Answering your questions

Hello, everybody. Welcome back to FreedomTech Friday, a live and interactive show taking place every Friday at 9AM eastern, 2PM UK time across the ungovernable network. Each week, we dig into the latest in FreedomTech, including Bitcoin, Monero, encrypted messengers, privacy tools, AI, and everything in between.

Essentially, if there's a news item tool or topic that can help you take back some control in today's digital Panopticon, then we wanna talk about it. This show and the topics we cover are powered by freedom.tech,

a daily news desk that uses AI to monitor hundreds of sources so that we can continue to bring you the signal every single week. I'm q and a, and I'm head of customer experience at Foundation. And as always, I'm joined by Max, the head honcho of the Ungovernable Network, and Seth, who is COO at K Wallet.

Freemtech Friday is live and interactive, and we would love to have you help steer the conversation by commenting live, asking questions, boosting the show, or just sharing it with your friends.

Top support from last week's show with Max and Seth covering the intersection of AI and Freedom Tech comes from... Well, we actually have two. The first one comes from a friend of the show, Aitmithrande, who boosted 0.05003 XMR.

And he said, Max was asking about entropy calculations. The BIP 39 word list has 2,048 words, which works out to 11 bits of randomness per word. Six words gives you 66 bits.

A random password using upper and lower case letters, numbers and symbols has 94 possible characters to choose from, which is about 6.5 bits per character. So a password of 10 truly random characters is roughly as strong as six random seed words. There you go. The important thing to know there is that I was right. It's...

It has got more entropy because 10 random characters as opposed to, what did you say, five words or Six. Whatever it was, six words, like, 10 characters, six words. There's a lot more characters across six words, most probably.

Yes. Thank you for your support, Ape. Thank you, Ape. The the second top support came in no less than thirty minutes ago from expatriotic who boosted 0.06 x amount, juicy boost. Yeah. He said to send some love to bill and kione dot org.

The the letter from kione just dropped on the rage. I forget what number this is now, but... Seven. Number seven. And he... He's saying it was heartbreaking hearing

what these two innocent men are living through, never mind their wives and the financial component to it all. So if you haven't donated lately, do it. There is an XMR address that was set up by non Crayon eating Max, presuming he means Tanner Hill, and sends that or XMR to Crayon eating Max to sustain his his lavish lifestyle.

Oh, that's that's my... Thank you Thank you to to Q for letting us have more characters than X and R channel. You're very welcome. Thank you for your support and for continuing to raise the awareness for for Bill and Keone.

Very important topic for us to continue to to talk about. Thank you both for your support, Ape, ExPatriotic. Without further ado, I'm back from a short break. Max, Seth, how are going? Good. Good. Good. I am... I'm doing well. Excited for this. I love q and a episodes.

It's always my favorite part of speaking or anything like that. It's just getting to getting to answer people's questions and digging into ton of different freedom decks, so I'm looking forward to it. Good to see that y'all found somebody to do your math for you while while while you're away too. So Somebody feed better than we are. Yep.

Better than... Even than AI. So Yes. Yes. He is. Yeah. Good over here as well.

Very much looking forward to having the new notes from the inside seven on the feed. We have someone special, very special lined up to read that next one for any of the people who aren't so good at reading like myself. So that'll be up live. And, yeah, just to reiterate, make sure you go and donate

to look after the people who have looked after us and excited for the show.

Awesome. I am cognizant that the chat doesn't appear to be showing up. I have activated it, and I'm not sure why it's not showing up. But if somebody could give some of the comment live on Twitter something like that just because it might be not showing historical stuff before I enabled it or something, but help us out, guys.

There we go. Thank you, Seth. Perfect. So if you have already said good morning, Sebastian, Jordan, Renekie, Bon, good good morning to you all. Thank you for for commenting. I didn't intend to ignore your chats. I just forgot to turn it on before you posted.

Yeah. Drop them again for for your moment of fame. Yeah. Let's get into it. I I also love these ones for two reasons. Number one,

it gets you guys involved. And number two, it makes my preparations for the show very easy. I just have to copy and paste some questions. So thank you for doing all the hard work for us, ungovernable crew. We we do appreciate you. And thank you to Max and Seth for holding the fort last week while I was away.

I did listen back when I was in the airport on the way back from Portugal. And, yeah, it tickled my fancy. I'm I'm I'm very much loving the the AI chat at the moment. But... I'm just shocked you listened to us when you didn't have to.

Say again? That's I'm shocked you listened to us when you didn't have to. That's true Cautious. True friendship, true dedication. I don't what that You'll never get enough of both of yours, dulcet tones, nonstop.

We should do it. We should do, like, a a twenty four hour one so I can just tune in. Put... Have you guys sent me to sleep or something? Well, you... You've got, like, 550 episodes or whatever it is that you can just listen through.

That'll let anyone sleep. Yeah. Yep. Problem solved. Yeah. The problem with that is I'm on half of those shows, I hate listening to myself.

Oh, yeah. That's fair. That's fair. We talk all about AI, and you can't even figure out how to just trim out yourself from all the episodes and make them better? Yes. Yeah. Claude, if you're listening, get on it.

Yeah. So no guest today. This this show is is all yours, ungovernable crew. We we put a call out earlier in the week, and you did not disappoint. Thank you all very much for pre submitting questions.

Had loads across x, noster, some boosts covering everything from local AI rigs, multisig setups, you name it, we've got it covered. We are never gonna get through them all, but just know that we now have a backlog and we will continue to to pick them off.

And some of them might actually turn into full topics as well for for, like, a full weekly show. So we are not ignoring you if we don't answer your question today. And also just a worth of reminder that if you boost live,

you go straight to the top of the pile and we'll talk... We'll answer your question straight away. Well, thank you all to people pre submitting your questions as well. Without further ado, I'm gonna continue the the AI thread with the first question. I'm gonna pose it to you, mister Seth for AI, first and foremost.

