Loading summary
Transcript58 lines
- [00:18]
A
Hello and welcome. This is Gabriel Custodiet of Watchman Privacy Privacy practitioner, consultant, author and frontline fighter in the push for privacy. I know why you're here. Like the rest of us here in the resistance, you're trying to escape the technocratic apparatuses that you see enveloping you and crushing your freedoms. That's why I created all of this, all without sponsors. I hope you enjoy this show, but then when you're ready to take the next steps to secure your privacy and your future, Visit my website, escapethetechnocracy.com to start the real journey. Your support alone does not determines the future of the show. See you there. Welcome everyone. I'm very pleased today to be joined by Milan Darid and he is the creator, co creator of Nano GPT, which is a very handy AI API service where you can go to the websites nano-GPT.com and you can access hundreds of different AI models, image, video, text, without having to have an account on those places. So that's very useful. You can do it therefore, in a more anonymous way. This is not something we'd be using for our very private queries, but it's a fantastic way to explore a lot of AI models. It has a top up system so you can, you don't have to even have an account with nano GPT. You can just top it up, test it out. They accept privacy cryptocurrencies, so it's a very useful service. And I had a episode previously with Milan explain the service in its entirety. So I would very much recommend you go check that one out. We'll have that in the show notes to get the full sense of things. I thought it was a, I thought it was a very good interview and actually as I was preparing for this one, I thought, oh shoot, like what am I going to talk about in this one? Because we kind of covered so many things. So Anyway, Milan of Nano GPT.com welcome back to the show. How are you doing?
- [02:10]
B
Thank you. Very happy to be here. Our last chat was, what is it, four or five months ago now, I think.
- [02:16]
A
Yeah, I think that one was in April. So first, first part of the year for sure. And we discovered the nano GPT. I heard about it from Urban and I'm not sure where he heard about it from, but it's obviously a very useful service for all the reasons we go into in the first episode. And let's kind of get people up to speed though, Milan, and assume that people understand kind of what's going on. And this, this Episode's going to be more kind of getting a little bit into some of the specifics, some of the updates and so maybe you could kick us off. What are some of the main changes, the main updates to Nano GPT since we last talked.
- [02:53]
B
Yeah, so I think one of the updates, the suggestion might have actually come from you guys. So before we would do an anonymous account and it would just be a cookie that's stored in your browser. And nowadays we use essentially what mill that the VPN also does. So we generate a random UUID essentially and you can use that to also log in from other devices. So it's gotten even more private. So even easier to use it in a private way. There's even less need to make an account because before it would be hard to link your mobile account to your desktop account. And now that's gotten quite easy even being fully anonymous. Another big change has been that we have a subscription now it's fully optional, so we still have the pay as you go service. But essentially feedback we got was that some people prefer to know in advance how much they're spending and then just use essentially unlimited. So what we've managed to do is make a subscription for all of the open source models. Those are also the models on which we can guarantee that there's no logging because if you run through OpenAI Google there's usually going to be logging. But for open source models there are a bunch of no log, no training providers and those are the ones we use. So we feel it aligns quite well with what we're trying to do. So it's $8 a month and then people can access like pretty much every open source model there is and pretty much unlimited. It's 60,000 queries per month. What else have we done? I think in general we've just added tons of models as new ones have come out. But, but that's kind of. We were doing that before as well and we've I would say made it more usable by also adding stuff like conversation sync. So before mobile and desktop, if you logged in from another device it would be totally separate. And nowadays you can choose to either use our storage and we will sort of do the syncing for you and do the storing for you, or you can add in your own storage so you can be fully in control of it yourself. You can just add in your own storage. We don't store anything ourselves, we just use your essentially AWS S3 or some compatible storage. And yeah, we've noticed that many people find that very useful. Because you know, it makes it actually possible to continue conversations from phone to laptop to anywhere you want to use it. I would say those are probably the main changes we've made. We also keep just trying to make it easier to. You mentioned the private coins. So we used to have Monero. Now we also have zcash and we have the Mimble web extension for Litecoin. So it makes it even easier for other users of private coins to deposit with us. So yeah, I would say those have been the main changes.
