Latest / AI Meets MedTech / Building Trustworthy AI in Healthcare
Transcript
- Anna Nesterova: Welcome to AI Meets MedTech. My guest today is Joanne Skiles, AI and cloud engineer with vast experience working through different industries, including healthcare. And I'm so excited to have her on board today. Thank you. Welcome to AI Meets MedTech. Joanne Skiles: Thank you for having me. I'm really excited to be here. Anna Nesterova: I'm excited about our conversation because I have nothing to do with AI and cloud engineering and I have so many questions for you, Joanne. But before we jump into my questions, if I didn't know what you do, can you tell me in the audience, what do you do? Joanne Skiles: So I get like long-winded. So, but let's let's do like a one sentence is I build systems for AI models that make them actually useful in the real world versus just that answer in a chat prompt with chat GPT. So my Anna Nesterova: Mm-hmm. Joanne Skiles: entire system, my entire work has been building these sophisticated implementations that utilize models, either custom models or or something in the back end and make those actual AI products that you use. And at the moment I am a strategic advisor and I'm also a professor. So I teach this to our future generations as well. Anna Nesterova: Well, that's awesome. So I hope you can teach me something today. I'm looking forward to it. Because you've worked across multiple sectors. You've worked in finance, entertainment, and healthcare. And all of your projects were related to AI. So what was the biggest differentiator for you when you were working on healthcare projects versus others? Joanne Skiles: Well, healthcare products are the fun ones, but they're also the Anna Nesterova: Okay. Joanne Skiles: scary ones because when you mess up, there's a person at the end of it. And I think that's like the biggest differentiation, right? and and I guess to go back is like I've been doing this for 16 years, AI being either at be pre AM AI is more machine learning, data engineering too than our sophistication, sophisticated AI. But If you think about it is entertainment, you I worked for a mouse. So entertainment, something went wrong. It was, no, my, you know, fast pass isn't working as well. and it so the stakes are low. And then you go into finance, which has regulation. Same with medical that has regulations, but the regulations for finance is that you get a fine. So if you mess up. You do something wrong, you get fined. you lose people's trust, which is not good. So that's a higher stake. And then we have medical where clinical decisions are being made and there's an actual person's life being affected. And I think that's like the biggest difference that I would say that we go with AI, medical AI, but also like the reason why I've left healthcare, health tech, and then went back to health tech because it was the more rewarding. in it what is what I would say. Yes. You can't get out. Anna Nesterova: So once you try it, it's you can't get out of it because you already feel the the impact. But I but what you said makes total sense. And I think even from my experience working with different companies, there is a bit of a gap in understanding of the high stakes between people who are building of the technical aspect of the tool or product versus the real world application and workflows. And I and I did see that gap myself. And I'd like to ask you, how do you change the way you build systems when you have those high stakes? Let's say you have a team of engineers and they come from, let's say, projects related to entertainment, and now you're working on a healthcare project. How do you really change that? Can you change that quick quickly? Joanne Skiles: Quickly, no, no, Anna Nesterova: Yeah. Joanne Skiles: no. I I and maybe maybe I'm a little old-fashioned in my beliefs in some ways, but I think you really do need to understand who is using your system, right? Anna Nesterova: Mm-hmm. Joanne Skiles: and also who's impacted by your system. So my Last world of health tech was at a radiology AI company and I shadowed radiologists. So like I followed radiologists along to learn how do they work. they say a lot of words I don't know. but but the point being is it's sitting there watching how they lurk at Actually, I don't have one with me anymore, but like I actually had a power mic, and I would practice doing dictation as well, reading radiology scripts to understand what they are actually doing. And I think that's the only way, really, to make a good system, be it AI or not, is to truly understand your industry that you're building for. And so that's not fast, right? And that's kind of the mentality that I've also but all my engineers as well is like you can't really build something, you can solve a problem, but you're not gonna solve the problem if you don't understand what everybody what what the actual user is doing. Anna Nesterova: Yeah. And I also think it's a difference in the mindset and the way how people think. I find that engineers in general, they're p problem solvers and they have point A, point B, and then find the solution. But sometimes in medicine, it's not the straightforward line or it's not the same for every patient. So there's such a variability in healthcare and in medicine that sometimes engineers maybe have a hard time on grasping on. And where where do you see the biggest challenges in trust when it comes to working with healthcare professionals? Joanne Skiles: