# 149: I Am The Harness  ## Episode metadata - Episode title: 149: I Am The Harness - Show: Mostly Technical - Owner / Host: Ian Landsman and Aaron Francis - Episode publish date: 2026-09-16 - Episode AI description: A Las Vegas trip sparks conversations about intimate tech conferences, travel disasters, and a first ride in a driverless Zoox. They explore Solo’s universal AI agent, burning tokens across multiple machines, and building a tiny offline classifier for terminal states. Other topics include Outro interface fixes, PostShiba’s email ambitions, and why higher-effort AI reviews uncover better code. - Duration: 01:16:26 - Episode URL: [Open in Snipd](https://share.snipd.com/episode/c7d48ec8-4a05-4418-988d-32385472bc33) - Show URL: [Open in Snipd](https://share.snipd.com/show/7ad45fb4-8aa6-49c2-83cc-bc2327d14b03) - Export date: 2026-09-26T19:19:50 ## Snips ### [Why CabanaConf Works Better Than Big Conferences](https://share.snipd.com/snip/fcb893aa-b524-454f-85be-b33d405944cf) 🎧 11:34 - 13:52 (02:18) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/cc3368b0-389e-4adb-90c4-323aaf855205" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - CabanaConf works because a small, trusted group can mix friendship, teasing, and candid business conversations. - Aaron Francis proposed turning Laracon into dozens of poolside cabanas, but Ian Landsman said cabana culture requires an existing relationship. #### 💬 Quote > Cabana can't be your first date. Cabana has to be your 12th date. > — Ian Landsman Ian Landsman on the conditions that make CabanaConf work #### 📚 Transcript **Ian Landsman:** This is my favorite shape of of conference where you just go somewhere and sit around with your friends right that is that is the platonic ideal of a conference so it's hard to hard to do with more than about you know 10 people so larikon's not gonna just be a hey i'll just come and we'll all sit around all day. It's kind of hard to pull that off. But when you have a cohesive group and you can just sit around and make fun of each other and talk about business, that's the dream. That is the dream. **Aaron Francis:** All right, here's an idea for you. Since you're now in a position to implement such an idea, **Ian Landsman:** the **Aaron Francis:** Aria, big casino, got like, I think there's close to 30 cabanas. Each one could hold 10-ish people. That's 300 people. If we rent out the whole pool area for a laracon, the whole thing could be cabanas. People just wander cabana to cabana, and it's a cabana laracon. I **Ian Landsman:** appreciate the spirit of the request. It's a terrible idea. It's a terrible idea. It's a terrible idea. Bring the cabana to the people, Aaron. No, it just wouldn't work. You know, you got to have some sort of pre-existing relationship to cabana just right. You know what I mean? Cabana can't be your first date. Cabana has to be your 12th date, you know? But maybe it's awesome. Where you know the people and you can **Aaron Francis:** rib. Everybody gets assigned. No, no, no. Everybody gets assigned a cabana. That's why everybody's got the day in the cabana. That's their base of operations. They're meeting nine new people, right? It's going to force that interaction that people are always like, I don't know, I got to go up to people. It's always tricky in a regular conference. No, you're there. You're in cabanas. There's a thing of spicy margs sitting right there that you can bond over. I'll **Ian Landsman:** take it under consideration. I wouldn't be optimistic about it from your perspective, but I'll take it under consideration. Thank you for bringing this up. Yes. Let's **Aaron Francis:** go. As the vice chancellor **Ian Landsman:** of marketing, it is under consideration. --- ### [The Best Mastermind Does Not Feel Like One](https://share.snipd.com/snip/cd29108f-d5cf-4ac4-8213-99e779f35469) 🎧 13:55 - 21:37 (07:42) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/6fb40cba-dbbd-4d5e-a758-71dee327571d" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Trusted friendships create an informal mastermind without forced presentations or scheduled problem-solving. - The valuable advice may surface months later, when a remembered poolside conversation suddenly applies to a new problem. #### 💬 Quote > We're doing the best part and that's why we're all here. > — Aaron Francis Aaron Francis on the value of informal peer groups #### 📚 Transcript **Ian Landsman:** So, did you get anything out of the conference besides the poker? Did you have any takeaways? Do **Aaron Francis:** I have any takeaways for myself? So, I try to approach it not super business-y, but we had some outro discussion, right? There was some outro discussion. So I think I maybe have an outro thing in there. We'll see. I don't, for myself, have a ton of takeaways. I enjoy, we gave some advice to Caleb Porzio. Kind of excited about that advice. I hope he takes that. That's one of the general business things I think could come out of it in general. But yeah, I don't know. I don't personally focus on the business-y end of it too much, although there is a discussion. You get it for the sweet hang. I'm there for the sweet hang. So that's not really my goal. But yeah, I do think there was an outro thing or two that was kind of interesting. And then the last day, there was just John O. Nolan and Justin Jackson and myself. We sat by the pool in the morning while everybody else was flying home because their flights were later. We had some like general business discussion about like kind of employees and management and some things like that. So there's a couple of things in there that was that were cool, too. Yeah. So a little bit of this and that. but yeah, I would say I don't have like a, I didn't like have a run home and I got to implement thing X type of takeaway, but you know, it's always good. There's always, it's always just, you know, there's just stuff flowing into, you never know what you're going to pick out as you're having other business discussions. Yeah, so like maybe six months from now, it's like, oh yeah, I remember we talked about that in the cabana in the context of somebody else's issue. And maybe that applies to me. So always on the lookout for those. But what about you, Mr. Chancellor of Marketing? **Ian Landsman:** Well, for you, I do feel like the big takeaway was the people want outro and they want it shipped. There were a few unnamed holdouts that were like, never ship because then you'll have customers and customers are the worst and you don't want customers. That was true. That was kind of the main advice from several people. That was the main advice from a minority contingent, which was very funny and very wrong. So, I mean, I guess if you don't want customers, a great way to do that would be not shipping. So in that regard, mission accomplished. But I do feel like the overwhelming advice was just get it out there because we like it. So for me, you know, it was interesting this year. This is my second year. And this year I went as a, you know, a W-2 employed person. I took my PTO and everything. I submitted my PTO to one Taylor Otwell. He approved, **Aaron Francis:** luckily. **Ian Landsman:** And he approved it. Not before asking the chat, should I approve this PTO request? And everybody said, no, make him work. So it was interesting because I didn't have a lot of like, hey, I'm trying to, you know, I'm trying to launch this new product. I'm trying to make like, I'm not, you know, looking for my own business advice. But I will say, it is a very good time to be Laravel Holdings Incorporated. It was nice to get because these aren't all Laravel people that came to this. Right. And it was nice to get