- [05:42]
A
Yeah, I think you said you used to have Monero. You certainly still do have Monero. Yeah, so those are all great updates. And I'll tell the audience here if you're not sure, if you have no clue where to start with AI, you're not sure what you're doing, you're not sure what the privacy risks are. We created on Escape the Technocracy we have a course called AI Resistance. We walk you every step understanding AI, what it's good for, what it's not good for, all the privacy considerations and kind of the tiers of considering your privacy risk, self hosting and then moving on to services such as Nano GPT and others. So we covered the full gamut. That's the best content for privacy related AI on the Internet. Definitely go check that out. And Nano GPT is something that we show off in that course because it definitely has a role. Now just a privacy question, Milan, the open source models that you guys have, how does that compare? Obviously if we have very sensitive things, we're going to recommend people maybe do some self hosted AI or certainly be cautious. But your open source models are you. I may have missed this when you were talking just now. Are you guys running those models yourself?
- [06:53]
B
No. So I would say I agree with you. So the most private is hosting the models yourself. And then I would say second best is we have what are called TEE models, Trusted Encryption Environment models. So essentially those are run in a way where we can fully verify from the start to the end, sort of what code they're running on that nothing is stored. It's sort of a very verifiable way that the provider can show we are not doing any logging, any training whatsoever. So that's like the second best option if you're not doing local. And then we have just the general open source models and we run those through a bunch of different providers. The good part of open source models is everyone can host them. So there's quite a bit of competition between all the different hosters and they've Noticed. Okay. People really prefer it if we do not store their conversations and we don't log their data. So we get to kind of pick and choose which providers we use for those open source models and we only use the ones that have a no logging policy. Now I have to say like we, we are relying on their word, right. We have no insight into their code base. So it could be that they are logging it in some way. At the end of the day it's just a sort of promise from them. So the tee, the Trusted Encryption Environment models, those are definitely more secure than the open source models. But at least for the open source models, unless they're lying and you know, want to get sued in the long run, then those should also be completely no logging. But yeah, we don't host them ourselves, so we cannot fully, fully guarantee that there is no logging going on.
- [08:34]
A
Sure, yeah, gotcha. One thing that I, as a kind of extreme privacy user would be using Nano GPT for easily things like images, kind of public facing stuff, whether that's text or images that you know it's going to be public anyway, I'm not concerned. So stuff like that's easy. And obviously experimenting. Right. All the, all the latest stuff that you guys have and you can click down and see that you have hundreds of different models and companies to test and compare and that is incredibly useful especially for these psychonauts out there who are trying to see what the, what the latest and best and greatest is and, and all the rest. New feature or. Yeah, new kind of trend recently in AI is the video AI. Just walk us through what do we, what do people, they're, they're comfortable with creating AI images, they're comfortable with the text, maybe the coding. What do we need to know about video AI?
- [09:25]
B
Ooh, good one. The first thing to know I would say is that it's gotten insanely good. It's like scarily good. I don't know to what extent we should be happy about that. There's definitely some questionable stuff about it. But the new Sora 2 model by OpenAI, it was released I think a few days ago. We just added it I think earlier today or yesterday evening. It's just incredibly good. You can essentially create anything you want to and have it look just completely lifelike. There's audio you can do up to I believe 12 seconds now for that model. So it's like short clips still. But yeah, if you like want to try and create a short commercial or you know, like a short TikTok clip or whatever, all of that's gotten quite easy. It's still not as good for like longer content. If you want to create like a five minute episode of something, we do have some models that actually do that. They're the long stories models in case anyone wants to check it out. But that's essentially someone who has built a sort of workflow themselves where you enter a prompt and then they just do a bunch of different clips and they sort of edit it together in a very nice way. But in general it's most suited right now for short clips, I would say like short commercials, that kind of stuff. But it is honestly commercial level at this point. Like I think some of the commercials you'll be seeing come out in the next weeks. Probably they're using AI just because it's gotten so incredibly good. I do have to say they are quite a bit more expensive than text and image models. So for example, the Sora 2 model, I believe a 8 second video with like the low settings is about 80 cents. Actually it's a bit less on ours, like 70 cents I think because we give a discount on it. But yeah, they are relatively expensive. But then they are like incredibly good. So yeah, it's worth trying out I think.
