LINUX Unplugged - 551: AI Under Your Control

Episode Date: February 26, 2024

Corporate AI is a hot mess, but open-source alternatives can be open-ended chaos. We’ll test some of the best ways to get local AI tools under your control.

Well, boys, I've been chewing on something long enough now that I think I'm ready to talk about it on the show. I've been trying to think about all this AI stuff that's been just thing after thing. And all of the discussion around safety and then the news that we're going to talk about in a little bit with Gemini. And I feel like perhaps I've stumbled onto a hard truth. I want everybody's feedback, but we're always talking about user consent and how much it matters. But I don't think we're thinking about it in terms that regular people care about. I think user consent matters and hijacking your intentions as a user is wrong. Telling me something like this isn't appropriate is hijacking my intentions.

Enforcing a particular worldview, I think that's unethical. I mean, I agree on some of the things that they're trying to protect, like racism. Those are bad. But so is cramming a for-profit company like Google, cramming their ideals and morality down my throat just feels like the worst case sci-fi dystopian future that we worried about. And we have these corporations that are setting themselves up as like arbiters of reality, of what's right and wrong, when morality should be set by the people, not corporations or governments. And the more I think about this, I think this is just going to be a hard reality because all these companies are so risk adverse.

Yeah, they're not really set up to be speaking truths. They're there to set up to be to incentiv you know, incentivize their profits and say things that won't get them in trouble. They're not necessarily doing a good job of that, but that's what they're trying to do. But like if I run it on my system and I have control over it, it's on my Linux box, those aren't my priorities.

And I think if this is a hard truth that we kind of accept, then really the future for interesting AI that's going to be pushing the limits and for good and bad, and we'll see stories around both, I think that's going to be pushing the limits and for good and bad and we'll see stories around both I think it's going to be open source I think it's going to be the open source stuff and that was really the motivation for today's episode Hello, friends, and welcome back to your weekly Linux talk show. My name is Chris. My name is Wes.

And my name is Brent. Well, hello, boys. Coming up on the show, you know what they say, if you want it done right, you got to do it yourself. So this week, we're going to see if we can tame the moving target that is the current AI stacks out there that you might want to try. We're going to see if we can tame the moving target that is the current AI stacks out there that you might want to try. We think maybe we have found one of the best ways possible to deploy these AI tools under your own control on your box. So stay tuned and find out for that. And then we're going to round the show out with some great boosts and picks and a lot more.

So before we go further, let's say good morning to our virtual lug. Time-appropriate greetings, Mumble Room. Howdy. Howdy. Room. Hello. Howdy. Say hi, Sam. Hello, Brent. Hello. Hello. Hello, and a big good morning to our friends over at Tailscale, tailscale.com slash linuxunplugged.

We're playing around with these tools. We're deploying it on our tailnet. We can all access it by name. We don't have to worry about security vulnerabilities in these brand-new projects because we never put it on the public internet. I don't really deploy anything on the public internet anymore. It is the easiest way to connect your devices and services to each other wherever they are, regardless of the platform,

and it's fast. Go try it on 100 devices and support the show at tailscale.com slash Linux Unplugged. All right, so this week, I think if we could have one overarching goal for the show, So this week, I think if we could have one overarching goal for the show, I would love to at least get the listeners thinking more seriously about open source AI tools and less about things like chat GPT and co-pilots and things like that. We're going to talk about the news that's come up this week in just a little bit. But also the tooling has come a long way for the open source stuff, just on the web and communities, the stuff you can install. And also some of it's gotten complicated and some of it's broken. And I think the other reality we're sitting with is these commercial solutions, they're so risk adverse that it's almost reducing the usefulness of the product, right? It's embarrassing, really. It worries me too, because that's just more of the, you know, it's already an imbalance because only these giant companies seem to have the

resources to sort of train the foundational models. And of course they're the ones applying all of these filters that end up harming the actual productive use. And so if you can't say anything, I mean, let's not even say offensive, but just sort of creative, weird,

out of the ordinary, then they're left with the only sort of access to these tools that, you know, they can tune down those filters internally, presumably, and then they can make the interesting stuff. That's a good point. Yeah, they're the only ones, really. They become the gatekeepers, essentially, of this. And this is why I sort of feel like we have a limited time to win the race here in just a moment that there are several factors that are working not necessarily together, but the result is they work together to limit the usefulness and utility of the open source stuff. And I want to cover all of that in a moment. But first, we have a couple of events that are just around the corner.

We are three Sundays away before we hit the road, boys. Okay. Three Sundays. I'm sure you're totally ready, Brent, right? Because, I mean. Oh away before we hit the road, boys. Okay. Three Sundays. I'm sure you're totally ready, Brent, right? Because, I mean... Oh, yeah, yeah. Always. You know me. I'm early on the spot. I'm not worried about that at all. He committed on air. Yep. Multiple times for weeks in a row. And I always end up there.

Yeah, that's true. It's true. It's just how much that I stress about it is really it. We are getting so excited about the very first NixCon North America. We can't wait to bring you coverage of that. That is co-located with Scale in Pasadena, California. It is the Linux Expo number 21 for Southern

California. It all kicks off March 14th. We're hitting the road just like a day or two before that. If you want to sign up for Scale or go to NixCon or go to them both, get a Scale pass and use our promo code JBFTW JB for the win to save 50% off and uh and then join us at lunch Saturday at 1 30 p.m at the art house we want to see you there and then of course go grab fountain.fm because they are creating a

totally upgraded live experience for us in app and on the web oh they're working on a cool web app that they've prototyped for some of the live value for value music streaming events. Okay. That can toggle between audio and video. You can use it in the app or you can use it on the web. Nice. Well,

I want to see this. Oh, it's going to be great. So we're going to have a whole batch of live streams, one on the 12th, the 14th, the 15th and the 17th. And you don't have to really worry about any of that right now,

but we're looking forward to it. And I think depending depending on how all this goes we're kind of going to be defining a new live experience for the show and for future events so i think it's going to be kind of a make it or break it trendsetter for us we're kicking off the first event bigly bigly as wes always says that's right okay well that's kind of all i had announcement wise uh you feeling good i am feeling good Did you decide you're flying back? I am. Yeah, for the work, right? For your hashtag J-Job?

Yeah, get back to the regular things. Yeah, that's right. We're going to take Brent maybe somewhere special then just to make you jealous. Dang it, I knew this would happen. I'm going to have to fly back to then visit for that part of it. What are the options? Brent, you know what I'm going to be advocating for. I don't know how we're going to make it work.

That's what you said last time. I'm going to be advocating for the coast. Oh, this is exactly what you said last time. That never happened. I know. I'm aware. But we haven't booked any of our return Airbnb-zles. Bring it on. You've got flexibility. We do. Maybe we just do like hotel or something. I don't know.

