148: AI Psychosis Day  🔗 **Source:** [148: AI Psychosis Day](https://share.snipd.com/episode/b5f45463-77fd-49e0-aa97-30820e3e4704) ## Highlights - **Aaron Francis Builds An AI-Generated Lost Civilization** - Aaron Francis is using AI swarms to build a fictional lost civilization intended to become an illustrated physical book. - Agents maintain canonical lore, generate Blender models, and create non-canonical narrated stories through ElevenLabs for world-building playtests. Transcript: Aaron Francis First and foremost, it is an experiment. Okay. So a while back, I saw, you know, all these open math problems were falling by the wayside to AI. And I was like, I have AI. Ian Landsman I can do this too. I can play this game. Yeah. Aaron Francis So I talked to Claude about what, you know, open math problems there were. And I quickly realized, no, I can't do this at all. There's like less, less than a 0% chance I can like play in this area. And so then I started talking to it about like, hey, I've got this, I've got this meta harness that can coordinate these agents and these swarms. And like, what are some fun things to do that are now possible, uh, with many agents working together that wouldn't have been possible before? Cause that's the whole math thing. It's like, well, let's just throw a compute at it. And this wasn't possible before now new things are possible. And so after, you know, an hour or two of talking and pushing back and accepting, and we landed on basically a fictional world. And so I am, with the help of a whole swarm of agents, building out an entire fictional world. I will not tell you the details of it, but it is going to be, I think, somewhat of a novel concept. And the final artifact will be something that you can read. Of course, there will be like, you know, micro interactions and stuff, but it's primarily actually targeted to be a, a physical book, like a giant, beautiful, illustrative physical Book. Um, and of course this weekend I was like, Hey Astro, what's blender? I see that everybody's doing Blender. We have this whole corpus of, it's like building up canonical facts about the world over time. And so now we have this whole corpus of canonical facts about this fantasy world and kind of how it, you know, fell. They never put you out as a fantasy guy. Ian Landsman Interesting. Aaron Francis I don't know. I don't know if it's pure. I don't know if it's fantasy fantasy. Like there's no. Maybe it's more fantastical than fantasy because there's no like warlocks or dragons or anything. But it is basically the story of a lost civilization with a few, you know, twists and turns. And so I fired up Blender. I didn't fire up Blender. That's too generous. I said, what's Blender? And had the agent download it and start using it. And now it's building out 3D representations of all the canonical details that we have. Oh my goodness. Look at this guy. And so Blender's cool. Never used Blender in my life. Still have it, to be fair. But once you build up all these like 3D, like basically sources of truth, you can render them in different ways while still keeping all the details correct, right? Because that's the problem with image generation is you're like, hey, take this picture of me and turn me 45 degrees to the left. And then suddenly it turns into the average white man. And you're like, well, that's not actually me. You lost a lot of details in your distribution. And so having the 3D models, it can do the wireframes or do different ink and wash styles, but all from the base shared geometry and so this weekend i've just had it ripping on that's why I moved it to the mac studio i don't know if you saw my tweet about transferring of consciousness i transferred it to the mac studio because that is an m2 ultra with 192 gigs of ram and like Turns out blender works a little bit better over there than on my MacBook Pro. So I didn't mean to yada yada the best part, but that is my art project where I am sending generous amounts of tokens for exactly zero ROI. But you know what? To the listener, turns out you can just have fun. Like that is possible. This is new to me, first of all. I love this. This is fantastic. This is new to me. I don't know just had fun. You know, I think you got to print an illustrative book out of this, you know, maybe there's something there. Ian Landsman Yeah. I definitely will. Be the big name in AI writing. That'd be cool. Aaron Francis And I wouldn't hide it at all. It would be like, this is a collaborative experiment with swarms of AI agents. And can it actually, with, of course, much of my guidance, can it actually produce something that people want to discover? ([Time 0:07:09](https://share.snipd.com/snip/cd5b0ccf-fbc4-4dc4-bfb2-5d44a93c75d5)) - **World-Building Works Best At Human Scale** - Building a world from societal facts lets Aaron Francis test whether its people feel compelling at human scale. - AI-generated bedtime stories expose weaknesses in the world’s lore while making historical research unexpectedly engaging. Transcript: Aaron Francis I have actually, so all the canonical data is basically facts and like lore and like historical influences for each particular society. Like there's like, you know, the Levites, the priestly caste influence, the house that is, you know, revered, all that kind of stuff. And so I've also, I got a little bit bored just like layering facts upon facts. And I was like, I kind of want to enter this world and I want to feel like, is this world working like at a human level? Like, obviously, we're layering up all these societal things, but at a human level is it interesting all right and so i had it uh i had it produce a skill whereby it creates a non-canonical Story like it picks one of the it picks