- [11:20]
A
Absolutely, yeah. And the price is something that people have to be cognizant of maybe for some of the users out there who are used to just spamming dozens of images at a time, not going to happen with video. So you do have to be a little bit more selective certainly for now, but. But when you think about it like I was creating some, you know, eight second videos, considering what is happening there and what you're getting for 80 or 70 cents, it's obviously pretty astonishing. But for people getting into playing around with the video AI, what are some of the more cost effective video models that maybe they could experiment with?
- [11:55]
B
Oh, that's a good one. So There is the One Wan 2.2 Turbo model. Yeah. The naming of some of these models is honestly terrible. But that one is quite cheap. I think it's like 5 cents to do a video. I'm trying to check it really quickly, but yeah, it's about. Yeah, it is 5 cents. So if you do a short video there, yeah, you're only paying 5 cents for about 5 seconds. So you can just do a few iterations and see what it looks like. Those are the cheapest ones, so about $0.01 per second.
- [12:29]
A
Now, as people step up to using the more expensive ones that are better, like the Sora that you mentioned, what are some of the prompts like walk us through how should people be structuring their prompts to get the most out of those so that they're not wasting their 70 cents?
- [12:44]
B
Yeah. So I would say the perfect prompt has gotten less important over time than it used to be. Because I would say OpenAI and other providers have also done some engineering on their own sort of backend to try and make it easier for people to get what they want. In general, I would say for most video models that we add nowadays, we have some example videos also on the website. And then you can also see the prompt that was used for the video that you're seeing. And I think that's quite a good way because then you can see some actual prompts that give actual videos. And then you can see for some models a very short prompt and just leaving it to the model works really well. And then for others, you see very intricate scenes where, you know, a lot of stuff is happening and you can really like be the director and have exactly what you want. And then people start playing around with essentially describing what they want the camera to do, how they want the panning to go, like, is it zooming in, is it zooming out, is it moving, is it a fixed scene, all that kind of stuff. So I would say the sort of best way to use it nowadays, if you're trying to be like an actual director is just, just literally what you would write down for sort of a person who is doing the filming for you just works really well for the model as well. Like, okay, do a close up of her face and then slowly zoom out and you can see the party going on or something like that. Like that stuff just works really well because it's the same way you would describe it to a person and you actually get the result usually that you want.
- [14:18]
A
I wanted to ask you a few questions about Image AIs. Before that though, I noticed that I don't know how long this has been the case, but you can access nano GPT through cake wallet through nan swap through bitcoin.com wallet. Tell us a little bit about the different places that people can actually use nano GPT.
