AI Sub Slave VS NanoGPT with Milan

Hello, and welcome to Freedom Tech Friday, a weekly live and interactive show hosted on the ungovernable misfits, X, Noster, YouTube, Rumble, and Twitch feeds. We go live one hour every Friday at 9AM eastern and 2PM UK time. You can catch up later on the Ungovernable Misfits feed.

On Freedom Tech Friday, we like to cover the latest news, trends, and anything related to Freedom Technologies. That could be anything from Bitcoin, Monero, encrypted messengers, privacy tools, and everything in between.

Essentially, if there's a news item, tool, topic that can help you take back some control in today's digital panopticon, we want to talk about it. My name is Max. I'm joined by my good friend Seth, who is the COO at Cakewallet, and I'm not joined by my robot friend.

He's decided to take a shitload of holidays. He's never working, driving around, living the good life. This is a live interactive show. You can help steer the conversation by commenting live, presubmitting questions or topics, boosting the show via podcasting two point o apps, sending tips, and sharing the show.

Top support for the last show was from welder Ian, who sent 2,222, and he said, great show, boys. Thanks for your support. Without further ado, let's dive in. How are doing, Seth? I am doing pretty well. Like you said, our, our resident robot ran away this week,

but, apparently, he just takes the summers off. It's very, very European of him. Offense, Milan.

No. I'm good, though. I'm excited for this one. Nana GPT has been something that, we've been we partner with them at Cake for a long time, and I've used them quite a bit off and on. And, yeah, I'm excited to to chat through it today. So it's gonna be a good one. Plus came on because Milan's came is on. So

I feel like a I feel like a human, not a robot.

So today, as Seth has just teased, we have a special guest, Milan. He is the cofounder behind NanoGPT, a project that's cutting edge AI models, making cutting edge AI models accessible in a truly user friendly way. NanoGPT combines the best of both worlds, a subscription option that gives you relatively unlimited monthly usage,

plus a flexible paper prompt model. That means instant access to hundreds of the best text, image, video, audio models from one clean interface. No heavy commitments, strong privacy defaults, and seamless crypto payments with Bitcoin, Monero, Lightning, and more.

It's an elegant idea that's been gaining serious traction. The infrastructure has evolved fast from early integrations and API access through handling massive model variety and abuse challenges to becoming the go to platform for people who want real flexibility without the usual subscription traps and data harvesting.

And the team have positioned it as genuinely privacy first, crypto native alternative that's actually usable for everybody, creators, and tinkerers. Milan, welcome to the show, mate. Thank you, and thank you for that far better description of NanoGPT than what I was gonna go with.

NanoGPT wrote it for us. Ah, great. That's that's amazing. No. Thank you for the clean interface because we always worry about our interface. So nice to hear that description. But, yeah, happy to be joining.

Yeah. Thank you for coming on. And we were chatting before we went live. Just everyone in the team using AI more and more in a variety of different ways. All of us doing things that we just were not capable of six months ago. It's exciting to have you on because

a lot of us are leaning into using these, like, subscription models, giving away information because there's always that pull of, like, I want the best. I wanna be using the best. So I can't run this thing at home on the hardware that I have, and there's this constant battle of, like, do I give away all my privacy and all my data

but have the power of this thing? And it seems like what you guys are doing is maybe making that shift a little bit easier. Yeah. Oh, sorry. I was gonna say we try to at least make it possible to use the top models with slightly better privacy. Like, if you wanna use the closed source models,

then at least don't give them your name and your payment information, ideally, not your personal details. So, yeah, we we try to we try to improve it at least. Okay. Thank you, Jordan, for putting such a long intro here.

My reading skills are poor, but let's do the next little bit here. Today, we wanna get into how NanoGBT actually works and what the team's been shipping. But first, let's see set the scene for anyone who's been heads down in their own projects and hasn't tracked the absolute explosion in AI models over the last eighteen months.

In 2026, the landscape has split cleanly into two camps. The big proprietary players, OpenAI, Anthropic, Google, who keep dropping frontier models behind their paywalls and open source or open weight crew that's catching up ridiculously fast.

Can you give us the current who's who, best in class for coding, general chat, images, and video, and split it between the big boys and the open source options that actually compete today. And for the hardcore listeners who want to run things themselves, what's the realistic setup needed to self host something like Kimi or GLM

without it melting your desk? That's a tall order. Okay. So the the top models today, there's a few ways of looking at it. Like, there are a bunch of bench benchmarks. You can also check these on our website if you want, like, the sort of full breakdown.

But I think most people would agree that the latest GPT 5.5 and Claude Opus 4.7, so OpenAI and Anthropic, those are sort of the most general all round leading models, and both are closed source.

I think for programming, that is what 90% of people use, probably. They're also by far the most used models on our website. So Claude Opus 4.7 is by far the most used. We publish, like, leaderboards of what is used most, and I think Opus 4.7 is used, like, five times as much as, like, the number two.

So for general usage, Gemini 3.5 Flash, which came out yesterday, is also or two days ago, I think, is also quite good because it's really fast. So OPUS 4.7 is expensive and a bit slower, and Gemini Flash 3.5 is very fast and a bit cheaper.

So those are sort of the three top models at the moment from the the top closed source labs. And open source is a bit more split, I would say. So GLM 5.1 is the one that's most used on our service in terms of or in terms of open source, and it's really good. Also, for programming, it's really good. It's probably

Claude Sonnet 4.6 level, so, like, a few months behind the the actual top models. They usually say it's, like, three to six months behind the open source models. So GLM 5.1, DeepSeek v four pro came out recently. It's also really good. And Kymi k 2.6. Those are sort of the open source contenders.