one of the areas it enters into that area and it creates like a you know life and blood story right and then uh it sends it off to 11 labs for narration And so it turns it it turns it into an audio story and i'll be honest i've been listening to some of these audio stories as i fall asleep at night oh my gosh it's very it's very compelling And so i feel like artistically it's working. And so it's kind of like a way to play test the world is say like, hey, we've got all these facts. Go in here, make up a story. You can add things that are not canonical, but you can't like reinvent canon. And so go in here and write a story for me. I like the human scale, not the society scale. And then it turns it into audio and I listen to it. It's so freaking fun. I mean, people love the go to bed things like this, like an infinite amount of stories in a world you enjoy. Ian Landsman We got to turn this into an app, get them on it. We're turning this into money now, Aaron, right now on the show. Aaron Francis So it is, uh, it is a freaking blast, I must say. And so now like I've had it, I've had blender render out, uh, like many of the structures where these people live. And now I'm going to have it start rendering out basically like character studies of like, all right, so there's this like kind of secretive priestly caste. Let's design their vestments. Like, why don't you pull from, you know, why don't you pull from the Middle Ages and from the stories in the Bible about the priestly garments and pull from this part of history as well? And so it's really fun. I'm actually learning a lot about the different societies, like lost societies throughout history because I'm like, where can we look? Where can we look for inspiration for X, Y, Z? It's like, well, you know, the Mesopotamian people, blah, blah, blah. And I've never heard of those people in my life. Tell me more. So I'm telling you guys, it's the golden age. It's the golden age. You are deep in AI. Oh yeah. Ian Landsman Deep deep in AI. Of course. That's crazy. You know, this is how, this is how Tolkien did it. He started with like, the foundation of the universe. He did AI? Yeah, that makes sense. He started, like, with the language. He made the language because he was, like, a linguist. And then, like, you know, the stories came after he had built the whole background and all that stuff. Aaron Francis C.S. Lewis and Tolkien were friends. And Lewis wrote, you know, a thousand books and Tolkien got lost in his one universe. And it's hard to say who was the better, you know, author or producer because of that. But I remember hearing like C.S. Lewis or somebody talking about how Tolkien just got stuck on the first thing because he invented the universe. And meanwhile, Lewis is writing Narnia and Caspian and then all the other religious stuff. But that is honestly very fun to be like, hmm, how can I invent a whole society of functioning like people, trade agreements, religious rights, all of that kind of stuff. ([Time 0:12:35](https://share.snipd.com/snip/b6b81f56-3704-4100-ad19-a420f2c6d94a)) - **Use Detailed Specs To Run Agents Overnight** - Write detailed plans or specifications at night, then let agents execute them while you work or sleep. - Aaron Francis reviews results in the morning and uses weekends for more hands-on iteration. Transcript: Aaron Francis Yeah, so, I tweeted recently about how I'm doing solo work these days, given that my W2 is beyond full time. Like it is, it's a lot. And so the way that I'm doing solo work, if you missed the tweet, and Producer Dave can put it down there somewhere, is that I basically spend, you know, an hour or two at night coming up With very in-depth, like, plans or specs. Maybe that's an hour on one. Maybe that's an hour on, like, three or four. And then I hand it off to agents running inside of Solo so that it can do all the spawning and to-dos and all that stuff. And then I check on it in the morning, see how it did, go off to work, and then we repeat, repeat. ([Time 0:18:51](https://share.snipd.com/snip/658e78ae-09b8-4a3e-b959-82fd7e1dd3a4)) - **Linux Makes Solo A Distributed Workspace** - Solo’s Linux expansion became practical through Omarchy, Tailscale, and remote access to a ThinkPad. - Linux’s fragmented desktop ecosystem creates platform-specific work, but cheap headless machines could make distributed Solo instances valuable. Transcript: Aaron Francis But one of those things I've had, obviously, in the back of my mind, and has now come to the fore of my mind is Linux support. And I don't know if you know the guy called DHH, who's the creator of Rails. He is his new shtick is Omachi, which should be pronounced Omarki. So Omarki is this Linux distribution. And I bought a while back, like six months ago, I bought, you know, an IBM ThinkPad for like, you know, 600 bucks or whatever. And I was like, all right, I'm going to figure out how to put Linux on it. Six months later, no freaking clue, haven't touched it. And then Omarki comes out and I'm like, oh, this is supposedly very easy. Turns out it is very easy to get installed. So I got it installed on this ThinkPad, which takes me back to my accounting days. But fortunately, I've plugged it in and put it in the closet so I don't actually have to see it. And this comes back to another tweet I had, which is TailScale. TailScale's amazing. So now I have all these computers hooked up. So from, from my Mac book, I can say, Hey, go in, go into the Omarky machine, get it all configured, get the repo pulled down, get all the dependencies installed, start working on it. And so because it's on this tail scale network, I can also VNC into it. So I can actually go and look at the screen. And so right here from this desk with my MacBook, no