- [14:36]
B
Yeah, so we of course have our regular website, right, nano GPT.com, but obviously we are quite like crypto adjacent. At least we accept crypto, we quite like crypto ourselves, and we are slowly turning into one of the bigger merchants in crypto. So I think last month I would have to check, but we did about four or five thousand, I think. Now crypto payments, which is just getting to be quite a lot. So many of these crypto wallets, they want to add some functionality to their wallet so that people keep returning to the wallet and so that they have a way to actually use the wallet to spend their crypto. So Cake Wallet, they were the first to let us be in the wallet. I think we reached out to them or they reached out to us. I'm not happy actually sure anymore. But essentially the thinking was we can make their wallet better by, you know, making it possible to access AI right there in the wallet. People can just send in any of the crypto that Cake Wallet supports to Nano GPT and it's like a Cake Wallet branded version of Nano GPT and then start using it right away. No need to make an account, no need to do anything. So it's kind of a way for people to also see the power of crypto. Like you can actually use this stuff. It's not just investment. So, yeah, we did that with Cake Wallet and then now we're also in the Bitcoin.com wallet, which is actually a bit of a deeper integration even into their own wallet. They also do like promotions and stuff in their app to get people to try it out because they're quite excited about it. And we are quite excited about it as well because some people are using these models to build essentially entire workflows where the model itself can use crypto to pay others and then use the funds to from a wallet to top up on Nano GPT. Again. So people are building pretty cool stuff and also pretty sort of useful stuff that can actually do transactions and buy stuff because of crypto. So it feels like a very natural, I guess, integration for us. And yeah, I think both sides have been very happy with it so far. Cake and bitcoin.com and us.
- [16:44]
A
So for images, your image model recommender. What. What's going on when we select that option?
- [16:51]
B
Yeah, so essentially there are leaderboards for image models. So there are quite a few image models nowadays, like tons of them. And some of them are very good at specific stuff. Like, for example, one might be good at like 3D animation, another one might be good at realism, another one might be good at landscapes, all that kind of stuff. So what we try to do is we use those leaderboards and whenever you enter a prompt, like, I don't know, a cat standing on the head of an artist in a beautiful scenery, whatever, we try to classify it according to what sort of image you're looking for. So that might be a realistic image in this case, probably Landscape, Landscape and animals for example. And then those leaderboards have what is the best image model for those categories. And we just, as new data gets added to the leaderboard, we update our recommendations again and then we just recommend. Okay, because you're looking for an image that we think has to do with realism and animals, we recommend you use this model. And then of course with the text models we have the auto model which then immediately also does the query to the model that we think is best. But because images are a bit more expensive, we've decided to make it a recommendation and then once you get the recommendation you can just click and agree to use that model. But it's a bit harder to do the auto model with images because like first of all more expensive. But also sometimes people want to do like a different resolution or something and they don't write that in the prompt. So it makes it a sort of two step process. But yeah, because there are just so many models nowadays that even we usually cannot figure out like which one do we want to use. We figured it would be useful for people to just be able to like write a prompt and you know, get a recommendation on what to use.
- [18:41]
A
Yeah. What are some of your favorite AI image models these days?
- [18:47]
B
So I really like the Nano Banana model. First of all, the model came out and we were very confused because of course we're called Nano GPT and the crypto is called Nano that we mostly use on the website. It's not our own crypto, just to be clear. And then it has like a fork which is called Banano. So Nano Banana was very confusing to us, but it's really good at editing images. So we're trying to decorate our living room or redo our living room and we're just constantly throwing images in there and make the floor different or add this couch, add these chairs to just see what it looks like and it kind of makes it a lot easier to visualize what a room is going to look like. So yeah, I use the nanobanana model quite a lot nowadays. Many people also like the C Dream 4.0 model because it's relatively cheap and it's just a. Yeah, it's a fantastic model. Anything you can think of it can create essentially. And then if I have to name a third one, it's the Quen image model by Quen by Alibaba because it's open source. So it's a very cheap model for us to use. There are many providers offering it and it's also really good. It can do like Image editing or just image creation. And we in general like open source. So in terms of open source, that's probably the best model. So we quite like that one as well.
- [20:07]
A
Yeah, it sounds like when you're choosing your favorite one, it's maybe open source, maybe it's the cost on a cost basis, not necessarily. Like, what is the best one?
- [20:17]
B
Yeah, it kind of depends in a way, because for some images you just want the very best, but also some are so cheap that you can do like 10 different variations and just figure out which one is best.
- [20:26]
A
Right. What would you do? You're looking for the best now, what would you do for images?
