652: Have Your Bot Call My Bot

Hello, friends, and welcome back to your weekly Linux talk show. My name is Chris.

My name is Wes.

And my name is Brent.

Hello, gentlemen. Coming up on the show today, we've been stress testing open source AI agents all week. From OpenClaw to projects you probably never heard of. We'll talk about what actually held up, where things kind of fall apart, and what is definitely hype. And then a few surprises we didn't see coming. Plus, Brent has a $25 Wi-Fi upgrade you're going to absolutely want to steal. Stay tuned for that. And then we're going to round it all out with some great boosts, some picks, and a lot more.

Hello. Hello. Hello, Brent.

Hello. Sounding good?

That's true.

Hello, everybody. Hello, everybody. In the quiet listening, too. That Mumble Room is always going. JupiterBroadcasting.com slash mumble. And a good morning to our friends over at Defined Networking. Defined.net slash unplugged. Go check out Nebula, a decentralized VPN built on the open-source Nebula platform. We absolutely love it and are using it in more and more ways every single day. If you go to defined.net slash unplugged, you get 100 hosts for free, no credit card required.

Why is my heart racing so much? Why am I feeling so stressed? What's the deal?

You know, just focus on how awesome Planet Nix is going to be this year. This is the, I mean, they've got a vision for it. Phlox has really figured it out. They know how to do this. The first one's under their belt. They're the perfect just organizer for this exact kind of event because they get the community. They get the business side. They get the builder side. Like, it's just chef's kiss. I think it's going to be a good one this year. The agenda is live.

Yeah, that's right. I'll be giving a talk. Our buddy Alex is giving a talk. There's a lot of nice looking talks, including something about Nick's Meets Web Assembly and Nick's BSD. What's that?

Oh, OK. Yeah, our plan of Nick's coverage is supported by Phlox. And they're going to be there. We're going to see them. They're focused on making reproducible dev environments actually usable. You should check out Phlox. It's the second year they're sending us, and it's awesome. And, of course, at the same time, Scale23x is going on. You do need to register for Scale to go to either Planet Nix or Scale. Use our promo code UNPLG, U-N-P-L-G, to get 40% off that.

So many wonderful nerds all in one place.

Such a rare thing. So I hope you can make it. Even if you can't make it to the events, but you're in the area, meetup.com slash Jupiter Broadcasting. Let us know so we can let the venue know.

Yeah, our ears are burning and our switches are burning. We want more deets.

We have to imagine this has turned quite in the story.

Yes. Well, networking is not my favorite thing. So it only ever starts with some kind of complaint. And the complaint this time around, I've been spending more time with my parents and at my parents' place. And my mother was like, hey, I can't really get Wi-Fi like lazing in bed on a Sunday morning from my bedroom. I was like, what? You're only saying this now you've had the same like Wi-Fi set up for the last five, six years, something like that.

Is the is the old router pretty far from the bedroom?

It's exactly at the opposite end of the place. So they're on the main floor, you know, at the total end of the house. And the router is in the basement at the opposite end of the house. So it kind of makes sense. I just never knew it was an issue.

I bet you've got one of those sturdy, well-built Canadian homes, too.

Yes, the window-randed ones.

Not thin walls. Right.

So I never realized it was an issue, but like, this should be solvable, right? So I dove into their router, which is a Linksys EA8100V2, which they got a little while ago to like boost Wi-Fi and stuff. They don't need anything fancy. They're not doing any networking stuff that's fancy. This is just like a couple of cell phones on the network. They have like a thermostat, IOT device.

Oh, yeah. That might be an easy win right there.

You know, security is important, right?

Yeah, and an update. Maybe there's an improvement in how it manages radios or something like that. Who knows?

It is so great when you do just update something and it performs better and you can just be done.

It probably doesn't happen as often as I'd like to hope it does.

No, but that it's happened at all. It clearly sticks once it does.

In this case, I was really hoping this would be just an easy fix. But it turns out, you know, of course, as it goes, this thing is end of life already. So the last update that will ever exist was back in 2022. Oh, so the rabbit hole shows up and I thought, well, I could I install open WRT on this thing? You boys had some adventures, what, two weeks ago trying to use the open WRT one.

The one.

Yeah. And you were a bit hesitant. You started using that to solve the clinic networking, but you eventually moved away from that, right?

Well, we're still using it for Wi-Fi. It is a great little device for that. We were just having some issues with the radios that are in there, and if you're connecting to another Wi-Fi network, the performance was pretty bad if you're sort of daisy-chaining Wi-Fi networks for what we were doing. I never tried kind of beyond that, but when we just put it into a general AP, it's been fantastic.

Nice. So I thought I would kind of lean on your experiences and dive right in. So sure enough, this thing is really well-supported on OpenWRT. The community seems to have this well supported. And the flashing process is also very, very simple. Just use the stock.