matter, actually, even if I'm at home, I can pull up the Linux machine, the desktop, look at it, play with it. And there were maybe, there were maybe like 10 to 15 little things that when I started playing with solo on the Linux machine was like, Hmm, this is not right. And so I've just been working through those. And I just overnight ran like the end to end performance test on Linux. It found a few things that were like Hyperland, whatever that is, Wayland, whatever that is, specific. Yeah, who freaking knows? And so it's working through those last like platform-specific problems. But we're ready to go on Linux. So we're at least ready to go on Omachi. I don't know. And this is what scares me is all the other Linux distributions. It's basic. Linux is like the JavaScript ecosystem where it's like, what if instead of doing one thing, you could do anything? And so everything is different. Like, oh, shoot um but it works like it totally works on linux i've got it it's actually on linux right now inside of solo it is fixing all the linux things so it works well enough that like The orchestration loop is working fine over there and the cool thing is i can just swipe on my mouse, you know, right, right, right. And then I'm on a Linux machine on this computer. So that part has been very cool. I got to figure out the final packaging and releasing. It is already built. I just haven't like exercised it yet. And then after that, people that are on Omachi can start trying it. So I'm pretty excited. Yeah. Ian Landsman Well, it seems to be a little like people are into the Omachi right now, right? They're talking about it. So if you can inject yourself into that conversation, I feel like that will probably be useful and get some sales and some interest brewing over there. I'm sure, yes, you will find out, I'm using this desktop on this OS it's not working and like whatever. There's going to be a million of those, which will be quite exciting for the agents to crank away on. But yeah, I mean, I just like this as like a new path of like, as this is growing, I still am not a believer in this is the year of desktop Linux, by the way, like I don't believe that's true, But there could be enough energy over there to uh get you some sales let's go with that yeah and it so it at least regardless of if linux people will pay which is the question we've had on Aaron Francis Here before it completes the story like it makes me feel good like i'm a i'm a completionist i like to collect things and to have all three platforms feels good. Especially in the grand vision where Solo can be headless and you can connect to from not even through Tailscale or VNC or anything. I could connect on this machine to a Solo instance running on another machine. Then Linux makes way more sense because you might have, you know, eight cheap Ubuntu boxes that run builds or something. Yeah, just cranking on tokens. So it's fun. ([Time 0:19:46](https://share.snipd.com/snip/1f88c0ec-02a0-498b-8527-38edf29246a6)) - **Ian Landsman Culled 4,000 Horse Photos** - Ian Landsman shot 4,000 horse-show photos across three days using expensive camera gear. - He rapidly culled each day’s files to roughly 100 in Lightroom before editing the keepers. Transcript: Ian Landsman I don't have much for this week, but I did take 4,000 photos of horses over the last three days. 4,000 photos. Yeah, a lot of photos. 4,000 photos. I was clicking away. Click, click, click, click, click, click. Aaron Francis What is your media strategy to be able to handle 4,000 photos in one weekend? Ian Landsman Oh, it's terrible. And there are more weekends than one in life. Aaron Francis So you must have a lot of photos. Ian Landsman Well, this is the big weekend. This is my daughter's first big, this is like a big, fancy, high competition level event. So this was her first time doing that. She got the new horse. They're still working things out. So it was quite a weekend. We had a little life lesson area where like the first day went very poorly. Yeah. Of course, we refused jumps and she almost got thrown off and whatever. So we had the adversity. And then her and the trainer worked on some stuff. At the end of that day, just on the side, they got sort of straight. And then the next two days went very nicely. She finished third and fourth in a variety of events. Sounds like the hero's journey to me. It's exactly the hero's journey. So that was very nice. It all worked out perfectly in that regard. Like you'd hate to have the adversity at the end. So it was adversity at the beginning. We worked through it. That was great. My adversity was like walking seven miles a day, carrying a bunch of camera gear and uh taking tons of photos of cooking all sunburned here so yeah so i like taking the photos you know i Got all the camera gear this is like one of the things i do with the camera gear it's like i got the fancy camera gear in part so i can take these specific photos so very cool like in the barn Aaron Francis Area i got some nice photos with her and her buddies and her you shared some privately with us and they turned out amazing the ones you posted are also good but you know the close-ups with Ian Landsman The kiddos face and the horse it's just they turned out great yeah so it's fun and yeah so every night it's like i knew i'd have a bunch of photos so it's like every night like the job didn't Stop i come home i to process that day's photos because I don't want to be behind on the photos. It's terrible when you're behind. So go in one day at 2000. And each of the other days had 1000. So process the photos. But they were 2000. My niece was all who also does this one some big award thing, blah, blah. So I took a bunch of photos of that award process and