But if I'm being honest, like, I really like open source, but they are not competitive with closed source at the moment, sadly. I don't know if it's just a sort of money issue because it's harder to make money from open source than the closed source models. But, yeah, they are slightly behind for the text models.

So, yeah, I think that's sort of the the current standings in the the text and coding. Largely, the sort of top models for text, like, that you would use for general usage are also the best for coding, so there's not much of a difference there.

Mhmm. Yeah. The sorry? No. I was just gonna say, so the the general chat, obviously, the coding side of things, it's really like this frontier models. But what you're saying is that even for, like, general chat, they're still four to six months ahead.

And if that is the case and that's what I've experienced as well personally is, like, I tend to lean on those because they just it's so much better. And then Yeah. Four to six months doesn't sound like very long because in a lifespan, it's not that long. But, actually,

if you're building and you're competing and you're growing a business or an idea or whatever it is you're doing, four to six months is enough to, like, completely fall behind. So it's it's a real lag, especially when things are moving this fast.

Yeah. Yeah. It's so for general usage, like, if you're just chatting and asking sort of the the capital of France, that is not gonna matter much. Right? Like, all of the models will be good at at answering those kind of questions. But if you look at sort of medical advice or programming or just anything where the quality matters,

it's the same as in real life. Right? Like, if you can have a doctor that's, like, decent or you can have the top doctor, then, well, you're gonna use the top doctor, like, 99% of the time.

So, yeah, I I think what you're saying is true. Four to six months doesn't sound like a lot, but at the same time, why would you use something that's four to six months old? In general, you don't.

Like, the advantage then is it's open source, so you can have better privacy and just open source feels better, but you are giving up some quality right now, sadly. Yeah. I mean, what what's the like, I know in in software, open source has the huge advantage of verifiability.

People can actually be building on top of it and making sure that what is supposed to be happening is happening. They can modify it and make changes themselves. But, like, when it comes to AI models, and maybe we're jumping a little too deep too quick, but, like, why why should it matter that they're open source?

Is that something that, like, you or NanoGPT as a team think is a a vital thing that we have to keep pushing on that, or is that not as big of a deal when it comes to AI models specifically when they're not something generally that we're tweaking and modifying

or building on top of. Normally, we're just using the model. I'm kinda curious your thoughts on, like, open source versus closed source, and how important is it in the the AI model world versus, obviously, the software world where it's critical.

I I think it's more of a philosophical thing. Right? Some people are gonna say it's extremely important because we need to make sure that everyone can have access to these top models. It turns into slightly sort of a moot point because

DeepSeek v four Pro, everyone has access. Everyone can access the weights. But, realistically, like, 0.01% of people in the world are gonna be technical enough to actually run it and also have the hardware to do so. So I think from a philosophical standpoint, it's it's sort of a

a more like, it's a beautiful thing thing to say. Like, we need everyone to have access to these models. But I think, realistically, that's already not the case also with open source models. Where it does matter, I think, is just making it more cost competitive.

So the reason, for example, we have a subscription for open source and not for closed source is there's sort of a monopoly on GPT 5.5. Right? Like, you can get it from OpenAI or from Azure or, like, a few other places, but their prices are fixed.

Whereas with open source, you have a bunch of different hosting providers all competing to try and, like, run the model slightly more efficiently so they can offer it to you more cheaply.

So there, I would say it is quite important in just bringing down the price, increasing the efficiency, and that also makes it more accessible to everyone. So if given the choice, I would prefer open source or open weights as it more commonly is in in AI.

But, yeah, realistically, most people are not gonna be able to run the models locally themselves anyway, so they're gonna go through some provider. But just, yeah, open source, we can offer more cheaply, which I think is also important for access.

Yeah. And I think that will be the the big shift long term is when you can actually host these things yourselves, it will matter much more if they're open source or open weights. We're talking, like, on device models on your phone or on common, like, commodity hardware and not having to have a massive GPU rig or massive amounts of RAM.

Think it will be will be more critical then, but the hard part is if no one supports it now, will it will it exist

when we actually get to the point of self hosting it? So yeah. And it's also kind of you can already host models locally and use them, and they're, like, usable. But then again, you get to the sort of three to six months behind or, like, slightly less intelligent, slightly less top of the line than the ones that you can access via API.

Right? So you can you can do it locally, but just it it's gonna be hard to have a model that's as smart as the top model and have it able to run, like, on the edge on even on a MacBook, right, on a strong MacBook or on a phone.

It's most of these models are really big.

Is is there something to be said for, like, some tasks, like, say, image generation and things like that where with my workflow, I might be using, like, a prompt, and then I might regenerate twenty, thirty, forty, fifty times to get something close to where I want it.

If I'm doing that and I was running that locally, in some ways, it doesn't really matter if it's slower. Mhmm. Like, if you start if you start to have things automated to some extent where it's like, right, let's say for this show, well, you know, take the show name,

take the show notes, then spit out some ideas on image generation, what things you know, what sort of things might we have. And then overnight, just keep regenerating. Choose your top 100 that you think are good, and then I'll review in the morning.

In some ways, like, the speed isn't so important because it can do it while I sleep. There's some, like, use cases like that where you think, actually, it does make sense to to run things on your own hardware and, like, you don't have to have such a a an amazing setup.

