AI, Burnout, and Everything After

What we cover

• Linus Torvalds’ apparent reversal from “AI is 90% hype” to supporting Sashiko for kernel patch review, his pointed stance against gatekeeping AI-assisted contributions, and the Software Freedom Conservancy’s new AI guidelines that try to support, not force, and not shun any position on AI use.
• Three core team members stepping back amid AI-contribution friction and burnout, and the crew’s debate over whether Jellyfin should focus on stable “plumbing” and let the community build bespoke clients, add user-requested features directly, or pursue a Nabu Casa-style funding model to sustain development.
• Geoff’s first steps in deploying Hermes himself after last week’s episode, Alex’s story of having Hermes diagnose a household voltage drop with TeslaMate and Home Assistant data, Hermes vs. Turnstone for deterministic workflows, and the limits of local LLMs for agentic tasks versus using frontier models to generate skill files for them.

First up, we trace Linus Torvalds’ apparent reversal on AI, from calling it “90% hype” in 2024 to now backing Sashiko, an agentic code review tool for kernel patches. We go into the Software Freedom Conservancy’s new AI contribution guidelines, and debate whether their “don’t force it, don’t shun it” stance is a coherent policy or just a way to avoid picking a side. Geoff also shares how he’s been using AI to build his own tools, including a vibe-coded podcast recording app and a custom Jellyfin client for offline, transcoded playback.

Next, the crew unpacks a leadership shakeup at Jellyfin, where a desktop app developer left the org partly over pushback on his AI-assisted contributions, and two more core team members stepped back citing burnout and personal circumstances. We discuss whether Jellyfin’s volunteer-only model is sustainable given the current flood of new users and AI-generated contributions, and debate whether the project should double down on stable “plumbing” and let the community build bespoke clients, or start shipping the features users have been asking for. Adam and Alex also make the case for a Nabu Casa-style funding model that could give Jellyfin the resources to avoid burning out its maintainers again.

Finally, the panel dives deep into their real-world use of Hermes agents. Geoff walks through his Outline-connected research assistant and automated morning briefing, while Alex shares his experience of using Hermes to correlate TeslaMate and Home Assistant data to diagnose a voltage drop at his house. Adam contrasts Hermes with Turnstone, a more deterministic, judge-monitored framework built for software development, and the group closes out talking through why local LLMs still struggle with context and how frontier models can be used to help make local agents viable.

Tags: Linux kernel, Linus Torvalds, Sashiko, Software Freedom Conservancy, open-source Jellyfin, self-hosting, home automation, Home Assistant, TeslaMate, Hermes agent, Turnstone, local LLMs, agentic AI.

Links

• Linus and the Linux Kernel:
  • https://www.gamingonlinux.com/2026/07/linux-creator-linus-torvalds-puts-foot-down-on-anti-ai-comments/
  • https://www.tomshardware.com/software/linux/linus-torvalds-rebukes-anti-ai-stances-in-the-linux-kernel-code-review-process-says-linux-is-not-one-of-those-anti-ai-projects-creator-embraces-ai-as-just-a-tool-and-clearly-a-useful-one
  • https://www.phoronix.com/news/Linux-Is-Not-Anti-AI
  • https://sfconservancy.org/llm-gen-ai/llm-backed-generative-ai-recommendations.html
• Jellyfin Shakeup:
  • https://www.reddit.com/r/jellyfin/comments/1v1murx/more_jellyfin_project_departures/
  • https://forum.jellyfin.org/t-project-leadership-changes
  • https://github.com/andrewrabert/jellium-desktop/issues/109
  • https://github.com/jellyfin/jellyfin.org/pull/1947
  • https://github.com/jellyfin/jellyfin.org/pull/1943
• Hermes:
  • https://github.com/turnstonelabs/turnstone

Transcript

Geoff: Hello, and welcome into another episode of BitFlip. We’re here to talk about the pragmatic side of infrastructure and all things nerds and whatever else we want to talk about. So this week, we’re going to talk about AI. AI is what we want to talk about. So why don’t we start off by talking about the Linux kernel? And I think Linus changed his tune on AI. He said in 2024, I think that AI was 90% hype. I think he said something very different recently, didn’t he, Alex? 00:00

Alex: Yeah, so in 2024, as you said, AI is 90% hype. That’s Linus Torvalds’ exact words. He then also told the kernel devs who don’t like AI in 2024 to fork the kernel or leave. It’s very dramatic, isn’t it? But recently he has adopted a tool called Sashiko, which is an opt-in agentic AI code review tool for kernel patches. And this thing claims to catch 50% of bugs in patches that have passed human review 00:28

Alex: with a 01:04

Alex: 20-ish percent false positive rate. Now, I can only imagine the amount of pull requests and code reviews that the maintainer of the Linux kernel has to deal with on any given day. Particularly, if you’ve seen the charts about GitHub lately, with the volume increase thanks to agentic coding that they’re having to deal with. So it’s kind of like fighting fire with fire at this point, right? 01:05

Geoff: Kind of. I mean, I thought it was because I think I think also recently there was a comment by the, you know, line by the the kernel devs saying like the security mailing group is just getting overwhelmed with A.I., 01:33