you know, 'cause I am an engineer for everybody to understand. I am an engineer. I have a PhD in computer science, sixteen years of software engineering, and then that just happened, the side effect was AI in my life. and I like to joke about it being that way. I'm like, Yeah, the side effect was AI. and yes, my life is I see a problem, there's a solution. I don't understand why you can't figure out that problem. I've softened over the years. Maybe with children. That's kind of what made me soft soften over the years. And I've also realized things are complicated. So we go back to trust is in a way actually the human psyche. Like I'm not a psychologist, but but I think it's like that. Anna Nesterova: Yeah. Joanne Skiles: You know, you earn trust incrementally. So that relation That interaction that you have with a physician, clinician, or anybody, that those little moments build trust. and so the only way is if you're working alongside, asking questions, not acting like you're a no-doll, not being who I was when I was 22, that that would be really the most important because you're gonna build it by little interactions, the same way with AI. AI with clinicians. you you're going to get a result back from the AI, like let's say with a radiology image and it comes back with suggestion. And if it's wrong, it gets filed into this box of this is why AI is stupid. Right. And so the trust is always going to be ruined. So I I guess what I'm saying is trust is incremental. The only way that you're going build trust with engineering teams or with AI in general is by slowly working together and making sure that you understand that it's something that builds over time and that it's really easy to destroy trust. So Anna Nesterova: Yes. Joanne Skiles: the moment you do and and this is the this is the biggest Silicon Valley engineering issue that happens is for engineering wise we have this mentality of you built, you know, build fast, break fast, you know, just you ship it out, it's broken, ship something out and fix it. That doesn't work in healthcare. You can't ship something out and let it be broken because then the trust is completely destroyed. No one's going to trust your software because tomorrow you could break it. Anna Nesterova: Yes. And like you said, stakes are high. And that's why I think physicians and healthcare professionals they're so skeptical now. Cause I see it from my end of of this, how sometimes healthcare professionals are skeptical and what they say about new technology, how they say, Yes, the engineer is coming in and tells me that's working and then, you know, as you just mentioned, and then something is off, maybe there's a bug. But for the physician it means that, you know, he or she cannot make the proper decision. Joanne Skiles: Yeah, and it on the Yeah, and the problem is like on the engineering side, we look at it and we go, well, it's just a tiny little bug, can't you work around it? But that's Anna Nesterova: Mm-hmm. Joanne Skiles: for example, if you had a bug where radiologists come in and they need to read a report. Now you need to re-read a report for a stat case, someone who's in the ER. And you have a tiny little bug that caused where the stat case won't go to the next step. Anna Nesterova: Mm-hmm. Joanne Skiles: Okay, it's a tiny little bug. I'll you know what, why can't you just print a why can't you just print it and hand it to the person? No, like now now the radiologist goes, okay, well this application is garbage and you're telling me to print it to hand it to the person. Yeah, that's the easy step, but now I have a person who's in the ER that actively has a problem and you are blocking my flow. And that's where the trust just completely erodes. And so what could be a tiny little, tiny little issue is like, okay, it's just not going to the next step. But if you do these things, you can get it to the next step. As much as it seems simple on the engineering side, it's like, no, this messes the whole Anna Nesterova: Yeah. Joanne Skiles: flow. We have people that are bleeding in the e har. I mean, it that that's where it comes from is understanding that even though it seems really minor, it's not minor. Anna Nesterova: Yes. And I I love what you're saying because it's exactly what you know, wh how I feel sometimes when I see it in a hospital setting and what type of comments I received when I was working for companies with some sort of a AI or software solutions. And what do you find is the biggest misconception about AI when you talk to healthcare professionals Joanne Skiles: I guess it depends on it depends on like where they they stand, right? So Anna Nesterova: Mm-hmm. Yeah. Joanne Skiles: I think there are people that are looking for AI to do everything. And and they think AI can do everything, but it's a non-deterministic model, a non-deterministic meaning that it's just it's it's just a prediction of words that comes out. And so the answer can change every single time that you run through the model. And it's going to be good at doing administration work. It's not going to be this thing that's going to maybe one day. I'm not saying that Anna Nesterova: Mm-hmm. Joanne Skiles: not, I'm not saying never, but it's not going to do everybody's job. So as soon as people come in and think it's going to do everything, and then they start seeing that it can't do what their everything is. I d I guess that's the biggest misconception is like the misconception of expecting it to be Star Trek. I