blah, blah, blah. So that was made that day longer and even more photos. So, yeah. So what I do is I download the photos. So I'll view photos like 80 to 90 megabytes. So it's a big, heavy photo. Download them all on the computer, crank through them. Onto what computer? Onto the MacBook Pro. So this I need the horsepower. I can't use my little, I bought a little MacBook Air that I use now since I'm on the orbs. Because you're working. Yeah, I don't need the local development anymore. No local development on that machine. But I tried it. I was like, maybe I can get away with this. It doesn't have the horsepower. Aaron Francis What's the hard drive space on the MacBook Pro? Ian Landsman That one has a four terabyte. So that's got a big boy. Four terabyte? Yeah. Aaron Francis That's like an $8,000 machine. Ian Landsman Right. Now it was $5,000 four months ago, but now it's an $9,000 machine. So that's why I got it. I didn't even really want it, but I was like, I know all these RAM prices. They're going to jack everything up. Let me just get the big machine now. So I got a big machine. Now, if I knew I was going to go to this little machine, I probably would have got like a Mac Studio instead, but whatever. So the MacBook Pro is fast enough with all the disk and a bunch of CPUs and 64 gigs RAM. It's sufficient for my needs. So when they were cranking photos and some people save every photo they ever take. I'm not a believer in that philosophy. I don't want all these crappy photos. I don't care about them. I'm never going to go look at them. It's just more stuff to deal with. So I go through and do a first pass. And every day here, I basically knocked it down to like 100 photos. Oh, wow. So I took one first pass. Yeah. So I take the first pass, get down to like 100. Are we in Lightroom? Where are you doing this triage? Lightroom, all in Lightroom, the new Lightroom, not the classic Lightroom. Because they offer both still because people freaked out when they updated Lightroom. So now they have both. And so it's just kind of annoying because there's some stuff in new Lightroom. New Lightroom doesn't have everything classic Lightroom has. And sometimes that's annoying because I wish it had some of those things in Classic Lightroom, but it doesn't. But whatever. I like the New Lightroom better overall. So I use that. Turn through. That goes pretty well. It's pretty fast. The thing with the horse photos is you're taking a lot of photos of like, okay, the jump is happening. You just got your hand down, your finger is on the trigger, and you're taking 20 shots as the jump occurs so that you can just pick out the one where like the horse is right and the kid is right And all the stuff's right. And sometimes none of it's right even with that, but whatever. So that's what you're doing. So you have a lot of like, okay, there's only one good photo in here. I can instantly get rid of the other 30. Okay, fine. So do that. And then I edit them, which takes much longer and like go through and crop and colors and whatever, tweak all the stuff. And then that's kind of it. So I did that every night. ([Time 0:26:09](https://share.snipd.com/snip/64a60118-326b-4b46-af45-d71b3c33bfad)) - **Match Camera Gear To The Shooting Problem** - Choose camera gear for the actual subject rather than maximum specifications. - Expensive lenses help with fast-moving horses in difficult light, but a $600 lens is sufficient for controlled 1080p talking-head video. Transcript: Aaron Francis So back to cameras this this super lens that you got yes i saw uh uh i saw eric barnes first of all taking some amazing photos i don't know if he's your protege. He is my protege, yes. But one of the things I saw you two talking about was, like, using this absurd lens for a webcam. So when I get, I've signed the lease for the new space, by the way. So I'm getting a new space. And I'll have more space. Could I use an absurd lens for a webcam? And like, would it look awesome? It would. Ian Landsman I mean, this one that we were talking about there, I think it was actually with Wathen too, was in on that one. But that one, I don't know if would be actually perfect for a webcam because it goes 50 to 150. So 50 is still pretty, you know, fairly telephoto for a webcam. But, I mean, you could push it back and do it. Yeah. So you could. There's some other ones. I would get a Sony. You need to upgrade to a Sony. Aaron Francis I would definitely say you have to do that. Upgrade? What's wrong with Canon? Canon's not a downgrade from a Sony. That's bad, man. Ian Landsman The focus is so much worse. The Sony focus, listen, there's 8 million hours of me on this podcast. I've never been out of focus even for one second the entire time because the focus is perfect. I'm out of focus for like five seconds an episode. That's not that bad. You're trying to be absurd here. You're talking about buying a $4,000 lens just because. So I'm saying maybe buy. I'm not talking about that. I didn't know it was $4,000. Maybe lens and a $2,000 camera and have perfect focus for the same amount of money. That's what I'm saying. Aaron Francis So the Sony cameras do better autofocus, like their algorithms better? Yeah. Ian Landsman I would say it's much better. But you're a Canon guy, so it's also tricky. You get the batteries, you get stuff. Aaron Francis Yeah, I got the batteries, the cables, all the lenses. So super specifics aside, if I get what millimeter range is good for like a this is absurd kind of webcam, provided I can push it back as far as I need to go. Ian Landsman What do I want? You've got a lot of space. You could go like the other absurd lens I have, which is a 300 millimeter 2.8. And what is that buy me being further away? Just better background blur? Yeah, like