Alex: Yeah. 01:46

Geoff: you know, prompted and A.I. coded, you know, patches. 01:47

Alex: You can’t 01:50

Adam: you 01:50

Alex: ignore 01:51

Geoff: So 01:51

Alex: them. 01:51

Geoff: it’s interesting to see. No, you can’t. But I mean, it’s on the same side. You don’t want to have yourself flooded with them where you can’t sort out what actually is useful and what’s garbage. 01:51

Alex: So Torvalds this week has sided heavily against people who want to filter out Sashiko’s output before reaching the Linux kernel authors, citing Software Freedom Conservancy new AI contribution guidelines. Now, Linus has an exact quote here of Linux is not one of those anti-AI projects. And if somebody has issues with that, they can do the open source thing, fork it and just walk away. A complete 180 in two years on his position previously. Keep your ethics where 02:09

Alex: they belong in your personal life. Don’t try to enforce your ethics on others. Now that is Linus Torvalds as we know him. There’s nothing wrong in what he said. Ethics don’t belong in the code base. Politics don’t belong in the code base. That kind of thing. His stance here is simply mechanical. It’s to say the quality of the code is what matters. Your feelings, to put it frankly, your feelings don’t matter. Am I wrong? 02:42

Adam: Oh, and he’s nothing if not pragmatic and honest. Like, that’s the thing I love about Linus is that you’re, you don’t have to wonder what he’s thinking. And 03:23

Alex: Yeah. 03:33

Adam: he’s also, this is not the first time that he’s changed his opinion on things or kind of come back to the table to a question and said, you know what, that’s that’s changed now. So I really appreciate the fact that we’re not pretending that this isn’t the way that the world is moving to, you know, right now. And 03:34

Alex: Yeah. 03:54

Adam: I think he’s he’s more than most people very aware of how ubiquitous these tools are being used at this point. 03:54

Alex: The AI landscape of 2024 was very different to that of today. I’m in a hotel room right 04:05

Alex: now, hence my surroundings, and I heard a stat today at Developer Relations Con from somebody saying five years ago there were approximately 400,000-ish software developers in the world. And now with AI, there’s what, 100 million? I don’t know. Nobody 04:15

Alex: really actually knows the number now because anybody that can prompt Claude Code or Codex or Pi or whatever, Anybody that can type a prompt in is a developer now. And so you end up with people building bespoke software, which is great, but you also end up with a whole bunch of people now looking for finding and submitting bugs to major projects. And I think we’ll touch on this slightly in the Jellyfin news that we’re going to come onto in a moment. But it’s… There’s this balance between gatekeeping of traditional software development being a craft versus the pragmatism of dealing with the modern landscape, like all the CVEs. I know Unraid have had to deal with a few, for example, lately, 04:35

Adam: Hmm. 05:22

Alex: where you’ve had to light a fire and light the beacons of Gondor to get things out in a timely fashion because responsible disclosure is kind of dead. Without that pragmatism from the top, from Linus, from Torvalds, the Linux kernel would be in a worse state right now, wouldn’t it? 05:23

Adam: Absolutely. No doubt, I would say. And, you know, what are you going to do? Bury your head in the sand? Like, honestly. 05:45

Alex: Some people 05:54

Adam: Oh, 05:54

Alex: want him too, I guess. 05:55

Adam: I know. But I just 05:56

Adam: think, 05:59

Adam: again, that that’s. people who are actually a lot of the people working in industry, maybe not on the kernel itself, but all over these different projects, like it’s not being talked about, but everybody is, is using these tools because effectively you would be crazy 05:59

Alex: Yeah. 06:17

Adam: not to because of the force multiplier and how powerful it is within educated hands with people who know how to be, you know, senior level developers specifically is what I’m talking about. Um, so yeah, I wonder if he’s being so pragmatic just because of who the majority of the actual people who are submitting, you know, uh, code into the kernel are versus, you know, if it was just a bunch of script kiddies who suddenly are pushing all these, these patches and such, uh, he may not have the same opinion, right. Just because of the overload from, from all of that. But, uh, I think it’s going 06:17

Alex: If 06:59

Adam: to work 06:59

Alex: the 06:59

Adam: out. 06:59

Alex: code’s good, it shouldn’t matter who you are, right? 07:00

Adam: I agree. 07:04

Geoff: Yeah. 07:05

Adam: I agree. But again, just in the crafting of the code and the usage of the tools, you can get very different outcomes. 07:05

Alex: Yeah, for sure. 07:14

Geoff: It’s actually interesting. So you mentioned the Software Freedom Conservancy. So they put out their guidelines for the use of AI. I kind of read through it, and the first three points I kind of think are interesting. So the first one says, the FOSS community should support, not just tolerate those who outright reject LLM-generated AI systems. The next point says, every FOSS contributor deserves self-determination regarding LLM-generated AI. No one should be required to use these systems under duress, basically pointing to the employers who are saying you have to use it. And the last one is FOSS projects should not shun contributors who choose to use LLM generated AI systems. So it’s kind of one of those don’t shun people for being against it. Don’t force people to use it. But then also at the same time, don’t object to people who it just I don’t know. It seems like it’s trying to, like, you know, have their cake and eat it, too, to me, at least when I was reading through it. 07:15