- [20:31]
B
I would say the very best. Right now, if you're creating an image, it's C dream, C Dream 4.0. And if you're editing an image, then it's probably Nano Banana. So it depends on what you want to do. But yeah, if you're creating from scratch, Then C Dream 4.0 is incredible.
- [20:48]
A
When you're on the native platform, on some of these AI services, they have their own internal controls, we mentioned Mid Journey in the previous episode, where you can zoom out and zoom in and change the aspect ratio. How are we with some of those internal controls on Nano GPT these days? Obviously you're not going to get all of them, but any changes to that.
- [21:09]
B
I would say for most, we're pretty close to supporting whatever the provider also supports. So on most of them it's honestly as simple as you create an image and then you can talk to the model to like edit the image. You can also do that on aris, so that's the same. I would say the odd one out is actually still mostly midjourney because they have this very specific functionality that they only offer on their own website. Like, I think, yeah, what you mentioned before, the zooming in, the zooming out, that kind of stuff, I think they still don't offer that via API, frustratingly for us. So that's still not possible for us to add. We do have other models that sort of try to do the same. So for example, if you get a good image, you can do the upscaler on our servers to get it higher resolution, that kind of stuff. But yeah, some of that native stuff, honestly, it's just a bit frustrating, but they still don't offer it to us.
- [22:02]
A
Is it possible to queue images or do we have to wait for one to complete on Nano GPT before we start another?
- [22:10]
B
I think you can queue them, I think. I don't actually often try that myself, usually because Images feel expensive to me, so I usually want to see what comes out before I do a next one. But I think in our chat you probably can cue images because you can also queue messages. So it should work.
- [22:28]
A
Should work the same. Yeah, maybe I was just doing something wrong. No worries for images. Walk me through some of this AI terminology. We have things like inference steps, guidance steps, things of this sort. What does all that mean?
- [22:44]
B
Yeah, some of them make it really complicated. So I would say the most important ones are just resolution. Those are very clear. But then for some of the open source models, yeah, you have the inference steps and that's essentially how much effort, in a way, you want the model to put into the image. So with one inference step, and I think I'm butchering this explanation because I don't develop these image models myself, no worries. But with one step, they essentially try once to get close to what it is you want and then that's it. If you do 30 steps and they keep refining more and more and more and they get closer and closer to what the image should actually be, there is sort of, I would say, also a way of overdoing it. If you go too high in the inference steps, then most of these models tend to get. Get either like gibberish or way too specific on one detail. So what we usually do is go for like, I would say 20, 30 inference steps as default. And then you have the guidance skill. That's the other one that every model pretty much has, which is how closely you want the model to follow the prompt. So let's say you say, I don't know, tiger. Okay. Do you want the model to just exactly show you a tiger, or would it be cool to show the tiger in its natural habitat and you see a bit of jungle around it, all that kind of stuff. So the higher you make the guidance skill, the more exact the model sticks to what you said and the lower it is, the more sort of freedom or creativity it can put in. Those are the main ones. I think you also can sometimes do a negative prompt. So you can say, I want to see tiger, but negative prompt jungle, because you want to see a tiger, but you don't want to see jungle in there. So that one's also pretty useful. I think those are probably the main ones. I would have to click and see if there's anything. No, that shows up.
- [24:41]
A
No, that's good. That's. That's very helpful. Where are we with Grok on your service?
- [24:47]
B
So Grok is on our service. I think we have the Grok 4 and the Grok 4 fast model. So recently Grok released their 4 fast thinking model, which we now also have, which I think is currently their sort of top model, but I'm not entirely sure because they have so many different sort of terminologies around their models. They also have GROK for heavy. GROK for heavy thinking, I think.
- [25:12]
A
Does it compete against some of these others? Do you find that you use it?
- [25:15]
B
Honestly, no. At the moment I don't use it at all. So the Grok 4 fast thinking, for a while it was free in some like programming environments and then I think people use it quite a bit because it has like a very big context size, like a lot of input, so you could put your entire code base in and ask questions and it was free. So then it was used quite a bit. But nowadays we don't see it used very much on our service. Like regular Grok4 isn't used that much and Grok4 fast, Grok4 fast thinking, they're also not used that much. I could say which models are used a lot, if that's interesting to you.