called that. The misconception is expecting it to be Star Trek when in reality it's like, no, let's let's calm down. It's it's it's a it's an admin. Yeah. Anna Nesterova: Mm. I f find myself thinking that way. I actually w the more I work with certain AI tools, the more I think that one day it will be able to do everything one way or another. can you tell me more about what you just said? Like where is that line that we see now? I know it may change, like you said, it may change in a couple of years. But where is that line? You said it's really good with administrative. tasks, but it's not really good with with what, because to me it seems like it's getting better and better and better in everything that I do. Joanne Skiles: It's it's getting better and better and better, but it's also only as good as the model that you're using. So Anna Nesterova: Okay. Joanne Skiles: for example, if I actually a good example is the tool that we're using right now. it has a speech to text so when we're talking it's going to transcribe our audio that audio is trained for conversational languages so like the conversational Anna Nesterova: Mm-hmm. Joanne Skiles: language and it knows if we started talking really sane medical terms it's not going to be able to it it will get the words wrong it'll start writing hippopotamus instead of writing I I'm not a radiologist so now my mind's blanking on the verbs but just imagine right so That is a model that's trained towards conversation. If you have another model that is trained with the data that is needed for radiology terms, it will do really well transcribing a radiologist that's dictating a report. It will be awful at transcribing a conversation. And I think that that's the example there is that it's going to depend on the model and the data that the model is trained on. Because I'm using the example of radiology because radiology has huge data set, and radiology has been using AI for longer than a lot of people have known that it's been using AI. Anna Nesterova: Yes. Joanne Skiles: I mean, we started they started experimenting in the 80s, so like so we have, you know, radiology using AI for a long time. So as we and they have a lot of data, right? But if there's areas that are more niche, which means that then there is less data, less knowledge about it, that's where the models will fail until we get more data. So I know it's a little bit of a vague-ish answer, but that's what I'm saying. It's like we we need more data. AI is only as good as the data that is being given Anna Nesterova: What you feed it. Joanne Skiles: and what you feed it. Yes. Anna Nesterova: Yeah. And I think that feeds very well to my next question. I really wanted to ask you about that bad AI. That term is now being used more and more. I see it on internet. I see people saying it. So what does it really mean, the bad AI? Is what you just said? It means that the the data that w it was placed into it was bad or Joanne Skiles: Yeah, I I don't there's exactly. there isn't bad AI. There is bad implementation. So the first step, which is my job, is I help people. You you you implement a system, you make sure you have the production guardrails around it, you make sure that you actually write a system that can handle the fact that AI models are non-deterministic. Again, meaning that they do not do the same thing every time. That means you have to write guardrails around it to handle the fact that it will go off the wall and hallucinate. That's just what it does. And you have to handle that. So there is bad implementation that does not do the correct guardrails to prevent the issue. And then there is bad AI, which means narrow models that are trained. So they're training or they're deployed without correct evaluations, or maybe not an actual use case. So you could overfit your model, which overfitting means that you train on you train only on Harry Potter books for writing. And therefore it has been overfitted to only write like you know. J basically write Harry Potter books. So it will do good on Harry Potter books, but all of a you say, Hey, can you you know, get some, you know, Lord of the Flies in here. It's gonna be like, No, no, they live happily ever after. That's right. Well, I guess there's not happily ever after in Harry Potter, but besides the point. that that's the point there. So there is poor models that have been trained poorly, and that can be bad AI. Most of the time, though, we're talking about bad implementation. Anna Nesterova: Okay. Joanne Skiles: That would be the bad engineering that gets built with it by not creating the proper ways to protect your model. Anna Nesterova: And from your expertise, from your experience, when do you consider a model being trained well enough to say that it's a good AI and good system? Let's say it's been trained in a very specialized area, but who decides if it's now it's trained well enough to go and let's say being implemented in a real world scenario in a hospital? Because you, like you said, the you will see hallucinations no matter what. But Do you measure it by percentage? Certain percentage of hallucinations Joanne Skiles: Yeah. Anna Nesterova: are fine, acceptable. Joanne Skiles: Yeah, you have a certain percentage that is around a allowed and actually depends on what the model is solving for, right? So Anna Nesterova: Okay. Joanne Skiles: like my the example that I go with a lot is if you have a model that screens for cancer. And Anna Nesterova: Mm-hmm. Joanne Skiles: let's say it's a cancer that occurs one out of a hundred times. So if you have a model and it's trained and it always says no cancer, well, that means it's right 99% of the time. Missed a hundred percent of the actual cancer diagnoses. So Anna Nesterova: Cancer. Yeah. Joanne Skiles: that means that it's actually failing 100% of time. So actually how the model is evaluated. So going back to the evaluation, which is when it sits in the lab and it's it's like I've been trained on all these use cases. Now here's a new set of cases. And let me see how well I perform. What's my recall? What's my accuracy? Those are very important. There's other metrics as well that have to be used when you're evaluating an AI model. Once it gets the threshold that matters, because in that example, accuracy, people say, but it's 99% accurate. Accuracy doesn't matter at all in that type of model. Anna Nesterova: Mm. Joanne Skiles: It's gonna recall rate is more important. So that is the threshold to say, okay, this is good enough to go out in the wild. That means that you need to actually show your lab results. So who gets to say that a model is ready? Well, I guess anybody could say their model is ready. Anna Nesterova: Mm-hmm. Joanne Skiles: Anyone could say it, but if you don't have the proof, then And I and I think that's actually something I would say cautionary for people going forward is to understand we're gonna have a lot more AI models coming out. And especially in a medical world, it's asking, okay, well, what is the metrics for this model before actually using it? We're gonna have to we're Anna Nesterova: Yeah. And who Joanne Skiles: gonna have to know a lot more about that than we needed to know before. Anna Nesterova: Yeah. It you would need to collaborate with healthcare professionals a lot to justify Joanne Skiles: Yes. Anna Nesterova: saying that the model is ready, I would imagine. Joanne Skiles: Yes. Yeah. Anna Nesterova: Which I mean makes sense, it's logical because otherwise who else who would double check on the reality Joanne Skiles: Mm. Anna Nesterova: of answers? This is all super fascinating and I think many people really don't think about AI this way. I still find that people think AI is the chat bot they're talking to or a chat bot that gives them answers. Even even for physicians, they use some systems right now, but I feel like it's more of a helper system to maybe schedule things or maybe put together like you said, put together the the transcribed material for them. But would you you what you're describing is much deeper and I think it will change. Actually, a lot of functions in medicine in the future. And if we I wanted to jump into the future talk with you, how do you envision the future with AI in healthcare? What do you think we will start seeing more and more? You already said we're gonna have a lot of new models coming out, but do you think we will see, I don't know, maybe a backlash and Maybe people say it's enough is enough. We have good enough models and we'll stick to them. Or you think people will keep pushing this? I I don't know, but I wonder it. What do you think? Joanne Skiles: Yeah. I'll get like philosophical in a way, some so if we live where AI stays a little bit of a magic black box, I don't see the adoption of it really going forward because I think What makes a good doctor is kind of being a skeptic and kind of like looking for answers in the truth. So I do think that AI needs to be a little bit more transparent in the decisions that it is making. And that transparency is super important and explaining the reasoning of why they think a certain diagnosis is gonna happen. If a model isn't going to create that when it gets it wrong, it's gonna lose trust very quickly with doctors. So one is how the The AI models are gonna actually act. And that I that maybe that's why I also keep going back into more research because I want to kind of influence that world and be a little more transparent. The other is computing power. Like right now, I mean, I don't think it'll be any shock to any of the listeners. Is like everybody's betting on AI, they're building data centers that are. insane and I can't buy a new graphics card for my computer because the prices are insane. Like yeah that that's the depressing part. It's ruined my side hobby of building computers. So until we come up with a way to make that cheaper. Cheaper either with money or cheaper with the fact of like we're not having to use as much resources or we improve our chips our our computer chips until we figure that out there is going to be a a a bar that ai cannot run until we fix that. So that that is a physical that is a physics problem that we have to solve as well. So I think between those two, that's actually going to choose where our trajectory for AI is going to be. So that's my Anna Nesterova: Mm-hmm. Joanne Skiles: theory at this point is if we can fix our physics problem and we can fix our skeptic problem, you know, maybe one day we will have my joking Star Trek where we could just use a scanner and all of a sudden we know what's wrong with us. Otherwise, we're gonna be stuck probably where we're at right now for a little bit with a Anna Nesterova: Mm-hmm. Joanne Skiles: very niche models for specific problems. Anna Nesterova: Yeah, that makes sense. Can we go back to transparency for a second? You said that we should find a way to build transparency and be more transparent about AI models. I, as a skeptic