Adam: To me, that kind of sounds like it’s a tool. Use it if you want. Don’t use it if you don’t want. And don’t force people to use it. 08:07

Alex: Yeah. I mean, AI has gone far beyond being just autocomplete, you know, plus plus. You can genuinely create features with the models capabilities that we have now. So, I mean, I look at this just from a purely pragmatic standpoint, and I think, Imagine if he was still fully anti-AI. One of, if not the greatest open source project in humanity’s history to date. It’s got to be this beacon of hope, of just pragmatism, of just being sensible. And I think separating the ethics… Because let’s make no mistake, how these models were trained slurping up all of humanity’s creative works completely ignoring copyright law and somehow that’s fine for them to do. And the access to electricity 08:18

Alex: and water issues. I say both of those things. I’m never 100% sure about the water one. You hear an awful lot of mixed stories, so don’t quote me on that. But there’s a huge number of ethical issues when you think about AI use. I still use them anyway because they’re useful. 09:18

Alex: They are so useful. I’ve got to write a blog post or a video for work. I have all these ideas swirling around in my head. It would have taken me a week to collate them into a, I haven’t got that amount of time with all the projects I try and take on. So I go for a walk with the dog and I do a voice dictation for an hour or two. And then I throw that transcript into AI and say, what did I just say? Help me understand what other people might find interesting in this. The risk we run here is that an LLM is just a mathematical probability machine and so they are only capable of creating things that kind of pretty much conform to the statistical middle of the road, like it’s just naturally how they design. So in time, of course, we’re going to end up with this homogenization of everything that we create if everybody’s using them for everything. And so your taste as a human is still vitally important. You know, the Rick Rubin thing where the one thing that I have is my taste. As a human, your taste matters now more than ever. 09:38

Geoff: I mean, again, I am I’ve used AI to code several apps now. I mean, I’ve done the one we talked about in behind the scenes. The app we’re using right now to do our recording in is entirely vibe coded. 11:01

Geoff: I mean, I have tested I’ve tested the crap out of it and I’m still finding a couple of bugs. But I mean, from a functionality standpoint, it is quite good, in my opinion. I mean, I think you guys would agree. I mean, it’s not perfect, but it’s better than some of the other stuff we’ve used. Um, and then, you know, I’ve also been working on a Jellyfin client because I’ve never been able to find one that lets me transcode my media, like, you know, from whatever, you know, 4k video down to 720p for a tablet when I’m going on a trip. And then also at the same time, let me be able to view all my nice metadata offline. So like, you know, I know Jellyfin will let you do a transcoding, but when you go offline, you have to use like VLC or another app to actually watch the video. You can’t just watch from the same app. And I wanted to make something that did both of those things. And so I’ve been working with Claude and I’ve gotten pretty close to, you know, 11:15

Geoff: something that it works and it does a pretty decent job and it’s out there and if people want to use it, but honestly, it scratches an itch for me. So I think that’s kind of what I care about. I mean, I, for me, it’s like, yes, I want to make something that people might find useful, but if no one else finds it useful and only I find it useful. Great. Cause no one else wanted to make this app apparently. So. 12:07

Adam: He keeps saying it’s for him and whatnot, but we all love what I think I feel comfortable saying. We all love the apps that, that Geoff has created and the automations and the workflows. And it continues. I’m always excited when I see something new pop up, some sort of like little indicator in Discord of some new automation that he’s tinkering with. And I don’t know what it is yet, but I’m like, Ooh, what’s going on now? What’s, what’s he up to? What’s the, what’s the latest thing? 12:28

Alex: Geoff’s been cooking. Let’s see what’s for 12:55

Adam: Yep. 12:57

Alex: lunch. Yeah, it’s pretty great. 12:57

Geoff: We’ll talk about that more later. So why don’t we move on and talk about Jellyfin’s had some developments and they’ve had some shakeups, I think, 12:59

Alex: Yes, 13:08

Geoff: recently 13:08

Alex: they 13:09

Geoff: with 13:09

Alex: have. 13:09

Geoff: their leadership core. So I think I saw that the first thing I saw that caught my attention was the Jellyfin desktop app. was no longer in the Jellyfin organization. It had moved out. And the developer posted a thing saying, have to rename the app, you know, have to, you know, change it. And, you know, he’ll give more information later. And then the next day, he posted kind of, you know, basically a background of what happened. And it was basically, he’s been using AI to improve the Jellyfin desktop app to kind of get through some of the bugs. And… These are kind of what he was saying. He was saying that there was some pushback basically in the organization and people were not supportive necessarily of his use of AI to do that. I’m not sure that was entirely the reason why he left, but that was 13:09

Alex: Well, 14:02

Geoff: at least one of the reasons why he said he left. 14:02

Alex: it’s way more nuanced than that. So there is a quote on GitHub from, I think it might be Claus 14:04

Alex: Vium, might be his name, talking about how it’s much more nuanced than that. Yes, the AI stuff played a factor, but Andrew Rabert’s entire position was that he just didn’t feel welcome contributing to the project how he wanted to contribute anymore. And honestly, he’s volunteering his time. He’s perfectly within his rights to do so. 14:10