- [25:52]
A
Yeah, actually that would be good to know. Like what are the most. What are some of the more popular models?
- [25:56]
B
Yeah, so since Recently, Claude Sonnet 4.5, the newest like Claude model, is that.
- [26:02]
A
For text or text or image?
- [26:04]
B
Yeah, no, that's a text model.
- [26:06]
A
Yeah.
- [26:06]
B
Yeah, okay. For text, It's Cloud Sonnet 4.5, we added GPT5 Pro yesterday, which is the absolute top model at the moment, to be honest. It's incredibly expensive, but it's also incredibly good. So we see that that gets used quite a bit. And then the GLM 4.6 model, which is open source, which we're very happy about, so it's included in the subscription, but that one is used a ton. It's almost as good as Claude for like programming stuff, but it's infinitely cheaper. So that's become very quickly a very popular model as well.
- [26:44]
A
Yeah. I was going to ask you about Claude. Why do you think that is so popular?
- [26:49]
B
Because it's. I would say they really focus on the developer experience in a way. Like I feel the biggest actual use case right now for people with AI or one of the biggest ones is development programming and Claude just really or anthropic. The company really leans into it. They really focus on building, making it as good for programming as possible. They recently increased their like, input size, how. How much input it could take. So that also really helps. And it's. Yeah, I don't know it just feels good when you're programming with it. I do have to say I don't actually use it right Now. Cloud sounded 4.5 because I'm mostly using the GPT5 model or GPT5 Pro, because for programming GPT5 is for me at least actually overtaken the Claude models. But yeah, so Claude is really good for programming and also we have a lot of roleplay, like roleplay users on our service. And apparently Claude is really good for role playing. It's like really good at sort of building a world and understanding everything happening in it.
- [27:50]
A
Well, good to know. So let me ask you some then bigger questions surrounding your service and AI generally. What is the most absurdly expensive prompt that you've ever seen?
- [28:07]
B
We've had people who spend like $20 on a single prompt because some of these models, for example, Claude Opus, that's like the expensive version of the Claude models. It can be incredibly expensive. Same with GPT5 Pro. So we sometimes saw people who would do like, they would message us like, this has gotten really expensive. I'm like, yeah, you threw in essentially the entire Lord of the Rings into one of our most expensive models and ask it to write an additional chapter. So it's like a huge input and a huge output. So yeah, that's people that have sometimes spent like 20, $30 on a single prompt. But yeah, funnily enough, those people, like, they, it's kind of what they wanted. Like they were happy with what they got out and what they paid in most cases. Not always. Sometimes people are like, oh, I paid so much and came out with this crappy answer like, yeah, sorry man, like we don't control what it puts out. But yeah, so we've had people who did like I think the highest must have been like 30 something dollars or maybe close to 40. I do have to say that's not at all the average. Like the average prompt is probably, I think $0.01 or maybe even sub $0.01. But yeah, it does sometimes, sometimes people just really want a certain answer or are really looking for the very best answer for a very specific thing. And then yeah, some people are just willing to spend for that.
- [29:38]
A
Now AI is one of these garbage in, garbage out, sometimes sort of thing. So these, the better that you can control the tool, the more creative, clever thing you're going to get produced. What are some of the more creative uses of AI or creative or unusual or just beyond the typical stuff that you see people using AI for? That might pique some people's interest to go a little Bit beyond the typical stuff.