Yeah. No. I think that's a good point. Right? For image generation, I also think it's more common to run the models. It's mostly it started with, like, stable diffusion or SDXL, and then people would sort of tweak as you do in open source. Right? This happened way more with the image models than with text. So people would make it more

like, make it better at anime or anime. I don't know how you pronounce it. Yeah. Yeah. Or give it, like, certain styles. Like, you might fine tune a model so that it fits with your branding, for example.

Mhmm. That happens way more with image models, and, also, image models are generally a lot smaller than text models. So, yes, for image generation, it's more doable for sure than for text models.

I think that the sort of disadvantage again here is text models or sorry. Image models have also gotten really good when you use them via API. So, like, mid journey is not open source. GPT image two, not open source. Nano Banana, not open source. And those are the ones that if I was going for quality, I would default to those models.

Yeah. I use Midterm. It's incredible. It's absolutely incredible. Like, it blows my mind every day what I can do with it. The only thing that pisses me off is, like, if I wanna do something jokey about Christine Lagarde or, like, she's like a monster or do you know, something that's, like, not even bad,

but it will get banned. It's like, you can't do that. Moderators have stepped in, and it's really trigger happy with stuff like that. So Yeah. It does put a cage around the creativity completely unnecessarily, which is actually fucking annoying. But it's outside of that, it's amazing.

Yeah. But that's actually a good point, though. Like, they are generally more censoring and more restricted than running it locally. Right? Like, the local fine tunes. So, for example, if you wanna do not suitable for work generation, there are not many of the public

APIs that are gonna support it. Right? Whereas, locally, you can do that. So I think quite some of the fine tunes are for for that or for, yeah, like I said, certain styles, but also just generating images of certain people. Like like, as you say, if you want Christine Lagarde, if you want Elon Musk. This

is one of the reasons I think people quite liked the x AI image model because it was not restricted, maybe overly unrestricted, but it wouldn't really push back on this. You could have, like, Donald Trump fighting Vladimir Putin or something, and it would be like, yes. Sure. Here you go. Here you have your image,

which, you know, like, there's a bigger discussion to be had there, I think, on what should be sort of watermarked as this is AI and this is not. And but yeah. So there's definitely a case to be made more so for image models that running locally has advantages for sure. Yeah. Mhmm.

Are there any creative or unexpected ways you've seen people using Nanite GPT at the moment? And so many Yeah. Go on. Yeah. No. I'm thinking so we have quite some role playing users. So they sort of construct a world with entire, like, lore books

and everything around it. So think, for example, Game of Thrones world, and they are a character in that world or dungeons and dungeons and dragons, I think it's called, with entire sort of random generation in there and whatever, and the models will be either the game master

or play other characters in there. I think that's quite interesting. We also know this is an older example. I think I've mentioned this before. But someone who is running a polymarket operation where, like, you have one leading model which can

contract out to different models, give them, like, things to research, look into this, look into that. It gathers all the information, and it essentially places bets. I guess it's not legally called bets on Polymarket, but it places, like, things it will things it will make a profit on.

And then if it makes a profit, it can buy more compute on NanoGPT so that it also has access to a crypto wallet so it can, like, send transactions. It it accesses multiple chains. So that was pretty wild to hear that he was doing that.

Yeah. It's it's many different things for many different people. I also think, in general, the vast majority is just programming, coding, agentic usage, like OpenCLO and such. And in our case, role playing is quite big as well.

But, yeah, it's a lot of cool stuff. We have, like, a community projects channel in Discord where people post the stuff that they're working on, and it's always quite cool to see. So yeah.

Yeah. I think there's there's a lot of unique things that Nano g p GPT enables. I feel like we've maybe jumped in a little too fast too and didn't let you more kinda intro, like, why y'all created NanoGPT.

Yep. I'd be curious. Like, I I know it's it's similar in some ways to other people in the space, y'all but have a very different flare on it, crypto being central to what you do. But I love to hear a little bit more about, like, what was the gap y'all saw?

Why did you create NrJPT, and what really does set it apart outside of just the the broad support for different different models versus if you're using Anthropic or something, you're just getting their their set of models. Yeah. So very early on, I back then, I think ChatGPT

it was only ChatGPT, right, at the start sort of, and they only had the paid option. You could do a subscription. There was no real free model, and I kind of got addicted to this immediately. I was like, this is gonna change everything.

And I wanted to have more people try it out, but not everyone wanted to do a subscription, which makes total sense. But then OpenAI also has an API that you can use. So it's sort of a machine way of sending in a prompt and getting an answer back.

But that was kinda hard to set up for most people. So I figured, okay. I'm gonna make this easier for people. They can deposit, like, 5¢ or something in Nano into this Telegram bot,

and they can ask a few questions. And they get to experience, like, what this is, what is Chechi b t, how can I use it, all that kind of stuff? So that's how it started, and people would deposit, like, 5¢, 10¢. I think I had a hard cap of $1 at the start because I had no clue. Like, I was just having fun building this.

And so I started as a Telegram bot. I, at one point, was like, okay. We need to expand this, and then my cofounder reached out reached out to me on Twitter. And he's like, man, we could turn this into something much bigger.

We're both into crypto, so it just very much made sense for us to do crypto payments. At the start, we didn't have credit cards. We just had Nano. We started adding in Bitcoin Cash, Monero.

And our thinking was kind of many people wanna use AI. More and more people are gonna, like, use it in different ways. We wanna make access as easy as possible. And both of us having a crypto background, we wanted to make it easily accessible to everyone in the world. Not everyone has a credit card,

and we wanted to safeguard privacy. So I have, like, my own privacy backstory. I used to work at the central bank, and I was also very much into crypto. So I was trying to keep those worlds separate by being anonymous on Twitter, ended up getting doxxed.