if you want to go crazy with the background blur. But there's other, I mean, for webcam, this was just, they were being somewhat farcical. This was not, I would not recommend this for webcam. Aaron Francis My thing is doing the farcical thing sincerely. So let's talk about that. I would probably go a different route if you wanted to do that. Ian Landsman And what I would do is get like, on the Canon, believe they have like a 35 1.2 so if you did that you'd also get the crazy background blur um and it would give you a more normal web cammy type Uh millimeter range and so that's probably what i would do or you could go 51.2 you'd have to push it back a little bit um and that would make you a very shallow depth of field. Now, the only problem with that, with these cannons, that in the video especially don't focus as well, is that 1.2, there's literally just going to be your eyes. It's like the tip of my nose. Yeah, so you're going to be in and out of focus a lot if it doesn't really keep up. Now, maybe cannons' newer bodies do a better job, so you could get a newer body, 50 1.2 or 35 1.2. But I don't know as much about the cannons like the newest stuff especially so maybe the newest bodies are as good as uh as good which is is it the more like if i did like a 18 millimeter that's Aaron Francis Where i get like pancake face and then if i get like a 200 millimeter i look thinner you'd be very yeah exactly all right 200 it is I'm not doing pancake face I got enough of that going on but Ian Landsman Even like a 50 millimeter will give you a fair amount of that without having to push the camera like so far back to be ridiculous yeah so something like that it depends too on like your setup Like I have the screen in front of me so I have to have the camera like up over um to go like watch your angles there and stuff uh but if you have like more to the side it's different but um yeah I don't know but the camera gear is fun the camera gear is like half of why I like doing it it's like you get the gear you sell the gear that's every hobby yeah exactly it's like you collect The stuff as we were saying do you want to ride bikes or have bikes well 50-50, you know? Yeah, exactly. So the other thing Eric was talking about, if you did want to buy something, what is this? They might have a Canon mount for it. There's this new company called Viltrox that makes really good, high quality, but cheap lenses. And so he got an 85 millimeter 2.0, which so Eric was showing off his new photos, done with that lens, which is much better for, he was doing like high school football so it's like not not Good light and things like that so you want to have a wide aperture so you can collect a lot of light so he had been using like an f4 lens which zoomed which is handy but is not as good in low Light so that's why he got this it's like 300 bucks for this lens but the Viltrox lenses, if they have a Canon mount, you could get something there too. That's a little bit cheaper. And they're quite good, actually. So a lot of options out there, a lot of options. But for the horse stuff, it does push. The good gear pays off for the horse stuff because talking head video, you're not actually getting much out of the good gear. But for a moving animal that's jumping things in various types of light, like the good gears, where you're getting more of your money's worth if you can, you know, I'm justifying it. Obviously, there are cheaper ways to do some of these things. But you are getting the most out of the lens and what you're paying for in those kind of things. Aaron Francis So you're telling me four grand for a headshot in controlled lighting may be a little bit overdone if it's going over the internet 1080p anyway? Exactly. Ian Landsman I'm using currently a $600 lens. Advice disregarded. There you go. ([Time 0:35:02](https://share.snipd.com/snip/a3c85234-acd4-40c1-94eb-c9e96f9b0a81)) - **Shipping Outro Is Harder Than Building It** - Outro’s remaining obstacle is not development but the operational burden of bugs, refunds, support, and release management. - AI accelerated implementation while simultaneously creating more work to coordinate and maintain. Transcript: Aaron Francis Can I give you some outro feedback right now? Oh, yeah. Give it to me. I want optimistic updates when I click. Yes! I was just thinking that is done it's like it's a full second lag from mouse down or mouse up to complete uh it's i was like thinking about should i go type this into the ai right now to do it Ian Landsman You should you should get orbit on that because i need that. It's very annoying. I totally agree that has to happen. I think I did have that at one point, but there was some other issue. But I've got those other issues worked out. So now I think if we do it, it'll all work correctly. So yeah, that's going to happen. Before our next one, that's going to happen. Aaron Francis Let's do, it's not on the list. Let's do a little detour. I think you're going to start giving this, you're going to start inviting friendlies into outro on, I think it's July 9th. And so that's coming up. Ian Landsman It's coming up. I got a year though. How are we? We never specified the year. Aaron Francis Ah, bamboozled again. How are we doing on outro? You've been Orbin. I've been like with the new orb, Psychosis, you're making progress. Where do we stand? Ian Landsman So I'm mostly still focused on a bunch of help spot stuff between SOC 2, a new UI for the admin area, help spot, some other new modules. So a lot going on. That's when most of the Orbin is going. But I have done a little bit of outroing mostly on things which I would never release until these things were done. So that's good. We got the AI working faster. Did you add anything