Honmaru project, again, I apologize to anybody whose name I butcher today. They they said... They posted on x and they said, local models need vram that small teams and solo devs cannot afford.

But handing your repo to a cloud AI is an op sec risk. What is the answer? And you're not allowed to say more money. Yeah. I'm not allowed to say more money. I I think the main thing I would push back on there is the actual VM requirements for a useful model. Like, yeah, you're not gonna get Fable level performance on your 16 gig

MacBook Air. Like, it's just... Obviously, there are limitations, and the more hardware you have, the better intelligence you can get. But the caveat is that there are really, really good small models now in the, like, the 27,000,000,000 parameter variety.

Specifically, there's Quinn 3.827 b, which will fit on... Really, you would need something that has, like, 24 gigs of RAM or b RAM or more. So like a MacBook Pro with 24 gigs or more would be excellent for that, and that is, like, a killer model for coding especially.

Really, really works well. And if you if you already have, like, a a low level cloud subscription or something, you can use Opus 5.5 to orchestrate and have it work your local model to do all the current work, to make the most of your limits there. And then there's actually even a new one that literally just dropped yesterday,

which is a a further quantization of that model, which only needs, like, 12 gigs of RAM or vRAM. So any 16 gig GPU can easily fit it. If you have 16 gigs or more of RAM, technically, you can fit it. If you're running like a a MacBook that's a relatively recent m series chip and 16 gigs or more of RAM, it should be pretty performant.

But, obviously, just like everything else, the the performance and the intelligence scales with how much hardware you have to throw at it. But, honestly, for the vast majority of tasks people do, even those

27 b models would be sufficient for all your agentic stuff and dev work as long as you have something larger to drive it or you really know what you're doing in terms of coding at the same time. Can I add to this as well on Maharu project?

I think we covered this in the last episode that we did last Friday, but financing, especially in The US, is a good option because, as we said, it's kind of... I think it was, like, $30 a month more to do the financing on a top spec Mac studio as it is as a quarter max subscription.

So I wouldn't write that off just because you're a smaller team and maybe you don't have the money to put up front. That is another option that you could look at.

Yeah. Yeah. And those are kind of insane how cheap those are at this point. And like I said, like, if you're a solo dev, you don't need a a Mac Studio Ultra two fifty six two fifty six gig probably. Like, even just getting, maybe a higher level MacBook Pro than you would get normally.

It has 64 gigs of RAM or something. So you can run all your tasks and have a model that takes 16 or 20 gigs of of RAM at the same time and not not have issues. Mhmm. I was just asked if you could just quickly repeat those those model names that you mentioned, Seth. I don't know how you couldn't follow. They they just flow off the tongue.

Quinn 3 point 827B? Orca sac dash 227B? Like, come on, guys. Come on, guys. Just listen harder. These

these AI guys are almost as good as Bitcoin as at naming them. Yeah. Our own our own lingo now for sure. Yeah. I, I dropped the name of the Quin one, which is the kind of the base model that is the most commonly used, but then the the smaller distillation that just came out, I just also dropped. Well... Okay. So explain to me this.

It's Twitter chat. I have a link from Twitter. I wanna post the link from Twitter into Twitter chat. Mhmm. I can't.

How does that make any sense? Bring it up again, mate, Elon. He'll be doing some something, won't he? Yeah. I will drop the name of that one though, at least. Hopefully hopefully, that's not censored. There we go.

Thank you, sir. I'm gonna try and group these so we can have sections. So sticking with AI, Max, I'll let you hit this one first. And I'm pretty sure there's probably gonna be some sort of gray area answer here because there there's no kind of black and white.

Maple AI versus local LLMs. Like, I know I know you lean more into the kind of privacy side of of AI usage, so I'll let you hit this one first. What's your takes?

Yeah. I think Maple quite a lot. I still use Maple quite a lot. I do find I run through my limits pretty fast,

and I find the same with Nano. Like, if I'm doing anything heavy use, I run through, like, a quite a lot of money quite quickly. So I would think, like, if I had the hardware and if I took my own advice and someone would actually give me financing, I would buy the hardware and run it myself because I think you can run through

more more money more quickly. And, also, it's never gonna be

as good as having your own hardware. That'll be... If that answers your question.

The the... They both have their pros and cons. I mean, obviously, like you say...

Oh, I know what Seth's gonna... The the Maple AI side is you you don't need pretty much any hardware. You just use your existing machine and you're offloading the computer elsewhere in a very privacy securing way. But like you say, you're gonna hit your limits pretty quickly. The flip side to that, obviously local LLMs is that you

need some form of hardware. Now again, that could just be your existing laptop and depending on how much RAM you have and storage and stuff, then that obviously limits the the capability of the model that you have.

But I'm a firm believer that the gaps do seem to be closing the models

or the capability of smaller models like Seth just alluded to is is improving, and I think that that trajectory is going to continue. So I think this is gonna be much less of a of a difficult question to answer, especially in like twelve months time. I think there's probably a a place for both.

If you need to do some big, you know, heavy computational task that need frontier level, but you also need the privacy, then definitely there's a place for something like Maple AI, which is obviously gonna cost you, you know, you're paying an API cost essentially for for your tokens.

But, yeah, I think there's place for both. Seth, I don't know whether you've got any anything to to add to that or to kind of contradict us.

No. I I mean, I think the only thing is something that I've been harping on a lot, which is, like, for most people, the best of... It's it's the best of both worlds. Like, combine the two. It's not one or the other. So, like, something like... I know Maple's a little bit harder to use

directly via API. I think that's gotten easier since I last tried. But if you use any kind of harness that can use multiple providers together,

you could run, like, one of those smaller models we were just talking about if you have a a lower level MacBook, but one that's good enough to run them or something like that. And then you can use the Kimi k three via Maple to have a perfectly private orchestrator,

but you're not having to use your API tokens, and that's a lot of money to do everything through that model. And I think that that would be... I mean, that's really the ideal scenario for almost everyone.