Like, they were contacting my work to try and get me fired. So I have a strong appreciation for privacy myself. And it just made sense to us that, like, it's sort of become very normal that all of your information is gathered up by Google, by Meta, by all kinds of parties, and you just have no choice really. And

AI is kind of like that on steroids because you pay with your credit card, and then you start typing about, like, your deepest fears, your medical issues, your girlfriend who's mad at you for something you've done.

And all of that is like, I think it's very dangerous and scary that all of that gets concentrated into these very few parties. So, yeah, we we saw it as everyone's gonna wanna use AI, but, also, we need to make sure that people can use it in a way where we're not gonna have a massive honeypot of information just lying around,

where they're gonna be able to target ads very specifically at you or, like, even worse, the information leaks and everything about your life is known. So, yeah, we we kinda saw a gap there of AI needs to be more accessible, and it needs to be private.

And the way you do that is start from crypto basics. So you can pay in crypto. You don't need to make an account. We still have this in place today. There's no need to ever make an account.

We kind of prefer if you don't make an account. We kind of prefer if you pay in crypto because Stripe charges its high fees, and crypto is just more efficient. So to us, it just makes complete sense. And we started from this privacy angle, and we've kind of, in a way, also made it more usable for those who care less about privacy

so you can make an account. For the longest time, this was not possible. You could just not make an account. Everything was stored only locally and sort of if you cleared your cookies, it was gone. Turns out some people don't like that when you get to the sort of more

normie people in a way. So we started adding in you can sign in with Google. You can sync your conversations either to, like, an encrypted storage that we offer. You can put your own password on it so it, like, encrypts it with that. You can even put in your own external storage.

So we've also made it more accessible for those who care less about privacy in a way, but we try to do it in such a way that even if you don't really care, you're gonna get better privacy than what you normally have and just make it as easy as possible for people to get the best possible privacy.

So, yeah, that's that's kind of the backstory of NanoGPT. Back then, it was this little hobby project. And, also, like Seth, props to cake because I think we had the website up for, like, a month, if even that. And we had, like, a few crypto supported, a few models,

and I reached out to Vic for something else. And he was like, man, yeah, let's get this into cake. Like, if it feels like this aligns really well with what we are trying to do. And that was kind of our first big break.

So, yeah, very grateful to Cakewallet. It's still our fondest integration that we have. So yeah. Yeah. I love that. I know it's been one that that our users love and something that we need to make a little bit more native as well because there's the like, there's just so much unique benefit to leaning on crypto

for this type of service. Like, there's so many things you couldn't do with a credit card in terms of user experience, in terms of even though, like, the agentic flow that you mentioned earlier with the guy using Polymarket, like, the ability to actually pay for this stuff

without having to give an AI agent your credit card, which is nightmare fuel. Yeah. Like, there's so much that that enables, and I know that's been core to what y'all have done from day one. I mean, you NanoGPT.

Nano is the name, which is a a cryptocurrency that's been around for a long time and very payment centric. How has the transition been from, like, you said, just Nano in a in a Telegram bot into, like, a full fledged service that now supports Bitcoin,

Monero, other cryptocurrencies as well? What's that transition been like? Has there been any kind of, like, learning curve or things that stood out or surprised you during that process of kind of growing what you accept and and seeing how users respond to that? Yeah. No. There has been many

surprises and sort of crypto works very well in theory, and then you start implementing it and you run into all of the the sort of issues that there also are with with implementing it.

We started with Nano because we both like Nano, and it's perfect for payments. But, obviously, like, Monero is very well aligned with what we do, so it's one of the first coins that we added.

And our thinking was very simple. We wanna give people the ability to use whatever model they wanna use for whatever they wanna use it for. We also wanna give people the ability to pay however they wanna pay. Right? Like, we don't really care that much.

We like Nano. We like Monero. But if you wanna use XRP, for example, we try to also support that. No one uses it, but we support it. So our our thinking is very broadly speaking, meet the people where they are. Right? Like, have them spend in whatever way they want to spend.

And if that is Monero, then great. You know? But if people don't really care that much about their privacy and they have some Bitcoin, it's already better than using a credit card. Right? So we we try to just, in terms of crypto support, pretty much anything we can think of.

If it sees no usage, then we're, like, not gonna try as hard to to maintain it or to keep it up. But most of the coins that we've added have been, like, quite successful. I've seen quite some usage.

We release payment statistics every month just because it's it's more of a personal thing really than business wise because we really just like crypto, but we don't wanna push it down the throat of our users.

So we usually do, like, discounts if you pay in crypto or you get, like, a bonus on your top up or whatever. But, ideally, for us, we wanna have more of a circular ecosystem. So we want more places to also accept crypto.

So our thinking with the payment statistics is if you're a merchant who is starting to accept crypto, one of your questions is gonna be like, okay. Which crypto, and how do I do it, all that kind of stuff. So we try to give some information on these are the coins that are actually used,

and anyone can reach out to us anytime to ask, like, how do you implement this? How are you accepting this coin or whatever? So we try to, I don't know, raise some awareness for, like, using crypto for payments and which coins are used.

But, yeah, been a been a fun journey starting with just, I think, one coin too. I think we have, like, 20 now, I guess, and, like, a lot more if you count all of the ERC 20 tokens and Solana and whatever.

But, yeah, it's a it's been a it's been a long process to get them all added. Do you, off the top of your head, know what's being used the most? Is it Monero? Is it Bitcoin? Like, it seems often it's Monero being used most places.