But I would love to take the coast. I think one of the prettiest places that West Coast is one of the prettiest places in the world. It's or something. I don't know. But I would love to take the coast. I think one of the prettiest places, that West Coast is one of the prettiest places in the world. It's really something. We'll take pictures though, Wes. We'll take pictures. All right. So you probably heard the news this week about Gemini just being embarrassingly bad. As the Register put it, Google halts Gemini image generation to fix a white balance.

it, Google halts Gemini image generation to fix a white balance. They go on to say Google has suspended the availability of text-to-image capabilities in its recently released Gemini multi-model foundational AI model after it failed to accurately represent white Europeans and Americans in specific historical contexts. And their sub-headline, which I thought was really on point, quote, big tech keeps poisoning the well without facing any consequences for its folly. Well, I say today here on this show, they face the consequences, gentlemen. Today, we say this far, no farther. The line shall not be crossed. We are going sovereign with our technology. And I was excited to see Stable Diffusion 3 came out. Big fan of Stable Diffusion image generation stuff. Big fan of the Stable Diffusion image generation stuff. I'm not so sure, though, the hosted version is really going to hold up.

In their announcement, they write, We believe in safe, responsible AI practices. This means we've taken and continue to take responsible steps to prevent the misuse of Stable Diffusion 3 by bad actors. Who bad actors are, though? I'm not sure. It doesn't say. I guess they define who the bad actor is. It's Brad. Oh, yeah, of course. Well, that's definitely one of them.

They continue. Safety starts when we begin training our model and continues throughout the testing, evaluation, and deployment. In preparation for this early preview, we've introduced numerous safeguards by continually collaborating with researchers, experts, and our community. We expect to innovate further with integrity as we approach this model's public release. So all of the talk right now really is leading with responsible AI practices and safety, perhaps as they should be, but it's definitely the focus. And I thought commenters on Hacker News

nailed this when they were talking about how it's basically making the products less useful. Their sensitive nature is causing the rest of us to just want to look for something that we can self-host and run because I get offended by them projecting some intentions that they manufacture onto me. Yeah. Well, and I mean you just – you're kind of left out. you're on to me yeah well and i mean you just you you're kind of left out you know it's not hasn't been a perfect system but uh by and large you kind of get to decide you know if you're going to repost something if you're going to share it if you're going to use your speech to have it and you're you're up to that and you're it's sort of implied like well you could do anything with this even though you might just want to make a silly limerick you know right there in the app and then

you'll never use it again it kind of feels to me like this is potentially going to create another divide that we've seen in, you know, computer privacy and data privacy. Like those of us who are technically sound enough to host our own stuff will have far more privacy than those who just don't have those skills. And this feels like it's headed in that exact direction. It's also alarming. The point that the reg makes is actually well taken. It's like they keep screwing it up, but because we have no other choice,

or no other viable choice, they really feel, they face no consequences for their folly. I mean, you think about something like this that had to ship, how they miss something like this in testing. And we were sharing screenshots this morning, I mean, you think about something like this that had to ship, how they miss something like this in testing. And we were sharing screenshots this morning, I was, because throughout the weekend, people are just trying various different things that you think would be easy for it to answer, and it just can't answer them. Sure seems like it is kind of like, you know, end of the pipeline. Like, you know, they're not necessarily being able to fix or address the systemic problems in their data sets.

And then at the end, they're kind of just like, oh, well, we will apply some little fix to try to tweak it to make sure it's like sufficiently diverse or whatever the, you know, like correct for other things that they want to make it generally more accurate. But then you're doing it so late in the process that you get crap results. Wouldn't it have to understand to not answer? I think it depends on what layer this is being applied. But your point taken is it must realize the truth in order to generate something that is either the opposite of the truth or to say it can't answer it because of moral reasons. Like it must have some context awareness then of the actual question underneath. But then you're right. The presentation layer, they're restricting.

than of the actual question underneath. But then you're right, the presentation layer, they're restricting. Well, I just mean like what is implied or learned from the data sets that are being selected. And then they're just trying to apply things at the end that are going to be very coarse grained and are not at the level that a giant data set has. Mm-hmm, okay.

And I think at the same time, as we have this sort of ridiculous, over-safe, we had an example on Coder Radio. Somebody wanted to know the absolute fastest way to do something in C-sharp, and the LLM would not respond because it was too risky and dangerous to do something just for speed and not for, you know, good code quality. Couldn't tell you the answer. And that's where it feels like it's another level. Like, I think probably a lot of us, especially for hosted for hosted services like i expect there to be some level of this you know like if i'm just like generating obvious hate speech i could see and it probably being reasonable for

the general public of something that's like um we're not cool with this from our product for you like and if it's an edge case it's fine it's like a reasonable limitation but right now it feels like every other thing that you ask these these machines do, you get back a little slap in the face. I mean, sometimes it's even like a parody of a brand name or something. It's like, oh, you know, we can't really let you speak. Hello, the Burger King. Yeah. Or like, you know, public figures, which an artist could whip up and it would be totally free speech. It's an important part of our public discourse.

Yeah, right. And then at the same time, you have a group out there that Politico has labeled as AI doomsayers that are funded by billionaires that are ramping up lobbying right now as we record. So two different nonprofits, the Center for AI Policy and the Center for AI Safety, funded by different tech billionaires, have begun directly lobbying Washington to address what they perceive as existential threats posed by AI. So each nonprofit has spent around $100,000, north of $100,000 on lobbying efforts, both of them. Specifically, they've spent that money in the last three months. And they're drawing their funding from different organizations throughout the industry that are tied to AI.

And the part that I love is, like Brent's good buddy, Sam Bankman-Fraud, both nonprofits have ties to the effect of altruism cause. Which has been an absolute, I mean, I just think it's been absolutely disgraced with folks like Sam Altman and others. I mean, I just think it's been absolutely disgraced with folks like Sam Altman and others. And these folks all have what, in my opinion, is a God complex where they believe they have to save humanity. So they create problems and then they panic about them and pretend like they're the only ones that can save us from them. And that's how they get off as rich billionaires is saving the world from these existential crises that they're very funding created in the first place. Previously, efforts to influence Washington on AI were conducted through going directly at staffers and think tanks. That was really the approach they took before. And we got some executive orders here and we got some action from a couple of senators, but not much.

Nothing really formed any kind of law. But so now they're going with the full on lobbying effort. Nothing really formed any kind of law. But so now they're going with the full-on lobbying effort. And essentially what they're pushing for is in order to work on AI projects, you need to be licensed. And you need to be essentially re-licensed every couple of years to make sure you're not creating dangerous AI. And then the projects you're working on need to be checked and audited by a group, an industry group. And they even want to have like restrictions around like the power and size of the models and whatnot is what they're advocating for.

And they want this implemented at a law level from Washington, D.C. here in the States. You know, I'm not against like any kind of regulation for this kind of stuff. I'm sure at some level, especially as the scale picks up, we'll need some. But boy, do I like it does not seem like a we're long enough into it to really know what the effective regulation should be and then be just the i don't know that our current system is really set up to like it's all going to change a lot so like whatever we need to do the qb get something reasonable and then something nimble that you would actually update as these things develop and that i have not a lot of faith in and something that didn't require

millions of dollars of lawyers to nursemaid through the system, which it would seem these systems do require that, which almost means you're just. Can we not lock in the incumbents? Right. You're just locking in the incumbents. Right. And you're making it particularly difficult to use open source AI tools in business and in business development and software development. You know, if you want to run a little model on your little GPU, that's fine. But don't go building something that's going to be running as part of a software as a service. It's the sort of same thinking that reminds me, I don't know, the folks who get prosecuted for hacking when they download a file that was in a public S3 share.