Geoff: So the next day, actually, I think it was two days later, Joshua Boniface, one of the other co-founders of Jellyfin, and Anthony Lavado also asked to remove themselves from the leadership team of Jellyfin. Lavado had been there basically since the beginning and making, you know, that he’s part of the core team. He made the iOS app and mobile. But they both kind of said that they were leaving because of personal reasons, I think. Anthony was saying that basically his life, his life had changed over the past seven years, which is, you know, to fairness, seven years dedicated to a project is remarkable. 14:35

Alex: That’s as 15:11

Geoff: But, 15:11

Alex: long 15:11

Geoff: you 15:11

Alex: as 15:11

Geoff: know, 15:11

Alex: we 15:12

Geoff: he 15:12

Alex: did 15:12

Geoff: was 15:12

Alex: self-host 15:12

Geoff: saying that, you know. 15:12

Alex: it. You know, that was a long time. 15:12

Geoff: Yeah. 15:14

Adam: Mm-hmm. 15:15

Geoff: Yeah. And then Josh was basically saying just pure burnout and mental health. He just couldn’t deal with the influx of new users and some of the demands 15:15

Alex: Right. 15:25

Geoff: that were being put. So. 15:25

Alex: Yeah. So I wonder how much of this is because of the new agentic development paradigms that we live in. If you’ve been running on 15:26

Alex: an eighth of a tank of gas for the last couple of years, and then now suddenly you’ve got all of the Jellyfin… Sorry, all the Plex users coming in. You’ve got all of the agentic workflow script kiddies, as you might put it, Geoff, earlier on, coming in with all of their vibe-coded Jellyfin features. I’ve read some commentary as well over the last couple of days that since they did a big database schema migration last year, Jellyfin has never been the same. I haven’t noticed any issues with it myself, but I know other people have. And so I just wonder if all these things like cumulatively, like on their own, each of them isn’t a big deal, but cumulatively they just add up to be like, ah, I don’t need this stress in my life. Joshua Boniface, as 15:36

Alex: the project leader stepping down, at a time like this when there’s an open goal with Plex, screwing the pooch again. It’s the worst possible time for Jellyfin to have a leadership crisis. But I remain optimistic because we just don’t know what’s gonna happen yet. The project hasn’t announced what’s gonna happen. We don’t know. Really, their constitution isn’t even fully baked yet in terms of what should happen. Officially, the existing leadership team 16:28

Alex: should propose a new candidate and then anybody who’s a member of their GitHub organization should get a vote on who the leadership team should be. But that constitution isn’t fully baked as I understand it. Honestly, we don’t really know what’s going to happen. I don’t think the Jellyfin project knows what’s going to happen either. Let’s hope it’s a good outcome and we really need it to be. 17:05

Geoff: Yeah. No, no, we do. Because, I mean… Going back to when we were talking about Jellyfin, my one… I love Jellyfin. I think it’s one of the greatest things. My one major issue that I have with Jellyfin is I feel like sometimes they focus on what’s useful from a development standpoint, like the schema upgrade and the stuff they’re doing, not necessarily some of the features that we as users want. Again, going back to my comment about having to make an app that can do transcoding locally, transcoding to make a local copy that’s smaller. 17:29

Alex: But in this age of personalized software, I just need a stable platform. You provide me the library management, the metadata management, the boring stuff, the plumbing, and then any old idiot can come along and vibe code whatever bespoke wrapper that sits on top of that that they want. The clients are the hard part, so outsource that. 18:03

Geoff: I guess that’s a fair position. I mean, again, I guess I was able to do it. So I guess anyone could, but at just the same time, I see other projects like, you know, Home Assistant adding features constantly. And I know apples to oranges a little bit in terms of, you know, project size 18:39

Alex: Totally different. 18:41

Geoff: and all that, 18:41

Alex: totally different. They’re 18:42

Geoff: uh, 18:44

Alex: both open source projects, but the similarities end there. First of all, Nabu Casa is a real company with real employees and a real business model. Jellyfin is run by volunteers, completely free. They are giving up their life. They could be doing literally anything else and they choose to work on that software for us for free. They owe us nothing. The contract is just totally different. 18:44

Geoff: I was going to say, but my thing is Home Assistant has not always had Nabu Casa. Even when they didn’t have Nabu Casa, they were still releasing features all the time. 19:13

Adam: But is it possible that that’s because it was the hotness at the time? I mean, they’ve been sponsored by Nabu Casa 19:23

Alex: Yeah. 19:29

Adam: for quite a long time at this point. So it’s kind of impossible to actually say if that would have cooled down over time. Possibly. The other 19:29

Geoff: True. 19:38

Adam: thing is like Jellyfin by design was unpaid. They had opportunities to get funding, frankly. And they, by design, did not want that. It was the total… antithesis of being even possible to enshittify. Right? So um there’s positives and negatives that go along with that so yeah so it’s kind of like but from that perspective also when we’re saying like you know these guys weren’t getting paid they could have done anything else that they wanted with their time. Yeah, they could have gotten paid though. So I really think, especially with that sort of a person who is purposely saying no to be funded, that, you know, they’re, they’re going to take actions based upon their feelings even more so than like the average person who’s seeing this as a career that they need to rely on, I would say. 20:02