It's Monero. Yeah. Monero has been the most used coin, I wanna say, for about ten months now. So, of course, when we started, it was Nano, which was used most. Then at one point, Monero overtook it in usage and kinda never gave up its top spot.

So most months, it's either Bitcoin or Nano that is in second place for most use. Mhmm. And then the other one's in third place. And then the rest are quite a bit smaller, to be honest. So stablecoins are used quite a bit. Litecoin is seeing some usage. I could actually pull up the statistics because they should be easy to find.

But the vast majority of it is is Monero, which I think kinda makes sense just because of how well aligned we are with privacy. Yeah. Let me see what else we have. Yeah. Litecoin USDT, TON, like the Telegram coin, Solana, Zcash,

Zano, also kind of privacy coin. So, yeah, we we we support a lot of different coins. If anyone wants to see the actual statistics, just check on the first of the month our Twitter account. Like, we post them on the first of the month for the pre the previous month every time.

So you can go back through all of them and see sort of how it's been progressing. And you said sometimes people reach out and wanna use a certain coin, and you take that feedback and make some changes.

Are there any other wins or moments that you've had from feedback from users as you build this thing out? So much. I I wanna say our entire website is built from user feedback. So people sometimes ask us because we've grown quite quickly

the last, what is it, one and a half years. We it's gone from, like, this fun hobby project to a full time business that's doing well. And they ask us, like, how? Because we don't really do marketing or anything. And it's honestly just from listening to users. Like, users have so many good ideas on what they want added.

So I think even, like, video models. At first, we were like, ah, who's gonna want video models? And users were like, yes. We do want video models, so we added video models. Then users wanted audio models, and they wanted voice, like, direct real time communication.

They wanted the API to be able to build it into their own stuff. Then people really like their privacy. So the TEE models, like the trusted execution environment, it's a longer thing to get into as well. People suggested those to us, and they're like, maybe you can add in those models.

So I I would say, like, 90% of the things that we add come from users having an idea of maybe it would be nice to have this, and then we just add it in as quickly as we can. Same with, like, the recently, we added a local model. So you can in the browser, you can just hit install, and it will download,

like, a small local model in your browser, which you can chat to, but also you can use it for, like, memory, which was another user suggestion. Like, it learns from your conversations. Okay.

So for those kind of things, it's nice to have a local model so not all of that information keeps being sent out every time. So, yeah, local model was the newest addition that was purely because of what you were Is that what you were showing me before we went live?

Yeah. Yeah. Should we do a little demo? Can we can we do a screen share? Yes. Always scary to see if it works when you wanna demo it. Let's see. Okay. I'll go to the website, and then I'll share my screen.

Here we go. Share this window. Does this work? Can you see NanoGPT? Let's have a look. You are sharing your entire screen. I can see I think it's just sharing that you are sharing. Thank you. That's weird. It worked earlier. It's a tiny That's really that's really weird. Sharing and retry, see if it works. Yeah. That's

Otherwise, you can just see how I We tested this as well. It was I was gonna say, this was just four as well. Okay. Let's try again. Share this window. Let's see. Right. Add to scene. Are we getting it now? There we go. Alright. Okay.

Yeah. So this is, like, private browser, fully new session. So, normally, when people land on our website, they can use the free model. So it's selected here, and you just go hi, and we will hopefully give a response back if this works well.

Yes. We do. But, of course, like, all of this uses some remote model. Right? And some people prefer to have it locally. So what we did is if you search for local in the models list, you get this entry, which runs QUEN 2.5, and it's a 500,000,000 parameter model.

And you should be able to just hit install. Of course, I misclicked. Ah, there we go. And it will, like, explain to you it's gonna be a download so that people who are on mobile on data don't suddenly get hit with, like, a massive download.

And what it's, like, gonna do is just download the model from Hugging Face. We picked a very small model because we want this to be usable on essentially every device. So you if you have, like, a MacBook, you can run a slightly bigger model, which we will add in as well.

But this is a model that should work for everyone, and it takes like, I have quite quick Internet, hopefully, so it shouldn't take too long. Let's see. Always takes longer when you want it to go fast, of course.

Essentially, this is our way of trying to make it easy for people to also have some local model running. Because usually setting it up involves, like, some coding knowledge or some technical knowledge, and this is just hit install.

So now it's installed, and then you can select what you wanna use it for. So by default, we try to, like, get you to use the local for everything, like generating conversation titles, using it for your memory, all that kind of stuff.

So I just save it like this. Yeah. And then here, you can see I can select the model, and I will ask, what is the capital of France? To just give it an easy question. And so this is, of course, free because all of this runs fully locally.

Mhmm. And it's a it's a not the smartest model. Right? It's an older model, and it's kinda small. But for, like, these sort of simple queries and for running it for, like, the global memory that we do, which, like, scans a conversation and adds in information that it thinks is relevant for all of your chats so for example, like,

you live in The US in a certain place. It can sort of save that information so that the other models know some background information about you. Oh, this is fully optional, of course. Can I can I just stop you there? Because that that's just made me think. So

whatever you have locally on that thing will then share itself to the other models that you're using so that it has some, like,

ongoing idea of, like, certain themes and things that you might be doing. Is that Yeah. Is that what you're thinking. So Ah. It would get quite annoying, for example, to have to explain every time, like, I don't know, I am male.

Yeah. So certain basic information can be quite useful for the model to have. So what we do with this global memory is every time messages are sent, it can sort of suggest additions to the memory bank that it has on you.