with the AI today? I didn't. I just clicked buttons like an animal. Oh, you just typed. Okay. You're usually my AI guy. So I see you there waiting for it to finish, and that was annoying me. And so the AI is now much faster. New model, new architecture, what? New switched over to the open AI models but more so I think partially because Luna is better. That's probably the main reason I would say. So I can use Luna for more stuff that I was using like Sonnet for and it's faster and better and cheaper. So that's good. And then did re-architect that. This was still all Sol, not Astra, but Sol went in there with a little fable guidance, I believe. And they fixed up the speed quite a bit of some extra calls and various stuff that was in there that didn't need to be there. So that's good. And some other just loading stuff was dramatically improved speed-wise. So we're getting there. I did have this, when you check a thing is done, it showed a little check and it had an intentional delay. And that was annoying me. So I removed that, but there's still a delay because it's making the round trip. So yes, we need to have the optimistic removal. So that'll happen soon. There was also, oh, that's the other thing it did, was there was like a, there was like some weird stuff if we both like were checking things nearly simultaneously. So now it does a better job of avoiding that scenario, which I still think ultimately there'll probably be occasionally times where that happens. But I just didn't want it to be like, it felt like it happened a fair amount and I wanted it to be all faster so that you can have that. You've accidentally built a distributed system and now you're reaping the benefits of it yeah nightmare scenario so uh yeah so those are things like the AI could just go off and do pretty Much without a lot of input I feel like when you're not doing UI stuff at this point like the AI can kind of go off and do things uh mostly so that's so I've done that stuff but I don't know you Know there's a chance this just ends up being our little playground this is just our little place we build stuff. Aaron Francis Why? Ian Landsman Jamie's on me to release it too, but it's like there's no good time to release it. I don't want to deal with releasing it. Jamie stopped listening a while ago because we got technical fast. Aaron Francis So send her this segment. Jamie, help me. Help me help Ian. This is crazy. I know. I thought about maybe possibly I hired like a part-time type person. Why? It's good now. Ian Landsman No, not even about finishing it. Just like, it's like the thing I don't want to deal with is like there's going to then be bugs and people are messaging and they need help or they want a refund or something's broken. Like that whole stuff just takes so much bandwidth. Aaron Francis Don't you have a whole company that does a lot of that stuff? Ian Landsman We're small and we have these huge help spot things we're working on, so I can't actually afford them to go off on any outro stuff right now, which is very funny because, like, when I started This last summer, we had loads of time. And I was like, well, all right, like, you know, between me starting it and then, like, once it gets to production, like, they can help or whatever. But now nobody has any time at all. So... Aaron Francis Why? AI is supposed to eliminate jobs and work. Ian Landsman How is that not happening? That's a theory, but so far it's just making lots of work. Aaron Francis So... Ian Landsman Shoot. Yes. ([Time 0:42:08](https://share.snipd.com/snip/99fb7626-ba16-43a8-8f57-3ad2dec45604)) - **Aaron Francis Trains A Tiny Terminal Classifier** - Aaron Francis trained a tiny local classifier to recognize whether terminal agents are idle, blocked, or awaiting permission. - Frontier models label terminal observations, while the resulting 537-kilobyte model runs locally in milliseconds. Transcript: Aaron Francis One of my soloterm.com, it's very good, problems is I am basically just a terminal. And so I try my hardest not to require a lot of setup. Right now, the only setup that you have to do is like, if we don't figure out which agents you're using, which we figure out like the top eight common agents, you have to add your agents. And then after that, you have to install the MCP into those agents. It's just one, it's a one click thing now. And so part of my, part of the way that I keep this low maintenance is I don't, you know, hand code heuristics for every 2E or CLI. I'm not like, ooh, this is cloud code. And therefore I can look at that on the screen. And I know that cloud, cause I can't keep up with that. Um, and I think it's bad. It's bad architecture and I can't keep up with it. But one of the recurring problems is does this agent, uh, require my attention? Um, which could be permission could be I finished this task. Am I allowed to do this thing? And it's not a real permission prompt. It's just a question printed to the screen. And so my solution or my pending solution to that is train a very small micro model that can input or ingest any terminal grid. So on the back end, I have, you know, a representation of what is shown on the screen because I am the terminal. And so can I feed that grid of cells to a very small local only model? I don't even think it's an LLM. I think it's just a model. Can I feed it to that model and get a structured score back of confidence of if it's idle, if it needs permissions, if it needs my input, that sort of thing. And so it is currently training right now as we speak on the Mac Studio. It is training the first run. And so what I've done is I set up Solo on the Mac Studio with this repo, and I basically said, great, start spawning all the agents you can find. Start feeding it