outside perspective on the state of like the hosting market and kind of like the state of I know we're not an indie business, but like the state of indie business or small sass or whatever. And it seems like Laravel has the mandate of heaven right now. It just seems like this is the place to be. People are very excited about some of the things we talked about that are on the horizon for Laravel Cloud, et cetera. And like being able to describe it all to people who are not in the ecosystem, like people in the Rails world, et cetera, just made me very, very excited because they totally get it. And they're like, yeah, I want to be a part of that, which is really cool. And so just made me feel made me feel very excited. And there were some good ideas for marketing some of the upcoming things. But on the whole, it was just like, hey, I made a I made a great call joining this company, which is a nice feeling. Yeah. Yeah. **Aaron Francis:** Yeah. Yeah. There was some interesting like kind of hosting world chat that was cool. That some of which I was around for. And then, yeah, it's like a good vibe check. There was like a vibe check element to it. And then, you know, it's a very Taylor thing. I feel like building something people want to be a part of. Right. This is one of his like subtle int skills i would say that's not so easily copied right is like how do you make an environment that's people are like jazzed about right so um yeah that was all cool all right so you got some good stuff out of it the cabana paid for itself got **Ian Landsman:** some good stuff out of it um yeah i I think other people maybe got more. We told somebody to quit their job. We told somebody else to blow up their business. Good stuff like that. known for a long time and people that you trust and then you ask a question and everyone's on the same page about it it's very i have to imagine very clarifying to be like wait all 10 of you guys see this and i don't well huh that's interesting maybe i should just take your opinions um which is the value of you know if you want to make it formal the value of the mastermind right you get these people that you know, like, and trust, and then you can get, uh, almost independent opinions. And that's the whole, like, that is the reason masterminds were invented. Um, you know, I don't want to call Cabana a mastermind because I don't want to, you don't want to over business it. No, no, it's just a bunch of dudes hanging. Yes. But it's a mastermind. I hate the sound of that. That sounds terrible. You know, yeah, **Aaron Francis:** it's sort of is. See, what I like about it is not the like most of the masterminds I've seen from afar are more like everybody comes with their they have to do a pitch or whatever. You got to bring a topic and i like the more it's like hey while we're in this cabana having some drinks if you have something that's on your mind then you should bring it up and we're all going to chat about it and everybody's willing to participate but it's not required that you participate um because also i mean there's like the element i mean this is what you know people in masterminds are always like and then we went skiing and that was the best part or whatever. It's like we kind of cut to the best part. Like the whole thing skiing, yeah. So we're doing the best part and that's why we're all here. We're having the social bonds of like, so that maybe I will have some real advice. I will have a moment where I need some true advice that's not just randomly happens to be when the mastermind is going on. But is that some future date? And now I feel like I have a bunch of people that I'm well connected to whose opinions I trust that I could ask at that time. Like that's, that's super valuable. Um, but yeah, I don't necessarily have to come up with something just because I have a 30 minute block scheduled and whatever, all that stuff. Uh, but yeah, it's cool. Um, so I don't know. --- ### [Ship The Product People Keep Asking For](https://share.snipd.com/snip/ae5110e4-a138-4d5c-b9b8-9970d1d8d803) 🎧 15:43 - 19:10 (03:26) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/00663a4a-b363-4aa1-b6ab-254bfca07f82" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman heard overwhelming support for shipping Outro despite fears that customers create burdens. - Aaron Francis also saw strong external enthusiasm for Laravel, Laravel Cloud, and the company’s broader direction. #### 💬 Quote > If you don't want customers, a great way to do that would be not shipping. > — Ian Landsman Ian Landsman on the danger of delaying Outro #### 📚 Transcript **Ian Landsman:** Well, for you, I do feel like the big takeaway was the people want outro and they want it shipped. There were a few unnamed holdouts that were like, never ship because then you'll have customers and customers are the worst and you don't want customers. That was true. That was kind of the main advice from several people. That was the main advice from a minority contingent, which was very funny and very wrong. So, I mean, I guess if you don't want customers, a great way to do that would be not shipping. So in that regard, mission accomplished. But I do feel like the overwhelming advice was just get it out there because we like it. So for me, you know, it was interesting this year. This is my second year. And this year I went as a, you know, a W-2 employed person. I took my PTO and everything. I submitted my PTO to one Taylor Otwell. He approved, **Aaron Francis:** luckily. **Ian Landsman:** And he approved it. Not before asking the chat, should I approve this PTO request? And everybody said, no, make him work. So it was interesting because I didn't have a lot of like, hey, I'm trying to, you know, I'm trying to launch this new product. I'm trying to make like, I'm not, you know, looking for my own business advice. But I will say, it is a very good time to be Laravel Holdings Incorporated. It was nice to get because these aren't all Laravel people that came to this. Right. And it was nice to get outside perspective on the state of like the hosting market and kind of like the state of I know we're not an indie business, but like the state of indie business or small sass or whatever. And it seems like Laravel has the mandate of heaven right now. It just seems like this is the place to be. People are very excited about some of the things we talked about that are on the horizon for Laravel Cloud, et cetera. And like being able to describe it all to people who are not in the ecosystem, like people in the Rails world, et cetera, just made me very, very excited because they totally get it. And they're like, yeah, I want to be a part of that, which is really cool. And so just made me feel made me feel very excited. And there were some good ideas for marketing some of the upcoming things. But on the whole, it was just like, hey, I made a I made a great call joining this company, which is a nice feeling. Yeah. Yeah. **Aaron Francis:** Yeah. Yeah. There was some interesting like kind of hosting world chat that was cool. That some of which I was around for. And then, yeah, it's like a good vibe check. There was like a vibe check element to it. And then, you know, it's a very Taylor thing. I feel like building something people want to be a part of. Right. This is one of his like subtle int skills i would say that's not so easily copied right is like how do you make an environment that's people are like jazzed about right so um yeah that was all cool all right so you got some good stuff out of it the cabana paid for itself got --- ### [PostShiba Turns Deliverability Into A Platform](https://share.snipd.com/snip/71d0646f-2a09-468f-8612-3680d88e14eb) 🎧 