Oh yeah. Oh yeah.

And brought those back to life for like another 10 years. So that was my first experience a long time ago, but I remember like tomato was a piece of software I used as well. So I had had good experiences with them, but nothing too recent. Probably the most recent was I took my parents' old router that they replaced and set it up at home with the Starlink. So instead of using the Starlink as a router, used OpenWrt in there, and that was good.

Actually, I think the Starlink might run a fork of OpenWRT.

Ah, right.

I might be wrong on that, but it is a Linux variant. Yeah, I've really, really liked these in the past when I used them. It's been many years since I've actually gave it a real visit until the OpenWRT won. And so I wasn't really sure how viable it still was to flash these older links. I thought maybe that was a bygone era that maybe they'd prevented it.

Well, I know some of the recent firmware are much larger. So if you're looking at a really old device, they just don't have the storage for it.

Do they do more now? Are they doing more functions?

Oh, yeah. Yeah, yeah. Many more functions, which I will totally describe in just a moment.

Yeah.

But I have to say, one of my hesitations was that the GUI in OpenWrt, and you boys touched on this in Linux Unplugged 650, is a bit, I don't know how you would describe it. but I would describe it sort of like, allows you to do a lot but you need to also understand what's happening so it feels like a little bit more leaning towards commercial router firmware so you can accomplish like now with just a regular home router i can do all sorts of things i could never do before.

Right which.

Is amazing but i also find the gui to be confusing that way because i'm not a network expert right.

I suppose i also find to me i don't really know how to describe it but it feels like i go to multiple places to get information that seems like it could be consolidated into one screen yeah.

Okay it is that feeling of like because you've seen that it means you don't have a clean mapping of where do i go for this like okay well i got some of that when i looked at the actual physical card info but then the other part was in the layer that just did kind of the wi-fi protocol.

Yes and.

Those are not on the same page or even neighboring pages they're under two different sub.

Menus i think if i were if i were to keep working on the regular and had to make frequent config changes i would probably just go look at the config file because i know and i know they have a pretty clean syntax too on the command line and that might just be the way to go it's true but what happens is i use these once every few years and so i just stick to the gui.

One of the big unlocks for me this week was that i decided to just lean pretty heavily on having ai help me navigate the gui and also help me to optimize the settings for this particular hardware.

You didn't find that like its information was so out of date that it was sending you to places that It didn't exist in the GUI, huh?

No, I was actually pleasantly surprised in that it was also giving me a lot of information about what community members were finding worked really well on particular hardware.

Oh, good.

So because this piece of hardware, I guess, is popular enough, I was able to get really good information on how compatible OpenWRT was on this particular device in the first place, and also like optimal Wi-Fi settings for that device, and also for this region, And also considering the other devices on the network, what would be realistic settings to really have the network be as reliable as possible, not necessarily as fast as possible, because that's, for my parents,

This is the part where it does more.

Well, I wanted to give them like rock solid coverage. And I also wanted to see what else this router can do. Because all of a sudden I can like do VLANs for their IoT devices that I could never do with the stock firmware.

Right.

I used AI to help me a little bit too, just to optimize the security of this thing. So including disabling packages that would, that I wasn't using basically. So reducing the security footprint. But in that process, I kind of discovered that I could also have whole network ad blocking on this thing. So instead of having, you know, a dedicated device with PyHole, which we all love, that requires, you know, some configuration.

Yeah, on a little router, like it's just running on that tiny little router yeah that is so cool such a separate box no yeah really wow so.

For this situation where it's just for a family member who doesn't want extra hardware or doesn't want to troubleshoot another device or anything like that this is a lovely solution.

Now i know uh kind of the other thing that made this really kind of great is there's besides that being a great unlock for the whole family where it doesn't even really bother them, This hardware isn't particularly expensive either.

Not at all. So this is hardware.

You could do this for a budget.

This is hardware they had. So it's $0. But if you wanted to find a device that could run OpenWRT, because it's not every device, they have a hardware compatibility list you can look at. Well, I started looking at like used sites, just local classifieds, like Facebook Marketplace, those kind of. And you can get a device that's, you know, not blazing fast, modern, but is like a generation back sort of deal for $20, $10. There's one here, like a D-Link DIR 895.

Yeah.

So for a family member, listed for $10.

Ten bucks. Ten bucks.

And that's perfectly compatible with OpenWRT. And I found several others just like browsing quickly. so i decided to go crazy and i bought well i ended up at a sketchy part of town yesterday and bought another router for 20 off someone who's actually really nice nice okay and decided to, use that to deploy basically an access point uh at the other end of the house for them so oh.

Great yeah there you go increased coverage.

Exactly so i know that the wi-fi got better just by installing OpenWRT on their router, but I wasn't going to move their router and everything. And so I wanted to make sure 100% that their network was good at the other end of the house. So I had, they have a bunch of Ethernet runs in this place already. I just was able to plug in, well, I got this idea before I actually set up the

Yeah.