Yeah. Yeah, it's information. And I think that's why I want to advocate to our listener base that even if you're not currently wrapping your heads around this, just stay informed a little bit. We'll give you some resources. I'm actively soliciting more from you out there in the audience because I feel like we have kind of a race to win here to a degree to get to a certain user-based size where they can't just completely ignore it at a legislation level. user-based size where they can't just completely ignore it at a at a legislation level if we got to a certain size adoption before it goes too far then it's like the genie's out of the bottle and so i feel like one of the things this show could do to help contribute to that is cover the tooling that is accessible to all of us and help you deploy in ways that are consistent

um that maybe are nimble like west, because the thing is changing really fast. So I think, you know, that's what we focus on next. Warp.dev slash Linux dash terminal. Yeah, that's right. Warp, the modern Rust-based terminal with AI built in, is now available for Linux. It used to only be out there for Mac OS users, but it's available for Linux on multiple distributions,

including Ubuntu, Fedora, Debian, and Arch. You can give it a try today at warp.dev slash linux-terminal. It is something that feels like they rethought the terminal for the modern era. If you work in the corporate world where there's Mac and Windows and Linux systems, you're going to feel like you have a tool that just keeps up with the rest. This is really slick because, A, it's built on Rust, so you know it's quick. But it has a really nice modern text editor built in. So when you're editing your YAML, your Docker Compose, your JSON, whatever it is,

you know, it's got a modern text editor built in. And it lets you navigate through blocks of output really quickly. If you forget a command, there's that warp AI that you can invoke that will suggest the right command for you. This is what I've been playing with and it is really handy. There's also just the ability to customize the UI. Yeah, I know you like to do that. You can create your own prompts too and have all the nice things set ready to go so you can recall those when you need them. It's a great user experience.

They have a collaboration feature in there. All of it is fantastic for developers or engineers who are working or living in that terminal. I've sometimes looked over at some of the other terminals on the commercial platforms and I've been a little jealous. But now, no, no, not anymore. Go check it out. Warp.dev slash Linux dash terminal. Now available for Linux, including Ubuntu, Fedora, Debian, Arch, Linux. And I bet a lot more coming soon.

You know how those things work out. Go check it out. Give it a shot. Support the show by going to warp.dev slash Linux dash terminal and give it a try today. warp.dev slash linux dash terminal. It's not all bad news, though. I mean, at the same time that these big corporations are kind of, you know, cutting their foot in public, they are also releasing free and open-weight models that you can go play with, like Google's new Gemma.

I mean, this is pretty nice to see it this soon after Gemini comes out and it's a two billion parameter model. Wow. Two sizes, the two billion and seven billion parameters available. Now, they do apply like all their safety stuff to it. So, you know, it's that's these different models you can grab. We've talked about this before. They all whatever institution makes them. It comes with with that institutional bias there is that aspect to it but wes why are they doing this why are they taking like these these models that are kind of behind all of this work they've done and releasing to the public is it is it like the like the github

idea where you make it free so developers kind of get hooked is it i mean i guess i don't really understand the motivation for all of these models coming out, especially when Google's trying to get people to pay for an outrageous amount for Gemini. Yeah, I mean, I suppose on one end it is to help mindshare adoption. I think because machine learning has been the area of academic interest for a long time and then has in the last few years sort of blossomed now to being really an applied science and industry um there is sort of a reputational aspect of publishing what you're doing like whether that be a paper and then a lot of times now it's a paper and the model behind it and maybe they just feel comfortable that there's enough you know special sauce that they can sort

of keep on their own plus uh it still takes resources and knowledge to sort of get this you know you and i can get this to work and tie it into stuff. But it's a whole other level to build into a nice service. And I hadn't thought about the historical kind of momentum around releasing this stuff too as sort of educational background. And speaking of that, Olamah, which we've talked about before on the show, has already incorporated this. So Gemini's Gemma comes out and then within a day or so, the Olamo folks are already incorporating it. I mean, the free stuff is moving really fast. Yeah, Olamo seems to be doing pretty nicely.

And you see they have a REST API. They've got, I think, chat GBT style API compatibility. There's also some bindings like Python bindings, probably JavaScript bindings as well. So it's getting easier and easier to adopt and build into your apps if you so want to. So I thought we should focus on image generation this episode. There's a lot of different tooling out there, but all of the hoopla around Gemini has been with image generation. And we've talked about easy stable diffusion before. It's a Docker container that you can get up and running pretty quickly. This week, though, I want to talk about Invoke AI. Invoke AI

is an implementation of stable diffusion that has, along with it, a great set of tooling and a streamlined process to generate images, pull in updated models or modifiers, and work to generate images, whatever you want, really, depending on the model you're using. And we've been experimenting with it on and off for the last few days. We be talking about that but the reality is there's a lot of ways to install something like this and i've noticed a lot of bad habits invoke ai in particular is guilty of this a lot of like just run this script and it just dumps stuff all over your system and running stuff and i mean i think also some of the, you know, just the way these, the history, maybe some of the coming out of academia,

there's a lot of specific stuff. There's different package managers. There's a lot of specific dependencies that you need for your particular hardware. Sometimes like sensitivity to which version is going to work and, you know, new ones won't or way older ones won't. You need this set over here matched with this one over here. Plus you're saying there's kind of a lot of stateful stuff where you're you're running scripts to get everything just so set up and then it's kind of based on

whoever happened to tie all those things together and then some of it's kind of brittle because it's dependent on certain like video acceleration libraries working and being there installed correctly and you end up with a bunch of stuff scattered plus then there's the web uis that go along with these things so you got a separate sort of javascript app probably that you gotta you gotta handle and build and then have a web server for see and then because you know you can't just trust them then there's not like a unified system for it everyone's gonna run their own thing and you know install apache or add to your nginx or oh god you're making my blood pressure go up just talking about it. It is really a mess. And then you also have like Docker containers. Of course, those can be loaded full of some stuff, but not have other things that you need for your particular hardware. Or

like if you deploy it on the Mac, you won't get access to hardware acceleration in some cases. So there's a lot of edge cases around the containers. And I haven't really seen any of this stuff shipped as like a flat pack yet.'ve been waiting for something like invoke ai and stable diffusion to just be a flat pack and you install it it starts a little web server gives you the url and like a little pop-up and you click that and it brings it up but so far that's not happening and so you have containers with their limitations sometimes they'll work great be exactly what you need and sometimes you're going to be execing into that thing and fixing stuff and adding stuff so it works on your machine. You got the blast and spray and pray method, which is what the Mac users seem to be just going for, which I don't understand. You

know, there's other ways you could do it. And then, of course, you could do it from scratch and you could actually build it, pull down everything. Man, is that a job? Yeah, right. You can kind of follow the developer path and be like, okay, if I was going to be working on this project, what all would I have to do? So as you can guess, we didn't want to do it this way. As you could probably guess from our tone, this is not how we did it. And we think maybe you shouldn't do it those ways either. We think maybe we have a much better way to do this. Determinate dot systems slash unplugged. Bring Nix to work the way you've always wanted with flakehub.com. Go register for the private beta and get secure, compliance-friendly

Nix and all the support you need at determinate.systems slash unplugged. So let's talk about flakehub.com. It's the all-in-one platform for secure, compliant, transformative Nix development, and it is made by our friends at Determinate Systems, the builders of the best Nix installer, including the massively popular MDM and Skiff mode support, which gets 20,000 installs per day, massively successful. And they have an advanced binary cache with sophisticated access controls that if you're working in a team is going to be mind-blowing for you.