22:17 - 27:59 (05:41) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/6cbe444f-6b59-4829-abc6-ea1153b43385" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - PostShiba turns Bento’s hard-won email deliverability infrastructure into a platform for transactional and customer-domain email. - Its deeper opportunity is enabling developers to build products like Resend or Bento without starting from raw email infrastructure. #### 💬 Quote > He's selling that underlying infrastructure. > — Ian Landsman Ian Landsman on PostShiba’s business model #### 📚 Transcript **Ian Landsman:** So Jesse and team are launching a new product called PostShiba, the transactional email platform that powers Bento's amazing delivery rates. Password resets, logins, receipts, emails your users are refreshing their inbox waiting for you to send. And you can even build Bento on PostShiba. So you build a whole email platform on top of PostShiba. It's all the sending power. Jesse's currently gathering interest in customers to test it out. So if you're looking for a replacement to SendGrid, Mailgun, Resend, or SES? Join the waitlist and he will reach out at postshiba.com. New from one Jesse Hanley and old from one Jesse Hanley, the double up, Bento, the email marketing and CRM platform built for the AI era. Send your product marketing and transactional email with Bento. Check out their new transactional email pricing with rates as low as $0.09 per thousand emails. All right, before we go to AI psychosis, we got to talk about this. Jesse's building, well, Jesse's selling, the underlying sending infrastructure that powers Bento. So it's always been in his terms and conditions that you can't build Bento on Bento, for example. Like you can't build a transactional email platform on top of Bento. Right. But now he's doing the hardest thing, which is trying to literally replace SES, SendGrid, Mailgun, etc. Right. Which I think is very interesting because he's like, he's having to secure servers and ASIN blocks and like get in good with all the, you know, the whatever they're called. Not the Postmasters, but whatever. Whatever. Gatekeepers. The gatekeepers. And now you could build your own transactional or marketing email platform on top of PostShiba instead of defaulting to SES, which is interesting. **Aaron Francis:** Yes. Yeah, so it's like, it seems like, we should get some clarification from, because this is interesting, because it's like, No, we'll just make it up here. This is fine. We're just going to make it up here. Let's go. It'll be more interesting if we make it up. But it's like, it seems like you can do both. Is right? Or am I misunderstanding it? Like, you can... What are the boths? That you could just do... It's a transactional email platform in the traditional sense, right? It's postmark. It's resend. Yes. It's right. Okay, I could just use it that way. But if I want to build my own marketing email service on top of this, that I could do that too. Like, I could build Bento on top of this, literally, or I could build ConvertKit on top of this, is the idea. It's not just, yes, we're a transactional emails but we will tell you infrastructure to do email stuff on top of this if you desire. Yes. **Ian Landsman:** So I think you can, I think you can just like, so Altro or Lara Jobs or whatever could just use PostShiba not to like build an email platform on top of. But then if you're, let's roll back, you know, four or five years, however long Resend has been alive. Resend to be, to simplify it unfairly, Resend is an SES wrapper, right? That's an unfair simplification, but it's directionally correct. But what Resend did was built developer experience and UI niceties, and they got super in tight with the React community, and so they added value on top of SES. Now, you could do that exact same thing using Poshiba instead of SES, and my understanding is Jesse has mastered the art of deliverability from running Bento in this way. So Bento, in my understanding, has never been an SES or SendGrid rapper. He's always at least had some part of actually sending emails from boxes, which just sounds like a nightmare which is his do the hardest thing um and so now he's selling that underlying infrastructure so yeah it's really interesting it's bold yeah oh it's super bold i think uh yeah i want to i want to edit his **Aaron Francis:** ad we got to edit his ad jesse we're gonna edit your ad because like i actually really like on the postshiba.com website, the first like under subhead is send from your customers domains. See like that to me, that's an appealing one, right? Like that's always hard. It's super annoying. So when you have a SaaS that like wants to do that, yeah, it's just a complicated thing to embrace so uh yeah it's gonna help me with that that sounds amazing i also don't think it's mentioned at least at the top of the home page but um he's also gonna do inbound which is something bento does not do uh which you know for the complete sort of sass package, I think is nice to have the inbound. So you're just dealing with like all your email stuff in one spot. Most of the time when you're sending email, you might in theory at least want to be able to receive an email. So I think that's nice to just have a single provider there too. You definitely don't want to try to receive emails without a provider. It's like total super nightmare. So yeah, so they're going to handle that too. So pretty exciting. --- ### [Ian Burned Two Thousand Dollars On Agent Reviews](https://share.snipd.com/snip/d0199208-20bb-48d1-baaa-911b43283bb6) 🎧 28:11 - 34:10 (05:58) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/48d43977-faab-494d-a8e9-6ab800593098" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman used several Tailscale-connected machines to run Solo’s cross-platform tests and review hundreds of commits. - The automated review loop consumed quota across four accounts and led him to spend $2,000 in overflow credits. #### 💬 Quote > The last money in the TryHard Studios corporate account is being spent on tokens. > — Ian Landsman Ian Landsman on Solo’s runaway automated review loop #### 📚 Transcript **Ian Landsman:** Oh my **Aaron Francis:** God, how? They **Ian Landsman:** put you on a plane and you burned it. I know. So I left my laptop in the room at CabanaConf, but the poor guy, you don't want to leave it doing nothing, right? Right, he gets that. It's got all that CPU just sitting there, so I had it churning. So over the past week or two, I've had it just kind of like nighttime churning on getting Linux ready to go. So I now have, I have an Omachi machine that I bought off Amazon, just like IBM ThinkPad, super cheap, crappy computer. And I've gotten it, on there, but then people were like, oh, what about NixOS? What about Ubuntu? I'm like, I don't know what those things are. So instead of buying new computers again, I just got VMs in DigitalOcean, and now I have a Ubuntu machine and a NixOS machine that are now on my Tailscale network. **Aaron Francis:** You're in the Tailscale. I love the Tailscale. Sponsor us. Eric, get Tailscale on sponsoring this show. Come on. **Ian Landsman:** Yeah, sales guy, get Tailscale. So now from my local machine, I just tail scale into i think from this machine into six other computers you know it's crazy i got windows i got a mac studio i got another mac book i got all these linux machines um and so i just had it uh and i also installed like vnc on these uh these linux machines so or i use vnc i don't know if it's installed or not i didn't do any of this um so that i can like pull it up and actually view solo running uh and click around and make sure that it actually works so i had a couple