Episode 4 - AI, Diet and Fitness
Transcription
Andy B: 0:01
So question for you guys do you have any New Year's resolutions for 2024?
Kiran V: 0:05
I do. I guess I always kind of have a one New Year's resolution, which is to travel. A goal I have is to go to a new country every year. So we're going to Thailand in January, so kick off the new year with the honeymoon trip. I'm excited.
Joe C: 0:25
Nice. I love that. I might steal that, but I have one similar to actually start going to, or try to go to all the national parks and an explicit goal to go to one this year, one that I haven't been to before.
Andy B: 0:38
This year I had set a goal of just exercising 100 times, just 100 times, just like approximately once every 3.5 days do something that counts as exercise real wishy-washy, and I crushed it. They did it like 111 times or something. So I think I'm going to do that again because it was a really low pressure, like do whatever, just do it 100 times. And my little planner I use a Hobonichi planner has a sheet called the 100 where you just get to count something 100 times. I was like moving my body 100 times.
Joe C: 1:10
Nice, nice, that's a lot. I think 100 times.
Andy B: 1:35
Welcome back to AI FYI everybody. My name is Andy and we're here with a brand new episode for the new year. How to Lose 50 Pounds with ChatGPT. What are we going to be covering, Kiran?
Kiran V: 1:45
This is one of our two-part feature episode. Last time we talked about mental health and AI, and today we're going to be talking about fitness, exercise and diet and how AI has been playing a part in our applications.
Joe C: 2:05
Maybe we start by talking about dieting a little bit?
Andy B: 2:08
Yeah. So when we started researching this episode, Kiran and I split up diet and fitness because obviously there's a lot of people with weight loss goals in the new year. As the girl of the three of us, I feel like I should have covered dieting because, the Lord knows, I've been on my share of diets it's just ruined women's lives as far as I can tell. So I started researching about AI diet apps because I know that I've heard some radio app ads for the app Loom saying they have AI generated weight loss plans and I was like I thought Loom's goal was to get everybody to eat like 1200 calories. And anyway, before I get too into it, I just want to go out there and make some statements. Tell me if you disagree, but I'm a big fan of the idea that dieting doesn't work. There is no peer-reviewed data that restricting calories and trying to control the shape of your body is effective at all in medium to long term, and you are the size that you are and I want everybody to like themselves how they are.
Joe C: 3:12
Andy, what's the craziest diet you've tried?
Andy B: 3:16
I can't logically or responsibly share it on a podcast. Absolutely not.
Joe C: 3:20
Fair enough.
Andy B: 3:21
But I was a feral college girl. That's all you need to know. Yeah, so the basic premise of if people go dieting is you eat less and then you are smaller. Right, it's calories and calories out is what people say, even though the data is showing that that's not even quite true. But a lot of people want to become physically smaller and occupy less volume, and we put a lot of value into that as society, for better or for worse. So there's a lot of reasons and a lot of money to be made selling people diets. All right, so like think of your brands, like Weight Watchers, Atkins. A lot of people have cashed in a lot trying to sell people's strategies for getting smaller and, of course, the next generation of these is all AI based.
Kiran V: 4:18
So is dieting primarily to lose weight, like a function of weight, not like my blood sugar is too high or I have high blood pressure, or like like allergies.
Andy B: 4:32
So obviously people need to watch what they're eating for a variety of reasons and like, but the main money and there's apps out there. Like my, one of my friends has a lot of GI problems and she uses a food tracking app what it doesn't track calories at all, it just tracks her symptoms, like did she have a stomach ache, did she have nausea, etc. According to what food she was eating, because it's trying to deduce, like the relationships. And you know I have family with heart disease who have to eat low sodium diets and so they use some trackers. But a lot of the apps that have made a lot of money and that are like on so many people's phones are specifically for counting the volume of food that you eat. It's like nutrient value as pertains to your bodily size, usually for cutting, losing weight, sometimes for gains, gaining weight, and that's kind of where I'm seeing the new generation of apps come out. I made notes for sort of three things. One is I thought it was going to be like the good, the bad, the buggy, but really I don't know what categories there and I wanted to share with you guys and get your takes. I don't know if I trust my initial take on all this stuff. So I go to TikTok and I type in AI diet and I just start browsing to see what are the kiddos talking about, what's happening in the world, and the first thing I found that I thought was really fascinating was people using AI photo editing apps to Photoshop themselves at different body weights as motivation for their weight loss plans.
Joe C: 6:08
That's very interesting. That's almost like when we were talking about art therapy in our other episode. It's sort of like vision boarding in a way. What did you find out about that practice?
Andy B: 6:22
So there's an app called Remini and they mostly market themselves as like they will give you professional headshots, for better or for worse. One of the things they do whenever they make professional headshots of you is they make you smaller. I think that says a lot about how these AIs are trained and like what that means. I would be really annoyed if that happened to me. But people are going in and basically eyeballing what they think they would look like at certain weights and then they're even posting that on TikTok and they're saying like this is why they're treating it as their, why how good they will look. And I don't want to poop on anybody who's dieting. Beauty is not everything, but it's not nothing either. It's your body. Do whatever you want. But I think that was just a really creative use of the technology and I'm sure that the people who made Remini probably weren't thinking like oh, I'm not set out a diet motivation and weight loss motivation app. But that seems to be one of the leading use cases when I was searching this of their application was people trying on different clothes and trying on different sizes for themselves as motivation. So I thought that was relatively harmless. I wish that the app didn't make people slightly lighter skinned and slightly skinnier every time I made them professional. I think, that's you know, hot bullshit.
Joe C: 7:46
It is very interesting that this use case or this way to use AI arose from something that like didn't intend this to be its goal, and it speaks to like what we really want for ourselves, like in our hearts or like inside of us. Like that so many people just want to see themselves skinnier, that, like they find a way to do it by using this tool that isn't meant to do that, assuming that's why people are using it. I mean, we're making a couple assumptions here, but it's fascinating.
Kiran V: 8:20
I mean, they have like all those filters on, like every you know camera app, right, it's like oh, the skinny filter, the better skin filter the whatever like I think it's, it's so deeply ingrained into our existence at this point that it's. I mean, I find it kind of sad and disappointing that we put so much emphasis on that. But you know, I think everyone does have their own motivations to do things and it's like you know, do what you want to do. But I think it is scary that so much is being controlled now by features and applications and not like my existence in the real world.
Andy B: 9:03
There's also a lot of like when you record videos like, for example, zoom. If you're on, zoom has a beauty filter. Why zoom has a beauty filter, I don't know. I recently had a dermatological procedure that left little burns on my face. I turned that beauty filter all the way up to 11, like I was. I needed it, but that's kind of weird. And then you know you've got makeup influencers on TikTok trying to sell you foundation and telling you how it changes the texture of your skin, but they've got a filter on that changes the texture of their skin. So lots of AI applications as pertains to the human body that are off topic for this episode. But I thought it was really interesting. That was some of the first things you'll pop up. If you type in AI weight loss, you'll see people basically photoshopping themselves Is that as a verb, not a brand name into different sized bodies. To see what would happen Brings the idea of catfishing to a whole new level. That's all I'll say about that. So I thought that was, like I said, fairly harmless, Would you agree? eople putting in their own picture?
Kiran V: 10:07
I mean, I think from a data privacy that's what you're like, that's totally fine I think from a mental health perspective potentially a little different.
Joe C: 10:16
I agree. We hope for a world where people are fine with the way they look, but it's not the case.
Andy B: 10:26
Yeah, even if not fine with the way you look, don't be obsessive over it Like that's there's lots of stuff, there's 8 billion of us and there's about 8 billion problems on earth. Pick something, go fix it. You know, don't be so self centered. But that's my Eastern European coming out. I'm sorry. So I thought that was interesting. Go ahead.