They're also the ones behind that cloud-native Nix project we've talked about and just a plethora of open-source tools and documentation for the Nix community. And FlakeHub. FlakeHub is solving big problems if you're building with Nix. Nix caching traditionally involves a globally flat namespace. So imagine building software. It's like having a massive toolbox with tons of parts and tools in there. But finding the right one can take forever when it's a mess, and auditing everything that's in there that got pulled in is just,

well, it's a nightmare if maybe not even impossible. Traditional Nix caching is like having multiple toolboxes scattered everywhere, and each team member has their own set, and it's hard to share the tools or know who has which one and if they have the right one. It's not really elegant. Well, FlakeHub is. It's a single organized toolbox for your entire company, if you will. Imagine one identity-aware cache where everyone gets the right tools based on their permissions.

No more searching through that mess with fine-grained access control so you make sure things stay sensitive and they're in the right tools based on their permissions. No more searching through that mess with fine-grained access control so you make sure things stay sensitive and they're in the right spot. And FlakeHub has that globally distributed build cache. Man, has that been useful here when we've been working with AI tools. If you can use GitHub, FlakeHub will fit into your existing DevOps workflow.

And it's the only SOC 2 certified platform for Nix, and it delivers a low-friction Nix development experience that you've always wanted. Go bring Nix to work the way you've always wanted with flakehub.com. You guys, you've got to check this out. Go support the show and register for the private beta to get secure, compliant, friendly Nix

and all the support you need. You go to determinant.systems.unplugged. We'll have a link for that in the show notes too. That's determinate.systems slash unplugged. And a big thank you to Determinate Systems for sponsoring the Unplugged program. Determinate.systems slash unplugged. Now, if you've been following the show lately, you might realize that we have a favorite way of solving this problem. That's right. We wanted to find out if we could Nixify our AI tools. And we didn't have to go very far.

You know, sometimes it's easy. Nixify.ai is a place you can start. They are trying to just make things simply available, a large repository of AI executable code that might be impractical or hard to build or figure out yourself. They've got it. They've got it over there. You know, I think we saw it kind of float by. Some of our Nix friends were sharing it around sometime last year, but we actually hadn't had a chance to give it a try. And, I mean, yeah, it's got exactly what we want, at least if it works.

Here's how they pitch it. Each project is self-contained but without containers. By the way, when we say Nix, we don't mean NixOS. It would work on NixOS, but this will even work in WSL. You just need to get Nix, the package manager, installed. So you got self-contained applications. It's easy to run. It's like a one command to actually install and run it once you've got Nix set up. Their projects have support for NVIDIA and AMD GPUs. So it'll work with other. And like I mentioned, you WSL users are not left out. And I recognize there's more and more of you out there. Go get NixOS WSL and then you can run GPU accelerated AI workloads on WSL using these commands.

So you can get it right there on your Windows box if you want. Now, Wes, am I correct in saying that they're using flakes under the hood? Yeah, I think that is the primary method that they advertise. You know, like if you go on the site and you want to go get started, let's say you've got an AMD system, you do Nix run, and then basically a GitHub Flake reference that points you over at their Nixified AI slash Flake repo. And then you tell it you want either the AMD or the NVIDIA version, whichever project you're doing, like we were doing Invoke AI.

I want to ask you how you got it working in a moment. I did the spray and pray approach on a demo system. I'm not proud of it. That's why I don't recommend it. in a moment. I did the spray and pray approach on a demo system. I'm not proud of it. That's why I don't recommend it. But a flake is perfect for this kind of thing. I think you could think of a flake as, it's actually pretty simple. Think of it as like a building block for a Nix project.

Imagine a flake is a self-contained box and it has code instructions for a thing. And inside that there's a file named flake.nix and it tells Nix what to do. It'll say, here are your inputs. These are the other bricks you need to build the full wall. And here are the outputs. This is what it should look like. This is how things should be set up.

These are the parameters. And then you can combine these to create bigger and bigger systems. They're just inputs you put in your config file. And each flake has a lock file to ensure it's always using the same exact version of everything that makes it reproducible and reliable so that's really big from an audibility standpoint or maybe you're deploying an appliance or maybe you're just trying to get something that worked that worked on this machine to work just like it did on another machine and that's i think particularly where it's useful for these types of tools they've got a lot

of brittle parts and are moving quick. So you could think of like regular NICs is you're building everything piece by piece with bricks, and you're assembling those bricks and you're building the wall. With flakes, they're pre-made chunks of the wall with clear inputs and outputs. They're like sets of Legos that you can grab and set down in a whole set of Legos all at once. Yeah, you know, if you've ever done anything with, you know, like PIP or NPM or any of these sort of programming language specific installers and package managers, you know, they've got these lock files and they kind of let you manage everything and make sure you do things in chunks that you can comprehend and you've tested and, okay, I do want to

update them all for me or, you know, leave them pinned for now. Flakes and Nix kind of let you take that to all of your packages. You know, it extends that approach to the whole system and it lets you easily integrate with just GitHub repos, whatever Git URLs you want to use. And like you're saying, this stuff's all moving so fast. You might have custom versions of stuff, forked packages, different versions of the web UI.

You're not necessarily going to want to wait for that to get into a centralized Nix packages repo. Or any distro. Right. Yeah. With the inputs, the Nix Flake stuff goes and handles going and checking what is the most recent commit, grabbing that, writing it in your lock file, pinning it, and then getting you all that stuff onto your system and then feeding it into the rest of your Nix stuff. So all that is just taken care of

for you. Now, late in the night, we all tried our various methods, but Wes, of course, did the Nixified approach and got it working. Did you use a flake? Did you use a different installation approach? How did you get things working on Nix? No, I just went with the Nixified AI

stuff. I did see that Invoke AI has a flake.nix in their repo. Yeah. But it's mostly just set up for getting the raw dependencies. So it will get, like, handle all the CUDA stuff. I think it built OpenCV, so it took a little while. But, you know, thankfully that rig has plenty of CPU. Shout out to our audience that donated very nice servers.