of cabana days where i just left it in the room burning on, Hey, let's run the performance harness on Nix OS, run the performance harness on Ubuntu, which takes a lot. Like the, my, my performance harness takes a lot of work to like figure out what's going on. And then it like, it made a bunch of commits from all these different computers, and then it pushes it all to GitHub. And then I think what's put me down to zero, I'm out of resets. I'm out of quota across four **Aaron Francis:** accounts. I have three resets. Across four accounts. I wish I could sell you my resets. I would buy a reset right now for $200. **Ian Landsman:** So I think what's finally burned through all my quota is after all these machines finished their platform specific fixes, they all pushed up and I pulled it down onto just one machine, one discrete machine. And I said, I need you to review the last seven days commits because it's like all these Linux, you know, all these Linux and Windows fixes. I don't want to make a platform specific fix that breaks Mac OS or whatever. And so I split it up into tranches and had every commit reviewed. And then every defect that it found has to have has to be red-green tested, and then approved by AMP, a reviewer. And we're talking like, I think it came up with 240 things that had edge cases or needed to be fixed. So this has been running since like Friday morning when I, you know, I set it off before I walked down to the cabana and we're now it's now Tuesday and it's still working. Still going? And it's still going. And now I'm out of quota. So it's **Aaron Francis:** like actually doing anything. You think it's like not just stuck? **Ian Landsman:** Oh, no, it's actually doing something. I'm watching it. Yeah. It's got 226 done so there's only like 14 more um but alas I did just I did just load it up with U.S. dollar credits so that should last me four or five minutes um but yeah I mean it is very cool to be able to because it's all on tail scale, it's very cool to tell Codex like, hey, just go run this on the Omarky machine and see what happens. And Codex just like figures out just SSH over there and then I'll do this over there. And then all right, now I can pull that back to this machine. So that part's been really nice. But I got to get this. I got to get this review doom loop under control because it's costing me all my money, man. **Aaron Francis:** How much have you spent in overflow **Ian Landsman:** credits? Just this morning. So I ran out overnight, so I left it running overnight. And I woke up this morning and checked it and it was at zero. And so just this morning, is a business expense. Remember, this is coming out of TryHard Studios Treasury Fund. The last, you know, the last money in the TryHard Studios corporate account is being spent on tokens. I put a thousand into one account and a thousand into another account. Yeah. Okay. So you got to do what you got to do You got to do what you got to do. You don't make hay while the sun shines. You can't stop. What are you going **Aaron Francis:** to do? It's in the middle of doing stuff. You got to keep going. Yeah. That's what I'm saying. I love this. This is fantastic. You're blowing through the counts. You're using credits. There probably is. I imagine you're doing a lot of weird stuff, right? And there's a lot of files moving around, which I think eats up a lot of stuff. You know, like, there's probably some optimizations you could eventually come up with of, like, probably it's doing dumb stuff in certain spots where, like, it's not token-efficiently doing things, you know? But also kind of who cares, so I don't know. --- ### [Tiny Local Models Can Monitor Agent Terminals](https://share.snipd.com/snip/eabcdbb2-852d-4f3c-85cd-5b33c3914fdd) 🎧 40:02 - 45:28 (05:25) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/d1e0fcf7-619a-4d67-a39d-b34fc9219706" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman is training a sub-megabyte classifier to recognize terminal states such as permission prompts, input requests, and active work. - Real captures from macOS, Ubuntu, and Windows produced a model targeting 98% accuracy with sub-10-millisecond responses. #### 💬 Quote > There are going to be so many tiny on device under a megabyte, super fast, totally offline models. > — Ian Landsman Ian Landsman on the future of specialized local models #### 📚 Transcript **Ian Landsman:** Okay. It's going great. So to catch up the dear listener, I'm training a small classifier model to ingest a terminal grid and output a confidence score on the state of that terminal grid, i.e. is it working? Is it paused on permissions? Is it ready for user input? Is it asking me a question, et cetera? So I now am, this is actually, I think, one of my more clever ideas. I have turned local solo, so behind like an environment variable, so you have to flip it on. I've turned all my local solo instances into auto-capture vectors. So now as I'm just working throughout the day, every 30 seconds, it's capturing the grid and the diagnostics and writing it to a file. So my training data is becoming a lot more diverse because it's not all synthetic. A lot of it is actually real. And so then what I'll do is I go to the Mac Studio inside of Solo and I say, hey, can you pull all of the auto captured grids off the other machines and you know how i'm able to do that ian say it say how say it say the word okay can **Aaron Francis:** i say it say not sponsoring us tail scale tail scale yes by the reprobates who are not sponsoring us come on tail scale get on it so **Ian Landsman:** the codex on the Mac Studio reaches out to all the other machines, which is nice because it's mini platforms. So I'm not just capturing macOS grids. I'm capturing Ubuntu and Windows grids as well. Hold **Aaron Francis:** on. We need an explanation of a grid because like, I mean, you've used the term a few times, but I was thinking that is like the state when it stopped. But are you saying every 30 seconds, just capturing the whole screen, even if it's turning through stuff? **Ian Landsman:** Yeah, every 30 seconds, regardless of what the user is doing, capture the rows and columns of the terminal. That's what I call the grid, the rows and columns, the cells. Even if it's just **Aaron Francis:** spitting out its normal, hey, I'm doing this, I'm doing that, like all that stuff. Okay, so you're getting a ton of data. A ton of data. Every 30 seconds across all your terminals, that's going to be a lot of data. Exactly. And that's **Ian Landsman:** important because I need lots of examples where it's like not waiting on a user permission so that this classifier can understand negative outcomes and positive outcomes both. So then I tell it on the Mac Studio, pull it all, it tail scales, it pulls it all from all the machines and wipes them from those machines. So they don't pile up on, you know, smaller computers like my Mac book. And then, um, uh, Codex and Claude both label all of the grids. So they give it their blessed label of I'm a world-class model. This grid is representing a user permission request. If they both agree, so if Claude and Codex both agree with confidence, this is representing this scenario, then it gets labeled. So then we know what the ground truth is. And then all of those, labeled grids go into this model classifier training pipeline. And out the other side is spit an under one megabyte model that responds in under 10 milliseconds with 98% accuracy. Wow, that's what we thought last week. We weren't sure if it was actually accurate. It was like 60% It accurate. So, **Aaron Francis:** okay. So, like, now to the model training, you're just feeding it only the