Alex: But because it’s open source, the agentic world that we’re in now can go and look at every single nook and cranny and build any bespoke solution we want. Because what occurred to me whilst you 20:32

Alex: were talking there, Adam, was we look at the Plex model and we look at the Emby model. 20:51

Alex: We can’t point our agents at those sources and say, build a house on that sand, because we don’t know what’s inside the sand. It could just literally be a hollow husk of complete nonsense. We just have no idea. I’m sure the code isn’t that bad. With Jellyfin, it’s so beautifully poised for the modern era of software because it’s open source, because the code is self-documenting to an agent. And 10 years ago, the idea that I could have entertained writing my own Jellyfin client was preposterous. Honestly, there’s just no way I would have been able to do that. Now, I mean, look at Geoff. He’s writing his own Jellyfin client. You know, it’s… It is amazing, isn’t it? 20:52

Geoff: It is. No, it is. I mean, I guess my only thing is going back to the comments then about Andrew and the kind of opposition to his use of AI. they don’t seem like they’re too open to the idea of AI. So I don’t know. I just think, I think the Jellyfin organization needs to kind of do a little bit of soul searching and figure out kind of the direction they want to go. And whatever 21:43

Geoff: they pick, I’m not going, you know, I have my views, but it’s ultimately going to be their call. So, but I think they just need to kind of pick a lane and go for it. If they want to be, look, we’re not going to add a whole lot of features. We’re just going to make the product as best we can. Cool. Cool. Let us know that. But I mean, there have been feature requests open for years and they just haven’t done anything with them. And like, it’s just say you’re not going to say you’re not going to do them. That’s fine. Just say you’re not going to do them. 22:07

Adam: It’s 22:37

Alex: not 22:37

Adam: true. 22:37

Alex: making a decision, you’ve still made a choice. I think that’s a line from a Rush song. If you choose not to decide, you still have made a choice. So true. You are right with that, Geoff. There’s gotta be some… I think new leadership, if it is open to changing things like the public decision process, making an AI policy and sticking to it, whichever way they choose to go, whatever, I don’t care. I think 22:37

Alex: this has the potential in the long run to be a good thing. Because if Joshua and Anthony and Andrew were all completely burnt out and not contributing, I mean, maybe that manifests in the issues you’re talking about. Their inability to deeply care about those issues manifests in just not making decisions and just not pushing things like constitutions forward and decision frameworks and all that kind of stuff. And so maybe we’ll get new leadership, fresh blood that is like, Yes, we’re going to fix this for once and for all, you know. We’ll see. Time will be the judge on this one. 23:09

Adam: I really hope that, that that’s the way it turns out to so many people love that project. I 23:47

Adam: love that project. Um, if, if they were to have like, to have significant changes also, I think if they were open to funding now because of this huge changing of the guard, that would be totally viable. I mean, I would throw, I would throw money at that. No problem. 23:51

Alex: Yeah, me too. 24:12

Geoff: Yeah. Same here. 24:13

Alex: I think there’s an opening in the market in general. You look at the FUTO model, are you familiar with that, the company behind Immich? We can’t expect individuals to fund entire software projects like that reliably over the long term, I don’t think. Although it’s working out okay for FUTO and Immich right now, we’ll see how that looks in 10 or 20 years time, I remain to be convinced. But there is a model 24:14

Alex: for those of us that do respect these kind of things, like Nabu Casa has shown, 24:39

Alex: to support a company to steward important projects like this with, I don’t know, how much is a Netflix subscription these days? 20 bucks? 25? 27? I don’t know. I would pay five bones a month for access to an app that had good fit and finish and listened to me as a customer. Absolutely, I would. I don’t think it’s been done before, but there’s no reason why it couldn’t be done now. 24:44

Adam: I kick down money 25:14

Adam: every year to good old Thunderbird just because I’m sorry. It’s the only local email client that I feel is worth using because it follows the old tropes that just worked. And everything is trying to be new now and whatnot. So along those lines, if you’re 25:14

Adam: solving a problem and you’re making my life better, I will kick down some bones, no qualms about it. And I think generally that’s with Jellyfin, there’s so many people that love that project. So if that infusion helped with this, helped give people like dedicated time to work on this and whatnot, I think that that would be a good thing. Um, so long, I mean, you got to tread, tread lightly, right? Because yeah, I’m sure this is what Geoff is thinking like, okay, but then there’s the possibility of the enshittification. It’s just coming. It’s right around the corner. It’s possible, 25:36

Geoff: Yes. 26:11

Adam: but it doesn’t always happen that way. It’s just, if it really comes down to like the business model is the way that I look at it. And there, there are really smart business models that can work around some of those, those challenges. 26:12

Geoff: Well, if you have thoughts on Jellyfin’s direction and the ways it can go, you could definitely reach out to us. You can email us at contact at bitflip.show. 26:23