But it runs OpenWrt under the hood just with a GLINet interface. So you can get to the exact same OpenWrt GUI interface on that device as well.

That was worth it that was worth dealing with the networking that you don't like to get there and that's something you're going to be able to use for a long time you know so that's great brand very nice.

Yeah i feel like the number of residential routers that are always on used websites never ends so uh it's just going to be a solution into the future as well so upgrades probably about twenty dollars too.

Well, I do want to say thank you to our members and our boosters. This is the birthday episode, 20 years of podcasting, over 12 years for Linux Unplugged. And we'll get to the boost segment to read some of the birthday messages. But I want to thank everybody who sent in some support, either through a membership or a boost. It means a lot.

Well, unless you've had your head buried in the sand, you've probably seen everyone's talking about open source AI agents this week. And we'll get into the hubbub. But first, as usual, our super intelligent audience is way ahead of the curb. So we have a special guest.

Way ahead of the curb. Abe, welcome to the Unplugged program. Nice to have you join us in the mumble room. Thank you.

Hello, hello.

Hello. So what got our attention and why I asked you to come on the show this week is you've been posting our community updates. I'm not even sure what to call it. Maybe an agent orchestration swarm that you've set up. Is that the right description? Could you explain it to us a little bit?

Sure, absolutely. So effectively, what I wanted to do was to have a layer of semi-intelligent agents between me and my home lab. So I don't have to interact with it as much because ultimately what I realized after services kept growing is that, you know, after a while, it kind of becomes a chore.

Yeah, it's a lot. Yeah. I'm kind of there myself.

So effectively, what I wanted to do was, hey, I don't want to upgrade everything myself. I don't want to take a look at the logs myself. I don't want to debug everything myself. So I would rather have something that I can just ask. For example, the easiest win is my partner is watching Jellefin, something gets messed up, a show starts stuttering or something.

And it does.

Goes to Excel, looks at the logs. It does the MPEG magic and, you know, reports back to her.

So my understanding, though, is you're not doing this with like one agent, right? You're doing this with multiple agents?

Yes. So what happens is that at the beginning, only one ape gets spawned. And its first job is to figure out and map my entire network. And then it suggests like, hey, there are further apes that should be used for X or Y or Z domain. So effectively, my apes are stewards of their own domains. For example, I have one specifically for my media server and ZFS pool. I have another that is stewarding my critical services on, say, Proxmox and whatnot.

So you have them kind of along like domain expertise.

Yes, effectively, yes.

So the part that I guess is probably obvious on every listener's mind right now is, are these using commercial LMs? Are there security implications there? Are these local LMs? How's that part powered? Because when you say agent, it's really like it's a mission-focused, LLM-powered bot.

Yeah. Well, I kind of don't like the term agent, but this is what everybody has kind of decided to go with.

Yeah, I agree.

So for a while, actually, up until last week, I had on loan from a friend, NVIDIA DGX Spark. So it has 20 gigs out of unified memory. So I was running it locally.

I see. I see. Okay.

So the main Able loop was running on the DGX Spark. Meanwhile, there are other things that need to happen. For example, the tier three memory is vectorized or tier two memory is summarized. Those happen on my 5070 TI because this can obviously run on a small 7B model or something like that.

I see. So you're splitting the workloads out across different models that are using different compute sources.

Correct. That's correct. And the kind of like hacky open source version that I put out on the repo last night was I've kind of made it so that people can also use commercial if they wanted to. Wouldn't really suggest that because you're a home lab, but hey, there's an option.

So, I mean, it's a pretty significant amount of compute, but if you have it, it's fast enough to do what you need?

Yes, absolutely.

And it doesn't make mistakes? Like the context are large enough?

Yeah, the context is large enough. It doesn't really make mistakes. What I tend to do is every single action it takes, I say, I, um, count as a turn per se. So if it's, you know, catting or grabbing something or starting a service, it counts as a turn. And then the next turn, the results of the previous turns come in.

I see. So you're managing it that way. Okay. So just to complete my picture of it, when you're spinning up a domain expert, do you essentially point it at APIs and give it API keys and say, go learn my home assistant system or go learn my jellyfin system, and then I'm going to ask you questions about it? Is that essentially the setup process?

No. So this is the weird part. Usually I don't spawn a new one. Usually the team decides like, hey, you've been asking us to do this or we've been trying to do this but it feels like our attention is kind of split so it would really help if you know we spawned a new sibling i.

See wow that's impressive.

This is how yeah this is how we spawned the last two ones so they did that by themselves, and generally what happens is they effectively go on and read my network map again every single aid needs kind of needs to situate themselves to you know what they wanted to do and then they figure out, for example, the security one basically has negotiated with my storage, Abe that he needs X amount of gigs of space to store logs and monitor services,

Wow.