positives, right? I feed it both. You do feed it both. So you're feeding it, like, tons and tons of data now. **Ian Landsman:** Yeah. We have to have positive examples and negative examples okay so that it can differentiate between the two okay um and then there is a there's like a like basically like an eval set of grids that are not used for training but are fully labeled and then that is our test set that we then score the model on yeah so cloud and codex will label just hundreds of grids we hold out a certain portion of them train on the rest and then evaluate against our ground truth that it has never seen before um it's working fantastic. Wow. It's great. This is wild. This is the future of like the not leading labs models. There are going to be so many tiny on device under a megabyte, super fast, totally offline models that can just do these little things for us. It's incredible. I'm loving it. I'm having so much fun. Have **Aaron Francis:** you tested running it like on Linux and Windows and all that yet? Do you think it'll work there? It's going to be okay? Yep, **Ian Landsman:** it's going to work there. I know that it works with that data and it is compiled down in such a way that it will be embedded into solo and just called directly from rust and so i've had it you know i've said like hey put the model over there and see if it works and it works because it's just compiled to be non-platform specific man --- ### [Solo Becomes The Harness For Its Own AI](https://share.snipd.com/snip/9191bdc2-ea4b-45b1-8f2b-ab799eaedf56) 🎧 48:26 - 54:46 (06:19) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/b626cc32-6e18-4de3-802c-2f0aee88c7fc" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Solo’s Oracle combines a language model with canonical application operations, letting users manage projects, agents, and tasks through chat. - Documentation search keeps answers current without requiring Solo to be recompiled whenever guidance changes. #### 💬 Quote > I am the harness now. > — Ian Landsman Ian Landsman on giving Solo’s agent control over the application #### 📚 Transcript **Ian Landsman:** right the old self-hoisting um i also on a lark was like what if what if solo had a universal agent where you could just pop open a little like a little UI chat? So not not a terminal. You just hit the little whatever it's going to be a chat bubble or whatever. And then you say, hey, can you add this project to my Solo? Hey, what processes are running? Hey, how do I, what the hell is a scratch pad? How do I do X, Y, Z? And then this universal agent can control solo the application to add projects, inspect processes, answer customer questions, like, because it has access to the documentation. I don't like that one as much. I know, that's fine. **Aaron Francis:** I don't think you're going to be able to pull that off. That seems much harder. That's a much harder thing to pull off than the classifier. The **Ian Landsman:** podcast gods have smiled upon me because **Aaron Francis:** you said that. I **Ian Landsman:** have already pulled it off. This worked out great for me. That segment could not have gone better for me. I got a great idea. You'll never be able to do it. I've done it. Hall of Fame. Hall of Fame. All right. Tell me about this. Now I'm intrigued here. Okay. So I didn't train this model. All right. This is, uh, is an open source, open weights model. Quinn, whatever. But it's **Aaron Francis:** a real model. That's what I'm talking about. Yeah. I was thinking you need a real model for this. That's not just like a **Ian Landsman:** self-trained thing. I never said it wasn't a real model. So it appears you have egg on your face. And it appears, it appears, Ian, that you have egg on **Aaron Francis:** your face. I still have a problem with this anyway. Don't even worry about it because I feel like you could just use all the models that are in there and just use Claude Haiku and be done with it or whatever, Luna. Like, the user has models that are real and you can just use them. Perhaps. **Ian Landsman:** Yeah. Now, let me tell you why you're wrong. Okay. Give it to me. This is my favorite segment of the show. Yes, this is a good one. All right, let's go. All right. So this is primarily a user experience, user onboarding adventure, we'll say. So when a user downloads solo and they just fire it up, we do our best to figure out what agents they already have. So we're doing the most that we possibly can to make the onboarding as fast as possible. However, people still ask me questions that I'm like, this seems obvious to me. Obviously, I built it so everything is obvious to me. And if you want to win, you have to, like, be empathetic and understand, okay, I maybe have more knowledge. The user is not understanding this. That is a me problem, not a them problem. Right? So, now, you can download Solo. And there's a little, I think I've called it the Oracle, which is a cool name. There's a little Oracle where you can pop open this chat and you can ask it, hey, can you add my outro.fm project to Solo? And this agent, it's interesting because this is like the first time I've built a harness, right? else I've done is meta harness which is like let me put some chrome around the lab twoies yeah now I'm just doing raw model calls and I am the one who's like oh you want to use this tool all right well I need to parse that out use the tool and give you the response back yeah so everything is coming together because of the way I've architected this thing. So I've got this layer of like internal operations that's like create scratch pad, update scratch pad, you know, complete to do. All of these discrete functions live in this middle layer of the application. Then every ingress point, HTTP, MCP, CLI, they all have their own adapters into these canonical operations, right? So again, this comes back to engineering, not just like vibe coding. Right. Now with this Oracle, it basically has access to all of those canonical operations by just writing another thin ingress point that is now the tool call ingress point so i have this quin i've got this quin whatever 3.5 i don't even really know uh model running through uh running through So it's hosted on U.S. inference. It's not hosted in China or anything. So Solo makes this API request to Cloudflare, and it basically passes along, like, here are all the tools that are at your disposal, or if you need to, like, search tools, here's how you can do it. Passes all that to Cloudflare. Quinn gives me the response and inside it embedded our tool calls. Then the Rust code actually executes those tools, gives it back to the model. So now everything that you can do via MCP, you can just do in this one little chat box. So you can be like, hey, how many to-dos do I have open in this project? Well, you've got 40 in that project. Or, hey, can you close all of the agents over there? And so this is something that some people I don't think many people have requested this, but I think many people in their core want this. But some people have requested like I need a better way to just like triage and do stuff without opening a new terminal agent. And so this is my response to that. But this is also my response to onboarding because it has access to all of the web hosted API docs. So it can like, it can call search documentation. And then the rest behind the scenes goes out to the API to read all the documentation. So that means it's alive. It's always living. I don't have to recompile solo because it's searching the web for the docs. And so you can ask the Oracle, hey, how do I spawn a new agent? What's a good orchestration flow? How do I set up XYZ? --- ### [Use Data Structures To Make Bugs