four tokens, care to cone us for tokens, fake emulsification, directly like to maybe track any of those. There you go. Yeah.
I talk about big words like, you know, DCRs are dapper of sister rhinosthenes. Oh, that's that's a lot of tokens there. And then now we talk about entire compound diagnoses like central serious choreographie. What what a what a what a bunch of tokens and it's expensive and talk about drug names that this isn't an advertisement, pegs, set a coplan farm was like totally they designed these to be unrecognizable and it gets it gets expensive and epinims like show grinds or bond, hippo lendao right design English.
And so these syllables get shredded individually right. But it's also confusing to and I'm going to get to why this is important in second like D segd mech, Yag slt those are one or two tokens. ICD 10 codes also are tokens. They does not understand ICD codes right H 40 dot 11 dot X dot one token wise right into four six fragments from an authentic merc mix right.
So tokens is not also just a texting images have tokens. So if you have a picture of a retina that gets chopped up into little pieces like a 16 by 16 pixel, which is a really tiny part of the retina. So that's a token. So a fundus image is 4,000 visual tokens.
So you're sure to get in a point that tokens are parts of words or parts of images that the AI used to chunk things right in this process that we call tokenization. What does it even mean right why does it matter to anyone who's listening to this. So the first thing is context window. If you're typing to chat chat GPT Gemini or Claude, they make a big deal when they start comparing each other to how big of a context window they can hold right.
We have a 200,000 token window. Well, that sounds really big maybe until you start actually pasting in a Tribeca like you know an operative note or patient chart or a complex document you want to be you want to get addressed or definitely images right and those then the budget will climb accordingly. And you know we've been sort of talking about pricing and that's where this really comes in. So if you're a developer and you're using AI, you're going to be talking about tokens because every time you interact with a model you use and called the API which you've talked about which is the way you interact with it, get information in and out, they charge per token.
So it's really important for developers and as we see you know I think this is a good topic for later on as we talk about cost structures. I think a lot of a companies are going to be talking about token. So as customers we should understand what tokens kind of are. Third is accuracy when a model has ever seen the word.
FACO emulsification right which is probably four or five tokens and just to be clear like one word is about probably one and a third token like 1.3 tokens. So think 500 word document is 700 tokens kind of like just a rough math of it. But the word FACO emulsification for example is chopped into just not five anonymous pieces and that you're the LM just learns them piece by piece. That's why you brought this up last week.
That's why general purpose AIs they mangle they sometimes really mangle medical terminology with kind of we talk about weird confidence as if it's not like that doesn't everyone wants to know what are you talking about that's just a mangled set of words that don't actually make sense. Again it's gambling it is an algorithm to figure out what the next token is going to be so just review so tokens aren't exactly words they're like tiles like scrap like a couple a couple tiles of a word. It is not it is the it is the language not the English language or this managed language was Japanese language but the language of the LLM. And then we abstract onto it we sort of put our human we know we think it's speaking English but in reality it is just figuring out the next token.
Everyone wants to ask us to look under the hood and see what's going on so next time you sort of read about token and tokenization and charging per token. We hope you'll have just a little bit more insight I know I learned a ton researching this topic Scott any any additional insights that you have. No I just think about every time you prompt you know the longer your prompt hits the more tokens you're using but the longer or better your prompt is the more information you're going to get also right so it's a balance of you know it's it's a deposit and a withdrawal from checking out you know zeroes and dollars and no dollars of. How what's the fastest most efficient way to use your tokens to get the right answer right so I always think about it is how can I.
How can I summarize this prompt as most efficient as I can use at least my tokens now I'm in granted a lot of these you know from a developer side as you said this is a big deal but as us as I like a personal user I mean many of our prompt things carry a million tokens or 10 million tokens or all. 100 million tokens per prompt right so you that's their capacity well we're not I mean you can. It's where you know only if I'm uploading several documents at once will I hit a token limit it's good to know and then you get very easy for the work I think that both of us to. So they're not doing part of our coding as far as you know i'm not.
You know we. I am but i'm still not hitting the time I mean I haven't it's a rare rare occasion that I ever like max out my token limit on a process. Yeah no exactly but I think we're going to start living in a world where more and more people understand the cost of compute. Yes as as as it is just get some more common so you know we've all I was talking about data as a new oil but there's there's got to be gas stations and.
Well I think you hit on that you know we've talked about a lot of the last couple of weeks about these data centers right in these data centers are getting really really expensive. To make and maintain both from water and power more so even than space I was having this conversation with brothers and town and he actually builds data centers. And there's over 500 data centers being built in the United States currently with each data center having close to 8 million miles of fiber in each data center. And you know the water that it's taken to cool as I know we've covered this before but I think about you know I.