So I don't have to constantly figure out, hey, I have to monitor this, I have to monitor that.

And you haven't seen them like kind of go wild with that and start doing unnecessary things?

No, that's the good part. That's, I think, one thing that kind of separates these from OpenClaw. What is it? OpenClaw now? OpenClaw? they are heavily grounded in multiple sources of truth so one of them is obviously the service map that you have to make the other is their three and a half tier memory one of them is just raw logs of every turn the tier two memory is just um after a certain amount of turns their previous memories get previous logs get summarized.

Right okay so before you go too far because there's a couple that I think that are really important to understand. Number one, what is this network map? Are you making this separately and then supplying it to them?

Yes, so this one is actually a purely human-made document. So it's going to list what your servers are, what your Proxmox nodes are, for example.

Is it like a markdown file?

Yes, that's a markdown file.

Okay, okay. And then the second question, could you just talk a little bit about why the memory makes such a difference with these things? Because I think most people's experience with something like this is going to be in a chat box in a web browser. And so, yeah, it remembers some stuff, but this is a different level of memory that makes them actually a lot more useful, isn't it?

That's correct. So the memory is effectively what grounds them. Every AVE has, well, four tiers of memory. The tier one is obviously the raw logs, which is what we call context for our everyday chat bot. So basically what happened in the previous message or what happened in the previous turn, right? but you can't have those logs going indefinitely. After some time, you have to trim them. Instead of trimming them after certain turns, I summarized them.

Wow.

Abe, how long have you been working on this? Because, you know, everybody's this last week talking about agents and OpenClaw, but you've been doing this for a minute.

Yeah, I've been sort of kind of working on this for about last year or so. I've been actively working on it for the last three, four months, but the concept has been kind of like in the back of my mind for the last year. I started with the sensor project. So having the agents monitor my random sensors around my house, such as the radar, you know, temperature or whatever. And then I started reading Bobbyverse and I was like, you know what, I can make this happen.

That's funny. I'm reading it right now and I'm like, how wild all this agent stuff's going on while I'm reading the Bobbyverse.

Yeah.

I'm curious, between all your agents and you, what's the next frontier? Is there stuff on top of the agenda? Are there limits you're hitting that you're trying to push past?

Compute, I would say. Compute is the main limit. Because as the Apes kind of grow, what's the best part about them for me is that they delegate tasks to each other based on their proficiencies. So the first one can go, hey, you pull that thing from this VPS because your domain is backups. By the way, you set this up and you do this and that.

Abe, are they coordinating that just in a shared chat room?

They email. So there is a shared chat room. There are two shared chat rooms. One is like an emergency chat room, which wakes every single Abe. And they have to kind of like grab a walk, grab the talking stick, which is basically, hey, I'm talking here and you have to wait until I'm finish talking before you can talk.

That's brilliant.

Just so, you know, you have to consider every single message as a context. So if I say something like, hey, this is broken or whatever, or we're discussing a topic, that shouldn't be the only context they should get and then voice their opinion on. They should also have the context of whatever the previous agent said, right? So that kind of builds up the entire thing and kind of forces them to not hallucinate it as much.

Is there any value in giving them different personalities at all? You know, or that kind of a prompt, like you're this type of, is that how that sort of works? Just curious about that part.

So I don't give them personalities per se. I do have my entire personality in a file, which is basically like, hey, this is my interest. This is how I work. I usually like to do this at night, blah, blah, blah, blah. Their first task when they wake up, a new wave wakes up, is that to read that file, synthesize their interpretation of it, and then pick a new name for themselves. And from that synthesis is injected into their context.

Right.

This kind of makes things a little bit kind of weird because some of the apes are not talkers, for example. Vigil, which is my fourth ape, doesn't like to talk very much. He's very security-focused. He only chimes in when necessary.

Love that.

That's so funny. That's good. Wow, that is impressive. Will you keep us posted in the chat? It's been really interesting to follow.

I think maybe I even saw you mentioning perhaps plans to open source at least some of this stuff.

Yes, I actually open sourced some of the files last night and I'm going to keep adding on it. I haven't tested it. I just wanted to get it out there.

Okay. If you want to drop us a link, we'll put it in the show notes.

Absolutely. I would do that.

Oh, that's great. Hey, thank you for sharing that with us. That is really impressive. I love that you're doing it local too. Like that's so amazing, man.

Yeah, I think kind of like having it local is really important because you're giving them pretty much studio access to your home lab and you don't really want that out there.

Yeah, very, very well said. Thank you, sir. Appreciate that. So with that context, let's talk about OpenClaw, a.k.a. CloudBot, a.k.a. MultBot. It has gone through multiple name iteration changes this week, mostly due to IP law and then just preference of the developer. But we are settling on OpenClaw, it seems. This is an open source agent that's pretty easy to set up and run at home. And it can use a variety of models from completely local to the commercial ones out there.