Geoff: We would love to get your emails. And 26:32

Geoff: you can also reach out to us on Mastodon. So we are at bitflipshow on TechHub.social. So you can reach out to us there. I’m the one who is behind that one. We actually had one listener reach out to us and point out that, hey, you know, I’ve been gone for a couple of weeks. Your episodes, you know, don’t have episode number in front of the title. So I couldn’t know which episode was in which order. That was an easy fix for me to do. And it’s now done. So, you know, we love getting that kind of feedback and suggestions for ways we can improve the show. Um, we also want to thank our Terabyte Patreon subscribers. 26:36

Geoff: So we have Mauricio, uh, BadNetmask is on Discord. Who’s awesome. And Alex Taylor, uh, thank you guys very much for your support and all of our patrons out there. We really do appreciate it. You guys make this show possible. So I think we’ll move on then to our last topic, which is going to be Hermes. Because 27:07

Alex: Oh yeah. 27:27

Geoff: after we talked about it last episode, I’ve gotten on the Hermes bandwagon. And 27:28

Alex: Have 27:33

Geoff: I got 27:33

Alex: you? 27:33

Geoff: to say, 27:33

Alex: Yep. 27:33

Geoff: it is… It is awesome. So, I mean, I set it up, and I’ve been kind of playing around with it, because I had said something, I think I talked about, I can’t remember if I talked about this, like, I had come up with a… Actually, I did talk about this last episode. I had come up with a bot that kind of did, like, a morning briefing, could check my calendar, could, you know, do basic things in Home Assistant, but it was kind of crap. Like, it just was not… It wasn’t intuitive because I had to kind of stick to it was somewhat flexible in the things I could say, like I could say, what is the weather today? And or, you know, tell me what the weather is going to look like today. And it would be able to figure that out. But if I strayed too far outside the lines, it kind of would struggle. Yeah. Hermes just gets it. Um, so like I created, I recreated basically all the functionality I had in my thing in about an afternoon. Um, I get a better morning briefing now that actually has a whole lot more than I actually care about and looks better. Um, And for me, the real kind of the real thing was I know a lot of people hook their things up to Obsidian. I hooked mine up to Outline, which is, you know, get Outline, which is an app I use. 27:34

Geoff: I like it because I can view it on a web browser. I can get to it when at work. I can get to it more easily when I’m on the mobile. It has like an API and an MCP server and having a hook up to that and be able to be like, Here are some links for I’m trying to play Satisfactory, which is, you know, like Factorio. Here are some links to some guides, 28:38

Adam: Okay. 28:55

Geoff: you know, please, you know, catalog them all for me. And it can create a document for me that has all the, you know, links that I have or like, oh, this looks interesting for Bitflip, you know, throw this in a Bitflip folder and it can do that for me. So and I’m only scratching the surface. I mean, Alex, I think you were talking before the show that you had to do something for you because you noticed something at your house 28:55

Alex: I’m 29:15

Geoff: today. 29:15

Alex: traveling this week and my phone, my car messaged me, my Tesla, to say, charging anomaly detected or something like that. And I’m like, that doesn’t sound good. And so I run this self-hosted app called TeslaMate. And this app, it talks to the car somehow. I’m not entirely sure. When I got it, I paired some app on my phone to get some kind of a key for it, some OAuth key or whatever. Anyway, it can tell me how many miles I’ve done this week. It can tell me what my current state of charge is. And it logs everything to Grafana. and of course that kind of a data source, I think InfluxDB actually, and that kind of a data source is like gold dust to an agent. 29:55

Adam: What? 30:02

Alex: So I’ve hooked up my Hermes to be able to access the ZFS app data folders read-only of all of my self-hosted apps, and there’s nothing too terribly sensitive in any of those folders. And I said, right, go look through the database for TeslaMate and tell me, were there any voltage fluctuations today at my house? And it went and spelunked through TeslaMate and found out that the voltage had dropped to just, I think it was like 209 or 210 volts or something like that. So it was like, well, it’s a little low, but honestly, middle of summer, you probably had your AC running. Maybe you only got a 200 amp service to the house. You’re drawing 48 amps through the car. Maybe also you were running two or three air conditioning units. Maybe you were getting close to 200 amps. And then it went and looked at… I didn’t tell it to do this. It went and figured out that, oh, I’ve actually got the MCP for Home Assistant configured. I can go and read only look at all of your ESPHome devices and Tasmota devices that are hooked up to the 110 circuits in your house. And all of those were completely fine and completely stable. And so what that allows me to do is call the electrician armed with that data to say, well, only this 230 volt circuit has drooped. Maybe there’s a loose terminal or I’ll just go and check it when I get home at the weekend, you know. It’s just amazing. I could go through and I could go and check each of those smart plugs manually and look at the Grafana data and la la la la. I didn’t have to. For me, I’m remote. I just put a little message into Telegram 30:03

Alex: and asked Jarvis. Yes, I know. How original for my… for my home automation profile. I’m like, here’s the problem. Here’s a couple of data sources, go. And it just does it. And it uses open source technology. You guys, you guys, it’s like, it’s so good. 31:46

Adam: what back end are you using 32:10

Alex: For Hermes, I am hooked into a ChatGPT Pro subscription, which is very kindly sponsored by OpenAI. So yes, I’m a bit of a shill, I guess. But if I wasn’t getting it for free, honestly, I’d pay for it anyway. I use it all day, every day. 32:12