Probably even me, honestly, as I have an insane amount of hardware. I still can't run something like a k m a k three locally. So I think that that should be the workload that people kind of explore is combine the best of both of them. If you can use privacy preserving approaches for the orchestrator, that's even better.

Yeah. I'm obviously very excited about the way that the space is right now and and the trajectory of things and the way they're progressing. I had a bit of a heart take and I'm curious, probably you, Seth, whether you think that this is me, the equivalent of,

I guess, Odell's 200 k by conference day famous shout. And I might be setting myself up for failure here, but I genuinely think, like, in the next twelve months, we're gonna have, like, OPUS four six or slightly higher level intelligence available for mobile phones.

You said in the next twelve months? I know what Seth's gonna say. He's gonna say, get back in your box. What the fuck are you talking about? In the next year, I mean, I I don't think I would argue with that. Really? It's possible. No Like, if you just think about the pace

of this densification thing, we talked about this a little bit last week where there's this this new law that's called the densification law, which is basically trying to quantify how rapidly intelligence is shrinking in size in terms of AI models. And the the small models we talked about before are Opus 4.6 class.

Yep. And they're 27 b. They're, like, roughly 24 gigs is what you need to run them right now. High end phones already have twelve, sixteen gigs of RAM. I mean, I I could definitely see in another year we could get there. We we already have this smaller model that we just got that has

almost the exact same performance as the the full QIN 3.827 v model, and it's only 12 gigs. Like, I I I absolutely believe that we will be there within a year and have something that's OPUS 4.6 class on your phone. Now the performance,

will it be as fast as Opus 4.6 was when you were using it through Anthropix API? Mhmm. Probably not. But you'll have the same level of intelligence, and you'll be able to just start something, and it'll be running on the background on your phone all the time.

You could do a full coding task on your phone, which is kind of an insane thought. I know. I hope so. I'm here for it. I'm all about it. As I said at the top of show, guys, if you if you boost live, you go straight to the top of the pile. So we've got an add on sending in 0 point 001 XMR via XMR chat.

Max, I'll pose this one to you first. Ever considered using AI to translate ungovernable shows to make them available in other languages? Mister Beast credits a worldwide or...

Audience as a big reason for his success. That's a great idea. That is a great idea. I... I've not only thought about it. I've actually done it about

two years ago, a bit more than two years ago. Jordan will vouch for this. We had a long conversation about it where I was saying, I want this in every language. And I was trying to do it on 11 labs, and I translated a few into Spanish and Spanish audio.

And we were gonna do Spanish then... What are we gonna do? Spanish, then French, then Chinese, I think, something like that. But I was hitting token limits Yeah. So quickly that I did, like, half a show, and then I'd hit my monthly limits. And I was like, fuck. We got...

You know, that's just one language. Mhmm. So somewhere I've got... I'd have to dig it out on my laptop, but I've got me and John talking in Spanish for for a whole episode. I need to hear that. Yeah. But, obviously, that was to... Yeah. It was before I fled The UK two, two and a half years ago. Jordan will know.

I definitely like to revisit it. It's it's definitely something we'd like to do, and maybe now it's possible. Probably now it's possible. Yeah. I don't doubt that for a minute. I think

when we... I keep alluding to this uncovernable lab that we're we're working on. I think when we get that set up Yeah. That's definitely going on the list. I definitely wanna hear what I sound like speaking Mandarin.

I might I might have one of you in Spanish as well, actually. It was before Seth's time, but I might have some of you as well. But, yeah, I think, like... And then also thinking about which languages are... My thinking is Spanish,

French, Chinese, but there might be a better one. Like, Spanish is so widely spoken. French is pretty widely spoken all across Africa as well, and then, obviously, Chinese. But maybe there's some other options there as well.

Yeah. This is, the perfect task for local AI. So I'm excited for you to, like, to get that set up and be able to have this as something you just run-in the background all the time when you're not doing other tasks.

Yeah. Thanks for the nudge. And I will dig out that that audio because it'll be funny. Definitely. Yeah. Bond just posted... And I'll take this one. I keep hearing people talk about having multiple agents working together with different models somehow,

but never get how they're doing it. I mean, there's multiple ways you you... That you can cut this. The most common among which is like the the open claw harness or the Hermes or something like that where it's kind of, you know, a front end for you to plug in, you know, one or... Well, any number of models really,

generally speaking for for different tasks. So you might have an image model, you might have a mixture of experts, you might have a coordinator model, you might have one that's, you know, like Whisper that's for transcription, like to turn max into Spanish and things like that.

That's the most common way to do it. And and obviously, if you've got API services as well, you can you can leverage those within that system as well. So you could have like a hybrid of,

you know, your your chat GPT subscription, a Kimi via an API, and the local quen model as well. So that you can use all of them to their own strengths and obviously to minimize your costs to have like the the local stuff doing much more of the of the hard work stuff. But that's kind of the way that I'm

planning to do it when we get the lab set up. But obviously, we've got mister AI who can speak far more eloquently about this because he's literally being setting exactly this up for the past month or two months. So, Seth, what about you?

Yeah. I mean, you you hit the core details there for sure. I think the core thing is if you're doing this all the time, find a way to automate that flow. Like, there's a really good plugin for... I believe it's for Claude code and codex called PilotFish

that basically does this, where you can set different tiers, where you have what your large model is you're gonna use for orchestration and verification, and then your smaller model that you're going to use for for actually doing all the grunt work for writing code, for doing research.

And so you just invoke that skill and tell it what you wanna do, and it it just does the rest. So you really want something like that, so you're not having to, like, explain every time what you want, like, how you want to break out the models.

So I ported that to OhMyPy, which is a a fork of pi, which is a CLI harness that I think is just absolutely fantastic. So I can just do slash pilot fish, and then I say what I want, and it just work... It handles all the orchestration, sub agents, everything else for me without me having to do anything else.