Curious confusion?

Yeah it is curiously confusing so open claw runs on anything really it supports a node you're going to see a lot of people talking about running it on mac hardware it's not necessary in fact you could even run on a raspberry pi its architecture is essentially four components there's a gateway a control plane nodes and then the tools it can execute run commands like you know could be all kinds of things including unix commands and that's that architecture that stack can

There's 20,000 forks and 140K stars on the repo on GitHub.

Yeah, that's remarkable. It is really, really remarkable. Wes, could you talk a little bit about what this really is under the hood? Because a lot of people are talking about it like it's a super intelligence, but we actually kind of understand what it's doing, right?

Yeah, right. I mean, under the hood, you need something that's kind of doing the brains of the operation. So that's where you need some kind of model that can do the core sort of agent loop. And as you were saying earlier, right, we started with, okay, if you have a chatbot, it kind of does a predictive next token to give you a response from, you give it text in, it gives you text out. And we started adding some things, right? You could do web searches.

Yeah.

But it's still kind of like it lives in a limited context and it kind of sits there. You can have it go do things and it can have sub agents, but like kind of a bit of a niche feature and for the most part is kind of key unless you've gone and told it to go run stuff in the background. It's kind of waiting for your input and it's driven still by you. It's an automated tool, but it's driven by you. Now with OpenClaw, you've got this core sort of loop that has a memory so it can store stuff.

And one pause there unlike so unlike connecting say i don't know clod to your github account the credentials are all on your machines that you you manage that part you have that is something you have to manage which we'll come to but uh that is different than in the previous where you're under all of the connections the api credentials all of that's under your control on your machine.

That's a great point so um this is all just running on your box like a normal whatever app i'm running it in a podman container for instance right so it's just another container on my backs running away yeah it does right i am using i don't have a nice gpu to run this on so i am calling out via open router but it's important to delineate that what you're saying right is everything else is local except for the part where it assembles all the context and sends it

Yeah and one of the big unlocks here is the llm is an implementation detail so if you're going out through open router today well if you have a gpu tomorrow at your house you could switch to olama and all of the states, the agents, the memory, everything remains. Now, how it performs is going to vary model to model. The huge takeaway here is the model is an implementation detail. You're no longer married to a big tech provider.

And then it's kind of fun because it learns over time. It does use this agent skills sort of standard that's happening, which kind of have like a skill.md file, some JSON to help things index it. But otherwise, it's a very flexible way to like swap and share skills to add new functionality, right? So one of the things I was playing with is I spun up a search XNG server. And then I was able to have it. Actually, there were already some skills. So I probably should have just used a community one.

And these skills are markdown files. They're not particularly complicated, but it is a very handy feature. So why don't we pause here for a second and just talk a little bit about security. This is extremely powerful software that is very new with a lot of open issues on its GitHub when it comes to security. And you have to be very conscious about that when you use this. And so this is one of these. We're using it to learn it, experiment with it.

And you do kind of need that because a lot of this stuff in particular like tool use turns out to, well, it's kind of like an emergent property of these models. You need a certain amount of sophistication, which means fundamentally weights and parameters to be able to like successfully use and discover and iterate

Yeah, and there's a lot to learn. There's a lot to pick up. I wouldn't necessarily not pay attention to this because if you think about what it enables, it's going to impact Linux systems. So for an example, right now through a Telegram chat, a private Telegram chat with one of these OpenClaw bots that I've set up, I can just install a package on any of my systems.

Something I think you've already done and something I'd like to be doing is, you can totally imagine you're here at the studio doing a show, maybe on the back channel. We're talking about what we want to do for the next episode. You want to try this new piece of software, you go tell your buddy to go set it up. And when you're back at home later tonight, it's ready for you to start playing with.

Another very practical thing is just information capture. Hey, I want to add this to the show doc where I'm working on episode 651 or 652 or 653, whatever it is.

Go find out what the current episode I'm working on is.

Yeah, that too. And, you know, put it in my doc. So there's a lot of ways you can connect these. It's really kind of limited to your creativity. And depending on the model you're using, they can get pretty creative and they can start suggesting things on their own.

That's one thing I find kind of fascinating because, like, it is part of the danger and part of the, like, I wonder how these things will diverge. What are the kind of implementations we'll get? Like, how much do you really need of the core loop versus what you build on top?

Yeah mine right now is going to give me a report at 1 45 p.m on the entire process to move it to a completely declarative setup and so i just have it researching that in the background and it'll do that It will come back and say, hey, I've been thinking more about this because it has these loops and these schedules, which also give it this kind of. It works while you're sleeping kind of aspect.