Impossible](https://share.snipd.com/snip/37802c07-9964-48a0-a649-18396af42966) 🎧 01:03:49 - 01:06:41 (02:52) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/bad3a45f-d99c-45fb-95a6-709136666e28" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Use a data-structures-and-algorithms audit to replace magic strings, scattered checks, and fragile logic with explicit structures. - Ian Landsman uses the audit before committing feature work so bugs become impossible rather than merely patched. #### 💬 Quote > Make the bug impossible by introducing a named data structure or algorithm. > — Ian Landsman Ian Landsman on his DSA audit prompt #### 📚 Transcript **Ian Landsman:** So the DSA audit is data structures and algorithms. And the whole intent of the audit is look for places where we are basically using data structures, but not. So, you know, if you're like, this should be represented as a state machine, but we're not. Or this should be an enum, but we're just kind of having magic strings passed around. The whole intent of this is break down my application into subsystems, and then every subsystem look for simplifications or bugs that can't not just be fixed, be obviated. Make the bug impossible by introducing a named data structure or algorithm. And it's like, oh, well, if this was just a discriminated union, you wouldn't need all these checks here. And you're like, oh, well, that actually seems pretty nice. And so that's the intent of this giant prompt, which we'll link up in the show notes. But it's very good. Before I ever commit, like if I'm doing a feature spike before I ever commit, I run a minor DSA audit on just that new work. And it always comes back with, yeah, this would be better if this were a struct or an enum or whatever. And it just makes the possibility of bugs, like this specific type of bug, it just makes it impossible, which is fine **Aaron Francis:** yeah yeah no seems like super good it's always nice when these things uh works i gotta put it i just have it like i'm pasting it in right now but i gotta put it maybe i'll make it a sky do you have it just set up as a skill or do you just paste it i have it set up **Ian Landsman:** as a solo prompt template right as a prompt template yeah so so **Aaron Francis:** not a skill though so you're just using it on on demand when it. Yeah. Yeah. Um, which I mean, it is nice that way too, because like for this, like I did, this was in help spot. So it was like, I only wanted to do this on the changes I made for help spot six, which has all the new UI stuff. Like I didn't want to like just go through all help spot and like, I don't want to get into like doing 20 year history. Yeah. Like at some point, maybe we'll do that. But like this one, I just wanted to focus on the new stuff. And yeah, so it would obviously did that and that was fine. And so, yeah, so this way you can like, so I had to tweak your prompt slightly at the beginning of it, right? To like not go through the whole app. So did that. And then, yeah, it was great. So the AI, it's getting there. I also feel like this is one that feels like improved significantly with the better model. Like I feel like Astrid is probably a lot better at this than, I don't know, just something about the way it's structured. Like I've done these things before, somewhat similar. I don't know if I've done this exact, maybe I even did this prompt on Sol once, but I don't know. It's structure, it returned. It's like confidence levels, how it described each one just felt like really tight and good. You know, it's like, oh yeah, like it's, it's, it's got it. --- ### [Higher Reasoning Found The Issues Medium Missed](https://share.snipd.com/snip/71fa3b9e-1cd7-4b3c-a158-5fbec2fe608d) 🎧 01:06:44 - 01:07:33 (00:48) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/6d4329d9-640e-4cbf-a27c-9055717f2561" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Aaron Francis found Astra dramatically more capable after moving from medium to high reasoning. - The higher setting connected impacts across the codebase instead of merely identifying isolated issues. #### 💬 Quote > With the high it does that, and I'm like, yes, I want that. > — Aaron Francis Aaron Francis on increasing Astra’s reasoning level #### 📚 Transcript **Aaron Francis:** I did tweet this out that I moved from the astro medium to high and that seems like a market improvement in everything. Like I was trying to do the medium cause the SIBO guy said do the medium and the medium is much better. I don't know. I feel like there's a big difference between medium and high and high is just like significantly better. **Ian Landsman:** I agree. I agree. **Aaron Francis:** Is high your baseline? **Ian Landsman:** Yeah. I just want the best, man. Yeah. **Aaron Francis:** Like I wanted to catch things. Like I wanted to find issues. Like I don't know. With medium, it was never being like oh and then like I saw over here that relates to this like we should fix that too or we should do this thing or whatever and like with the high it does that and I'm like yes I want that I want it to know that like to look around and be like oh this impacts these other areas we should do that over there like yeah that's what I want like I want you in there looking around so make it better --- ### [A Zoox Ride Felt Like The Future](https://share.snipd.com/snip/707f492e-13c6-4f23-968c-cba45030f920) 🎧 01:07:40 - 01:09:58 (02:18) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/fbbbdf94-741d-4f87-adc8-77cc54d6f521" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman took a driverless Zoox from the Las Vegas airport and worked alone inside its four-seat, train-like cabin. - The seamless ride convinced him autonomous vehicles could carry most driving miles within a decade. #### 💬 Quote > I walked out of the Vegas airport and [...] I'm not taking an Uber. > — Ian Landsman Ian Landsman on choosing a Zoox over taxis and Uber #### 📚 Transcript **Ian Landsman:** I took a Zoox for the first time. Oh, you did take a **Aaron Francis:** Zoox. That's right. **Ian Landsman:** I took a Zoox from the airport to the resort, the casino. The Zoox is a four-passenger vehicle with no steering wheel no driver front seat back seat it's just like a miniature train car but it's better than trains because i still have my personal autonomy as an american which is nice um so i took a zooks from the airport and it's awesome was **Aaron Francis:** it good it **Ian Landsman:** was great so it's two seats on either side facing each other i was in there by myself which is the ideal um and you have you know you've got your little controls for like air conditioning music whatever um and i just got my laptop out i'm just working in this little train car all by myself uh and then you pull up to the resort and the double doors just kind of like slide open and you just step out and you never say anything. It was very interesting because I was like, I walked out of the Vegas airport and you know, there's the line of taxis. And I look at the taxis and I kind of scoffed myself. I'm not taking a taxi. Come on. This isn't 1999. And then I walked downstairs and I see the Ubers and I'm like, come on, I'm not taking an Uber. This is a 2025. And so I follow the signs to a Zoox and then I get in the car that is not a car and has no driver and has no steering wheel. And I'm like, obviously this is what I'm supposed to be doing. And I do, I think it is. I think it is the future. I think in 10 years, I think the majority of miles driven will be driverless. And I think many of them will be these like tightly packed trains of Zoox-like things following each other on the highway and peeling off and coordinating with