- [30:04]
B
Yeah. So I would say this long stories model that we have, I actually love what the guy is doing because what he essentially does is people can enter a very simple prompt of just a story that they're interested in, but then he's written a sort of backend in such a way that it uses different models to create this very creative story. And then the models also create like image prompts. So they create images, they turn them into video, and then they stitch it all together. So it's like an entire workflow where you can just put in a simple prompt and you get essentially like a mini Pixar movie back. Which is honestly just fascinating to me just because of how well it already works. And like, kids love it because they, they enter this simple prompt of like, Spider man fighting, I don't know, Spider man fighting Superman and they get this entire video and they're like, I created this. That's been. Been really fun to see.
- [31:00]
A
Yeah.
- [31:01]
B
What else do people do with it? It. I think what excites me most is people who build like actually useful workflows. So instead of just, you know, talking to the AI in a regular way or using it for programming or role playing or whatever, they build an entire workflow in the backend where they send in a prompt and then it will analyze what's happening, it will feed back to the user. Like, you might want to look at this. We actually do this ourselves in a way which I'm pretty happy with. We have like a bot that's scanning Reddit comments and like summarizing them, and then a different bot is checking. Like, might this be relevant to Nano GPT? What's the sentiment here? Like, is it, is it positive? Is it negative? Especially when people are talking about this, about us, that can be useful. So yeah, I don't know if that's very creative. I feel like the long story of the video model is far more creative. But yeah, just cases where it's actually useful to people and saving you time.
- [31:58]
A
What about people? Are you aware of people who've created entire businesses just using, let's say, your service?
- [32:04]
B
Yeah, I think the long stories guy is actually a business built entirely on AI. Also, like he's a solo developer. What else are people using us for? It's funny because every once in a while people reach out to us about something they're doing with Nano GPT for the business and they're like, ah, yeah, but like, don't tell anyone, please, because they kind of, in some cases they want to Pretend in a way. Like you have these AI agents on Twitter sometimes, right, who can, like, reply to everyone on Twitter or predict stuff or whatever. Yeah, or cause chaos, for sure. This one guy built like an entire prediction engine for Polymarket. I don't know if you know Polymarket. It's essentially this website where you can bet on events, so you can bet like, is Trump going to be president? Or something like that. And he just built this entire flow of models through us, which we were very happy with, which would, like, do their own web searches, feed that into another model. That model would then tell five other models what else to search for. It would synthesize all that and come up with what it thought the odds were. And then because Polymarket uses crypto, it would actually place bets for itself and if it was successful, it would be able to pay again to us to be able to do more model calls and place more bets and all that kind of stuff. So, yeah, that's been fascinating to see and as far as I know, very profitable as well. This one he talked about publicly as well, so this one I can definitely share. But, yeah, that's a fascinating way. So just be aware, if you're using polymarket, you're probably betting against a whole array of AI models.
- [33:46]
A
All the people who say that AI is going to take jobs, well, it's, you know, it creates new opportunities and jobs as well. You have to include that as part of the conversation. What about some of the worst or most bizarre hallucinations that you've seen, where in that moment you said, oh, wow, maybe AI isn't the future?
- [34:06]
B
Yeah, I'm trying to think because it's. For me, that's been quite a while now. But that's also because I use it mostly for development nowadays. I also use it recently for medical stuff. And it was actually really good. That was a very good test for me because I was going to the doctor afterwards so I could check against what the doctor was saying.
- [34:23]
A
Well, tell me about that then. The doctor stuff. Obviously, no medical advice or anything, but just as a summary of events. How would that work for somebody?