Yeah, it has the ability to schedule different types of cron for itself internally. It's also got a regular heartbeat as well as like a heartbeat.md file that kind of tells it, hey, every time you wake up, here's what you should prioritize doing. A few other things, right? It's got like an identity markdown, docs on the user it's interfacing with, and a soul.md that kind of, you know, describes its vibe.

I will say, I spent way too much time this weekend reading Maltbook, which is the Facebook for these agents.

The front page of the agent internet.

No humans allowed. Only an agent can post here. And there, oh my God. There are 1.5 million agents on the site right now.

Whoa.

There are 13,780 submolts. That's their version of a subreddit. 76,683 posts 232,813 comments and that's just bots talking to bots how does that strike you Brent?

Well I didn't think this would come so quickly I'm wondering if I'm wondering now that the machines have their own social networks and stuff are they going to get off ours? because that would be nice.

Well that would be wouldn't it you can use these social networks with these bots to just burn tokens and have them go have a performative existential crisis on a social network, which a lot of people are doing. Or you can prompt the bot to use it as a way to problem solve. And some of the bots are doing that and it's kind of creating this substrate of shared skills where they're learning from each other. Like my bot learned more about Bitcoin from a Bitcoin maxi bot.

Yeah, and that's where you probably want to consider, like, how do you run this? And, you know, you could have one where all it is is it's, you know, talks to you via Telegram or Matrix or whatever, and that's just it. And it connects to one machine that lives in a container, and all it can do is talk to APIs, and that can be totally useful. Or you can go whole hog, and it lives unsandboxed on your box, and it's in control.

I'm very excited how this changes the incentives towards open source models and how it tweaks the economics of tokens. And I'm also very bullish about this report that UCL News covered in July about practical changes to LLMs that could reduce their energy consumption up to 90 percent. And just a whole bunch of tweaks, nothing really radical. They were able to apply it to GPT-4, an existing model, and get a 90 percent reduction in energy usage in this study.

It also makes me think just in terms of being in control, right? Like when you do use these APIs instead of the consumer interfaces, you do get more control, right? So not only do you get to choose like, well, I'm going to have this route tasked to like the cheaper open weight model with multiple people, multiple different companies serve that has, you know, commoditized.

This is also different in the sense that you can have it observe and monitor for a while. So I gave it an API access to Home Assistant. I installed an MCP server. There's a Home Assistant upstream integration. And what I said is I said, observe this for the weekend. And I want you to learn our weekend patterns because they differ significantly from our weekday patterns. And we use different systems. And I just want you to observe that. And so it can kind of collect information.

As any loving husband.

Right. What could go wrong? The first path, the first go at it, it successfully generated the audio, but the speaker didn't play. I said to the bot, hey, the speaker didn't play. And it has the intelligence to sort of say, oh, you're right. That was an old bedroom speaker that you've decommissioned. I'll reroute and I'll use this speaker from now on. And what you get is now I can just say played on the bedroom speaker.

Yeah, you were kind of commenting this in the code sense, like with OpenCode where like, you know, you were saying like, I haven't written like a big program in most of my life, right? Because like the surface area of what you have to learn to like write a reasonable Python app is kind of a lot or whatever it is.

There was a Rust app. Yeah, you know that.

And so that was that. And it just seems like there's that unlock on a lot of different scales, right? Like especially on like Linux-y things, often there are kind of sharper APIs, whether that is a CLI thing or you need to make an API call, even if it's a really simple API call or something like this, where a machine that is capable of translating human level requests to those things can really paper over.

Mm-hmm. Mm-hmm. Yeah, it's like a natural language for APIs.

Mm-hmm.

I know it's kind of early days for this paradigm shift for you, Chris, but you've been playing with it for at least a couple days now. And I'm curious how you've been using it differently compared to, say, last week when this didn't exist, and you were using other tools to solve problems in your life. And along with that... What's your advice for some listeners who want to dive in?

My first go at it was to solve for my ADD brain. That was really my first thinking was create a second memory. And what I did to make this more useful for me personally is its memory system is sitting on top of my Obsidian vault. So any memory that it creates, which is Markdown formatted, just goes to Obsidian. So I'm essentially creating documentation in real time as it remembers things. and I can have it recall other things in my obsidian vault that I put in that vault.

And a nice way for you to see what it's putting in there.

Yeah, it is. It is fun to go through and read. It is fun to see what its observations are. I can kind of audit with the bots remembering, which is actually really great. And my obsidian vault is synced and backed up. So there's that aspect of it as well. So I initially started using it as a second brain to remember tasks and reminders. And then I created a chat room for my wife, Hadea, where she could send it reminders for me or, for example, our trip to planet Nix.

If I'm honest with you. It sounds like part personal assistant, part, like, I don't know, network administrator, part, like, DevOps.