whatever inputs you want. The goal is to capture a bunch of samples of different agents at different points their life cycle, capture the visible grid and certain underlying diagnostics we have from Solo, like The last time that we saw visible PTY bytes, the last time that the human typed any sort of input, that sort of stuff. So feed it, like capture all of that. And then what you're going to do, this classifier repo, it basically spawns a new codex, a fresh codex, and says, give me your score for these five labels. Give me your score and your confidence. And that's how we do the labeling. So I don't know if this is distillation or not, but either way, it's awesome. And distillation is my right as a U.S. Citizen for your diehard. So like I'm using the frontier models to basically label these observations. And then now I've got like 500 labeled observations. And so still pretty small, but I'm feeding it through this first training round, which is some Python thing. I don't know the first thing about those. And on the other side, something is going to pop out that I can consume from rust. And it should be under like five milliseconds to get a response back from a grid. And it's all local. And so I can just continually feed these grids into this local little model. And then that will help me surface back to the user. This terminal needs your attention. This one's stuck on permissions. This one's ready to go, that sort of stuff. I've also built in, just for development mode only, it's hidden behind a flag and then compiled out, auto capture. So now as I'm working on all these machines running solo, I can turn on this dev flag and every 30 seconds it'll just capture all of that diagnostic. So instead of these artificial L generate, not generate, LLM driven captures, it captures all of this from my day to day usage. And so we'll quickly get up to many thousands of samples, which will make the model more accurate, I believe. But it's currently training. ([Time 0:49:39](https://share.snipd.com/snip/4c8ce6b9-1253-4e47-be5f-fce29fa3c474)) - **Small Specialized Models Can Be Extremely Fast** - The classifier mechanically passed its tests and achieved sub-millisecond typical inference, but prediction accuracy remains unproven. - This illustrates how narrow models can solve focused interface problems without requiring an LLM. Transcript: Aaron Francis And it's like, yes, sir. Right away, sir. So we're cooking. Ian Landsman So you're going to train this first one. The first run of training is just on like the 500 it generated. And then you'll do new trainings. Yeah. I just want to see it work. You know, does it work? Yeah. How long is it? Did it estimate how long or how long? No, it didn't. Let me see if it probably shouldn't be too long with only like 500 data. Oh, shoot. Aaron Francis It's done. There we go. All right. Live on air. Let's do it. Feed it something. It says, yes, the local training smoke test passed. It took 11 minutes and 36 The model's 537 kilobytes. What? Ian Landsman That's pretty tight. Does it work, though? Aaron Francis I mean, there's not much text to, like, analyze. I guess not. Verification. Reloaded successfully and scored all 14 synthetic checks. So I guess it worked. Well them or accurately scored yeah probably let me ask so now that the model is trained the 14 synthetic checks did they all pass like is the thing working uh and how fast is it working Oh man we're doing it ai we're doing it live you gotta talk to the ai you gotta talk to the AI. Because when you type, your little fingies, especially being so old, your little fingies get tired, right? That's true. You can just talk, talk, talk, and then you get all of that context out of your head. Okay, so it says all 14 checks. That's not very many controls, all 14 checks passed, including valid four outcome scores. So the score comes back with four labels or heuristics and a percentage of each. That checked that the scoring worked, not prediction accuracy. Well, accuracy is kind of important, so I need you to check accuracy as well. What the hell did you check then? The AGI is here, Aaron. You set off all this work. That's not the point of the exercise. It says, I will benchmark it now so I can give you measured latency. All right, we'll come back to it. That's over on the studio. We'll come back to that. Oh, boy. Ian Landsman The AGI is taking your job. Silly model. What are you doing? Oh, my goodness. All right, so you train this model. Cool. I mean, this sounds great. I was a little pessimistic at first on this. I know you were. Aaron Francis I think it seems reasonable-ish. It seems reasonable-ish, and that's my sweet spot. I don't want to be too reasonable because then, come on, hey. I don't want to be too unreasonable because then I'll, you know, never be a success. But right in the middle of reasonable and unreasonable, that's where I live. Ian Landsman I'll be curious to see if this thing actually runs well across all your OSs and everything, you know, if it hits like security, macOS security says this thing shouldn't be running or Like, you know, I don't know, whatever, random stuff that. Aaron Francis One of these Omachi people that runs their machine on a potato, like, is it going to be performant over there? I don't know. Ian Landsman Yes. Stuff like that. Although at 500 kilobytes, if it actually ends up being accurate in that range, presumably it's not taking too much memory to execute. And this is all text you're feeding in, right? It's not the images. It's all. No, it's all text. All text. So it's pretty, everything involved is pretty small. So that's definitely working for you here. Holy crap. Aaron Francis It's fast too. Okay. I benchmarked the exact model using desktop Rust runtime on