So I think that's that's the ideal, is if that's something that you're regularly doing in your workflow, you do that. But I also... I don't use that for everything. For a lot of tests now, I'm using our smallest model that we're running locally directly

and just letting it do its thing. Sometimes we use a large model on its own, but nice to have that flexibility. So, yeah, I think that's that's normally the go to, but you can use that in Cloud Code today. And I think something a lot of people don't don't know is that

at least Codecs... I had... I think Cloud Code as well, you can connect to local local models. I've definitely seen people do it in Codecs a lot. I I don't know that I've seen it as much in Cloud Code, but it should be an option on both.

Thank you for that. I'm gonna make this the last AI question. X patriotic tip 0.01 x m r. And he... I'm gonna abbreviate this one. But he said, basically, what can I run on 32 gig of gigs of RAM Optiplex machine?

Seth. That's going to be very slow just because you're... It's gonna be a very old CPU, I'm assuming. Optiplex is normally, like, really, really old systems.

So hard to say for sure. You're gonna want some... Want to run something that's very, very small and very lightweight.

I I... You could try, like, the Quinn models we talked about because they would fit in the amount of RAM that you have, but don't expect it to be fast because you're running in a CPU on older slower RAM as well. So it it it is gonna... It's gonna be a little bit sluggish.

Okay. Alright. I've just lost all my questions. So here we go. Johnny Iverson. Max, I'll let you take this one first. A gun to each of your heads, which three hardware wallets are you choosing for a two or three multisig today?

Using Sparrow as the coordinator in our setup. Right. So he's making a presumption. And also, any thoughts on utilizing Liana for decaying multisig in a traditional two or three. So, Max, you hit the first one and then I'll take the the second one about Liana.

Now, obviously, I'll I'll just chime in there and say that this show is sponsored by Foundation where I work. So I'll just spell that out before people start putting their finger. Yeah. Only I would say Three cold cards. Right? Yeah.

Only the old firmware, though. Yeah? Yeah. Yeah. Yeah. Fucking hell. That snake's crawling back. Yeah. I think I would probably say seed signer and passport would be... I'd I'd, like, have probably two passports to seed signer.

I've not tested. And I say that because I've I've not tested any other bits out, and I only wanna say what I'm comfortable with using myself. I don't know if that's... Whether that's a fair assessment.

I also quite like that frost snap thing, but I, again, haven't tested it. And I... Can you use that? That's obviously... That's just one... That's just then using one Yeah. It doesn't work with Spark. Hardware supplier. It's a separate kind of system altogether. So for this question, yeah, seed signer and a couple of passports?

Yeah. I mean, I would I would go for passport prime, a... Probably like a Trezor Safe seven maybe. Yeah. That looks like it's got pretty good UI. Again, I haven't tested it in person, but it's got a fairly big screen.

And the third one will probably be maybe a Blockstream j plus, the bigger one. But again, I haven't tested that in person, but I'd definitely go for a minimum of of two different manufacturer devices. But obviously, you know, the whole point of multisig is ideally that you use three separate manufacturers

so that you protect yourself against what we've seen with with with cold card. Absolutely. Yes. I would be using Sparrow as a coordinator in that setup. It is the gold standard and nothing even comes close.

And then the the second part of the question, any thoughts on utilizing Liana decaying multisig? Yeah. Absolutely. I think it's great. I think that the the support is is growing. You know, we... We've got Nunchuk now.

I believe all Bitcoin are also adding mini scripts soon if they haven't done so already. I think it might be in public beta perhaps. But then you limit yourself with the types of devices that you're able to use that with because not all of them support support

mini script just yet. But having the ability to kind of bolster up your your security by having that decaying option where one of the keys, you you you know, the requirements to have a two or three or or a one or three after a time lock

can definitely be beneficial if you really fuck something up or even, you know, having an expanding one where it moves to a two or four where an additional key that's, you know, buried under a mountain somewhere becomes available for... To be part of the spending path. What? And for inheritance and

stuff like that. Say again? I said it's interesting for inheritance and things like that as well. Yeah. The the the problem I have with it right now, and it's a bitcoin network limitation is that the longest time lot that you can have, correct me if I'm wrong here Seth, is like eighteen months.

Something Yeah. Like Isn't a long time in the grand scheme of things for long term storage, like you blink and especially if you got kids, blink and eighteen months has gone by. Yeah. And the the requirements are then cycle all of your coins to then refresh that time lock is a bit of a pain in the ass, not to mention a bit of a privacy

headache, especially if you wanna kind of maintain the segregation of those GTXOs as well. Yeah. He's gonna move all of it.

Exactly. Now if you're spending more regularly, that's obviously easier to manage, but the nature that it's a two or three mini script multi six setup tells me that you're probably not gonna be spending from that very often.

So that kind of time lock has always just been, like, the thorn in the side of that whole proposal for me. I... And clearly, that's not something that's easy to fix because it's something that's enforced, I believe, at the network level. So that would be something that'll be nice to to see, maybe change or talked about in the future.

But, yeah, Seth, I don't know whether you got anything to add to this one before we before we move on. It's definitely generally beyond my expertise. I'm definitely not the the multisig guy here.

The only thing I'd say about a time lock is the only, like, maybe under appreciated advantage of that is it does force you to actually, like, use your setup every now every now and then, which is always one of my concerns, especially in these multi device, multi sigs where

people do what people did with all their cold cards. And they just... They buy it, they load Bitcoin on it, and then they forget what it is and how it works. And they don't deal with firmware updates, and they they don't remember how to work it. I think, like...

What is her name? Issa... Isabel, whatever her name was, influencer girl. She was a good example of that where she had no clue how to actually use her cold card to send the money when she found out about the the hack and the the vulnerabilities,

and she kinda documented her process of figuring out how to use her cold card. Because she'd just been told, use this, put your money on it. Yeah.