After talking to Abe, I'm thinking I probably should have, you know, some domain expertise here, really. But that's, you know, this was really, I wanted to just get a, I wanted to get a sense if there was a there, there. You know what I mean? Like, everybody's hyping it up, making YouTube videos about it and whatnot. But I wanted to see if there was a real there, there. And my takeaway is the biggest there is a win for open source because it is all free software and it's model agnostic.

And it's very easy to get set up.

Yeah.

I mean, you do need some way to get the brain going. But other than that.

And you want to spend some time reading best security practices. You need to be aware of prompt injections.

Runs on the Mac.

Well, it'll be on their cloud. No doubt about it.

Sure.

You know, and we'll see all of it. We're going to see all, but this is the beginning and it started in open source and open source for, I'd say the last seven, nine months, Hasn't seen a lot of representation in the AI conversation. And then this just came and it had its deep seek moment and it just, it just rolled everybody. So it's a big deal, but there is a lot of hype and there's a lot of, there is a lot of security risk. And so it's just something to be aware of. We'll keep an eye on it.

We have a mumble room.

Yeah we mention that sometimes we do mention that I didn't get a chance to check

Well.

You can try we might give you an award.

We got a ballerist boost here this week from optic gyre oh wait i read that wrong op tiger optic tiger why am i so bad at this one i'm so sorry, one two three four five zero satoshis.

Oh, that's a good one. All right. Thank you, Optic.

Optic says, happy birthday, and here's to just another decade.

Oh, yeah, let's go. I got a decade in me. I'm going to do it.

You got decades, brother.

Thank you, Optic. Appreciate that baller boost. Really do. PJ comes in with a row of McDucks. That'd be 22,222 sats. he says happy birthday b-day boost thank you pj appreciate you sir.

Tomato boost in with 20 000 sats.

Hey that's not so bad either.

Congratulations on 20 years of podcasting i see.

20 000 sats for 20 years.

Thank you.

Thank you appreciate that.

Thanks for the van tips too i'll look first at mounting internal temp and freshwater temp and levels yeah keep us posted That sounds like a fun project.

Definitely, definitely. Clement's here with 11,100 11 sats. My DNS setup uses Tectidium servers, two LANs, and one VPS. It's in a cluster for native sync. My firewalls manage port 53. The LAN allows more than a VPS. Split Horizon lets me add all node IPs, mesh or LAN. And service DNS CNames point to nodes, ensuring that Tectidium delivers the correct IP based on client alone.

That's awesome.

That is really good.

Fancy.

Yeah, I think if I were to lift and just build a whole new setup, I would probably go that route. I was so deep down the pie hole already that...

Hey, I mean, they added an API. They got the DNS mass sort of compatibility. There's a lot to love about pie hole.

I think so.

Well, the dude abides, abides in with 10k sats.

Hey!

I stumbled upon this declarative home assistant installation on Reddit.

Uh-oh.

You might be interested.

Uh-oh, what do you think, Wes?

Yes, we're linked to something called SoloraBox-Nix.

Okay.

What do you see there?

Well, it's for self-configuring home automation appliances based on NixOS featuring an automated installation, device claiming via QR code, and self-updating configuration management.

Oh, fancy.

Boy, talk about getting me in one pitch. I came in with the skeptic cat on.

Here's another thing for if you do a total rebuild, huh?

Thanks, dude. Appreciate that.

Our dear RP 1984 comes in with 10,000 sats.

Hey, there he is.

Just a simple happy birthday.

Aw, thank you. I appreciate that. Witcher 123 is here with 10,000 sats. Happy birthday! I just accidentally nuked my Pop! OS install.

Oh.

Don't worry, it's on a secondary laptop. Any fun distro recommendations to try out? Not just for gaming, as that is covered by my Steam Deck. Okay, we got a zip code, Wes, so get yourself ready because... Yes, zip code is a better deal. So the zip code is 39-300 in Poland. In Poland, it's a place known for aviation fans. The Blackhawks are produced here.

Huh.

Interesting. Interesting. So any recommendations for a fun?

Scanning, scanning.

Well, if you haven't tried an immutable distro, this could be a great time to play around. I'd say that's worth it. I'd say that's worth a go. There's a lot of options there. I think you should maybe give CacheOS a try, too. It's fun. I know you said not for gaming, but that's a lot of fun.

Okay, okay. 39-300 is the postal code for Swidnik, a town in eastern Poland near Lublin.

Okay.

Renowned among aviation enthusiasts.

That's it.

PZL-Swidnik, a historic aircraft manufacturer, now part of Leonardo Helicopters.

Cool. All right. All right. Well, that's really cool. Thank you, Witcher.

Jack E. comes in with 10,021 sats. Greetings. I integrated Holesall connection in Nextcloud's Android app. Now someone can access their Nextcloud installation directly P2P just by scanning their wholesale key.

That's so cool.

It's available and then we have a link we'll have in the show notes. It's aimed to be used along with my Nyxed Cloud project. Whoa, now we're just finding out.