Geoff: i’ve been from my end i’ve been playing around with OpenCode. They have an OpenCode Go subscription which your first month is like five bucks and i think it’s 10 bucks ongoing which is not bad um and i’ve been using DeepSeek. I think it’s DeepSeek. The Flash model just because it’s stupidly cheap 32:31

Alex: Mm-hmm. 32:47

Geoff: and i have not had too many issues with it i mean a couple times i’ve had to steer it correctly or you know this afternoon i was playing around with something 32:47

Adam: you you 32:55

Geoff: and it’s like oh i’ll go and i’m like this is for testing only don’t change anything oh i noticed that you you know you gave me a different url i’m gonna go and change i’m like no no don’t don’t change anything just just use it for testing so it is not perfect 32:56

Alex: You can do stuff like, I have this window in my house that I have a smart blind in front of it, and I have one of my speakers in front of this window as well, and it gets the sun until the sun is at, I worked it out with The Photographer’s Ephemeris, until the sun is at 73.3 degrees or something. So I just literally said to this thing, set this window covering and create an opening guard mode. So this prevents this window covering from opening until the sun degree is at 73.3 degrees north. By golly, I’m watching the football and it goes and does it. It would have taken me ages before. 33:11

Alex: First of all, having the idea for an automation in Home Assistant is a totally different thing from actually figuring out what the trigger should be, what the control surface should be. particularly things like a guard mode of like, if my wife pushes the button to open the blind, it ain’t gonna work until after the threshold has been passed. Like all of that stuff required like templating and it’s just too complicated. So I just don’t do it. But with Hermes, okay, it’s not really Hermes. That’s the incredible part. The incredible 34:04

Alex: part of Hermes is the self-improvement loop that it has, and the memory. It’s never starting from zero every time. It’s always got some context of you, your systems, and how they all work together. I can’t say enough good things about it. 34:29

Geoff: Yeah. I mean, I was able to feed mine into Scrutiny. So that way it can, you know, update my S.M.A.R.T. stats and let me know. Yes. Yes. I had notifications going on, you know, for Scrutiny and S.M.A.R.T., but it’s nice to be able to see on my table, you know, again, yes, I could go into Scrutiny every single time and look at it, but I don’t want to have to do that. So I get a, I get a weekly, you know, I have weekly cron jobs set up. Every week, “Hey, here are the updated stats on your drives.” And, “Oh, hey, that one drive is, you know, at 45 degrees. You might want to go take a look and figure out what’s going on with your airflow.” Yeah. 35:05

Alex: Are we getting an Unraid MCP server any time, Adam? 35:16

Adam: I’m going to just quietly sit here and not not to answer that for you at this moment 35:22

Alex: Okay. 35:28

Adam: in time. 35:28

Geoff: Well, you’ve been talking to us about, was it Tombstone? Turnstone? 35:28

Adam: Turnstone. 35:33

Alex: Wendell’s model, yeah. 35:34

Geoff: Yeah, why don’t you talk about that one? What’s got your interest with that one? 35:36

Adam: So with my background and working within quality organizations and working for a software development company, I care a lot about deterministic outcomes, especially for quality, right? And Turnstone has a specific focus on that aspect. It’s actually not… really well suited to be the general, I ask it, whatever I want to sort of, um, sort of a tool. It’s really focused on that. You can set up a judge, which is basically a second model that is monitoring your main model. Um, it’s, it’s really well suited to local, uh, local LLMs as well. And, um, yeah 35:40

Adam: so their their focus is really on software development and like enterprise level system administration sorts of tasks versus create me a a you know a day planner kind of kind of workflow um it’s really cool it’s really powerful i believe you can even do certain interactions with a Hermes agent as well between the two But it’s more of a manual setup for sure. Like I think to get the same sort of outcome. I think from what I’ve been reading, again, I’m just dabbling at this point and I’ve been going like, okay, I’m weighing these two things against each other and which one makes the most sense for me. 36:24

Adam: And I think as a general home user and somebody who’s going to have it interacting with a bunch of different sort of tasks and workflows that you’re trying to implement, I think Hermes is the right tool for that. it’s better at the self-teaching aspect with Hermes. And it’s less to actually set up. So that’s kind of my take on it. But if you’re into software development, like I said, if you’re working and you’re utilizing this in a business and you really care about the quality of the outcomes and again, the deterministic consistent outcomes of the tasks that you’re using these tools for, Turnstone is the one that I would go with at that point. 37:03

Geoff: One kind of thing I wanted to mention, you kind of talked about the local models and all that. I did try and run Hermes with my 3060 Ti with a 12 gigabytes, you know, with using Llama. It’s just, it doesn’t have the context is the problem. Like it was basically about after five messages, it had to start, you 37:46

Geoff: know, compacting the conversation and, you know, summarizing it. And I’m like, and versus I’ve been going on my DeepSeek, discussion i don’t think it’s compacted once at least not that i’ve seen so unfortunately it’s just one of those things where it just isn’t it just can’t local 38:04

Geoff: is great like i could definitely do some things but it just can’t quite keep up with what i’m doing with Hermes 38:23

Alex: So my old friend Chris at Jupiter Broadcasting did a session on this. I think they called it like Clanker Therapy 38:29