I think that would be a nice setup if you were explicitly trying to develop on it. Because you kind of get all the different Python libraries and the stuff you needed. And then you could just do the build yourself right in that environment. But if you want something that's a little more end user focused and packaged, that's where the Nixified AI version came in. I think they had some caching in place, which was nice. So you didn't have to build absolutely everything. Right. You can do the Nix run, but you might want to do Nix shell because there's more than one binary

for some of these. So in particular for invoke AI, you've got just the invoke AI command, which does like a command line version, targeted more advanced users, maybe for like scripting or automation or that kind of thing. But you probably want the invoke AI dash web command because you're trying to run the web UI.

Yes, right. And I think by by default maybe there's one it comes with but there's a whole bunch of different models this thing can use so you're going to need to do some configuration and most importantly you're going to need to actually pull down all that stuff so you're going to want to run some stuff before you get it all launched uh where they have a configuration command you can run uh it's even got a little sort of mouse enabled uh command line interface that you can run. It's even got a little sort of mouse-enabled command line interface that you can kind of click through either with tab and spacebar or double-click right there. Yep. Let's you select what models, some of the options, some of the enhancements further

on in the pipeline that you need. Yeah, it makes it, it sounds like it's a lot of work, but that part of it makes it really simple for anybody that's even new to this. Yeah, I didn't super know what I was doing, so I just kind of clicked some of the ones. I was like, yeah, that sounds good. And it, for the most part, just worked. It does take a little while because there's gigs and gigs of data to download, depending on how many models you choose and which ones. But it downloads it. You don't even need to restart anything. The web UI will pick up the new models. There's even a little button

on there to refresh. And now you've got InvoKI. And you've got a web UI that lets you generate images, and then you can play around with the different plumbing and the different models behind it and the different accelerators that they have to get everything tweaked just right, and you can start creating your own pictures. And, you know, what stands out to me so far about InvoKI

is you can tell it's kind of like there's workflows that they have in mind. There's all kinds of, like, you know, setups here so you can see the queue that's going on yes right um you can have like we could have no sorry to interrupt but just to something was really neat is we could share a central invoke ai server and we could all have different images being in the queue and i could see yours and you could see mine and they would just be generating in the order they were submitted yeah and you can you know you can set up different workflows on here. I don't know how, you know, it's not maybe

competitive with everything, but you can see that you're getting a little more sophistication than just like, you know, here's the input form and then go run the model for me. Yeah. So I'll tell you, obviously I use this. If you go to jupiter.tube, you can see how we use this. For me, it's like generating stock photography instead of paying for some stock photography website, I just generate my own now. But I also, I've used this now to generate backgrounds for my kids' devices. So they have like, you know, a custom background from dad. My wife has used it to generate like her perfect wallpaper for her iPhone. So she has like this photo that she just loves to look at that is something that she kind of created and then put it in there and created it for her iPhone. So she has like this photo that she just loves to look at. That is something that she kind of created and then put it in there and

created it for her. And there's just ways you can use it that are not necessarily, you might think of at first, but the more you have these kinds of things and you think, Oh, it could kind of be nice to have a little piece of art there. It doesn't have to be all that important. It just needs something to add a little visual splash.

It's just great for that kind of stuff. And they have a bunch of other things on here too. Of course, like one, I did not get a chance to play with, but I have it installed installed is a text gen which as you can probably guess is a web ui for large language models to spit out different stuff and it supports a bunch of the different models and i have that installed on my system but i just didn't get a chance to play with it but again using the nixified ai method it's so straightforward and because it's not NixOS dependent, you can be on Ubuntu and be using it

or you can be on macOS or WSO. I think that's pretty powerful because then you get a consistent experience across all your different systems. And of course, you know, it is still modifying your system. In this case, it's downloading models. It made a whole directory to keep a bunch of stuff and store its settings. But, you know,

it's not, especially in that case, it didn't have root permission. It was running as a regular user. So, you know, you's not, especially in that case, it didn't have root permission. It was running as a regular user. So, you know, you can make a whole separate user to run it as if you'd like. And then everything else, the stuff you built via Nix, that's all just in the Nix store. And when you go to collect garbage, when you're done,

you didn't like that model, you want to do something else, or you're just going to deploy it another way. It's cleaned up. It's cleaned up. I think our opinion is, is that it's easier to use. You get reproducible builds, and you have better organization It's cleaned up. I think our opinion is that it's easier to use, you get reproducible builds, and you have better organization of these really complex projects that can have a lot of pieces.

I have a question on what's reasonable hardware to expect this stuff to even be useful on? Because I know some of us are running it on fancy hardware, but can anybody just run this on their work laptop and give it at least a try? I mean, obviously the performance is going to be different, but. Yeah. How painful, Wes, would you say the image generation was on the CPU? Decently, decently painful.

Yeah. I mean, probably like four-ish minutes. Yeah. Versus the GPU, it could be a lot quicker. Yeah. So it depends on how impatient you are if you have patience right uh you know and it is nice like it gives you the little um you can set it so it'll show you the in progress sort of pictures as the model is refining and diffusing

um and uh in the command line or in your logs it'll also have a little progress bar uh so you at least get some pretty decent feedback especially compared to some of the commercial systems in terms of like, how long am I going to be waiting? But you are probably going to be waiting. I think, you know, there's a lot you can start with. I think invoke AI is a fun one. If you're listening to this and you're thinking, I do want to try some of this local stuff. I do want to try to help these projects grow. I would like to understand this a little better. If you get invoke AI deployed, the building blocks you use to do that will work for all these other ones as well. There'd be little individual little tweaks you might make if it's depending on what the software is, but like the method and the

approach and the documentation is all going to be the same at that point. So it's, it's a journey, but once you learn it, you'll have that path well-treaded. And I think it's, I mean, I think it's like, we're just going to leave it. We're just going to keep it. I mean, it's going to keep updating now. It's like, we're going to be current with the projects as they release stuff. And it's great for us. Like we've just set it and forget it. Collide.com slash unplugged. Well, you've probably heard us talk about Collide before,

but did you hear that Collide was just acquired by 1Password? It's pretty big news since these two companies are leading the industry in creating security solutions that put users first. For over a year, Collide Device Trust has helped companies with Okta ensure that only known secure devices

can access their data. And, well, it's still what they're doing, but now they're doing it as part of 1Password. So, if you've got Okta and you've been meaning to check out Collide, now's a great time.

Collide comes with a library of pre-built device posture checks, and you can write your own custom checks for just about anything you can think of. Plus, you can use Collide on devices without MDM, like your Linux fleet, contractor devices, and every BYOD phone and laptop in your company. Now that Collide is part of 1Password, they're only getting better. So go check out collide.com to learn more and watch their demo. It's K-O-L-I-D-E dot com slash unplugged. Watch the demo, support the show, collide.com slash unplugged.