Joe C: 10:47
Also, the design perspective. Like designers always say, the user's never the problem, it's never the user's fault, and this is an interesting like perspective to put on. This is like there are folks out there making products that are solving a user's problem, even though we wish the user felt differently. You know, I don't know, maybe there's a solution in not trying to fix the people and what they desire for themselves, but instead make technology that adapts to them.
Kiran V: 11:18
Yeah, that's like a much harder app to design and deliver, because then you have to like actually understand, like at a much deeper level, like what people are trying to really do.
Andy B: 11:31
But I don't think any app can change culture right? The reason we people want to be skinny is because we put a lot of value into being skinny. White people who are skinny have better economic outcomes. They've got better social outcomes. Like you no app, no AI can change that until we as a culture change that.
Joe C: 11:49
That's a great point. Is it inherent in us?
Andy B: 11:56
it's not to be more beautiful. I'm gonna contest that. But to fit a particular idea that we've shaped.
Kiran V: 12:05
What if? What if? There was a video calling service application that randomized your, your image, for the people on the screen, so I would look totally different and it was like the voice would be different, the look would be different, and every time it was just totally random and it was explicit, right, like you knew that this isn't what those people look like, but you're just like we have things that we need to get done and we're just doing those things.
Andy B: 12:44
Blind does this. There's the one where like people get put into like animal makeup and they have to date or monster makeup? it also seems like a black black mirror episode.
Joe C: 12:56
Yeah or Shallow Hal. Inner beauty
Andy B: 13:03
Anyway, point being, we like you just how you are people. But if you want to change something and you want to see what it would look like turns out, you can just use a headshot app and make yourself look different and see what it would look like.
Joe C: 13:20
Or Ozempic. Let's not get into that.
Andy B: 13:21
That's catty, that's so catty Joe. Everybody you're meeting Joe as he is when he's not on the podcast. So the next thing that I category that I want to get into that I actually do all the time and I think a lot of people do, and I think it's really cool use of the tech. But I want to call out some warnings. I call this the AI meal plan. So if you talk, go into tiktok or go to YouTube, you say how can I use chat GPT to lose weight or whatever, almost everyone will tell you like go get chat GPT to make you a custom meal plan to meet your goals. And I actually do this pretty often because I think the brainstorming is really awesome. Like I can go in there and I can say, like here's what I have in my fridge and I don't want to grocery shop until Saturday and like, can you just like tell me what some general things I could make with what's left? You know, like I got a half of a cucumber and two cups of bulgur. Give me an idea. Like so I think that's awesome. But I want to call something out. First before I get into that. Have you guys ever done this? Have you done any meal planning or meal prep or recipe stuff with chat GPT?
Kiran V: 14:35
Not much. Not with Chat GPT.
Joe C: 14:41
I want to go do that now. Maybe not even stuff in my fridge, but just like I'm about to go grocery shopping, I want to eat healthy this week. What do I buy?
Andy B: 14:50
Yeah, on the platform that I've been building for Verta AI, we have these things called starter kits so you can like build an LLM app, start to finish and build your own custom streamlet app. And one of the starter kits we have for people to test this build a meal planner. And we let people like, basically like pre, we pre write them prompts where they can input like what the size of your family, what are the dietary restrictions, what your cuisine preferences, and then they can pick which LLM they want to serve and modify the prompt slightly for themselves To make themselves a little app where they can put in those variables and it spits out a meal plan. So if you want to build your own meal plan UI, check out my little company, you can find it on our website but highly recommend it, it's really fun. My public service announcement is don't trust the math. So let's say you are trying to do a really calorie restricted diet or macro restricted diet. You're like you know you want to eat x-many grams of protein. Unless you explicitly tell the LLM to do so, it's just gonna start writing and at the end if you add up the numbers, it's not gonna add up to the total that you want. I actually I forget what I was doing. It wasn't meal planning, but it was something else where I was trying to get it to like make 200 of something and it kept giving me like 140 or 260 when you actually add it up. So you know, use this as a brainstorming tool, use this as an ideation tool, but just know that, like, don't trust it to get the math right. If you want to make sure that's doing the math, go ahead.
Joe C: 16:27
Yeah, just blanket rule: LLMs, generative AI; They don't know how to do math.
Kiran V: 16:34
Yeah, and this is actually an example of of a limitation, of why we you know why it's not actually general AI, right, because it it might understand, like the concept of numbers and like, oh, this is a number, right, I can categorize. But then to contextualize that, like there's numbers going on here and I need to use, like actual math that has rules to determine the outcome of whatever I'm trying to add or subtract or whatever it doesn't have the that ability to make that connection and contextualize that like this is a math problem and there's simple rules that I need to follow. It's just like trying to generate text, right, and like create something based on what you've asked it.
Andy B: 17:19
So one of the ways you could try to play with this is to make it more accurate for yourself is something called chain of thought prompting. Joe, Kiran, do you guys need me to explain this? Should I explain this?
Joe C: 17:33
I never heard of it. Yeah, what is it?
Andy B: 17:35
So chain of thought prompting is basically you can sometimes get LLMs like chats, GPT, large language models, to perform better by telling it how you would solve that problem. Essentially, so you can say, like I'm making a meal plan, first look at the ingredients. Second, break it into three meals over seven days. Third, add up the calories of each item and give me the total, then change the plan to make sure it's enough calories and then re add up the calories and give me the new total. And by telling it to take it step by step, it'll actually when It's a bit more deterministic, it's more accurate. It still gets it wrong sometimes, but you can get it to be more accurate.
Joe C: 18:23
Yeah, some LLMs out there have performance improvements, even if you just say like, what it is like.
Andy B: 18:41
Yeah, and for this particular just use case, like some people were also doing this with exercise plans and I did this with a running plan for me earlier this year because I used to run a lot and then I broke my foot and I was like trying to get back in the running a little bit but I wanted to like go even slower than any reasonable person would ever write up a detailed plan. I was like split, getting back to a mile into like the tiniest little baby steps, and I could keep asking Chat GPT like don't make smaller steps, make smaller steps and get something really discreet. So I think it's a super awesome tool. I mean it's you can get access to GPT 3.5 for free. I believe. this can give you something customized that but maybe have been unavailable to you. That feels like you got a custom plan from nutritionist or a trainer or a coach and that's something you can go back and modify and you can fuss with it however much you want. So I think that's really awesome. But again, I just caution people do not trust the math.
Kiran V: 19:48
So I actually have an example that I just did also if I can share my screen. so if you're watching the video this is the first time I'm trying this, we'll see you can see this- for everyone we can talk through it. So yeah, so this is a prompt I created on chat GPT. If you're looking at my left screen, ignore it. I only come to chat GPT to try random things that I thought of.
Andy B: 20:17
you shouldn't have called it out? Well, you will blur that out in the video. It says some wild stuff in there, everybody. There's some some really truly heinous things,
Kiran V: 20:28
Cherry screen
Andy B: 20:32
Okay, I'm gonna read Kieran's prompt. Give me a running and diet plan. Use the following criteria, I currently run two miles maximum. I want to be able to run 13 miles. I am such and such pounds. I want to gain such and such pounds. I want to improve my VO2 max. I cannot eat selfish. I did not know that about you. Give me a running and diet plan over three months.
Kiran V: 20:56
I mean. So I've actually done a lot of like. I've taken exercise classes, weightlifting classes, so I've done like a fair amount of fit personal fitness. So, while I'm not an expert, I will say like I have a fairly decent understanding of my body and what I can do and how it responds to different training programs. And, looking at this, for me based on the inputs that I gave, it is like a pretty solid Workout plan, right, like it's like week one to four, building endurance, and has, like each day of those weeks, you know what do you do Week five to eight, increasing mileage. Week nine to twelve, preparing for 13 miles, right. So this is me putting a couple you know criteria in trying to get a prompt or trying to get a workout plan to how do I get to run 13 miles at a time and Does that's?