I do believe that tokens and that kind of stuff are going to have a cost and right now that cost is really dirt cheap and it was really expensive not too long ago and now it's dirt cheap. But it's more and more you got to remember that less than 5% of the United States uses some form of LM on a daily basis other than just asking a question to really use it now if we increase that to 50% of the United States just the United States alone is starting to use. You know detailed complex AI I think we're going to see the cost of tokens go up and we're going to pay for what we use. Yeah I'm not sure that's a bad thing but totally agree totally agree I think that's coming so what do you have in terms of our fires side.
Yes you know we you and I were talking with this earlier this morning during a call and I've had this conversation with my son is going to do IP in the field of AI for law school and and I've talked to this conversation a couple other people lately about well who's accountable you know and I think I mentioned this morning we'll probably have an article coming out on that about this exact subject about who's accountable in health care AI. You know there's a couple of big lawsuits going on in the state of Pennsylvania right now or they were just launched on Monday and I'm sure we'll be addressing those in more details more details come out but it makes me wonder like as we go right now I wonder like in our current world without AI if I make a mistake as a provider. The net result is it's probably going to come back and get me you know it's probably going to be that's what we carry malpractice for or whatever because we're responsible but now as we start depending on AI and I think there's two sides of this rayon maybe we'll go down side one is who's responsible for when the AI gives us information that's not correct. We'll go down that way first and then maybe we'll come back to what happens when providers don't take the advice of AI and the AI was correct and the providers wrong but we'll get there maybe in a minute so you know I was thinking about right now the current thinking about that rayon is you know there's three people that are this let's call it ecosystem of AI that might bear and nobody knows.
There's no legal precedent for this because AI is moving so fast the courts can't possibly keep up with this. But you know there's three parts of the shared ecosystem there's the developers and as you know much like you rayon being somebody who's developing some of these AI products you know I think about all the time if I incur bias in the system or I don't have a system that has great integrity in times of how it works. I'm responsible for that right I mean I'm the one of the developers. That's part one part two is then you have the health system saying hey no everybody you need to follow the AI well are they mandating that what is that responsibility curve.
Look like and how do you prove that health system said you should or should do it and then last but not least like the clinicians themselves are they choosing to follow it even if it suggests like you did a great job and it's you know it makes upwards right a fake or piratoma loosens right we all know there's no such terms that but it might throw something together like that we kind of clinicians go wait that's not right. But maybe there's something more subtle that our gut says we're not right but we go yeah but it's AI and it is right most of the time and we follow that course of action and something bad happens. Who's responsible is it the developer is the health care system that said you need to use this is the clinician for not knowing. I don't know the answer to that right I mean it's that we've talked before about you know algorithmic bias and discrimination that's a developer problem but if there's no transparency clinicians can't know that there's an algorithm discrimination problem so we come back to are you going to make it transparent.
I don't know what's your thoughts on that I mean I. I think back and forth about this is a tough topic and this is a great fireside chat because there are no clear answers I don't think anyone has come up with an answer this is all stuff that is will be prosecuted and in courts and will develop precedent and all these things but I think it's for a big picture right. We have to we have to think about well in terms of the AI whereas whereas the risk of what applications being used some are low risk and some are high risk and in particular with with the AI system which ones are autonomous meaning there's no oversight it's not designed for any oversight and which. No, for example in I care we already have that was a 9229 code and in AI screening for diabetic retinopathy there's no I'm it decides on its own the algorithm and there.
Interestingly my understanding is that the companies like say digital diagnostics or you know these other they actually carry their insurance now they design the algorithm so it is 9 I mean it will and there's different things about what you're missing are you going to miss what's your rate of missing someone who actually has diabetes and classifying that as normal and what's worse or versus someone who's normal and classing them as diabetes well you know and there's different ideas on that. If it's a screening test or not we don't have to go into all the details there but I think that's one part of it I think for assistive technologies where there is designed to be human oversight well I think you know my sense is that the physician will be will play a larger role for any issues that come up because those systems are designed the indication is with physician oversight. I think where they where this becomes really tricky is where there are AI superhuman systems and we've sort of talked about that bucket before and here I'm talking about for example the retinal photo raft that will predict cardiovascular screening or Alzheimer's suppose you do this on a patient and flags them as having you know cardiovascular disease or flags not as not having cardiovascular disease. And patient has a you know has a has a my has a heart attack you know the next month or something you know who's responsible and you better believe that that can happen and the reason that is I think more difficult it's because that's a superhuman capability I can't look at the red now and tell you what your cardio which whichever is that heart attack next month you know but the that's the promise of these AI systems that they're going to be this predictive power of some degree of certainty in the future and so what I'm going to do is I'm going to do that.