the studio. The small prompt took 0.1 milliseconds. The P95 is 0.12 0 milliseconds. A build log, so a build log having lots more scroll back would be 0.8 milliseconds, P95 0.85. And then super dense output plus a lot of scroll back is P95 8 milliseconds. So an order of magnitude slower, but still 8 milliseconds. And it says loading the model takes 1.7 milliseconds one time. And then prediction timings include feature extraction, which accounts for almost all the work. There's synthetic benchmarks, yada, yada, yada. So it works mechanically and is fast. Accuracy is the next thing to establish. Yeah, I know. Dude. Oh. All right, we'll check back in with AI. Yeah, we'll report back. But it is fast and it is small. ([Time 0:54:25](https://share.snipd.com/snip/14b087e0-de15-43fb-82ec-9903f226211e)) - **AI Model Rankings Depend On Capacity** - Aaron Francis finds Astra powerful for subtle bugs and performance work, while Ian Landsman attributes model perceptions partly to provider capacity. - Anthropic’s strongest model may be better, but restricted availability makes OpenAI’s broadly accessible model more useful. Transcript: Aaron Francis I've used all of it. There's no more Astra left. If you ran into this, this model is at capacity that's my fault oh boy yeah it's good man it's really good it loves it loves python and it loves to stop and ask like great we're done can i do This part now and i'm like yeah i told you to do all that so that that feels like a minor regression yeah but in terms of things that's found and like the actual, you know, I can't judge the Quality of the REST code. But in terms of like finding subtle bugs, finding performance improvements, it's very, very good. I can't speak exactly to how much usage it uses, because like I said, I did burn through banked resets across my four accounts. And so there's an actual chance I'd use like 10 quotas worth of Astra so far. But it's powerful. I've kept it running on either high or medium. I've only dipped into like extra high or max once or twice and then thought, I don't even think this is needed. I saw that Tebow tweeted that Astra on low is the same or better as Soul on, I think it was extra high. Yeah, was it high? I think it was high. So it's very good. And, you know, to broaden the conversation a bit, I don't know what Anthropic does at this point because I feel like they're losing the mandate of heaven. I know that they're embedded in enterprises. I fully agree with that. I'm not talking about that. I'm talking about the models specifically. It does seem like OpenAI has started to run away with it just a little. Ever since like Opus 4, was it 4.8 or 4.7 that people started to be like, myself included, this isn't as good as the old Opus. I think 4.5 and 4.6 were awesome. And then somewhere in 4.7, 4.8, they kind of like started to get the clod speak and started to do a lot worse. And I just feel like OpenAI has got it right now. Ian Landsman So I disagree, but for different sorts of reasons. So like, I think people are misattributing this idea to like the models are worse. But I don't think it's really that at all. I think it's that like they lack capacity. And OpenAI went out first and got tons of capacity. And Anthropic didn't, right? So they've now got, they stole Elon's capacity and they got capacity from Google and whatever. But I still think they're like pretty light on capacity. And so that's forced them to make Fable, which everybody has for months been like, Fable's way better than all the OpenAI stuff, than everything else. Fable's better. It's clearly better. But you couldn't use it like you use Sol, even when Sol... I think Fable is better than Sol, but it's like I can use Sol like unlimited. They're giving me basically like unlimited tokens every week or I can like use fable and be out of tokens in the first day and then have to like you know drop down and cherry pick and all Stuff so that's where i think it's like if they could have released fable as just hey use it it's it's just it's just now the top model but it's like a normal level of token usage in your account. I think that everybody would be like, oh, Anthropix has been on top for two months now, and now they're going back and forth. Maybe Astro is better than Fable, fine. But they're going back and forth. Whereas it's had the perception that that's not true because everybody moved to Sol, because Sol gives you unlimited tokens, and Anthropix is restricting the tokens on their sort Of top-tier model. And so that creates that perception amongst the X $200 a month world of that part of it. Aaron Francis But I feel like Cable is really good. Are you saying that warfare is primarily an exercise in logistics and this is no different? Yeah, I think that's true. Ian Landsman And I also think like who they're each serving is quite interesting, right? Like OpenAI is, I think, much more trying to serve the X crowd. And this is where they've kind of won is like the normies who have a free chat GPT account. Yeah. And like the people, the prosumers who have got their $200 a month account. And then Anthropics won the enterprise of like, we got in there and like got all these five-year deals with enterprise. So it's like, well, I only have so much capacity. Am I going to give it to these people who are basically stealing tokens from me or am I going to give it to IBM? Well, I'm just going to give it to IBM, right? So like, so I think the capacity is actually a big part of it. I don't think it's quite as clear on like the model capability because I think Fable's been out a long time and Fable's better than those OpenAI models, except maybe Astro, who knows, I don't know. ([Time 1:00:36](https://share.snipd.com/snip/3948a4bc-d16b-4bf5-bf3a-1692037803ff))