Joe C: 21:52
Does that say run 50 miles while eating a bucket of shrimp?
Kiran V: 21:57
No, I'm just, I'm just kidding. And so. So yeah, I mean it's like pretty solid and if I did this plan rigorously, I am pretty confident I would be able to run 13 miles.
Andy B: 22:12
And what I really like about this is not like an article right Like, for example, it tells you long run, a long run. But right now you could send another message being like what is a long run? What do I need to know to be successful at long runs? And it'll just keep answering your questions forever. It's patient.
Kiran V: 22:27
Yeah.
Andy B: 22:28
Yeah, there's a.
Kiran V: 22:29
Diet plan and it's like you know, I indicated that I want to gain weight, which is actually something that I'm. That I saw in another app, which is what's interesting, but I want to gain weight. So it talks about, you know, having a caloric surplus. You know, protein, carbs, etc. Again, I'm sure I could prompt this To give me a day-by-day breakdown. So it gives me a sample meal plan and it says like breakfast scrambled eggs, greek yogurt with berries, lunch, grilled chicken, you know, some nuts, dinner salmon, sweet potatoes and then some snacks really white people food. It is right, but again, like at, because I'm talking to this thing and it I'm prompting it. I could say you know, give me a diet of Mexican food, or give me a diet with more chicken and less beef. Like you can go as deep as you want, you know, with this, which is really cool and there's so many thoughts I have on like diet and exercise and applications and AI right now, but like this is essentially like has solved Whatever you problem you might need when it comes like diet and fitness.
Andy B: 23:42
Well, hold on until chat GPT can meet me at the field at five, and I think I don't want to disappoint chat GPT. I better get out of bed and jog on over there.
Kiran V: 23:54
Exactly the problems I have so that is the buggy here, which is really interesting, at least in my research. Humans are the problem to our own fitness and diet and help, not the tools that we have at our disposal.
Joe C: 24:11
I think that's. I think that's mostly true. But there's a huge barrier to get a human to help you with training and dieting and that's cost. Like I would love a nutritionist, I'm not gonna, I don't have the money right now to hire one, but now that I know I can have chat GPT to like make me a meal plan, I'm one more person who has access to that kind of information. But yeah, maybe that means it won't put actual nutritionist out of business, which is great.
Andy B: 24:41
What I also really liked is I made myself this like exercise plan for getting back into yawning running. But I told it like I really don't want to run more than one or two days a week because I Got my yoga practice and I got other stuff and like time is limited and it's not. I don't know like I would feel kind of embarrassed trying to tell like a running coach that I want to get serious about my running again but also I only want to run one day a week. But I don't feel that shame telling that to chat GPT and there's memory. If you have it set up in open AI, you can like go back to the conversation. So I would like go and like do an exercise or something would not go to plan and I'd say like hey, like you told me to do my yoga class yesterday, like I'd scheduled, but actually didn't go to the yoga class yesterday. I'm going today and it would like reshuffle it and Like again it's, it's patient, it's on demand, it's free, it has basic memory when I did do something. It would like cheer me on a little bit because I told it to be my cheerleader. So it send me like cute little messages. I Found it really useful. Very cool I still use it yeah.
Kiran V: 25:41
Yeah, yeah and so Well, a couple things. The first thing I do want to talk to is so I used this. I prompted chat Gbt with these inputs based off of I was trying out another app called jogger.
Andy B: 25:57
Let me see how it's spelled J, oh and I have one more weight loss diet thing that's nuts that I want to get into, but I accidentally started talking about exercise.
Kiran V: 26:05
Sorry, okay, let's do that first then, and then we'll come back to this.
Andy B: 26:09
Okay, so what most people think about when they think about dieting is Restricting calories, eating less, losing weight, and there's a couple big names in apps in this world. Right, there's an app called my fitness pal, which is owned by somebody. Now there's there's an app called I don't know. They all have kind of the exact same sounding names, that but point being, you meticulously track everything you eat and Then it adds up for you how many calories are in it and other nutrients that compose those calories and that tells you you've eaten too much or too little, against the plan of you gaining or losing weight.
Joe C: 26:50
I used my fitness pal. Yeah, I liked it, but the hardest part of sticking with it was the manual labor of measuring things out. In putting them Like that's sort of the problem I need solving is like Can I just can? It know what I take out of my fridge and like put it in there.
Kiran V: 27:07
Yeah so I Track it for and I got I paid for a year subscription Because I was like I'm gonna do this. I paid for a year subscription and then I got like six days into it and I was like this is so annoying, I'm not gonna do this anymore.
Andy B: 27:24
Yeah, I fully. That's too much for me at these. This point in my life, I'm over counting calories but, there's a whole bunch of apps coming out right now and I'm gonna deep dive into one of them that basically try to solve the problem you guys have, which is you want to do this but, like who's gonna like, scan the barcode and weigh everything to get an accurate measure right.
Joe C: 27:48
If you're lucky that there is a barcode, that was like. The only thing that made it like Significantly easier is barcode scanning, yeah.
Andy B: 27:56
So there's these apps coming out when you search AI, dieting and tick tock or other platforms. That the whole premise of them is you make, you play the food, you take a picture, you eat it and your calories attract Cool very cool right. Yeah, I'm about to roast this poor little startup. I'm sorry, folks, if you're listening to this, no beef, no shade, just Holding up a mirror to you. The one that I went going deep on because they're paying for advertisement on tiktok is a company called SnapCalorie (https://www.snapcalorie.com/) and the first thing I did is I went to their website, which is like almost nothing on it because they're a mobile first company. They got a blanket website for SEO and the first thing I wrote in my notes is the fuck is wrong with their blog. Because if you go on there, I first went in there trying to be like how are they doing this? Tell me more about the computer vision. I want to know about the performance of your algorithms. Right, like, how accurate are you actually calorie counting? And Instead, the first article I found on their blog that caught my eye was can you take melatonin with Nyquil (https://www.snapcalorie.com/blog/can-you-take-melatonin-with-nyquil/)? And I was like what does that have to do with calorie counting? And it's just like SEO, optimization and Okay, whatever. I Was my mistake for thinking that the blog would have anything useful. I Get it. So instead I go over to the Apple store description of their posts and they got this big, big description In the description section of Apple and at the top they have a quote saying Stop obsessing over every calorie and shift your focus to your overall health goals. That's a neat idea, not the point of calorie counting, which is famous for getting people obsessive right and like, leading to eating disorders. So, like you, just scroll down to the bottom of their very long description and it says this app is not for anyone with an active eating disorder, an eating disorder history or an eating disorder family history. This app is also not for anyone with body dysmorphia or any predispositions to body dysmorphia and our body dysmorphic disorders. Isn't this every single person on earth?
Kiran V: 30:16
This app is only for perfect people, and then why do you need to count your calories?
Andy B: 30:22
Stop obsessing. So here's this app that will give you the most precise, most accurate calorie count. Are you kidding me?
Joe C: 30:31
I'm not. I'm on board if it works. I think the disclaimer is weird. It's like why I even mentioned that, probably because of extreme cases, but I Think those extreme cases are really common, yeah, really common, yeah, it's also like ranges, right, like body dysmorphia could range from like my feet are too small to like you know some way more intense things.
Kiran V: 30:55
So yeah, I don't know.
Joe C: 30:58
I don't know either, and it does like your. Is calorie counting the way to go or you should? Should you just forget about it? And if that's the case, then like this app isn't really helping people.
Andy B: 31:11
Yeah, I mean again the Shout out to the folks at one of my favorite podcast maintenance phase Aubrey and Michael, if you guys could become our podcast mentors, fairy godmothers, please adopt us. We would love that. But yeah, there's no data that dieting works. People still do it. You can do whatever you want, it's your body, but we know we're data oriented people.
Joe C: 31:36
It doesn't work you know I had a doctor. He told me that I should aim to have like a third of my plate be green, which was interesting advice because it was very broad and like just sort of a general like just make sure third of the plate is green with every meal. Maybe this app would be more useful with something like that.