myself, can see a problem and maybe a conflict of interests for a company to even admit they have an issue because they really want to raise funds because they are right now under high pressure of raising funds, being the first ones, going to the market, proving that the system is working, their model is working. I find a bit of a Problem with that because companies may hide certain things, may not be as transparent as they need to under that pressure. How do you think? Well, maybe you know how it looks like in real world. Have you seen that? Or do you have the same type of let's call it a fear? Because I do have a fear, because again in healthcare you have very high stakes. think it's giving Joanne Skiles: Yes. am afraid. I am I am in a way afraid if we keep hiding things. Because really they there is a way to find the balance between giving explainability, transparency to a model without giving away the secret sauce of your model. Like there is a way to do it. showing reasoning, showing how something's thinking, for example. go back I always go back to imaging because I feel imaging is the best way to like imagine something anyways. Is if you have a a a scan of a chess and there's a a CAM model that actually shows you a heat map to show where why the it determined where pneumonia was in the chess. That exists today. So you can utilize tools that do exist today, give the heat map and If you feel that you need your model to stay locked down, then there you go. You can show proof, you can show the thinking without having to give away your models as secret sauce. But yeah, there's a little bit of trust that needs to happen, but I think there's trust that needs to happen with anything, like it comes back to the fact that. And I'll be a little more cynical in this sense, is that anybody could build something if you're building like a software as a service product, anybody can build it today with with enough tokens from Claude. That they can build it. So what stops them from taking over whatever system that you work on, whatever company that you're at? It's your reputation and your trust that others have in you. It's not just the person who comes up and goes, Yeah, got I got a new Facebook. Everybody join my Facebook. Anybody can build Facebook. Anna Nesterova: Mm-hmm. Joanne Skiles: So what what stops it from happening? And Anna Nesterova: Yeah. Joanne Skiles: that is my answer to those companies instead, is that if anybody could build your model, then what stops people from going to your model? Are you maybe evaluate that part of your company a little bit? Anna Nesterova: Yeah. Yeah. And I also think the ethics component is very important. Whoever keeps the ethical standard high, I hope long Joanne Skiles: Yes. Anna Nesterova: term will win. That's my philosophy. On the personal level, on the technology level, I think you have to keep your ethical standards high. Joanne Skiles: Y yes, yeah. And you have to keep it. And I think part of it is if us as consumers start asking for proof, that also is going to you know, the companies that choose to give it, those are the ones that people are gonna go with. They're not gonna go with the ones that hide it. Anna Nesterova: That's true. But again in healthcare it's tough because consumer at the end of the day, if we're looking at patients, very often they have no idea Yes. Yes, yes, yeah. Joanne Skiles: Yeah, so our consumer in this case would be the doctors. The doctors, the clinicians, the admins of the hospitals need to say, Hey, show me your evaluation re metrics. What are your metrics on this model? I'm not interested until Anna Nesterova: Mm-hmm. Joanne Skiles: I see that. Anna Nesterova: Yeah. I have an interesting question for you that came up about ethics. I know that there is a growing demand for specialists in AI ethics in different companies. In your advisory work, do you suggest companies to have someone on board who controls the ethics of AI? Or maybe you call those people differently. So what what's your suggestion for that? Joanne Skiles: I think I mean there's nothing wrong with having someone. I've never actually thought about the the having like a role around the ethics. I mean that does exist at every company, right? Your your HR ethics world. So yeah, I guess having someone specialized to there makes sense. I think it's also something that people in general, we're not gonna get away from AI. As much as we wanna it. there are many people that I meet that want absolutely nothing to do with AI is what I'm saying. And like as much as you want to say that, it's n it's not gonna happen is here for now and unless, you know, the real price of AI is so insane that everything just disappears and then we'll have another problem because we'll have a class based world where some people have AI and some people don't. Putting that to the side. But To be well versed in AI, you should probably also be a little well versed in ethics. So I think everybody has the responsibility to learn the ethics. So I didn't think about that role until you mentioned it right now, but it does make sense for every company probably to have that role. Anna Nesterova: And how do you see the over reliance on AI in the future when it comes to medical decisions? Do you think we will get to that point? Is there a bit of a danger in over reliance? Joanne Skiles: there's a huge danger in over reliance, right? Because if you start offload I mean, it's the same problem with in general for AI is