DCA to it, and just don't ever worry about it again. And that is fine until it's not. So that is maybe an undersold part of the the time lock forces you to do that regularly, but it should also just be part of your your self custody hygiene, I guess, or is it a better term? If anyone if anyone missed it, I had Kevin Loeck on,

I don't know, maybe, like, six months ago, and we did go into... Or he did go into how long and and why it is the way it is and etcetera, etcetera. So if you go on the website, you'll be able to actually search through the text and find out exactly how long it is. But eighteen months sounds about right.

Alright. I know I said we'd finish with the AI question, but I am gonna include this one because he's asked it twice over the course of a couple of weeks. And for whatever reason, we've not got around to it. So

Seth, this one's for you from Status Quan on x. And he's asking, how do you judge a local model's quality is, quote, fable quality? Is it just go on the speed or the output? After you settle on one, what's involved in switching or upgrading? And can you keep multiple loaded so that Tasker uses x,

then y, then zed? There's a lot there, but pick apart what you can. Yeah. I think probably the most useful part of that is how do you judge if something is better or worse for your workloads.

Because that is really hard. Like, we see this every time there's a model release, whether it be frontier, open source, where everyone sees the benchmarks, they go, oh my god. We now have a model that beats everything. The world is over. And when you actually use it, it's usually not quite as good as the benchmarks say.

Like, there is definitely a a plague of benchmarking right now, where everyone's actually training with the benchmark data in mind, and so they're they're kind of optimizing around those use cases.

So in reality, I think, like, the the thing that I have started to do is I actually have built, like, an a b test framework that constantly looks through the traces of the work that I've been doing with a given model, and it builds realistic workloads to test against a new model.

Obviously, there's also the vibes part where, like,

you should also use it and see if it fits your worst... Your use case. When we're talking a local model, that can mean tearing down the old model to spin up the new one, which maybe you don't actually want to do that, or at least not for a longer time. So I think both is best, but having something...

And again, this is just something you can you can prompt pretty easily of, like, a b test this model against GLM 5.3 flash that we were running, see the traces on this machine, see what I've been using the model for, and then build some real world benchmarks around it.

And any higher level model is really good at that. I wouldn't like use a super basic model for that, but that's a... I I think a good approach for me. To quickly touch on the last one, you can't just have a bunch loaded

because each one has to take up whatever amount of RAM is necessary for it. And usually, you're not gonna be running a bunch of small models that can coexist together in RAM. So usually, it would require tearing down the old one and then spinning up the new one, which is a five to ten minute process. It's not intensive, but...

That's a... Maybe a TLDR kind of for that. Thank you very much, sir. Alright. What have I got next on my list? Oh, this is an interesting one. Shielded bitcoin slash private transfers on bitcoin layer one. I've seen this dotted about on x, and I'll be honest, I haven't had much chance to

dive into it. So I'm hoping that you have, Seth. And this question comes from CrashNeon. So I have not had a chance to read through the proposal yet.

From what I have seen, it is... Like, I've glanced at it. When I say I haven't read through it, I mean, like, I haven't sat down and, like, actually read through every word and dug into the concepts. But from everything I can tell and from reading other good cryptographer feedback on it, it is interesting and novel, but very experimental

cryptography and very expend... Experimental tooling. It's not something that's like, we're gonna be in production on main net in three months or something. Like, this is probably multiple years out if it even proves itself out.

It's not as straightforward as it's it's kind of put forth as, which makes sense. It's a company who wants to start probably raising around this or keep raising around this. So Right. They need to push that. But it does look very interesting, just very, very experimental at its core. Is it something that needs

network level changes to to No. No? Okay. So it could, in theory, work today.

Yeah. Yeah. And that is that is the cool part, is that it does not require a hard fork to be able to function. I'm sure, like, everything else interesting on Bitcoin, it would be easier and better if we had covenants or if we did a hard fork, obviously, or a soft fork.

But, no, that is the the interesting piece is they've they've found a way to do this essentially as a side chain without making changes to Bitcoin, assuming everything that has been proposed and researched pans out in real life.

Yeah. And what I could say is you've gotta, like, peg in and peg out of it, and there's just a, in my head at least, there's, like, a little alarm bell going off where it's like, it's new. It's untested.

If it did happen, you gotta peg in and peg out. Like, how safe is this thing with all the stuff that's been going on? Like, or how safe would this thing be? I don't know. I I stick to my my basic ways of getting privacy at the moment, at least.

Okay. One to watch for sure. But like I say... Always like say with a lot of these proposals is that we'll come back in five years, I guess, kinda like what you said, Seth, and see if anything comes to comes to fruition.

Alright.

Quick one for me from Noster Gang is asking, when is the Noster sign up coming to Passport Prime? We are just waiting on some changes to the communication layer to open up like USB connectivity and Bluetooth and stuff like that to SDK or third party applications.

I have the app ready and primed for that point. And as soon as the SDK is able to speak the kind of communication language, then it'll be published and it'll be probably published from my personal repository.

So keep an eye on that. I I won't like to put a specific time on it, but if I have guns in my head, like say within the month, at that point, most of my apps are also in the same are in the same boat as well. So you'll see a bit of a of a flood of q and a based apps coming out.

The passwords one is also a very... In this very similar situation where we just need to open up that USB connectivity to to non foundation made apps as well. So sit tight and, yeah, there's gonna be a landslide once that is is ready. So thank you for keeping an eye on it though.

Max, how does the start nine... The new start nine, the 0.4, which we seem to be long awaited for quite some time compared to the new Umbrella, who also had a release this week. And if you had one node just for self hosting and you could only choose one, which one would it be? Good question. It

always used to be start nine until my laptop nearly set on fire. The the battery, like, warped. And Yeah. I would... Actually, it was at Black Coffee's house, and they set his house on fire.

And then that was decommissioned. And then since then, I've actually been using Umbral.

I have to say, I would personally stick with Umbral at the moment. I know it maybe isn't the the fan favorite of the more cypherpunk people, but I fucked around with start lines a lot. I fucked around with a lot of other nodes over the years. And the only node that I've never had a problem on ever is Umbral.