Geoff: Clanker Therapy yep. 38:36

Alex: And so these local models aren’t terribly good at the reasoning side of things. So what you’ve got to do is 38:38

Alex: do what’s called like a model fusion. where you use the intelligent frontier models that do have giant context windows and can synthesize 20 data sources all at once and have tokens out the wazoo. but have them create skills and all the documentation that a local model needs so that the local model doesn’t need a ton of context. It just needs a couple of MD files and scope it into a dedicated profile, a dedicated agent profile. So again, you’re keeping the context as small as reasonably can. And you can set rules to say like, don’t auto-update, don’t self-improve, all the rest of it that can really help this stuff. In my experience, the local models, they’re not as good. Just straight up, they’re not as good. They’re improving all the time. But if you couple them with these frontier models for training, that’s pretty compelling. 39:28

Geoff: That’s interesting. I’ll have to give that a try. Yeah. I mean, the other project, I kind of alluded to this. I kind of feel like I have to talk about it. So I did spin one up on the Discord server because, you know, there are some behind the scenes work that we have to do, like creating the next show doc after the episode is done recording, creating the session, you know, the recording session in this app we use for, you know, chatting and all that. And there’s just reminding us of the next episode and reminding 39:45

Geoff: us of things. There’s just a few things where I have to remember or someone has to remember to do it. Be nice to have it be automated. So I talked to Hermes and Hermes, we’re going to give it a go and see if it actually does what it’s supposed to. But if it can make my if it can make life easier for some of us and take tasks away that we don’t have to do the menial jobs and remember to do things and. It would be so nice. 40:12

Alex: Yeah, well, Hermes, give it a try. Write us an email, let us know what you think of it, all that kind of stuff. What else should we talk about before we get out of here this week? 40:36

Adam: I still have a Hermes question. 40:47

Alex: Oh, 40:50

Adam: So 40:50

Alex: go on then. 40:50

Adam: this is 40:51

Alex: Last 40:51

Adam: for both 40:51

Alex: one. 40:51

Adam: of you. So understanding that it’s not necessarily tied to a specific backend LLM, and I know you can switch between your backends with Hermes. You do have to pick one though, correct? It doesn’t do what you were talking about, Alex, where you can plumb them together and kind of, can you? 40:51

Alex: You can switch mid-conversation, yeah? 41:12

Adam: Yes, but, 41:14

Alex: You literally… 41:15

Adam: but, but, what you were saying where you’re using the foundation models to pull down and, you know, create these skills and such and the tools for the local model, that that’s not something that’s inherently part of how Hermes works, correct? 41:15

Alex: So what you would do 41:31

Alex: is you would set a profile, it would set a default model for that profile, which could be DeepSeek, it could be Fable, it could be whatever. And then there’s a slash model command that lets you on the fly in between different messages switch models. And then it will do its best to transfer the context from the existing conversation across. You probably shouldn’t, it kind of sometimes works, sometimes doesn’t. What you want to do is with the frontier models is create the documentation effectively in the skills files, in the SOUL.md files 41:50

Alex: and all these other things, right? And then that’s what gets passed to the new model when you load that into Hermes to say, right, now I want to talk, instead of talking to Fable and I want to talk to DeepSeek, it then will, DeepSeek will go, well, I don’t know anything about you right now. And it will load those MD files in after you send it the first message as part of that first prompt that it gets sent. 42:10

Adam: okay i understand i 42:38

Alex: So 42:39

Adam: think 42:39

Alex: you can switch mid-conversation if you want to. 42:40

Adam: I think Turnstone is set up, and again, this is just from reading, but I believe it’s set up to do some of that work more dynamically. I think it’s inherently set up to be like a multi-model use case versus that being something that’s an option. Again, with like the judge model, Wendell put out some content on this too. He was using like the Gemma model as the judge and then another model as the actual worker so to speak so it’s that multi-agent orchestration within a single instance versus um i think the way that that generally Hermes works it sounds like it’s a little bit more i have to be thoughtful in order to to utilize it that way with the foundation model doing x task and and the the local model 42:43

Geoff: I think don’t quote me wrong, but I think you actually can have like multiple Hermes agents and have them interact with each other and 43:30

Adam: You can, 43:38

Geoff: do 43:38

Adam: for 43:38

Geoff: it 43:38

Adam: sure. I know 43:39

Geoff: that way. 43:39

Adam: that. 43:39

Geoff: Cause I think I, yeah. Cause I think I’ve heard Chris talk about that, you know, that he has like, you know, Data and Lore and all the other, his, all the other bots, you know, chatting and, you know, figuring out what’s going on. So it’s definitely something you can do. Um, but tell us what you’re using Hermes for. I would love to hear if you guys have got crazy uses for Hermes. Cause I, I think it’s awesome. And I think there’s some use. I’m probably utilizing a 10th of what it can do. So again, contact at bitflip.show, reach out to us. We love, we really do read all your emails and we like to get them. So please keep sending them. But for now, this has been episode 11. I’ve been your host, Geoff. 43:39

Alex: I’ve been Alex. 44:17

Adam: And I’ve been Adam. 44:18

Geoff: Thank you, everyone. 44:21

Adam: Bye-bye. 44:23