Andy B: 32:00
Well, you know, like one of my health things is, I like to eat a lot of vegetables and if you've been to my dinner parties, you'll notice that, like, I try to make sure that, like, more than half of the plate is mostly plants To me. That's seen. Anyway, you know what we don't need me to get on a Little podium and talk about. Just eat a lot of vegetables and you won't have room for chicken, but that's, you know, that's not useful. Anyway, so their inscription, this company that I'm going to continue to kind of roast Snap calorie, is your smart AI nutritionist. Ask questions and get immediate coaching and feedback from our AI model. Use our photo calorie counter to log a meal in seconds and get an accurate estimate of calories, macros and micronutrients. We leverage modern AI along with human supervision to make expert level guidance accessible to everyone. So highlights, low lights. You guys probably immediately thought of the same ones that I know, the same ones that I did.
Kiran V: 32:59
Well, it's not actual AI, it's just humans labeling data as it?
Joe C: 33:06
oh, possibly. Yeah, my two thoughts on it are okay. Well, first, andy, I noticed on their website that they claim it's twice the accuracy of someone who's doing calorie, calorie counting, counting manually. So that's interesting. I'd love to see the numbers. The second thing is there's good. There's a bowl of soup on their website and I'm like, how do you know? I Don't think computer vision, that is that good that you can't tell me there isn't like half a stick of butter in the soup versus no butter, like it's soup.
Andy B: 33:42
Yeah, you know, like they don't know what's in the soup.
Kiran V: 33:46
Anyways, unless they watch, you make the soup.
Joe C: 33:49
Well, that's true.
Andy B: 33:53
So, like I thought, the fact that they say we leverage modern AI along with human supervision to make expert level guidance, like they're doing, human in the loop, which we're big fans of, right, we we think that humans regularly augmenting AI, q-aing it, it's good.
Kiran V: 34:08
And.
Joe C: 34:09
I they. Well, we should have mentioned. We worked with customers that had a product that you could put food into and it would detect the food and then understand how long to cook it for. Is that right, yeah? The oven people, the oven people, and that's similar like they were confident they could look at a plate and understand well enough what's on it to know how long to cook it.
Kiran V: 34:33
Did they ever actually get anywhere with that?
Andy B: 34:35
Yeah, you can buy one. Oh, I don't know how well the AI is, but I'm sure I can tell you if it's burning. So they also. I was digging around and their stuff's not. They're fairly newly launched. I found their crunch our country's. Yeah, crunch, crunch Tech crunch, tech crunch articles and stuff. So they launched in June of this year 2023, but Basically they looks like they every single prediction. They also have human verified currently, which I think is the responsible way to launch this. But even for their whole premise is that people are bad at eyeballing how many calories are in something, but no one's gonna have the patience to weigh everything. And then that nutritionists are better than normal people and that the AI model is better than the nutritionists, and they said that they have cracked it and it's all research-based. So I was like let's see the receipts. Please Show me the research. I know how to search archive. They have a single paper (https://arxiv.org/pdf/2103.03375.pdf). We'll link it in our show notes which these guys released. I'm just gonna pull it up so I can make sure I don't miss quote it. The paper was released in June of 2021 and it was Sponsored by Perception Labs, which is the company that built this, and Google research. So one of the employees there worked at Google and Essentially, the paper describes a process by which they are estimating food volume and food contents Using two streams of data. They built a data set that could be used to train a multitask model that can do this performance. So I'm gonna go through this, hopefully, pretty clearly, so people understand your phone. Most phones have just a camera on them. Cameras take 2d Images or video right, so there's no depth perception. So if you have a picture of a bowl of soup and you're looking straight down on it, it can be actually difficult to estimate Is it a shallow bowl of soup or is it a deep bowl of soup. But knowing the depth of the soup rate is the volume of it is really important. They built a data set where they had a buffet I'm guessing it was at Google and they would add items to a plate and weigh them so they knew exactly how much was on there by weight and take top down inside camera views as well as lidar sensing. And the paper is about them building a training data set that was then able to Predict total calories, nass in grams, fat grams, carb grams and protein grams, as well as detect the Ingredients.
Joe C: 37:49
Is that to say, the research involved a lot more sensors than just a single camera photo.
Andy B: 37:57
Not quite. I'll explain in a second because I got surprised by something. There's a picture in the paper of like they've got like a plate of food on a scale under, like a box with a camera on top, and the food is again really suffering from white people. It is so the data set they built has egg whites, chicken scrambled egg, just olives, and they have a architecture of the multitask learning stuff, a multitask that basically they have the 2d image, they put it through a feature map and they end up calculating and calorie, mass, fat, carb, protein.
Kiran V: 38:35
I'm getting hungry.
Andy B: 38:36
Yeah, I know me too, yeah so here's the problem I start having. This is the only paper proving what they're doing and the paper talks more about the method for Constructing the data set to prove that it would work and it's more accurate than other, like open source food data sets and getting you more accurate data for doing this type of task. But like not. If they look at their charts of accuracy, I copy paste it one into our notes but human estimations and model estimates as a percent of the total calorie for the respective plate. X axis label for each plate includes ground truth, caloric value. Whatever, it's a little chart on the x-axis is some sample plates and then shows the accuracy of the model and nutritionist and not just so. For example, one where the model performed much better than both people. Non-nutritionists performed with looks like a 50% error rate, nutritionists with like a Little bit worse than them and the model had almost like less than a 5% error rate. That was categorized as plate 8, the total contents of which was 77 calories. What's 70? Like one cracker On their big chart of. Look at how much more accurate that is that people. The biggest plate They've estimated was 335 calories, which for me is barely a snack, and the model is also off by more than 20%.
Joe C: 40:06
I see.
Andy B: 40:09
So, and then, as you mentioned, they used LiDAR. I did not know this, but did you guys know that iPhones have LiDAR sensors in them?
Joe C: 40:20
No, oh wow, what a world we live in.
Andy B: 40:23
Yeah, so some newer generation iPhones and possibly some other front brands do have depth perception in the camera. And so what is it? Lidar?
Kiran V: 40:32
though, or is it like a simulated, because they have multiple cameras?
Andy B: 40:37
Can you look it up real quick? I really don't know. One of the test plates was eight calories and I'm thinking that is that one baby carrot.
Joe C: 40:48
That's like an. That's like an extra long salary stack.
Andy B: 40:52
And it's, yeah, it's. That's like people are 60% inaccurate at Estimating calories. For that, yeah, if you asked me how many calories was in the carrot, if I said 12 instead of 8, I would be like 50% off.
Joe C: 41:06
Yeah, even like human, verified calorie counts are all over the place, right like packaged food.
Kiran V: 41:17
Isn't it?
Joe C: 41:17
all still just some like estimation.
Andy B: 41:23
And so on. There when they did a press release, they said the team's able to reduce the caloric error to under 20%. There are other apps capable of using AI to do photo-based meal tracking, but none of them help with portion size estimation the most important part to reduce error. So like they are in there and I found a reddit thread (https://www.reddit.com/r/Myfitnesspal/comments/14l83pr/anyone_tried_snapcalorie_yet/) and I'm gonna read you guys some of the comments. But they're fighting for their lives to be like we're research based and we're so accurate. But then there one paper that they link that's not even on their website. I forget how I found it is like just a tiny toy experiment with datasets, like I. I Again, I don't want to disparage anybody's work, but we're talking about test sets of 10 images, right, like it's nothing. They got the paper some, and Then they're giving patting themselves on the back for this AI generated calorie counter thing. So somebody on the are my fitness pal subreddit. I'm gonna read you guys this this is my dream is become one of those podcasts where we get a thousand followers just for reading reddit threads. So Somebody asked has anyone tried snap calorie? It's uber expensive but seems pretty convenient if you don't do an annual subscriptions 30 dollars a month illness.