And, of course, like, we have the private settings or the sort of full automatic. So you can have it suggest things to add to the memory, and then you manually add them. You can also turn this on entirely, or you can have it on full autopilot where it just constantly

every message, it scans and it, like, adds things in, and it learns about you over time. But yeah. So we figure, okay. A local model is quite a good idea for something like this. Right? It it's kind of annoying to have to send it all to Google every single time, for example.

So because that's one of the most annoying things I find, like, heavy user of Claude and GBT is I I abuse them so much that I run out. Like, the chat just doesn't work anymore. It's like you need a new chat. You like, we have no more memory here. And then I constantly have to, like, upload about Max. Like, he hates Emdashes,

be direct, blah blah blah. Like, how to deal with me, like, and things that we're working on, all these, like, separate folders to remind the thing that I've already been dealing with, all the things that I've already told it, and it's fucking annoying.

This is a really that's a really good, like, clever way around it, actually. Yeah. Yeah. So we have kind of two memory options. So this one is a global memory, so it spans across conversations.

So, normally, every single conversation is completely separate. We do not link between them in any way. So, also, the sort of model that gets your question doesn't know anything else about you than, like, the hi that you just sent.

And global memory sort of solves across those different conversations. And then we have a different thing which is called context memory. We're not good with naming. We have different kinds of memory, apparently, which kind of extends the the context window of a model within the chat. So, like, where you would normally run into a limit of

like, the model cannot take any more input, we can use context memory, which sort of expands the context by feeding it into a different model. And then, yeah, it's a complicated thing, but, essentially, you have a model that selects the most important parts of that conversation

to feed into the other model. So, like, if you have a very long conversation with a model, it might not be relevant anymore what you said, like, I don't know, five messages in. But the first message might still be very important, right, because you set sort of the baseline for everything.

So what it will do is it will take that first message and some important stuff from the messages in between, but not every single word in the entire conversation.

So, yeah, that that was another user suggestion that we that we received because people were annoyed about, like, the the window running out and having to restart chats. So now you can have sort of infinite chats that just never forget the most important things.

Yeah. That's brilliant. Yeah. Yeah. So it's but, yeah, the local model, we wanna improve on it further. This is just sort of a first version of it. We wanna make it so you can select what sort of model you install.

Because, like, I have a MacBook. I can run slightly stronger, but our thinking is always it needs to be accessible to the most basic devices as well. So then you have to think small again.

But, yeah, it's I think it's quite a cool thing to to improve on. It's it's obviously not the the most profitable thing profitable thing for us if people run the models locally, but, like, we genuinely care about people having privacy. So we think something like this is is a net positive to have.

Yeah. And it it shows the ethos too of wanting to really enable people and not just how do I maximize profit from this. There there are plenty of AI companies seeking maximum profits and doing whatever it takes to do that. So Yeah. I definitely appreciate that.

I think, like, on the on the cost side, I know everyone thinks AI is just insanely expensive, and also other subscriptions have kinda gotten crazy and usage limits and all that. Like, what what's the expectation for people jumping into Niner GPT in terms of cost?

I know it's gonna vary per model, but maybe just a general idea of of what kind of a a normal user would expect, and then any tips or tricks to to minimize that or to to optimize what models you're using to save a little bit if you don't need something for a specific just kinda, like, yeah,

tips and tricks and ideas around around what the cost is like. Yeah. So we have many people who do the subscription. So the subscription that we have is to open source models, like all of the top open source models, and you pay $12 a month, and then you can use 30 to 60,000,000

tokens per week. It's gotten a bit more complicated because we wanna keep the price low, but the models keep getting more expensive. So for some models, for, like, the new very big models, our costs are also higher. So then we give, like, slightly less allowance in a way.

But that is the, like, by far the cheapest way to use these models because you pay $12 a month, and I would say 95 percent of people never hit that limit. So that's one way to do it. What many other people do is they put, like, $1 into the website, and they start with the open source models because those are relatively cheap.

And that actually lasts you, like, quite a bit because the open source models are just very cheap. Where it starts getting more expensive is when people use this for programming, for example, because usually, have, like, a big code base that you are feeding into the model. And how the pricing works is you pay per

token, and a token is essentially a word. That's a sort of bad way of explaining it because it's not one word. It's one token. But the more words you are putting in, the more tokens, and you pay for every token.

So what we see is there's a very big difference between those using it for questions and for, like, back and forth typing manually and those using it agentically in OpenCLOB for programming, all that kind of stuff, there, the costs can really jump.

What some people do for programming is they will have an expensive model like Claude Opus 4.7 or GPT 5.5 as a sort of architect and, like, orchestrator, and it will like, it gets used to write up the plan and to divide it into small parts, and then those small parts get handed off to the cheaper models because

it's, like, simple enough that the big model can verify afterwards. Like, yes. This was done correctly. Quite some coding harnesses or, like, agents also support this. They have, like, different tiers of model that you can set.

So, yeah, we we try to help people a bit that way, but it's some people just wanna always use the best, and and some people wanna always use the cheapest. So we also have you can choose the model yourself, or you can use the auto model, which routes to the best model for your task.

And then you can set premium, standard, or basic, I think, are the names. And then basic only uses very cheap models, normally use normal uses, like, slightly more expensive. And if you go for premium, you get, like, OPUS 4.7 and such.

Mhmm. So we try to give people some ways to control their spending that way. You can also set hard limits, of course, on on, like, your spending on your API key. If you take the subscription, you can compare how much you would have spent on NanoGPT if you did pay as you go instead of subscription.