Well, conference season is coming up quick, and we have some updates specifically around Texas Linux Fest, which is happening April 12th to 13th. Chris, can you catch us up on those updates? Yes. Well, mainly the time slots and the schedule are now up on the website. So if you've been trying to look at what's going on, the talk titles and the speakers have been filled in as we have them, and more will be coming up there as they get confirmed. And it sounds like there's at least some planning in the works for a party Friday night. Oh. Potentially. And there may be others.

They tend to be a little more kind of just planned at the last second. And I think the good news for anybody that's considering going is we do have a hotel group rate link. So I guess normally like this hotel is like 320 a night. And Texas Linux Fest group rate is going to be 250 a night. So it's not bad. And you're right there. You know, you don't have to go far. So that's good.

And again, that's April 12th and the 13th. And then just shortly after that, Linux Fest Northwest 2024 is the 26th and 28th of April. And the update here is the schedule is posted. You'll see some familiar names on there. I believe there's a Wes Payne on there. And if you go to Wes's talk on Sunday at Linux Fest, just stick around in that room because then we're going to do Linux Unplugged after that right there. It's going to be a whole day of fun. Yep. So you can just hang out. Pretty great. Shout out to the core contributors. They

are participating in a membership program that finances the show directly by the listeners and gives them access to additional content. There's an ad-free version of the show, but I think the real winner is the bootleg version because the pre-shows have just been bangers. Some spicy pre-shows recently. Some spicy couple of pre-shows in there. Maybe it might be worth checking out just for that.

You do get a little of the sacred sauce. And now it is time for the boost. Thank you everybody who boosted in this week. We got a whole big batch of Fountain FM feedback and I, well, with Wes's help, bundled it all up and sat down with Nick and we went over all of it. And if you've been boosting from Fountain,

I think nearly all of them, Nick has been replying to directly. So go check your Fountain FM app. So I'm not putting all of it in the show because some of it's getting handled directly. But I will put some of it that I have answers for. And our baller boost came in from User38 who sent in 51,000 sats. Hey, rich lifestyle! They write, my original donation was intended to be 150,000, but I had some problems with Fountain.

Maybe they could fix the custom donation feature, which was broken. Yeah. Yeah, really. Come on. You still managed, even with that bug. We appreciate it. You managed to become our baller this week. Hybrid sarcasm comes in with 43,000 sets.

I hoard that which your kind covet. I would like Fountain to allow me to build playlists from subscribed show feeds, not just individual show episodes. I like to listen to my podcasts in groups, you know, like tech, news, dad stuff, etc.

And just can't be bothered to peruse show episodes individually themselves. I want to know what dad stuff he's listening to. I don't have any dad stuff podcasts. Would you boost in and tell me what you're listening to? Give me a recommendation. So I wanted to put this one in the show because I talked directly to Nick from Fountain about all of these. And this one he said is doable today.

You can tag the episodes and then you can build a playlist based on the tags. So that is how you solve that problem. Sarcasm, it is available to you already. And I wanted you to know that. Thank you for the boost. Ian clearly boosted in a mega row of ducks. Things are looking up for old McDuck. For the Docker Compose versus Arian discussion,

I ended up on the Ansible Docker Compose for Ansible vault. I didn't understand secrets when doing it with NixConfig. I get that the config says what file you need, but doesn't store the content or secrets. Am I missing something? We've done this. We've had definitely configs where we refer to a secrets file. What are we doing different?

Well, I mean, secrets in Nix is a whole topic. So that depends on the approach that you want to do. There's the simple, or maybe it's not simple, depends on what you think. But there's the approach described here that we've done There's the simple, or maybe it's not simple, depends on what you think. Okay. But there's the approach described here that we've done and we did in our example Nix cloud setup where you just manage secrets separately.

Whatever you'd like, you could use Ansible Vault, you could use, you know. Which we did link to, so he may be able to refer to that. And that just means you need a separate process that's going to set secrets up where you want them and then the Nix stuff is just set up to refer to those and expects them to exist. All right.

Or you can deploy stuff in Nix. There's a few different options. I mean, you could just embed them in there if you weren't worried about that. But you do have to be careful if you're worried about security or you're on multi-use machines, because depending on what method you use, you probably don't want your plain text secrets in the Nix store, because anyone can just go and read that. But two popular ones, there's AgeNix, as well as Sopsnix, S-O-P-S. And then there's a bunch of other projects as well.

So depending on the approach, the complexity, what your actual needs in terms of deployment are, there are a bunch of ways to do it in Nix. But that's not something that we built in or planned around our solution. And of course, Ansible and Ansible Vault is a time-honored way to do it. Yeah. I mean, if it works. Sounds like a nice setup. Thank ian appreciate the boost sir lurks a lot comes in with 2048 sats hey sir nice to see you so i like hearing your experience with vr headsets like the quest chris but i wonder how well they will work for people with poor eyesight do you wear glasses inside the headset or does it have optics to adjust to focus for each eye? I don't think it has individual eye focus, but you could wear glasses.

I know some people do. For me, I'm, as I age, becoming nearsighted. And so since the lenses are just like right in front of your eyeballs, for me, it's fantastic. I, in Immersed, also realized that you can adjust some of the encoding quality and things. So I didn't actually have everything turned all the way up to like its pristine quality. It's looking even better than it was for last week. I used it for prep on the show. And one of the things that I've also done now with my Immerse setup, which I can't believe I didn't think of this sooner, is I've replicated the monitors like I have them here at the studio. So I have a vertical screen here at the studio and I now have that same on that same

position. It's a vertical screen and I have the, and I now have that same, on that same position, it's a vertical screen, and I have the same horizontal screens. And so it's very much the same workflow for me, in immersed or in the physical world. And I'm liking that a lot. VT52 boosts in with Aerodux. Reusing NixOS.org search for NixOS options. Little pro tip, try searching over at mynixos.com. It searches across options and packages at the same time. So you'll find eg Nix packages slash packaged slash tmux

and Nix packages slash options slash program slash tmux in one search. That's cool. The UX is a little nicer too. You can browse among the options hierarchically instead of one flat list. Bonus, it also searches Nix packages and Home Manager. Yeah, this is a great one. I kind of forgot about it, but I have run into it a few times,

you know, just when trying to search for Nix options and not already at the main search site. It's great to have other options. Thank you, VT. MyNixOS.com. That is great. Thanks VT. Really appreciate that. I have a little pro tip to add to this that I've implemented this week. So I've been searching the Nix packages, um, through duck, duck, go bang implementation.

So you bang Nix packages and then you just search for what you're looking for and it sends you directly to the Nix package search search but i've one upped that this week on kd plasma and i'm using the plasma launcher which has the same functionality if you look deep enough in the settings and i'm going searching directly from my desktop it's amazing yeah nice yeah i need to use krunner more it is the best runner out there i feel like it's gotten a little pokey. I'm hoping with plasma six, which comes out in a few days, I'm hoping to fall in love with K runner all over again. Well, Gene Bean boosted in with a row of ducks.