Joe C: 42:45
That's a lot of money for an accurate.
Andy B: 42:48
It's a lot of money to pay. Nutrition is to fix all the model. Oh so top comment is I loaded it up and the UI looks like it's been put together by a bunch of middle schoolers.
Joe C: 43:01
That's the easy part.
Kiran V: 43:04
Yeah.
Joe C: 43:05
Tell him Joe make it look nice, make it look trustworthy.
Andy B: 43:14
Somebody and that same person posted again After they tested a few things. I've tried a handful of pretty basic items and the apps been off by a country mile. For all of them it's been wrong on ingredients. Can't tell the difference between butter and cheese on top of a piece of toast, hasn't gotten the food volume weight even close to correct. 20% margin of error or more can add up pretty quickly and seems to default the lowest common generic denominator for everything it does. So, for example, two small sizes of garlic chabat of a get, each with a thin size of gruyere cheese that I clocked in at 138 calories, was recorded at two slices bread. Two tablespoons of butter at 358 calories. So in this case, big difference, huge difference for $30 a month for it to tell you. Overestimate how much you're eating. You could starve somebody on accident.
Kiran V: 44:03
Yeah, that's a dangerous.
Andy B: 44:10
Yeah, so did you figure out, kieran, whether the phones have real lighter?
Kiran V: 44:16
it does so. The the black dot, this black black dot. That's a lighter sensor and it basically sends waves of light, and there's also one on the front, which is they use for face ID. Oh it's an actual lighter sensor. It's cool.
Joe C: 44:32
Very cool. Yeah, you would need that for More secure face recognition, wouldn't you?
Kiran V: 44:38
Yeah, it works up to five meters away too, and that's what they're using when they have like the AR. You know stuff, so you'll have like, if you're on like Crate and Barrel, right, you can say like, show me this couch in my room and you know it's using those yeah, lighter sensors for that.
Joe C: 44:55
Have we explained lidar in this episode? Who wants to?
Kiran V: 44:59
So, essentially, lidar is a type of camera. Cameras generally use different sensors to Absorb light coming in and make sense of it. What lidar does is it actually sends out Laser beams, essentially right like really thin beams of light, and you could have Lidars that go 360 degrees light, ours that are directional. So on the iPhone it's gonna send a bunch of beams of light forward and it's gonna catch the reflections coming back and, based on the difference in time that it receives a particular signal that's gonna be able to map out the Topology or geography of whatever is around you.
Joe C: 45:46
And I like sonar, your depth sensing exactly.
Kiran V: 45:49
It's like sonar, but with waves of light.
Andy B: 45:52
Yeah, basically, your phone is a bat. You know how bats go or dolphins go. Yeah turns out that's your phone now too. Yeah, so the developer of this App I'm currently roasting Again fighting for his life in the reddit thread is like hi, what are the developers of snap calorie here? We want to comment on some of the feedback. We expect the photo logging to be about 20% error rate on iPhone Pro phones they have lidar to measure volume for portion size and 30% error on non pro iPhones and Android phones. I Just want to call them out on this. I get that you expect some error. Okay, like I built, you know some models here and there. Flawless is impossible. If somebody is heavily restricting calories, let's say they're already one of the people vulnerable to eating disorders All right, who might really be interested in accuracy like and have you know control issues over their food, and they're already restricting to basically very little food, and then you overestimate how much eating by 30% you could give somebody organ failure.
Kiran V: 47:02
Yeah, yeah so actually dangerous.
Andy B: 47:09
And their defense is that when people eyeball I, I proportion size they get above 50% error rate on average. But here's my pushback on this you know you're probably eyeballing inaccurate and you will. If you still feel hungry, you'll eat a little more. If you have a robot that you're trusting, that giving you the illusion of accuracy, and it's telling you, nope, you've definitely had enough. Stop eating. You might actually not eat when you really need to be eating. And then he challenges the users who had a bad experience to do a estimation test with them and they'll send them $100 if they can do better than the average 30% rate. And I'm like that's not. Nobody's arguing as to how good people are. Good estimating guys, people are bad at it, robots are bad at it.
Joe C: 47:57
And then they link in this comment.
Andy B: 47:58
They say we published a top peer reviewed academic paper here, based on estimates, and that links to that paper. I told you guys where again the test set is only 10 images and some of these. So they're out here bragging about being get up. Get on the board, andy, get out you should get in there and start Telling people what's up.
Joe C: 48:22
No, no, no.
Andy B: 48:23
I challenged the guy arguing for this Wade Norris to send us an email if he ever hears this and defend himself or just get more training data and work on the problem like it.
Kiran V: 48:39
It is something that you could solve, but you need a lot more data, clearly. So, like you know, do that Exactly and hopefully, the human in the loop setup that they have.
Andy B: 48:48
They are collecting more data and plan to set up some learning loops to make this better, but I thought, this was just like. You know, you know the biggest eye roll like it's a great idea, but the text just not there yet. Most people on earth do not have the brand new $1400 phone with a lidar sensor in it. And for people who are calorie restricting and obsessive about it and have eating disorders, putting a little three line don't become obsessive. At the bottom of your app download description doesn't help, like um, if they just seem to be. Oh, whatever, you can make a lot of money selling people diet stuff. Right, it's a gold rush. If you can make people feel bad about their bodies, you can sell them a lot of stuff to make them feel good about their bodies. That's the premise of victoria's secret, as far as I can tell, like Mm-hmm. That's the thing. So, um yeah, that's the, the, the buggy, the ugly snap calorie. There's some other apps that do exact this. There's one called meal snap, one called avocado, one called calorie counter. All of you, I'm watching you. I do not advise people use you. You're not going to be accurate enough yet.
Kiran V: 50:07
Yeah, and I think the buggy is really like AI isn't helping us be better here. It's just Well. One is like is it even doing anything? But the thing that it is doing right is like not trying to improve us generally. It's just like oh, let me help you continue to do this thing that you probably shouldn't be doing, just faster or better, like yeah, yeah, exactly.
Joe C: 50:33
It's a problem for something that isn't a real sorry, a solution for something that isn't a real problem.
Andy B: 50:40
We talk all the time about how AI is a tool and who wields that tool matters. The developers of snap calorie you guys are a bunch of Muppets.
Kiran V: 50:48
Find something better to do Better to do if you've got like computer vision expertise.
Andy B: 50:54
Go go detect cancer, go find lost bathers, do something useful. Shame on you.
Joe C: 51:02
Lost bathers.
Andy B: 51:04
Yeah, you know, like lifeguarding and stuff I had when I was working in the scouts, I had to do something called lost bather drills. I don't know why this came to mind. It's really, really hard. When kids drown in lakes, they sink. When kids are drowning, they sink. Lakes are really hard to see through. So any tools that can. And drowning is a leading cause of death in children, by the way, it's like one of the top five.
Joe C: 51:25
So Lost bathers.
Kiran V: 51:30
Yeah, yeah, I never heard that term, but that makes a lot of sense.
Andy B: 51:33
If you're swimming in anything other than a crystal clear blue pool, like if you're swimming in the ocean or a natural body of water, it's very hard to find drowned kids, yeah, so anyway, that's my little roundup on diet apps. Um, sorry for the diversion into your field exercise, karen, but um, this is the kind of the major things that are kind of trending right now. Again, I think it'll be different in a few months, in a few years, but Be careful.
Joe C: 52:01
Thanks for that research. That was all really interesting.