So we're trying to, like, make it as transparent as possible to people what they would what they would spend. We show, like, average cost per message for all of the different models and sort of the the range of of costs for it.

But, yeah, it's at the same time, it's hard to answer, right, because you can go wild.

Yeah. It's one of the things with it being so so varied in terms of what models you can use and how you wanna use them that, obviously, you could you can spend, like, crazy or you can go really cheap if you need to go really cheap. And Yeah. Also, like we talked about before, crypto does kind of enable that tiny usage

pattern. If you just need a super simple chat app or you just need something to do orchestration or something like that, you can use something that's very lightweight and very cheap and Yeah. Don't have to pay for a massive subscription just to get access. Exactly. So we have some people who have built it into certain workflows.

I don't even know what, right, because we don't see. But they do, like, tens of thousands of prompts per dollar because they're just small prompts to cheap models, and they're probably doing some, I don't know, automated thing.

And then it can be super, super cheap. And then you also have people who are spending more than a dollar for a single message because it's, like, entire code bases into Opus 4.7 and everything in between. So I think the average user I don't know. I would have to look it up. But, like or the median user is better probably,

spends way less than a dollar per day. But then there are some, like, big outliers, of course. Yeah. With with things like if people are using this for, like, Opus 4.7 or these, like, the the more expensive models, does it work out, let's say, someone like me who might sit in front of Opus for, like,

five hours straight, just, like, annoying it with with changes to a website and building just back and forth and back and forth. Presumably, if we're using NanoGPT, we're gonna be spending more money because of the setup the way it is.

Is it so much more that it doesn't make sense for tasks that are extremely vanilla and boring and you couldn't care less whether anyone knows or not. Like, I don't personally care

if it leaked that I'm changing the orange to a slightly different orange on a website that's going live anyway in the fiat world, and it's the most fucking boring thing in the world you could imagine.

That doesn't matter to me. But is am I actually overlooking something there, or does that make sense? Because bask you know, different baskets of, like, boring and maybe I want some privacy.

Yeah. I I mean, it's always hard to answer this because it could be that you're sort of changing something on the website is very simple and then quickly done. And it's the only thing you're doing that week, and then it's, like, I don't know, a few dollars.

Or it could be that it, I don't know, ends up taking millions of tokens or, like, dozens of millions of tokens, and then a subscription is usually cheaper. Mhmm. So I don't know. I I find it very hard to say.

Usually, I think what people do is they put in a few dollars, and then they try it out, and they check their usage page where you can see exactly what you've spent. And then you can compare for yourself, right, like, whether a subscription is cheaper or using it via us.

I would say for many projects, it doesn't matter maybe to you, the privacy, but I think people kind of default into using the nonprivate way then. I think that's the the sort of dangerous thing about it because

you start using it like that, and then for another project, you you do the same thing. And it turns out there is some personal information in there or something you wouldn't want to have stored on on other servers.

So, yeah, it depends, honestly. It's a bit of a trade off everyone has to make for themselves. I also have a coding subscription, so I'm not against using subscriptions. It's just yeah. You have to sort of make that decision for yourself usually.

Fair. We have some quest we had a question just come through. I don't know why I can't share it on the screen, but it's from doctor I can see the chat. Can I use the Claude CLI in NanoGPT?

Yes. Well, so not inside NanoGPT, but, like, the Claude CLI, you can set your provider to NanoGPT. So that is true for Claude, Codex, OpenCode, KiloCode. Any coding harness or agents that you can think of, we will work.

So we have an OpenAI compatible API is what it's called. There is one industry standard, luckily. Well, largely, everyone abides by this for how you are supposed to send in messages and get them back out again.

And what that means is we align with that standard. So any sort of coding tool that exists or anything where you can use OpenAI, generally, it will also work with us. So, yes, you can use the Claude CLI. Many people do, actually, I think.

Talking about AgenTek AI and Claude Claude Claude bot and things like we were before we jumped on. Just wondering, as somebody hasn't set this up and and, like, taken that next step, you said you use it quite heavily. I know a lot of other people in our ungovernable circles are.

How does that link into your service and, like, say someone's let's take ungovernable as a as an example where it's like, okay. We have the same

things that we do every week to some extent. It's like, every Friday, we have this live show, and we need to prepare show notes, then we need to get a transcript, then we make clips, then we do image generation, you know, then we've made posts that go out, blah blah blah. We have a certain workflow.

How would someone like me who's not that technical set something like that up and link it in and use your service So it's all crypto native. It's all leaning on the different models that are gonna be best. Like Yeah. Like I said, image generation, I have my preferences.

But is this is this something we can link with with what you're doing? Yeah. Quite easily, I think. So, honestly, the best way to do it is probably to ask an AI model to set it up.

It's very cliche, but so what we've done is we have an agents page, which is it's just nanogpt.com/agents Okay. Which gives information about all of the different endpoints that we have, like, of the different APIs that you can use in a way that it's also very simple for any AI model to figure it out. So how it would generally work is

it would sort of for this show probably wake up every Friday sort of. You'd like, it would get a ping. Like, okay. Start this process. Mhmm. And then the process would be gather information about the guest, for example.

So we have different plugins, different sort of scrapers, you can call it or whatever, that can search Twitter, that can search Reddit, Facebook, or just general googling, web search, whatever.