Hey, regarding fountain, it's missing the dynamic playlists that are used in castomatic and overcast are the show stopper for me. For example, this show is in a playlist called Linux and Tech that contains the shows I subscribe to that fit that particular topic. The playlist automatically shows me the unplayed episodes for these collections of shows. This allows me to see what I've heard and haven't heard yet in a given topic. Tags, labels, and manual lists just don't cut it for me. Hmm. Tags, labels, and manual lists just don't cut it for me.

How many people are doing this where you're like you're play listing out the topic like Linux and tech for Gene being here? I just always have and I'm curious about you guys. I always just have like my main list of chronologically released podcasts and I just kind of scroll through that and pick what I'm going to listen to. And I don't ever create a playlist or anything like that. I think at most I might create like a, you know, per session, like I'm going for a drive and I want to keep up a couple, something like that. But I do have, I do have road trip playlists. What about you, Brent? Do you just stack your, your list or do you have playlists? Do you go all organized? No, I think I would love this because typically when I'm reaching for my podcast players, because I'm in the mood to listen to a certain type of content. And so I end up Chris, like you browsing through the list.

But I find that more to be friction than inspiration most of the time for me. So I think I would love this feature. LazyLocks comes in with 10,000 sats. It's over 9,000! All your talk of virtual displays has got me wanting to check out links in the show notes. That visor looks promising. I hope it delivers. Looking forward to seeing you guys at scale.

Yeah, Lazy Locked. Heck yeah. So visor.com, the deal I think is over now. But it was like 400 bucks for VR headsets that are just designed for working. Really high quality displays that look like sunglasses. So you're not wearing big old ski goggles. I think people like that. Master Rebo with nine thousand cents just following up on a telegram message about the comparison between quest 2 and the quest 3 if i look dead center everything seems fine on my quest 2 but if i stray just a little

things begin to look fussy can you tell me if the Quest 3 is like this as well? The Quest 2 is my first VR, so I don't really have anything to compare it to. That's a great question because you would think this wouldn't really be viable if you looked at this through the Quest 2. I tried the Quest 2. I saw what you're seeing. I tried it this weekend with my daughters. It's one thing for a game. It's another if you're working in the terminal,

you're composing an email, and you're reading text for hours. Huge difference, yeah. Yeah, with a 3D game, you don't really care. But when you're trying to read text on a monitor, big difference. So on the Quest 3, you have a much, much better field of vision. And I noticed, because I had probably, I mean, my screens were at my, I zoomed zoom my screens to test this for you at the maximum width of my peripheral vision. And at the edges of my peripheral vision, the monitor does kind of have like a little kind of a blurry effect to it at the very, very edges. But I'm talking like if you held your hands to the edge of where you can see them, that's where it starts getting blurry at.

But then the field pretty much from in there to the center is looking really sharp. And if you use immersed, make sure you mess around with the encoding options, and that also makes the text look better, I've learned. So you got to – the defaults are optimized for Wi-Fi performance, and they assume your Wi-Fi is crap. But if you've got good Wi-Fi or you hook it up over USB-C, you can step up the quality and the text looks even better. Bear 454 boosted in 5,000 sats. Well, hey there,

Bear. I've tried Fountain a few times, but just keep going back to podcast. And there are a few reasons for that. Number one, the home screen makes it feel a lot more like a social media platform than a podcast player. Let's take these one by one. So the home screen making it feel like a social media app than a podcast player, I agreed with too and was the biggest friction point I had with Fountain initially. Then I discovered like six new podcasts that I never would have found ever and they're my regular listens now and I've changed my tune a little bit. Then I also – I started following a couple of other JB listeners and they started following me. Oh, that's fun. And now I'm seeing their comments and what they're listening to,

and they see my clips and now it makes more sense. But bear, I definitely had to take, I don't know, man, three weeks of just being like, I really wish you would just open to my library. I really wish. And now when I launch fountain, I actually stay on the home screen because I know in the, in the library screen, it's just downloading my podcast anyway. So while it's downloading my podcast, I just scroll through the home screen, see what other listeners have been boosting or listening to. So I know it's not exactly what you want to hear, but you might just give it a reframe of thought. Bear says, number two, I find it really hard to see new unplayed podcasts.

Even when I play an episode, it stays on my episodes list. I would appreciate an option just to see new unplayed episodes. Yeah, or auto-delete maybe when you finish an episode. So they've added swipe to delete, that makes that a little easier. But yeah, I think right now by default, unplayed are still chronologically listed, they just are now marked as played. And bears number three, my biggest gripe is I really would like an option to make a good use of bigger text. The text in the app is just small for me. Right.

I mean, I wonder. I know on iOS you can pop up the zoom up. I don't do that on my Pyzelle, but maybe there's an accessibility option there. I'll chat with them about that particular. And Bear, I know you had some Albi issues too. If you hit me up on Matrix, I can help you troubleshoot that. That's good feedback though. And those are things we will be talking about in our next sit down next Thursday.

Presley Weep HD. I'm going with it. Comes in with 2,669 sats from Fountain. And he says, I got a math problem for you guys. It's the sats times 17 equals my zip code so he's doing a little math game with us did you bring that map oh you did this is why we're weeping in hd i make sense now uh yeah 2669 times 17 that would be uh 45373 and that seems to be a postal code in Miami County, Ohio. Ah, hello, Ohio.

With cities like Troy, Alcany, or Staunton. So they got a county called Miami. And they got a city called Troy. This is great. This is so great. I mean, you're just like, where do you live? I live in Troy. In Miami.

You know, Ohio. Thanks, Weep. Appreciate that. Boost. Difficulty adjustment, Boost-in, with 10,000 cents. It's over 9,000!

Love your work, gentlemen. A video option on Fountain would be a game changer. Little Birdie tells me that may be a possibility. I think a lot of the plumbing is there, and a lot of the underlying

libraries support it. We may get there with our LIT support for our scale trip. We shall see. So would that be taking advantage of like alternative enclosure setups? Yeah, you got it. So podcasting.to.spec has the alternative enclosure and in LIT too you can do here's my primary stream and here's an alternative

stream. You basically put like HLS or RTMP URL in there. And then the client player would just grab that and figure it out. So it could happen. Difficulty adjustment. I like that username too. Barnminer comes in with 5,555 sats. He says, tell Fountain I switched to Breeze Wallet for boosts since it's self-custody.

I actually did. And Nick's response was, yes, that is on the list. It's been on the list for a while. It's still on the list. Well, Network Rob sent in a Spaceballs boost. So the combination is 1, 2, 3, 4, 5. That's the stupidest combination I ever heard in my life! So I'm boosting from Fountain today for a couple reasons.

One, giving a little extra value back to my stream sets, and also, number two, scale! I know it's a little, but I know every little bit helps. Number three, keep up the great work. Love the value to value. And number four, well, since I'm behind on episodes, giving a small hello to future Rob when I get caught up. Eight lines of shimmering cement running from here to Pasadena.