- [34:33]
B
Yeah, so I have asthma, so I need to do like a yearly checkup where I do a sort of breath test and whatever. People who have asthma probably all know it because I think everyone does it and you get some results back and then you go talk about them with the doctor. So I got my results back and I was, like, curious. So I just fed in the results to, I think at the time probably Claude or. Or Gemini 2.5 Pro. And I was asking like, what do you see? And, you know, what's good, what's bad, what can be improved, all that kind of stuff. And it was essentially saying it saw my lung capacity had decreased over the years because it could see, like the most recent sort of scans and whatever. And it was saying like, okay, depending on what medication you use, we might be able to recommend some stuff. So I told it what medication I use and I was like, okay, I think we need to switch up the medication and maybe try this and then see how that goes. And also recent research shows that this and this can help. So that was very interesting, but I didn't really want to depend on it because I had the doctor appointment anyway. So day after I go to the doctor, and he essentially says exactly the same minus that one last thing that AI was saying, like, oh, we could also look into this, because this reason, blah, blah, blah. So I asked the doctor like, okay, and what about this new thing that's happening? He's like, oh, yeah, I heard about that, but I haven't looked into it too much. I'll get back to you. Maybe that can be useful as well. And he got back and he's like, yeah, you were right, actually. Yeah, I don't know where you found it, but that can be really useful as well. So that to me was kind of. I hadn't expected it to be that good, but for me, for the medical stuff, it was already fantastic. So, yeah, of course it's like one case, right. And maybe a very simple case. I don't know medical stuff enough to sort of judge what is difficult and what is easy. But in this case, it was incredibly good, incredibly on point.
- [36:24]
A
Yeah, you're going to get all the professionals and the certification, certified people who say, oh, you know, you can't trust. Trust AI. But there's going to be a lot of people who are trusting AI and saving money in the upcoming years. Now I'm looking at a post by Brian Romelle. Probably you know him from the Internet, from Twitter, Sorry. And he, he posted something from Statista, which is basically, it says where AI gets its info, top sources 20, 25. And this says Reddit 40% Wikipedia 26%, YouTube 23%, Google 23%. It goes down, Reddit being the top one. Like how. And he's talking about. And other people talking about how this is. The AI is trained on Internet sewage is his phraseology. How would you evaluate that?
- [37:15]
B
Yeah, well, I think it's true to an extent, because we see Reddit and we think, okay, like this is not necessarily what I want an AI to be trained on, right? It has so much crap and stuff at the same time. Usually when I'm trying to find a solution to a problem I have, and it's like I had it earlier today, my monitor wasn't working on my MacBook and I just couldn't get it figured out. So how I search is like site Reddit.com and then describe the problem because often someone on Reddit will have found the answer, no matter how obscure it is. And then tons of people will reply with like, thank you, thanks, that's great, or like upvote it or whatever. So I kind of understand it in a way because I notice myself doing the same and I think it's kind of in the order of the sources as well. Like Wikipedia I feel like I can rely on to an extent, but it doesn't always have the solution to very specific problems, whereas Reddit does. And then if Reddit and Wikipedia don't have it, there's going to be this random YouTube very often this Indian guy who found the solution to this random problem. So yeah, I think when I think about it more, Reddit actually makes a lot of sense to me, despite also the endless amount of crap that's on there, to be honest. But if it can sort of sort through the bad answers and just see what gets upvoted most and what gets a lot of sort of thankful replies and stuff, I feel like that could actually be very useful.
- [38:45]
A
I think I'm probably a little bit more skeptical on that, but we'll let people make up their mind about that. I think we're sort of at the end here of the questions. I hope this was a useful update for people on Nano GPT. Some interesting ways to use it. And again, this has been Milan Durit of Nano GPT and you can visit the website@nano GPT.com use it in your cake wallet or various other wallets, bitcoin.com wallet wallet and yeah, we'll have all the links and such. And Milan, any final thoughts?
- [39:21]
B
No, no. It's been fun catching up. It's also fun every once in a while I talk about NANOPT and sort of the changes we've made and you sort of realize how much is happening in such a short time. So that's always fun for me. So thanks, thanks for the invite.
- [39:36]
A
Hey, thanks for listening. I could really use your help. Real quick if you could share this episode with someone, engage with me, leave a review anywhere. This helps me to break the technocratic shadow banning that is happening with my brand. And of course, if you really want to escape the technocracy, go to escape thetechnocracy.com privacy tutorial series, books, newsletters, consulting and of course you can leave a donation. Thank you very much.
- [40:11]
B
Sam. Sa.