And at this stage in life, with the amount of shit I've gotta get done, I don't wanna be tinkering around and fucking about and having problems. I'd also say that start nine,

I destroyed my node twice and had to rebuild it from scratch on start line twice just from having a Monero node on there, and I wasn't the only person that had issues with that. And then I've done the same again on Umbral, and it works immediately straight out of the box. No problems.

So Umbral. Cosign that. I I think Umbral is a great product. Again, like I say, they... You know, I wish it was truly open source. They've got some weird kinda license.

But as a as a software product, well, I guess it's a hardware product as well. You can buy the plug in boxes. Mhmm. I think they've done a fantastic job with the way that the whole thing is packaged. The hardware looks good.

The software works really well. I echo your comments around it being stable. I've had one running on a pie for a long time and I know pies are shit. And I guess that's maybe a testament to to the way that it's architected. I don't know. Obviously, that's above my pay grade, but it's been super stable considering it's on shit hardware.

It looks good. It's easy to use. And then the new two point o release which came out this week is definitely worth checking out. Okay.

Seth, couple of quick fire ones for you. Let's do it. When is Cakewallet gonna work with RoboSATs? It's Spark Lightning can't take or create offers. Yeah.

I I wish I could say any specific time frame, but honestly, it's just probably not gonna happen, because hold invoices are... They're a very hacky way to do what they do, and they work. And, unfortunately, the only way to do that for RoboSATs, but most LSPs don't support them because they... They're essentially freezing liquidity,

intentionally. And we don't run the LSP, Spark does. So I doubt that it will be a thing, which is really annoying because I I also am a heavy RoboSats user or at least have been more in the past. So, yeah, I wish I had a better answer for that, but pretty pretty limited because it's not a...

It's not, like, the proper, quote, unquote, way to use Lightning, even if it's a good way to use it for the specific RoboSats use case. But that's why they have pretty limited wallet support across the board even for, like, full Lightning wallets.

So what... What's the fix there is you just run two Lightning wallets and you you send some out to do... Yeah. Okay. Yeah. Just use, like, a Phoenix wallet as your Yeah. Your RoboSATs one. And... I mean, the nice thing is it's it's Lightning. So I can just send to to Phoenix in a split second and

send from Phoenix out and Yeah. Yeah. Easy. Okay. Obviously, not as easy as I like it to be. I wish we could do it directly in cake, but... Yeah. And is asking when NWC, Nosta Wallet Connect for cake wallet?

I've gotta dig into this one more. I will... I'll make that a higher priority and do some some research on that today.

My initial understanding was that this would again rely on the way that Lightning works via Spark, but I think that there's a way that we can get this running regardless. So I'll I'll dig into this because I would love to have it. I know it is relatively niche.

Yeah. But obviously... Have a look at wisp client. That that has a spark wallet and also has n w c, I believe. W I s p? Yes. Yeah. Is it dark wisp, isn't it? With this too. Yeah. DarkWisp is for, like, the stripped down version for graphene. Oh, I have never heard of this one before.

That's all I've heard about for the last fucking two weeks. Really? Yeah. Okay. Yeah. Everyone trying to get me to to be able to accept Monero through through it. And I then spent the morning trying to work out how to do it, and is it safe and everything. And my AI is

having a meltdown at the moment. Haven't set it up yet. But supposedly, you can just, like, zap XMR. Yeah. Well, that was perfectly timed actually because doctor Givago on Nosto was asking me, and I figured out you as well, saying why haven't you set up your payment targets? And he asked the question. I had no idea this was even

even a thing. They they don't sound like...... Haunting me. They've It sounds like you've got in more than I have. But is it basically just, like, zapping but with XMR

on loss? As far as I can tell, yes. And I've I've been hounded for a couple of weeks and called out for a couple of weeks on this. I promise you I've tried this morning. I sat for about half an hour going through it.

It... I've been told DarkWisp, you can do it with, and I was also told Amethyst, you could do it on, but my AI is telling me that you can't do it on Amethyst and to use DarkWisp.

And then I was saying, well, should I put it on a graphene phone? And it said, yeah. But I would suggest using a separate environment because it's not sure how tested it is, and it's better than just sandboxing and blah blah blah. And I just went, not this morning. I'd like... If this is gonna open me up to potential attacks,

I just don't wanna be hacked. So I left it at the... At that. But maybe I'm just wrong, and maybe it's, like, being neurotic, the AI I'm speaking to.

Cool. Well, I I haven't dove in yet, but I will check it out. Yeah. I will do it is the answer. It's just I'm not gonna do it, and then it's in there with my other wallets, and then suddenly I get swept. So I'm like, fuck.

So, I'm just... So I'm gonna move slowly with it. But in theory, people can then just zap XMR, which is cool. So I'll try. Awesome. Bit of 21 on is asking, FettyMint on start nine, is it worth it? I'll I'll take this one.

It seems as though you would only use that if you wanted to kind of be one of the guardians in a federated Jamian e cash mint, which for most people is... They're not gonna do that, and they have no need to do that. To my knowledge, you don't need to run FerryMint on start nine to be a participant or to kind of connect to other mints.

So for most people, it's not worth it. But if you do have the the the desire to run your own mint or be a part of an existing mint as one of the guardians, then, yeah, of course, you're gonna need that software. But for most people, no.

Boo... Boost from Rev Rev Huddle. He said, where's the group chat to rent somebody else's compute for SATs?

Now, I know he's been a little bit facetious here because it's something that we we kind of touched on in a joking way a couple of shows ago. But Yeah. I actually... The reason I included this is that I do think that there is a potential for a market here.

And this is the part where Seth jumps in and goes, yeah, it already exists and it's called this. But there are people, looking at you, Seth, that dive in heavily to local AI that inevitably are not gonna be maxing that hardware out twenty four seven.