Kiran V: 52:03
Yeah, yeah, it is really interesting and I think it ties into the second part of this, which is the fitness apps and I think, um, you know, going a little beyond apps but also wearables. I think it's something that's really big. So I kind of I was also, you know, doing research on this and just learning so much along the way, as well as reflecting on my own experience. Again, like I mentioned before, you know, for me physical fitness has been a big part of my life for many, many years. I've taken, um, you know, weightlifting courses of Um, participated in many of those, you know, spartan races, and trained for different things. I've played all kinds of sports. So I've, you know, had personal experience with a lot of these things. And I broke this To kind of like go through it and to like two sections One is like hardware, one is software, because I think, um, Beyond that, there's so many different Thing, different ways or segments that we could break it up into. So I think that was, to me, the most sensible and I think, like you know, starting with hardware, right, it's like these are like wearables or like if you've heard of the tonal or mirror, or formerly known as mirror, currently known as, uh, lululemon studio (https://www.cnet.com/health/fitness/lululemon-to-stop-selling-studio-mirror-fitness-device-as-it-teams-up-with-peloton/), but there's a lot of these. You know, devices that we're being sold now to help us do different things. You know, track our fitness, um, you know. Track our sleep, um, track our workouts, um, you know, give us maybe some motivators. So, uh, you know, the best example I have of this is like dual lingo, which is not a fitness app, but that's the one that I've personally used, that where the motivation actually has helped me To continue using it, and all it basically does is like you're on a streak, you know, add one more day to your streak and it like does that? I've never used a fitness app that has successfully motivated me.
Andy B: 54:04
Um, I'm sure they're out there, so let us know if you, if you, use them, but, um, the only thing that motivates me with fitness is making plans with other people, not wanting to disappoint them. Yeah, I've learned yoga classes with joe.
Joe C: 54:18
Oh yeah, I think that's motivating as well. I do use a fitness app, that, um. I don't know if it motivates me, but I think it works really well for me, like logging and understanding, like my progress. It's called fit bod, um, and I've tried a good amount as well, and this one has really stuck.
Kiran V: 54:37
And what? What is it like? What do you do in that experience? Uh, you said like manually logging or yeah.
Joe C: 54:44
Yeah, you set a few things up in terms of like your goals and your schedule and like what equipment your gym has, and then, um, every time you go to the gym, it gives you a workout that uh, lets you sort of Say, yeah, this workout looks good. Or like, drop this exercise, drop the amount I'm doing it. Um, it'll do like common exercises, like different lifts and stuff, and it will um, uh, yeah, it'll sort of like be a system of record for for all that you've done.
Kiran V: 55:17
And are those AI generated workouts or are I be able to tell? Or maybe yeah?
Joe C: 55:24
If it's AI, it's probably very simple. I think it's more just kind of like hey, what muscle group have you like targeted, like not recently, like put that one up next. I wouldn't say it's very AI.
Kiran V: 55:37
Yeah yeah yeah, um, yeah, so I think. So, going back to the hardware piece, right, I think there are a couple different pieces. One is like sensors, right? So you know I have an Apple watch and I primarily Use it for fitness tracking and all I really do is there's like a fitness app and you can like indicate the workout that you're doing and it's like you know I'm doing a, you know, traditional strength training or outdoor run or cycling, whatever right, and then it has some metrics that tracks and essentially the things it's looking at is like heart rate. Heart rate, it can look at your oxygen levels, it has a ecg, so you can, you know, manually do an ecg. Again, with each of these things there's always the you know asterisk that you know this is not medical advice, like consult a doctor for any actual, you know, medical situations or scenarios, so. So there's like different types of sensors that are used to track fairly basic things like heart rate, oxygen. Um, there's Things like sleep trackers, which was actually interesting to me, but everything I saw indicates that none of the sleep trackers are particularly accurate and it's, and even knowing whether you are sleeping or not, it's just like really poorly accurate and so it's generally like a yes to me.
Andy B: 57:02
But, um, so I have a Garmin watch. I used a Garmin venue watch and, uh, I use it for sleep tracking sometimes and it'll like, um It'll, I thought it was quite accurate. Uh, it tends to give me scores saying like my sleep was poor when actually I feel very well rested, so I don't take it super seriously. But if I have periods of like going to the bathroom in the middle of the night or like waking up and rolling over, I feel like it captures that.
Kiran V: 57:35
Yeah, and I think those are fairly straightforward events to capture, right Like pose is really easy to like am I laying down or standing up? So I'll use things like that, but it's still like, am I laying down just to like lay in the grass and I'm not sleeping, or am I like laying down and that maybe is like how long have you been there? So there's like various things that it's looking at to determine that. But in terms of like accuracy measures, everything is just like. It's not particularly that you're not like it in the 90%ile of accuracy when you're detecting it. And again, it's like sleep and you do it every day and it's probably not something that you necessarily need to, unless you need to do like a sleep study, and that's when you need to go in a lab and they're actually putting electrodes on your brain to like look at your brain while you're sleeping. So we're not anywhere near that level of accuracy and for this is it definitely seems like more of like a. Let me just get a high level understanding of what my sleep looks like and then it'll look at like your breathing rate based on the oxygen levels in your blood. It'll look at your oxygen levels and then there'll be, like some derived components like heart rate variability. So how quickly is your heart able to change pace? So if I've done a workout, how quickly is my body able to get back to a normal heart rate after I stop working out? And the more healthy you are, the lower your heart rate variability means it's able to adapt a lot quicker. So these are all derived metrics from those sensors but still, like, none of this is really getting into using AI unless you're starting to actually have those, some of those derived metrics and again, like it's a simple math problem. So it's not like AI when you're computing heart rate variability. So then there was another one that I saw, bosch sensor tech and that they're trying to get a much more accurate wearable, but again using similar types of sensors. Where they're using AI, there is to actually use, make better use of the sensors. So, for example, if it's unable to detect a really strong pulse or heart rate, it's actually going to change the sample rate of and the intensity of the LEDs that are looking into your blood vessels to get that signal. So that's where AI is going to be used to improve the quality of collection and also improve the battery usage so that if it has enough signal, it can lower that sampling rate and conserve battery. So that's kind of where I was seeing AI within the sensors and this computing is done at the edge. So, like within the device, they have AI that is making these decisions. So again, you're saving energy on communicating with the server and having to waste battery there. So that's kind of like the extent of sensor research at least I got into. In terms of like wearables and fitness. I'm sure there's much more and there's more specialized devices, but if we're talking about just like a general consumer, I think this is kind of the extent. And again, like the Apple Watch or Fitbit, what is it? Whoop, no.
Andy B: 1:01:24
Whoop yeah.
Kiran V: 1:01:25
Yeah, whoop, and like Aura Ring. I think those are like the big mainstream wearables and again a lot of it is like simple sensors looking at very simple metrics and not doing a whole lot of complicated like AI at least that I saw.
Joe C: 1:01:42
It sounds like they're very limited by just input data right now and like the amount of sensors you can fit onto a watch or a ring or whatever.
Andy B: 1:01:51
But there's been some improvements. So I was an early adopter of like the fancy pedometers I had. This was like circa 2011,. Maybe I had a the original Fitbit one, which was like the Fitbit that they made as like a little pedometer that clips into your pants Like old school, and that was like much less. It was really expensive, all it did was count steps and it was not nearly as accurate. And now, like my, my phone has a snore detector, has a step detector, and when I wear my fancy Garmin watch and have my phone with me and I take my dog on a walk, the numbers are within totally reasonable error range of each other and they seem much more accurate than what we used to have. And this is only in 10 years, 10, 12 years has this changed?