each other and driving in **Aaron Francis:** tunnels. No way, Elon. You're way off. You're way off on that, Elon. A hundred percent. Way off on that. A hundred percent. I think there'll be more than now, of course. But no. We're building coal plants and stuff, man. We're going the opposite direction. I don't care. The electricity **Ian Landsman:** has to come from somewhere. I don't have a problem with coal. Coal's fine. Coal's terrible. But in terms of self-driving, I give it 10 years before the majority of miles are self-driven. --- Created with [Snipd](https://www.snipd.com) | Highlight & Take Notes from Podcasts # Snips updated after 2026-09-24T14:23:14 ### [Why CabanaConf Works Better Than Big Conferences](https://share.snipd.com/snip/fcb893aa-b524-454f-85be-b33d405944cf) 🎧 11:34 - 13:52 (02:18) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/cc3368b0-389e-4adb-90c4-323aaf855205" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - CabanaConf works because a small, trusted group can mix friendship, teasing, and candid business conversations. - Aaron Francis proposed turning Laracon into dozens of poolside cabanas, but Ian Landsman said cabana culture requires an existing relationship. #### 💬 Quote > Cabana can't be your first date. Cabana has to be your 12th date. > — Ian Landsman Ian Landsman on the conditions that make CabanaConf work #### 📚 Transcript **Ian Landsman:** This is my favorite shape of of conference where you just go somewhere and sit around with your friends right that is that is the platonic ideal of a conference so it's hard to hard to do with more than about you know 10 people so larikon's not gonna just be a hey i'll just come and we'll all sit around all day. It's kind of hard to pull that off. But when you have a cohesive group and you can just sit around and make fun of each other and talk about business, that's the dream. That is the dream. **Aaron Francis:** All right, here's an idea for you. Since you're now in a position to implement such an idea, **Ian Landsman:** the **Aaron Francis:** Aria, big casino, got like, I think there's close to 30 cabanas. Each one could hold 10-ish people. That's 300 people. If we rent out the whole pool area for a laracon, the whole thing could be cabanas. People just wander cabana to cabana, and it's a cabana laracon. I **Ian Landsman:** appreciate the spirit of the request. It's a terrible idea. It's a terrible idea. It's a terrible idea. Bring the cabana to the people, Aaron. No, it just wouldn't work. You know, you got to have some sort of pre-existing relationship to cabana just right. You know what I mean? Cabana can't be your first date. Cabana has to be your 12th date, you know? But maybe it's awesome. Where you know the people and you can **Aaron Francis:** rib. Everybody gets assigned. No, no, no. Everybody gets assigned a cabana. That's why everybody's got the day in the cabana. That's their base of operations. They're meeting nine new people, right? It's going to force that interaction that people are always like, I don't know, I got to go up to people. It's always tricky in a regular conference. No, you're there. You're in cabanas. There's a thing of spicy margs sitting right there that you can bond over. I'll **Ian Landsman:** take it under consideration. I wouldn't be optimistic about it from your perspective, but I'll take it under consideration. Thank you for bringing this up. Yes. Let's **Aaron Francis:** go. As the vice chancellor **Ian Landsman:** of marketing, it is under consideration. --- ### [The Best Mastermind Does Not Feel Like One](https://share.snipd.com/snip/cd29108f-d5cf-4ac4-8213-99e779f35469) 🎧 13:55 - 21:37 (07:42) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/6fb40cba-dbbd-4d5e-a758-71dee327571d" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Trusted friendships create an informal mastermind without forced presentations or scheduled problem-solving. - The valuable advice may surface months later, when a remembered poolside conversation suddenly applies to a new problem. #### 💬 Quote > We're doing the best part and that's why we're all here. > — Aaron Francis Aaron Francis on the value of informal peer groups #### 📚 Transcript **Ian Landsman:** So, did you get anything out of the conference besides the poker? Did you have any takeaways? Do **Aaron Francis:** I have any takeaways for myself? So, I try to approach it not super business-y, but we had some outro discussion, right? There was some outro discussion. So I think I maybe have an outro thing in there. We'll see. I don't, for myself, have a ton of takeaways. I enjoy, we gave some advice to Caleb Porzio. Kind of excited about that advice. I hope he takes that. That's one of the general business things I think could come out of it in general. But yeah, I don't know. I don't personally focus on the business-y end of it too much, although there is a discussion. You get it for the sweet hang. I'm there for the sweet hang. So that's not really my goal. But yeah, I do think there was an outro thing or two that was kind of interesting. And then the last day, there was just John O. Nolan and Justin Jackson and myself. We sat by the pool in the morning while everybody else was flying home because their flights were later. We had some like general business discussion about like kind of employees and management and some things like that. So there's a couple of things in there that was that were cool, too. Yeah. So a little bit of this and that. but yeah, I would say I don't have like a, I didn't like have a run home and I got to implement thing X type of takeaway, but you know, it's always good. There's always, it's always just, you know, there's just stuff flowing into, you never know what you're going to pick out as you're having other business discussions. Yeah, so like maybe six months from now, it's like, oh yeah, I remember we talked about that in the cabana in the context of somebody else's issue. And maybe that applies to me. So always on the lookout for those. But what about you, Mr. Chancellor of Marketing? **Ian Landsman:** Well, for you, I do feel like the big takeaway was the people want outro and they want it shipped. There were a few unnamed holdouts that were like, never ship because then you'll have customers and customers are the worst and you don't want customers. That was true. That was kind of the main advice from several people. That was the main advice from a minority contingent, which was very funny and very wrong. So, I mean, I guess if you don't want customers, a great way to do that would be not shipping. So in that regard, mission accomplished. But I do feel like the overwhelming advice was just get it out there because we like it. So for me, you know, it was interesting this year. This is my second year. And this year I went as a, you know, a W-2 employed person. I took my PTO and everything. I submitted my PTO to one Taylor Otwell. He approved, **Aaron Francis:** luckily. **Ian Landsman:** And he approved it. Not before asking the chat, should I approve this PTO request? And everybody said, no, make him work. So it was interesting because I didn't have a lot of like, hey, I'm trying to, you know, I'm trying to launch this new product. I'm trying to make like, I'm not, you know, looking for my own business advice. But I will say, it is a very good time to be Laravel Holdings Incorporated. It was nice to get because these aren't all Laravel people that came to this. Right. And it was nice to get outside perspective on the state of like the hosting market and kind of like the state of I know we're not an indie business, but like the state of indie business or small sass or whatever. And it seems like