So the bot could use all of those different APIs to gather information about me, and it will learn some of my background and whatever. And then probably,

it would set up some other model that would then sort of collect that information and feed it back to you. That would honestly be super simple to set up. We also have just like we have the auto model, we have, like, claw model. I think I I don't know if it's called claw model. We have, like, a few easy

to find models that we know are good with OpenCLaw and with Hermes agent, which is sort of the new version of of or, like, it's an alternative to claw, I would say. Setting that up, like, honestly, I I find it hard to say how difficult it actually is because I let AI set it up, which is very meta in a way. But

and then so if you tell it to to add in image generation, for example, it will, on that same agent's page, be able to see, like, okay. This is the API that I can call. I can see all the different models.

And in your case, it would probably use Midjourney or GBT image, whatever you prefer. But this I feel like this is kind of what we are good for in many ways, this agentic use,

because we have so much different stuff in one place. So for for agents, for claw claw bots, whatever you wanna call them, it's kinda nice that they can just have one API and they discover everything.

Yeah. They they discover image, video, web search, anything they can think of. And then if you have it set up a crypto wallet for itself and you fund it with some funds, it can also top up again if if it starts running out. But then, like, to be fair, I I think that's still a bit gimmicky unless it's also earning money somehow.

But, yeah, technically, like, people here could tip the bot. Like, it could post its address in the chat, and then it could use that to, like, learn more again about the guests or generate more images to go with the show.

The technical way of of how to set it up, honestly, it's it's gonna be way better to ask an AI model to, like, how to set it up than asking me. The only thing that I would contribute here is don't just blindly give it access to your email and to your local files, all that kind of stuff. Start it out in a sort of sandboxed

way where it only has access to Yeah. Separate computer ideally is what I was was thinking. Yes. Would be best. Ideally, yes. Yeah. Or something to do about, like, a a VPS, like, a virtual private server.

Okay. Yeah. And like you're saying, just ask the AI to set up the AI. It's it's weird because my workflow with anything technical for years now, for, like, the last nine years at least, has been have a problem as technical

message q and a. Q and a is like my AI. It fixes that robot fixes what I need fixing. And now he gets an easy time because I do just ask AI to do all this stuff. It's it's so good. I as I was saying beforehand as well, it's it's crazy to what extent we start relying on these models. Like,

we have one that's checking our Discord chat for, like, new ideas or for bugs, and it will do, like, a first

sort of draft of, like, we could improve it this way. We have others that are constantly looking for, like, new models, new providers, so they just refresh different websites. They're like, oh, something new got added. Mhmm. Like, it's so many things that just take work work out of our hands.

The way we see it, our added value in a way, is our communication with people. So we try to spend a lot of time in our Discord chat to sort of see what people want improved and what, like, what else we could do.

And all the support stuff, we always actually type the emails ourselves or the messages because we feel that's still quite annoying to get an AI reply. Yeah. Yeah. Yeah. But for just building and integrating and staying up to date with stuff, it's it's a magical time. It's amazing.

Okay. We're getting up to time now. That has absolutely flown by. I know Seth has a hard stop, and I just wanted to get to the other messages. We did have one from Otto. He tipped 015 He said, why sometimes the system is so unstable?

Sometimes for days, Kimi or GLM is unusable. Yeah. So I hope this is a lot better now, actually. But early on, when those sort of when new models come out, the open source models, there is usually a sort of overload in demand. And we try to spread it out over many different providers,

but I I a few weeks ago, it was still very slow. So it would take, like, sometimes half a minute or more to start getting your response back. So I don't know if that's what he was referring to. I think lately, it's very good. We publish our time to first token and TPS on the website. Like, you can see it in the models page.

And right now, I would say it's it's quite good just because we keep adding more providers. Like, we keep growing, and then more providers get like, pop up, and we add those in as well to be able to route even more.

But yeah. So that's just unfortunately, sometimes models still get sort of overloaded. Okay. Great. Well, thank you so much for coming on. We have now hit time, and I could have chat all day really about this stuff. We'll have to maybe get you on for a confab in a few weeks or something because this has been absolutely fascinating.

Thanks to everyone who's joined us live. Thank you for the questions. And, yeah, I really appreciate your time. Yep. Thanks. It was nice being here. Before you do, jump off, do you wanna just direct people where they might wanna contact you or, the website and that kind of stuff? Yep. So it's very simple. Nanogpt.

There you go. Nanogpt.com. Or on Twitter, my name is Milan, m I l a n. It's Dutch. You can probably more easily find me by clicking my name in this in this podcast somewhere. But if you if you reach out to NanoGPT,

it's probably gonna be me answering you anyway. So if you wanna reach out and give a suggestion or bug report, anything, just contact us anywhere. Brilliant. Alright. Thank you so much. Great chatting. Thanks for the time, man. Yeah. Thank you. Alright. Talk to you guys later. Bye.

Thank you for listening to Freedom Tech Friday. To everyone who boosted, asked questions, and participated in the show, we appreciate you all. Make sure to join us next week on Friday at 9AM EST and 2PM London.

Thanks to Seth, Max, and q for keeping it ungovernable. And thank you to Cake Wallet, Foundation, and my Nim Box for keeping the ungovernable misfits going. Make sure to check out ungovernablemisfits.com to see mister Crown's incredible skills and artwork.

Listen to the other shows in the feed to hear Kareem's world class editing skills. Thanks to expatriotic for keeping us up to date with Boost's XMR chats and sending in topics. John, great name and great guy, never change and never stop keeping us up to date with mining news or

continuing to grow the Mestadel. Finally, a big thanks to the unsung hero, our Canadian overlord Jordan, for trying to keep the ungovernable in check and for the endless work he puts in behind the scenes.

We love you all. Stay ungovernable.