Kiran V: 1:02:39
Yeah, and that's that's an improvement in the sensors themselves, not necessarily AI. Again, there is a little bit of AI being used to control how the sensors are used and that's where, again, you can get an improvement in quality. But at least from what I saw, there wasn't like this. You know AI, you know system within your device that's doing something so magical. Yeah, makes sense. Again, I think it's a you're trying to turn, like you know, an analog signal into a digital signal and there's only so many ways we can collect that data right. So those were like sensors again in wearables. Then there's other things like like, I would say like maybe machines, if you will like, the mirror and the tonal. That again are like fairly popular mainstream things that I saw with, you know, hardware and total that's actually, yeah, that's when I've used and essentially what it is. It's like a screen on the wall, similar to if you've seen the mirror, and it's just a screen that has these cable pulleys built in. And where it actually uses AI is in controlling the tension on the cables so it can adjust the weight mid-workout. I've used this, my parents have it, and I really didn't like it because literally, like I will be doing an exercise and mid exercise it will change the weight and maybe there's like different settings on like how interfering you know you could get. But at least when I used it they're like two or three times I've used it. I've always found that super distracting but again, maybe once you've used it enough it like gets a better sense of you know, your body and it doesn't have to make those intermediate adjustments. But that was again the extent of where AI is being used. In the tonal I was trying to find anything on like are the workouts AI or these AI trainers, like they're humans that are recording workout programs and then you know you're on a screen and you're just kind of watching that replay.
Joe C: 1:04:49
That's interesting Because it's I feel like that product is really marketed as like a very smart device and it still is, even without, maybe a lot of use of AI.
Kiran V: 1:05:01
But yeah and again, a lot of like. The tracking right is simple, like spreadsheets right. Essentially it's just the data. It's just keeping track of all your workouts. It knows, like, what was the weight that you used and this is, you know, last week, this week, next week, like it can track that. There's no AI involved in that, so I don't know. I thought, yeah, go ahead.
Andy B: 1:05:27
Question I thought like some of these had cameras like I forget what I felt connect is the name of one of these things from like five or six years ago, but a lot of them had like cameras that could see you exercising computer vision and then like give you feedback.
Kiran V: 1:05:44
Yeah, so that was the Xbox, the connect, and that was essentially like you said. It had two cameras and it was able to do pose detection and then they used it for a lot of games. And it's still there and that technology is there and essentially what it does is it has two cameras so that you have depth perception and then it's just doing pose detection. Right, that again is like a relatively solved problem, like we can have a camera point at someone and identify their legs, identify their arms and their torso and their head.
Andy B: 1:06:21
Fun fact, the three of us actually developed an object detection in video tool that did pose detection algorithms for the company that owns your favorite clock app. Just that's true.
Kiran V: 1:06:35
Yeah, way back in the day so, but I didn't see any examples of, again, broadly adopted fitness gadgets or applications that do that. I know there's a lot of specialized applications for very specific things, but I haven't seen like a general, like here's a workout app, take a video of yourself and we're going to assess it and remember, like reading something about that with Mirror, like way back in the day before it was part of Lululemon, but I know, and that was kind of like what it was being advertised to do and like oh, it will give you feedback on your workouts. But like I've never used it, I've never seen it, I tried to look online and apparently Lululemon bought the company called Mirror, rebranded to Lululemon studio, and is now shutting that down to partner with Peloton to where Peloton is going to be their fitness arm or do all of their fitness on the back end and Lululemon is going to like be the apparel brand for Peloton (https://corporate.lululemon.com/media/press-releases/2023/09-27-2023-211026323). So I think Peloton is another, again, really popular one which you know I've never used it, but in my experience I have a lot of friends that have used it and they primarily go to it for the classes, right? So there's, they started out with the cycle and they had cycle classes. Again, it's a human, like there's a person there, sometimes it's prerecorded, sometimes there's live classes and and now they have different types of workouts on different machines and things. But again, like, at least as far as I know, it's not AI generated content or people and there's not really any AI like hard AI in any of these systems beyond. Maybe you know small things on sensors etc.
Joe C: 1:08:36
I think it's very easy for marketing teams to build things as AI that maybe aren't. Ai Oris, whatever you said, soft AI.
Kiran V: 1:08:48
Yeah, so that was on the hardware side. Again, I was actually fairly disappointed with at least in my research, with, how little there is, and I think this kind of gets me to like my bigger point that I want to get to, which is like humans are the problem, not the technology. And I think you know, as we like now talking about software, you know I talked through that chat GBT example. That is plenty sufficient, as like if humans weren't the problem, if I had all the motivation in the world and the actual problem for me becoming fit was I didn't know what to do, right. Then chat GBT, like that's that's plenty, right, if I have that thing, then for the next three months I got my workout plan, do another one three months from now, change the focus, so like that's it, that's all you really need, right? And so the problem isn't people don't know what to do. The problem is people don't do it and they need, in a lot of cases, a lot of motivation to get through. And that's where things like a personal trainer, because there's like another human that is waiting for you at the gym at a certain time and if you don't show up, you're going to feel that shame from. You know our society, when you miss things or, you know, make other people feel bad. So I think that's the problem, right. And so now, like talking through different apps, it was actually very apparent that that is the thing that people are trying to solve, because, when I look at I just searched top fitness apps none of them had anything to do. Maybe they have some AI, but none of them were marketing AI. It was like there's like a Nike or an Adidas app and so that's like oh, I'm part of this brand community. There's an app that's like women for workouts, for women, right, and you're like working out with other women. There's one Chris Hemsworth, Thor from the Marvel series. He created one called center and again it's like it's about like personalization and then they have these apps, but then it's like it always brings in like community aspect. So like a lot, a big focus in workout apps if I have not, if I'm not searching like AI apps and I'm just searching like fitness apps is always centered about like community is primarily first, and then like the types of classes or types of workouts that they have.
Joe C: 1:11:27
That's a very yeah, that's a very interesting takeaway. Actually, the app I mentioned earlier has, like, a lot of sharing features that are meant to be like you know motivate yourself by sharing your workout with your friend who's also doing like whatever on the app.
Andy B: 1:11:43
Yeah, Well, isn't Strava one of the most popular workout tracking apps? And it's just because it's the workout app for gossips? That's all I know about. Strava is like the amount of times remember back at figure eight. Sometimes the engineering guys would sit diagonal from us. So for people who did not work with us, there was the product and design team and Karen was the engineer adopted into the product team as our prototyper and we had like the window spots and we put the engineers in the cave deep in the spot to code away and I'd walk over there about five times a day to make sure they were doing their jobs and tell them to type faster. And that's exactly what happened. No one can say it didn't and they would be like talking about who saw. What on Strava is that?
Joe C: 1:12:31
it was all about the social aspect Is that the running tracking app.
Kiran V: 1:12:38
They actually have different ones. So I actually use Strava and I use it for exactly that actually. So when I was living in New York me and my co worker this was when I was working on my startup office and so we all had each other on Strava and there were segments of each of our rides that overlapped, and Strava will actually break your ride into segments and tell you like, oh, you had a personal record on this segment today, and so then you can like compare with the and and yeah, it was just. That is like we always would see who could get down 9th Avenue between I think it was like 28th Street and our. Our office was on 43rd Street. So, like whoever could do that the fast. And it was like you're on a street with with cars and signal lights and we would set some like ridiculous times because we're like running through red lights and it was definitely dangerous.
Joe C: 1:13:33
But yeah, it was like I'm gonna say how that's like are still alive. Who's with us today? Still?
Kiran V: 1:13:38
Yeah, it was definitely dangerous and so, but it was fun and like that was. That was the motivation for me and why I became more fit through cycling. Not anything to do with AI, right.
Andy B: 1:13:53
Yeah, it was the best shape of my life when I, just after work, me and X and a body of ours would like go and lift weights together. And because two other people were meeting me at the gym three days a week and if I didn't show up then I was the jerk and I didn't want to be the one to start the we're not, we're bailing and just going to go eat sushi, like no, I didn't want to be the one to start that. So we got really into shape.