people people Some people use AI to like draft every single comment that they write on LinkedIn. I'm like, really, can you just write one comment yourself? Like the fact of offloading your thinking. And that is a huge problem that can happen in the medical world is that then we offload our thinking onto the AI. Then the question comes, when the diagnosis is wrong, when someone's life is affected. Who's at fault? You're not gonna like the answer, but if you're the doctor that signed it, you're at fault. So no, do not offload your thinking to AI. That would be the huge danger of over-reliance. Anna Nesterova: let's move on to something a little bit more fun. So it's what have surprised you in the I world recently. Can you share with me something from your world that surprised you? Joanne Skiles: I think what has been really you what? Actually, no, what did surprise me very recently was Fable. so Claude coming out with Fable and then the US government deciding to regulate it. And actually, I know that's like I was trying to make it lighthearted, but I'm like, it's still not lighthearted. But I think actually that's my surprise to actually see some regulation coming through, which means that we are thinking about the impact that models can have beyond just, this is cool and it's a fun toy. So I I do kind of enjoy that that it actually is being taken a little bit more seriously. I thought it was gonna take longer for for governments to actually take it seriously. So Anna Nesterova: So you find it's a good thing. Joanne Skiles: I it is a good thing. It's a good thing we need to start thinking about these things. I don't know if this one was necessary or I'll leave my opinions on a table there, but it is something that we do need to start having conversations about. And that's where I think it's a good thing. Anna Nesterova: Yeah. And if you would leave my audience with an advice, and my audience is mostly interested in healthcare and med tech, what would be your advice on skills that people can work on right now to stay, ahead or on time with current developments in AI? Joanne Skiles: Yeah. I would say don't be afraid of AI. So go and use it in low stake areas. So go in and play around with Claude or Chat GPT. Use it to help you draft, draft an email, draft a response, make a help you clean up a report. Use it lightly. and see How well does it do when you give it something very vague versus using your expertise because you do have expertise and being detailed with it? And you'll realize very quickly, I'm giv giving you the answer, you'll realize very quickly the more detail you give it, the better it is, and the more it sounds like you. that way to understand that it's not magic. There's nothing about it that is really overall complicated. It is just, I'll be a little nerdy, is just angles and predicting what the next word is. So it's actually just a lot of hidden vector math underneath it. whether or not you guys like that when you were in college, but it's a lot of vector math. And I think that's what's really important right now. Don't completely ignore yourself from from from AI. Play around it. Pretend to be an engineer for two seconds. Maybe, maybe make it build something for you and just see: is it, is it kind of decent? Okay. Don't share your local house with someone. They can't see it. But that would be my suggestion. Go and have fun with tech. Make mistakes in low-stake environments. That is the only way you're gonna learn anything. Anna Nesterova: I have a bit of a fear, and maybe I voice it out for many people. When I put details of me and the AI starts talking like me, like you said, it actually scares me that I'm giving away something proprietary to me, to an external system. And I'm sure there are people like me feeling the same way. It scares me. It Joanne Skiles: Yeah. I mean there Anna Nesterova: doesn't give me I mean, I I I'm amazed, but it scares me. Joanne Skiles: Yeah. Anna Nesterova: So what d should I do? Joanne Skiles: You know, I think I think that's asking a little bit about what you what what do you value a lot on privacy? Because now we're going into the privacy question here. And I'll go like everything that I do on the other side of that is that I have a I have a podcast editing tool that is runs only locally to edit podcasts. So that means it doesn't go anywhere. You could have local AI models that only run on your computer. It'll be a little bit slower, but it it will run. So it kind of, I guess that's the point where you have to say to yourself, is how much am I okay with this data going somewhere I don't know? And I mean it's the same way of post something on LinkedIn. what happens to my posts on LinkedIn? Well they're gonna train on your post too. So where are you okay that this information is gonna be taken by someone else? Because I mean that's the only I it's the sad truth. The sad truth is that as soon as you put it in somewhere that you don't own, you don't own where it goes. That's why, you know, like You hear the stories of like the the family that goes to Switzerland and they see that they're on a billboard over there because they posted a picture on Instagram, right? Yeah. The world that Anna Nesterova: Yeah. Yeah. Joanne Skiles: so that's what be very careful of your data. I think it's a good fear. I think it's a good fear, and it's just asking, what are you okay with leaving your control? Anna Nesterova: Yeah. Thank you so much, Joanne. Joanne Skiles: Yeah, thank you.