Thank you very much, Network Rob. Appreciate that catch-up boost. We look forward to hearing from you when you do get caught up. So excited about scale in NixCon. So, so glad we're going. I feel like we don't give enough attention to the other events that are going on, but this is the big one that kicks off

the year, right? And we only have so much space in our tiny little brains. That is probably really what the problem is. Ben the Tech Guy comes in with Road Ducks. Says Steam Link is already in the Met little brains. That is probably really what the problem is. Yeah. Ben the Tech Guy comes in with Rodeux. It says Steam Link is already in the Metastore. I mentioned I sideloaded it. And it does support proper VR, but only on Windows and Mac OS.

They're actively working on Linux support. But in the meantime, ALVR works pretty well. Okay. I didn't even look for it in the Metastore. I just got so in the workflow of sideloading APKs, I was like, I'll just go get it. I don't need no store. I don't need no store.

This is Sightel, former iOS user, folks. No, don't you think I would have defaulted to the store? Well, I mean, you're so excited about the ability to sideload at all. Yeah, you might be onto something there. Mr. Pibb boosts in with 12,345 cents. 3, 4, 5. Yes. That's amazing. I in with 12,345 cents. My issue with Fountain is I use an auxiliary cable to listen in my car, and the sound quality always gets distorted at the necessary volume. I use the iPhone app, and I'm always fiddling with phone volume and car stereo volume trying to get it

loud enough without distortion. I have the same problem on Cast-O-Matic. I enjoy the JB shows for the professional-sounding audio quality, and it's great with my Bluetooth headphones or a speaker, but just not with the auxiliary cable in my car. But if I use Spotify or Apple through the aux, it sounds just fine. Now, what could be causing that? Distortion when you're using certain apps, but not when you're using Spotify or the Apple

apps. Are these other apps, you know, normalizing, boosting the audio to a level where, you know, there is some distortion at those higher levels, but you just don't have to get to the higher levels with these other apps? So it's something like a, you know, a louderizing sort of. And it only happens when you're using an aux cable. That's interesting. Yeah.

If anybody has any ideas on this one, let us know. Also, remember, we're collecting your boost ideas for AI resources and tools that we should be looking at for the show. So please boost those in. Thank you, everybody who did boost in. If we didn't read your Fountain FM feedback, don't worry, Nick and I covered it,

and we're still collecting all of them. And check your profile. He's probably replied to you directly. We had 24 boosters come in, and we stacked 222,219 sets. Almost a row of ducks. A big old row of McDucks. So I'm going to give it to us. Thank you kindly. I'm going to call it a big old McDuck.

Appreciate everybody who boosts in. It's a great time to get a new podcasting app because we are working on new features. We've already rolled it out to our members. They're loving it. And they're going to be coming out to the main feed very soon. And some of them are going to start as we head off to scale. And we'd like you to be there with us and keep podcasting weird and decentralized. Podcastapps.com, Podverse, and Fountain FM, and Castomatic are really the favorites of the show.

You pick the one that works best for you and then boost in with a message and you can support us directly that way. No middleman, all using an open source peer-to-peer network. And thank you everybody who streams Sats to us as well. We appreciate you. And a big shout out to our members out there for using their Fiat Fund coupons to keep us going on the regs every single month. We really appreciate them too.

You mentioned that a lot of these large language models and whatnots are coming out with a paper. And I think one of the best ways to just wrap your head around what models have come out recently and what problem they're trying to solve and if you even need to pay attention

at any of that is probably the daily papers over on HuggingFace.co. And every day, and it is every day, it's pretty wild. I mean, every weekday, they're posting just one white paper after another. And it's, I mean, I don't know, Wes, it's a fantastic resource for just kind of watching the development of these. You must have taken a look at this earlier.

It's a, I mean, HuggingFace.co itself is a resource. I'm going to assume most people listening to this already know about it, but as a higher level recommendation, hugging face.co, if you're into any of this kind of stuff for the community around these open source models and data sets and yeah, they host a bunch of, a bunch of stuff like that. And then they have these daily papers and that's at hugging face.co slash papers. And these are the white papers that accompany the announcement of these large language models or whatever they're working on. And I think just you can look at the headline, you can look at the summary and know if it's something that interests you or not and pretty quickly stay on top of all this stuff. I love resources like this.

If you have any out there, let us know. Yeah, and you maybe take a look and see just how fast it goes from paper to something you can actually use. Yeah. Yeah. Yeah. I would say if you quit Reddit recently, you can go here. Even if you don't want to dive into the papers, just looking at the names of these projects is such a treat. So just scroll through the names. I'm enjoying that.

Yeah, I feel like this is our moment in history as a community to keep this stuff open and to keep the open store stuff viable, to keep the community vibrant. You guys probably saw that on the back of all this Gemini news, Google also announced though a massive, what was it, like $60 billion deal or something like that with Reddit to pull in all of Reddit's data to train AI. That cost and those kinds of deals, those are going to always be relegated to the absolute top, right? The big tech companies that can afford to write 60 bill or 30 bill or 10 mil or whatever it is, 30 mil. Yeah. I mean, JB tried to buy the Reddit data, but they weren't interested. Yeah.

It turns out they wanted more than a row of ducks. But like you look at things like Hugging Face. You look at Invoke AI. You put it all together together and we still have a very, very viable shot here. And I think ultimately if I could be a dreamer long term, I'd love to hear stories of listeners that are implementing this stuff in small businesses or in their workplace, maybe even large businesses. Because that's where I fear the most that we're going to get locked into corporate AI. And it's going to be this risk averse, not very useful, overly talkative, watered down crap tool. And the reality is, and the BSD folks have always been right about this, is a powerful tool you can shoot yourself in the foot with. And you can do good and bad with rm-rf. And I feel like it's the same kind of general principle with this stuff.

So if you're out there in the field, you're deploying this somewhere where you've got more than a couple of users using this stuff, boost a write-in and tell us how that's going. Yeah, are you relying on AI-powered stuff for your business? Would you? Do you feel differently about it if it's an open-source thing versus something you can only interface with by API? Yeah, I sure do. Yeah, I sure do.

It's just a wild new frontier. And we're really just kind of at the beginning. I thought we'd see how NVIDIA went. And depending on how NVIDIA went, we knew how much momentum there was in the industry. And NVIDIA popped to the high end for sure. I think there's still a lot of legs left in this. for sure. I think there's still a lot of legs left in this. But where it actually ends up ultimately and what kind of end result we get,

that is not so clear at this point. And we, I believe, still have a chance to shape that future. We'd love to have you join us live. You can get in that mumble room, which is popping this week. Lots of folks in there. We always hang out with them and they can pop in during the show or before and after the show is really when we just get to chatting. It's great to hang out. It gives us that live vibe. We've got details on our website for that, and we do the show on Sundays at noon Pacific, 3 p.m. Eastern. Links to what we talked about today, linuxunplugged.com slash 551 for the tools and the links and the resources, all that kind of stuff.

If you thought this show was useful or maybe somebody else should hear it, please share it. Word of mouth is the number one way podcasts get marketed. It's not really through anything else. It's not going to be Google ads, I'll tell you that. So I appreciate that too. Thanks so much for joining us this week. See you next Sunday. Thank you.