Kiran V: 1:14:16
Yeah grinding on the real thing. So so, yeah, so top apps generally focused around community types of workouts and none of it seems to indicate AI generated content or workouts, etc. And I think again that that to me seems like a solved problem. Chatgbt can give me a very sensible workout and I, at least, knowing myself and my past experience when working out, if I do the thing that chatGbT is telling me, I will have a great workout, I'll be fit, and that's really all I need. So it's like, okay, how do we motivate people? And so then we get into like personal trainers and again, you know personal trainers or other humans, but this is where you know we're starting to see a little more AI. So there's an application called fitness AI and it that is an actual AI assistant personal trainer. So like the AI is your personal trainer and it creates workouts, it will give, send you like motivational phrases and and that's kind of the extent of it, right. And what is the application going to do beyond that? Beyond like here, do these things? This is how your workout plan looks like, and and so, and, reading the reviews on it doesn't seem like it's particularly special, like it was a life changing experience with AI. It's. It's just like, yeah, it kind of understands what I'm doing, it can give me a trajectory of workout, but one of them said, like it takes it takes weeks for it to like really understand you. So it doesn't seem like the AI is particularly powerful anyways, and it's probably picking from a bunch of pre canned workouts, so also very underwhelming for an app that's like an AI fitness like all. That was the one that I found it's.
Joe C: 1:16:19
The interesting thing about that one, though, is it's trying to like, mimic a personal trainer more closely, like with the one off messages to you. Maybe they have picked up on the idea that, like, accountability is a huge thing, or like a motivating friend, and they can take that places in the future.
Kiran V: 1:16:41
Yeah, yeah. And then, and, and I think again, like every app talks about like personalized workout experience, personalized this, personalized that, and like the first thing I think about is like, oh, wow, are they using AI? Right. And so I actually tried to sign up for one. It's called Joggo and it's it's about like running, and there's option to also like get a diet plan along with your running plan. And so I started going through the onboarding flow and it asks, like you know, height, weight, like typical questions. But then it's also like how far can you currently run Right, what's your goal for using this app? You know, are you trying to lose weight? There's even one that's like are you trying to work on your mental wellness? So, like, there's some options. And then the last question was what's your target weight? And I want to gain weight, right, because for me, I've always had this like I'm too skinny, complex, like since I've been a little kid. So I'm like, oh, I want to gain weight so I can, like you know, put on some muscle and like, and the app literally wouldn't allow me to finish the onboarding if I had a weight above my current weight. So there's like the finish. There was a button that said finish and it was like disabled If the number wasn't equal or below my current value. I'm like what? So and then I, so I put it as the same value and they press next. And then it was like all right, sign up, pay $33 for your first month to get your workout plan. And I was like I'm not going to do this. But those were the same prompts that it prompt me in that workflow that I input in chat CBT. That's why I was like I can run two miles, now I want to get to 30. And it was like a totally sensible workout plan. And you know, if I were to do that and I do actually want to run a marathon one day like I would probably just do that and follow chat CBT like that's good enough.
Andy B: 1:18:43
And this is just a little tech hat coming out, but there's a lot of companies right now who are basically trying to resell you what you can get for free from chat, cbt. Some of them will figure out a way to make it better by adding safeguards and value, and you're going to be like, yes, this is much pretty UI.
Joe C: 1:19:01
Some. Of them.
Andy B: 1:19:02
Yeah, pretty UIs, some of them won't. So when you're in doubt, just ask. Ask one of the large language models chat, CBT, open AI. You probably already have access to it. Start there.
Joe C: 1:19:14
I will say the same goes for some humans trying to offer you their services. Chat CBT can just do it for better or worse.
Kiran V: 1:19:23
So I mean there there is more stuff in like software. Again, there's there's a lot of apps. A lot of the health one. I mean the diet ones came up as well, like my fitness pal Noom. But I think overall I'm fairly underwhelmed with the state of AI when it comes to diet and fitness and I think for me again as someone, that this is something that has been very important to me for a long time. My all I've kind of like my baseline now is like I wear my Apple watch and I'll track my workouts, just so I can like see my calories burned and and how long, etc. And that's it. And then, like I make up my own workouts because I have that experience, but otherwise I've used like those like those 30 day workout thing by Kayla or whatever, like there's these different people and they just like put up, here's a workout plan, like that's it, that's like all I've done, and yeah it shocks me how expensive they are, for how shitty the product is Like.
Andy B: 1:20:33
the calorie counting picture app is $30 a month If you go month to month. Your Joggo app is $33. Are you?
Kiran V: 1:20:41
kidding me, it's crazy.
Andy B: 1:20:42
I would use you on chat. You got the exact same stuff basically for free in a second, no on-boarding needed.
Joe C: 1:20:49
Yeah, very customizable.
Andy B: 1:20:53
Yeah, yeah, and no weird value judgment as to how much you should or should not be trying to gain or lose weight.
Joe C: 1:21:00
Yeah, I actually think, like taking taking a look back at everything we've discussed, you know, in this part on diet and exercise and then also on our our other part on um, like wellness and mindfulness, I'm a little bit more like, uh, happier with the wellness and mindfulness stuff because it's less of a like value judgment. It seems more kind of like let's just get you to understand yourself and to like know yourself better and you'll grow into a better person that way, whereas with the diet and exercise stuff it's like this is how you should be, or like these are the things you should be caring about that are just a little too prescriptive to be safe, or like personalized.
Kiran V: 1:21:44
Yeah, Like it almost feels like I need to know what I need and want for my body already and the app is just like a way of helping me log that. Um, I just I just haven't seen something that is like so overwhelmingly useful.
Joe C: 1:22:05
All right, shall, we shall. We wrap it up. Do we want to do like a round robin? Final thoughts.
Kiran V: 1:22:12
Sure, you want to go for it, joe.
Joe C: 1:22:14
I'll start. This is something I I always feel like we hit the end, and I reflect on the role of capitalism in AI. And you know, we we may have brought this up when we talked about like dating and and stuff, but a lot of these apps, the groups working on AI, are they sort of have a conflict of interest. They're trying to keep you in the app, but then also they're trying to make you better and more well in order to get you out of the app. Um, so how does that work? And you know, throughout our conversation, I've sort of questioned that a lot with what these apps are doing.
Andy B: 1:22:50
Yeah, the technology is nothing mind bogglingly cool. From what we can tell, the coolest thing that exists so far is the little app you mentioned in the previous episode, finch, that nags you into being your outdoor buddy as well as just chat. Gpt and LLMs being some conversational and having a wealth of knowledge that seems to be the two big innovations in health and wellness for 2024. And the rest of these apps are just smoke and mirrors.
Joe C: 1:23:22
Yeah, that's a good summary.
Kiran V: 1:23:24
Yeah, I agree, I think, overall, ai in health and wellness, well, health and wellness, well. So health, I think we haven't touched on and maybe that's I think that's a different episode. But in terms of, like, wellness, fitness, diet is fairly underwhelming and I think I think the reason is because a lot of it comes down to, you know, in the case of mental health, very open ended goals. I think, like be happy isn't can't really be a goal, right, like, so, like, I think there there's open ended goals. I think on the diet and fitness side, it's more about individuals not behaving the way that they need to be to get to where they want to be, and so so it's like as much, if you don't have, you know, the right AI, or no matter how much AI you have, if you can't change people's behaviors. And maybe that's where something like a AI personal trainer could be helpful, where, like, the focus is really more on like the motivation behind it to change those behaviors. But I, yeah, I don't see anything that's like. AI is the going to be the thing right, and I think it's. It's clear too, like there's not this one thing that everyone uses and it's a lot of. It is really on like the community you're in. So like I'm going to do these CrossFit apps or or the Nike app, because I have. I'm in that Nike club or whatever that is.
Andy B: 1:24:55
So yeah, and if you have a wellness or fitness or diet goal for 2024, we wish you luck in it. We hope you get what you want. You get the effect of it. If you want to be happier, that you're happier. If you want to be healthier, that you're healthier, whatever it takes. And no shade to anybody who uses any of these tools, because use what works for you.
Joe C: 1:25:19
Absolutely, and if you use anything interesting and have some insight into how AI plays a role, send it our way. We love to hear about it. And that wraps up our episode. Please rate, review and subscribe. You can find our podcast wherever you get your podcasts and we'll see you next time. Thanks,