And I wanted to get your thoughts to whether you think there is gonna be market for selling your compute, be that to kind of like a gated community or to be something out to a wider network where it's just to kind of like sign up to the the set for privacy API key and it would probably be significantly cheaper than

than Anthropic or OpenAI.

Yeah. He's a bit money hungry, isn't he? Yeah. I guess, an open question inspired by RevHuddle basically to to... For me to ask you how you foresee this playing out. Do you think there's legs in it?

I definitely think so.

I mean, I think it's something that we need because if local AI especially continues to take off, like, most people aren't gonna be running stuff on it twenty four seven. So I would love a way to be able to monetize downtime, essentially.

And obviously, Bitcoin makes that stupid simple in terms of payments and Lightning specifically in terms of micropayments, like, per prompt. So I think it's definitely something that needs to happen. I know that there have... There has been some work around this. There's there's a project called BitTensor.

I'm trying to remember what the specific underlying provider is, but it's basically like a... It's a little scammy in the way that it actually is. It's a it's a cryptocurrency in a blockchain and has, like, all these sharded blockchains and stuff. So it's it's unnecessary in its complexity, in my opinion.

But it's essentially that idea of people being able to contribute compute and earn directly for that. They use it for more than just, like, inference. They also use it for training and other things for providing services for that.

But, yeah, I I definitely think we need something like this. The the only other one that I've seen so far is something called Dark Bloom, which they're a provider for, like, OpenRouter and all these others.

And... But I think they're specifically for Macs only, where if you have a Mac and you have a local AI model, you can kind of attach it to to their dark bloom system, and you get paid based on usage.

I have a question off that question, Seth. You remember on Monero monthly ish, the recent one, we were talking about, could you use your compute while it's not doing anything to mine Monero, and would that be

a good way to use that compute or not? Did you look into that? Is anyone doing it? Is it not really the right thing to be doing? Would it be better to use it for AI purposes? Yeah. I mean, because the hardware you're using to run AI on is really better for AI specific workloads, I I would expect if there was a a good easy way to

to kind of sell it out, kind of like... What's that the mining thing that you can just pay to rent hash? Super fun. I want to sit on your s nine. Nice hash. Nice hash. There you go. Yeah.

Nice hash. If we had like that for AI, I think that would be that would be ideal. Or I could just say, like, hey, I've got... I have eight sparks. I use them during these hours. I don't use them during these. Prioritize my workloads, but let other workloads run on them. I think that'd be fantastic. The huge downside, though,

you have no privacy guarantees at all, what... Whatsoever. Not even a promise because each individual person is gonna have their own infrastructure and their own choices, and I could enable logging for all of your prompts, and I could train on all of them. I could publish them publicly to a Seth for privacy AI prompt

wall of shame. Like, I mean, it it would be something where it's totally under the control of the person operating the hardware. So that is by far the largest downside. I would like to think that there's some sort of middle ground of, like,

trusted people sharing compute to some degree. Like, you know I mean, yeah, uncle Jim. Families. Yeah. Uncle Jim families or, like, ungovernable

crew members or people like that where it's like, we trust each other enough. You're not gonna put, like, your deepest, darker secrets on there maybe or whatever, but you could you could use it for most things.

I feel like there's there's gotta be a middle ground there. Yeah. Yeah. I mean, I think the the Uncle Jim model that we've taken with so many other services has to come to this too. Like, if if you're running... If you have two Sparks and you're running something reasonable,

you can easily serve eight, twelve, 16 people at a time. Like Mhmm. Farm that out to your family, set them up with Hermes or or something, a Hermes account on your Hermes instance, and they get Mhmm. Local AI without having to do any of that. So... No. I think that's... That has to be kind of the first step.

And then from there, it would be nice to be able to rent out someone else's inference for cheaper than I would pay a direct API provider. Yep.

Thank you for that.

Oh, important question. Bubba wants to know who cuts my hair. Well, I I actually do it myself with a belt sander.

I believe that. I believe that. Straight in with a great and important questions there, Bubba. Thank you for that. Thank you, Bubba. What we got here? Alright. This is a good good one we can dive into. Probably took the rest of the show, to be honest. Cruz is asking Signal or Radar versus Keats versus Simplex. And

he also suggested it could could be cool to get a guest on the topic. I guess maybe let's just talk about the the rough trade offs between them all. Obviously, we know who's best placed to talk about this as they run their own signal fork.

Seth, what are your thoughts? Maybe the the the the key architecture differences will be a good place to start. Yeah. I'll be totally honest and just go ahead and say I have not really tried to use KEET, because when I first tried to use it, when it was launched, it wasn't good.

I just I just never circled back to it, but I have been hearing better and better things over time. My understanding, and y'all please correct me if I'm wrong, that's it's basically peer to peer using their, what is it called, punch hole or something like that, service that works kind of like Tailscale to let you connect to other people

directly and do messages that way, which is a really cool concept. Signal, I mean, I think everyone knows and understands that generally, but centralized server, but absolutely fantastic into an encryption algorithm.

And with radar, obviously, you could do Bitcoin payments while still sending signal messages, which is great. Main downside to signal is the phone number requirement. Still, it's not a problem from a privacy perspective now because

other people that you message don't see your phone number anymore, which was the the case for a very long time. But still, obviously, just a pain because if you wanna have multiple accounts, have to have multiple phone numbers, like me, and it's really annoying.

So they are actually exploring not requiring a phone number and instead of requiring a one time payment of something like $5, but, they're not gonna do crypto, it seems, at all. Not Bitcoin, not Monero, nothing, which is I don't understand how they could fumble that. I really hope that they do Bitcoin even if that's,

in a way, bad for radar because, like, I can't imagine a very privacy sensitive individual

paying signal with their credit card. Even if signal can't directly link it, which I I don't know exactly how they would do that, then the credit card company knows. I I don't know. I just... Like, I can't fathom that. They know... But for... We covered this on the brief, didn't we, Q? It was Oh, nice. I think as far as I understand it,

Signal know that you've paid them. Mhmm. But they then don't link the account. So there's a separation of, yes, they know you have a signal account, but they don't know which one it is, and then they can't see your messages either. So it's not perfect. Obviously, it'd be better if they took cryptocurrency.