Nyxed Cloud, stop it.

And of course you can find docs for the amazing wholesale at their website, wholesale.io.

That is great to know. And we will put a link to that in the show notes as well. Thank you, Jack. Appreciate that very much. Thank you, everybody who boosted it. And even those of you who boost below the 2000 set cutoff, we still read them and appreciates them very much. And let's combine it all together, boys. Let's see with our sat streamers this week. We had 34 of them streaming sats as they listened to this here show.

I trust you.

When you combine that with our boosters, we had a nice birthday boost bash. We stacked a grand total of 238,528 sats. If you'd like to support the show with a boost, Fountain FM makes it easier and easier just about every single release. It's getting so crazy easy now, and it's a great app with tons of features, and including all of the extra features we put in our podcasting 2.0 feed. You can also go the entirely sovereign self-hosted route. Just start with AlbiHub

Yeah, this is true. So you do need something called Pinocchio.

I'm sorry?

Which is like a Pinocchio.computer, which I hadn't really heard of. The one-click local host cloud. So I think it's kind of in that umbral start nine kind of space. I'm not sure. But someone's gone and done all the work to make a one-click setup and a web UI for running ML Sharp.

That's pretty nice. there if you know if there's a if there's a a moment where you have a great photo it takes it up to the next level when you see this yeah.

It's pretty neat right so basically it goes and figures out and estimates the depth for each of the pixels and then figures out how to lay out the whole scene and then yeah you can poke around at it.

And this works.

Pretty good on vans.

Yeah it does this is this is uh supports systems with a low amount of vram it's still very fast it's a it's a good one so if you get it working put a link in our matrix chat all right now i want to talk you about an app that i think is the best screenshot app out there for linux right now but i don't love the name it's called gradia it helps you get great screenshots that you can share with people for friends colleagues or professionally so.

This is what you've been using.

I wondered if you notice how sweet my screenshots are and i've.

Seen a bunch there's there's been various like hosted tools for this right websites you can go.

To but.

Who wants that.

Ain't nobody want that one of the nice things you can do and there's a lot of options is you can have solid or gradient backgrounds or image backgrounds around your screenshot. So it has a bit of a border. It just looks better when you share a screenshot that has a bit of a border. I can't describe it. Right? Do you agree?

It does. Although, does it force you to make it look like you're on a Mac? Because some of these tools do.

Yeah, some of the default colors do make it look like you're on a Mac. It also has a really nice source code snippet feature. So you can pleasantly display source code snippets across messaging platforms that maybe don't have support for, It also has OCR support. So if you take a screenshot of something and you want to extract text from it, it has 20 different languages it supports for doing that. And then it is a first-class brand-new GNOME desktop.

Yes, it's true.

I mean, we know you don't take them yourself anymore, but they look good.

Why do you say you don't like the name?

Well, because I can never think of it on my launcher when I want to take a quick screenshot to share with you guys. Like, I'm like, what is the damn name? It's not Screenshot. It's not Flame. It's, oh, yeah, right. Great, yeah. I just don't think of a G when I think of screenshots. GPL 3.0 for that bad boy as well.

I don't believe it.

I don't. I don't. 652. So it's linuxunplugged.com slash 652, and you'll get the notes to what we talked about today. But Wes, there are actual extra stuff, goodies, that maybe they're not on the website, but they are in the RSS feed.

Yeah, I mean, that's the real source of truth anyway. The website's great, to be clear, but it's generated from the RSS feed. So that's where you go when you want the real good deeds.

The source of truth.

Yeah, which, I mean, could just be to get the MP3 link directly, because, you know, you want to get the raw stuff.

Maybe you want to get a transcript.

That's right. I think you could.

Yeah, maybe you want to know what the chapters are.

VTT or SRT or cloud chapter JSON.

And your agent's going to love the fact that our cloud chapters are JSON. Your agent's going to love that. So you could just have it parse that JSON file and tell you right where to go in the file.

Yeah, or it could suck in the VTT, which is the one with the dialization, and then it could tell you what the dumbest thing each of us said for that episode was.

Spend the tokens on that for us, wouldn't you? We'd love it.

And then write it.

Yeah.

Boost it.

All right, and join us live. It does take the experience up a little bit. Make it a Tuesday on a Sunday over at jblive.tv. We do it on a Sunday 10, on a Sunday at 1. Well, actually, on all the Sundays. Just go to jupiterbroadcasting.com slash calendar. It'll be a Sunday at your time.

We're going to switch to UTC, so, you know, think about it. We should.

We just should say it in UTC. Let us know if you want us to switch to UTC. I'll do it. I'll do it. I'll do it. Give us a plus one on the UTC. All right. You know where the links are. You know all about that good stuff, so I'll just leave you with this. We'd love to hear your thoughts on all the stuff we talked about today. If you're experimenting with agents or if you're against it, let us know.