And so I think that's that's going to be a crucial issue I think anyone knows the answer but I think we are coming to the same way that these elements are being sued currently for you know really serious stuff like homicide you know they're being like for sure like giving you a psychological challenge bad information right information and so I think you know we will see we shall I don't know what I mean we change will we now have informed consent for patients to sign I mean saying that hey you know you're taking the same risk if there's a hallucination well boy that's risky unless there's transparency but then we're asking patients to be as informed about what risks are with the algorithmic bias and stuff as doctors are now I think the doctors really know where the bias is and we all have our own bias and we're talking about that before we all each and every one of you listening as well as you know we're talking about that. We have our biases right and we sometimes those are wrong bias sometimes those create incorrect solutions but here were there's a there's a trail right I mean here where there's an electronic trail to it not just the discussion in a room and I think that's one of the things we're starting to see from accountability you know is that this is one of the lawsuits that's going on in Pennsylvania is a scribe was an AI scribe misinterpreted something that was said and then another doctor actually took that information and prescribed off of it because it was misread and nobody checked it there was no human in the loop on that and so now they're saying well the AI scribe is the responsible party. I don't know I mean I I think there's there's so much we don't know right there's so much we we don't know about you know where's the audit trail who's going to be responsible but then on the other side because as I talk about around is that what happens when we know that I get so much of this right because it can look at all of the data that's out there across the spectrum of digital data and give a better answer way smarter than I could ever be and probably smarter you could ever be you know if we as a provider go yeah that doesn't sound right I'm going to follow my gut there you go that's my bias and make a wrong decision I chose to ignore what the I said and then am I at risk if something goes wrong because I made a gut right now that happens and it's pretty tough it's he said she said against other people but you know now and the I goes yeah well we've looked at 30,000 people with this and this is what we suggest and you say now I don't believe I'm going to do some different and the outcomes bad am I the provider at risk am I the provider at risk I think I am I think the definition of standard of care is going to be redefined so right now it's dead. So right now standard of care is basically what would a reasonable provider physician do in your circumstances in your basically same interest interest in your same geographic area right so standard of care should depends on geography now I wonder if the standard of care would now include it with the use of an AI assistive technology what would a reasonable position do with AI technology given it so it's so we're going to do that.
So prevalent now it's everywhere so the physician be using it and you have made a really good point about the auto trail because these AI systems are also interdependent in a way because there 95% many many of them are using quad GPT Gemini as backbone right in somewhere the other there you. Okay as you all remember we talked about this last week their LLM's creating answers not always reading what's already out there. That's right and so it may be using APA as a call various vendors and so it gets really messy very quickly and and where's the auditor who made the mistake or what system made the mistake or can you tell if there's no transparency. Yeah and then there's a black box problem to know i'm just so funny you read right into what I was going to say next now on the other side I also wonder does this if we step away from this as providers now who's responsible when an insurance denies something that AI says we should do.
And there's a bad outcome is the insurance company now responsible for denying prior authorizations are denying a procedure because of a financial reason instead of a medical reason I kind of wonder if we're not going to see this let's call it black box denials of insurance claims really get flipped on his head because I mean you know i'm sure you every day too every day I get some prior authorization or some we did some. We did some procedure they go no we're not going to cover it I'm like but it's it's it's the best course of action I mean it's the best course of action clinically standard of care whatever and they go note and I'm like well so when something bad happens. Since you said I couldn't do the standard of care are you responsible insurance company I think the answer should be yes but once again we're not there I think we're having a ice fighting a ice you know the insurance companies are going to have their own a ice. That's true but they get their data right I mean are they getting their data from true they can't get rumored because all the insurance club insurance companies collect now is your IC 10 and your CPT code that's not care that's the financials that's like how does there add to determine what's the right thing and wrong thing when all they have is two pieces of the over 500 pieces of information that happened in most exams right so.
We definitely don't we don't promise answers here but we we do have a lot of questions I think this is one of those topics far fireside chat where there's definitely a lot more uncertainty out there. We are I am so interested in a Scott you would love to hear from our audience on this topic get your feedback what are your thoughts on AI and accountability when it comes to I care reach out to us at my email address is really on our eha and I care diet Scott it's a C O T one T. At AI and I care dot AI and of course you'd see all our articles watch all our videos at AI and I care dot com look looking forward to speaking with you all soon Scott enjoy your vacation my friend. Thanks my appreciate everybody have a good week.