Mhmm. But there is that separation, which to me, at least, is still better than a phone number. Yeah. Yeah. Yeah. I do think it is better. It just seems like you're so close. Like, just just accept accept Lightning, accept Banero. Like Yeah. That just... It seems like low hanging fruit.

Yeah. I mean, lastly, real quick, Simplex, for those that don't know, it's a really fascinating protocol that uses two bidirectional channels for every chat. So when I chat with Q on Simplex, my messages get sent to him via one channel on one server, and his messages get sent to me via another channel on another server.

It has a lot of other really good privacy properties, very cutting edge quantum resistant cryptography as well.

Main downs, I love a lot about it. I love the multi user support. I love the anonymous accounts where you can choose to join as a new entity and everything. Alright. Very Like, really, really, really cool stuff.

The main downside for me is just the user experience has just not been quite good enough, especially groups are really painful.

But they have been doing a lot of work to improve that over time. So I I do really love the underlying approach. I love the simplex team specifically and know if Guinea well. So I'm I'm hopeful that that continues to improve and they figure out the UX side because I think the underlying technology is fantastic.

Did you find you, like, lose messages on that and stuff like that? Like, they just sort of don't appear or they disappear and things like that? Yeah. And notifications were always very flaky. I mean, I I will be quite honest. I haven't really used simplex in, like, a year because I just Okay. I just kinda

gave up. Mhmm. And I just... I can't keep up with all the messaging platforms anyways. Like, it's hard enough for me to keep up with just signal and telegram and email. So I just... I simplified. That's when I also dropped three month and and a few other more niche ones.

But, yeah, it was it was that kind of a problem, like, constantly losing messages or group chats don't load messages until I open the group chat or notifications don't come through until I open the app or just all those accidents.

But I know they've been steadily working on those. So I'm sure that it's in a better state now than it was then.

Nice. Make this the last one. No surprises here. This one's coming too. For fine tuning, is a Spark setup with CUDA better than Strix Halo or an m five max or ultra? So, again, leaning on the AI stuff is a hot topic right now. Mister AI, what do think? I I don't know exactly what fine tuning means,

so it's a little hard for me to answer this.

If you mean, like, optimizing your own setup to better serve whatever your favorite model is, I think sparks are a little ways ahead in terms of what's available because CUDA is just so freaking good.

Strix Halo, there's just not a lot of development around it, and you can't link them together. So I'm not very interested in a device that I would be strictly limited to a 128 gigs no matter what I did.

It's one thing that sparks have going for them is you can get one and then add another. Well, now you can't get any. But you could get one and get another and pair them together to expand how much storage you have or how much VRM you have access to.

I I expect that the new m five ultras are going to have an incredible local AI scene around them. I'm already seeing fantastic performance improvements from people who have just gotten them in the next... In the last two days. Mhmm. I I think especially with Apple

heavily heavily focusing on local AI, like in their keynote, in a lot of their marketing, on the page for any go to buy one, like, they're they're constantly talking about this use case, which is really fascinating for a lot of reasons.

But I think because of that and because they're actually at a good performance level now, I think we'll we'll see those exceed Sparks in terms of what the community and the resources around them are like in the long term.

But early days, it's going to be a lot more painful because you don't have CUDA and you don't have a lot of the stuff that's already been built around those.

Yeah. Absolutely. I'm I'm also seeing the same. My my ex algo's got me very much AI pill at the moment. I'm seeing all the same stuff. Everybody's coming out of the woodwork and doubling or tripling the the tokens per second and stuff. It's... It feels like the the levels of excitement feels like $20.17

Bitcoin to me. Don't know whether you kind of get that same sort of vibe. Oh my god. Yes. Yeah. So I've... Honestly, I've just been so bored with

Bitcoin and crypto for, like, a year now. Not building on it. Like, I love Yeah. Freaking love Cake Wallet and all the things we're getting to do. But, like, actually engaging with and learning about the space, it's just... It's so boring. It's so boring.

I'm just kind of... I guess I've hit that, like, old man needs to go live in a cabin in the woods. Hey. Come on. It's not been boring the last, whatever, eight, ten weeks has been a little bit more exciting.

Maybe not in the right ways, but it's fucking... It's it's definitely not been boring. I would maybe take boring instead of that specifically, but, yeah, we we got the feeling that all of the all of the Ethereum and EVM and Meme Coin folks have been... Because they they have hacks every, like,

seven hours. They have a $100,000,000 hack. Like, this is... That's just a day in the life for them. So we just... I do think that's really funny, like, just a quick caveat. I know we're over time, but a quick caveat.

Like, most most of the Bitcoiners I talked to, like, seem to just not know that this has been just a common thing in the Ethereum space and that, like, our hack is relatively tiny compared to all of that. And even that, like, the AI driven attacks started in Ethereum and Solana and all of these blockchains

where it's a... There's a lot more low hanging fruit. Like, we're not... I don't know. We kinda talk about ourselves like we're we're on the cutting edge. Like, we were the first to experience all this, and, like, it's definitely not true. It it definitely was interesting that something that was so core to Bitcoin, like, the Bitcoin Maxi

audience was part of that. But Mhmm. Yeah. It's been happening for a while. But, yes, to fully come back to what your question was. Yeah. It's it's tons of excitement.

People... It's also like combining the self hosting part, which I just absolutely love. Yeah. It's just so much fun. I I love it. Yeah. It's a nice intersection. Right? Mhmm. Anyway, we are at time. That was a very fun show. I do love these question ones where we just get to to interrupt you guys and just shoot shit a little bit.

So thank you all so much as always for getting involved, pre submitting your questions, commenting live. We really appreciate it. Keep the boost coming. We love your feedback. And as always, we will be back at the exact same time next week. And I hope you all have an awesome weekend. See you guys.