Laravel has the mandate of heaven right now. It just seems like this is the place to be. People are very excited about some of the things we talked about that are on the horizon for Laravel Cloud, et cetera. And like being able to describe it all to people who are not in the ecosystem, like people in the Rails world, et cetera, just made me very, very excited because they totally get it. And they're like, yeah, I want to be a part of that, which is really cool. And so just made me feel made me feel very excited. And there were some good ideas for marketing some of the upcoming things. But on the whole, it was just like, hey, I made a I made a great call joining this company, which is a nice feeling. Yeah. Yeah. **Aaron Francis:** Yeah. Yeah. There was some interesting like kind of hosting world chat that was cool. That some of which I was around for. And then, yeah, it's like a good vibe check. There was like a vibe check element to it. And then, you know, it's a very Taylor thing. I feel like building something people want to be a part of. Right. This is one of his like subtle int skills i would say that's not so easily copied right is like how do you make an environment that's people are like jazzed about right so um yeah that was all cool all right so you got some good stuff out of it the cabana paid for itself got **Ian Landsman:** some good stuff out of it um yeah i I think other people maybe got more. We told somebody to quit their job. We told somebody else to blow up their business. Good stuff like that. known for a long time and people that you trust and then you ask a question and everyone's on the same page about it it's very i have to imagine very clarifying to be like wait all 10 of you guys see this and i don't well huh that's interesting maybe i should just take your opinions um which is the value of you know if you want to make it formal the value of the mastermind right you get these people that you know, like, and trust, and then you can get, uh, almost independent opinions. And that's the whole, like, that is the reason masterminds were invented. Um, you know, I don't want to call Cabana a mastermind because I don't want to, you don't want to over business it. No, no, it's just a bunch of dudes hanging. Yes. But it's a mastermind. I hate the sound of that. That sounds terrible. You know, yeah, **Aaron Francis:** it's sort of is. See, what I like about it is not the like most of the masterminds I've seen from afar are more like everybody comes with their they have to do a pitch or whatever. You got to bring a topic and i like the more it's like hey while we're in this cabana having some drinks if you have something that's on your mind then you should bring it up and we're all going to chat about it and everybody's willing to participate but it's not required that you participate um because also i mean there's like the element i mean this is what you know people in masterminds are always like and then we went skiing and that was the best part or whatever. It's like we kind of cut to the best part. Like the whole thing skiing, yeah. So we're doing the best part and that's why we're all here. We're having the social bonds of like, so that maybe I will have some real advice. I will have a moment where I need some true advice that's not just randomly happens to be when the mastermind is going on. But is that some future date? And now I feel like I have a bunch of people that I'm well connected to whose opinions I trust that I could ask at that time. Like that's, that's super valuable. Um, but yeah, I don't necessarily have to come up with something just because I have a 30 minute block scheduled and whatever, all that stuff. Uh, but yeah, it's cool. Um, so I don't know. --- ### [Ship The Product People Keep Asking For](https://share.snipd.com/snip/ae5110e4-a138-4d5c-b9b8-9970d1d8d803) 🎧 15:43 - 19:10 (03:26) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/00663a4a-b363-4aa1-b6ab-254bfca07f82" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman heard overwhelming support for shipping Outro despite fears that customers create burdens. - Aaron Francis also saw strong external enthusiasm for Laravel, Laravel Cloud, and the company’s broader direction. #### 💬 Quote > If you don't want customers, a great way to do that would be not shipping. > — Ian Landsman Ian Landsman on the danger of delaying Outro #### 📚 Transcript **Ian Landsman:** Well, for you, I do feel like the big takeaway was the people want outro and they want it shipped. There were a few unnamed holdouts that were like, never ship because then you'll have customers and customers are the worst and you don't want customers. That was true. That was kind of the main advice from several people. That was the main advice from a minority contingent, which was very funny and very wrong. So, I mean, I guess if you don't want customers, a great way to do that would be not shipping. So in that regard, mission accomplished. But I do feel like the overwhelming advice was just get it out there because we like it. So for me, you know, it was interesting this year. This is my second year. And this year I went as a, you know, a W-2 employed person. I took my PTO and everything. I submitted my PTO to one Taylor Otwell. He approved, **Aaron Francis:** luckily. **Ian Landsman:** And he approved it. Not before asking the chat, should I approve this PTO request? And everybody said, no, make him work. So it was interesting because I didn't have a lot of like, hey, I'm trying to, you know, I'm trying to launch this new product. I'm trying to make like, I'm not, you know, looking for my own business advice. But I will say, it is a very good time to be Laravel Holdings Incorporated. It was nice to get because these aren't all Laravel people that came to this. Right. And it was nice to get outside perspective on the state of like the hosting market and kind of like the state of I know we're not an indie business, but like the state of indie business or small sass or whatever. And it seems like Laravel has the mandate of heaven right now. It just seems like this is the place to be. People are very excited about some of the things we talked about that are on the horizon for Laravel Cloud, et cetera. And like being able to describe it all to people who are not in the ecosystem, like people in the Rails world, et cetera, just made me very, very excited because they totally get it. And they're like, yeah, I want to be a part of that, which is really cool. And so just made me feel made me feel very excited. And there were some good ideas for marketing some of the upcoming things. But on the whole, it was just like, hey, I made a I made a great call joining this company, which is a nice feeling. Yeah. Yeah. **Aaron Francis:** Yeah. Yeah. There was some interesting like kind of hosting world chat that was cool. That some of which I was around for. And then, yeah, it's like a good vibe check. There was like a vibe check element to it. And then, you know, it's a very Taylor thing. I feel like building something people want to be a part of. Right. This is one of his like subtle int skills i would say that's not so easily copied right is like how do you make an environment that's people are like jazzed about right so um yeah that was all cool all right so you got some good stuff out of it the cabana paid for itself got --- ### [PostShiba Turns Deliverability Into A Platform](https://share.snipd.com/snip/71d0646f-2a09-468f-8612-3680d88e14eb) 🎧 22:17 - 27:59 (05:41) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/6cbe444f-6b59-4829-abc6-ea1153b43385" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - PostShiba turns Bento’s hard-won email deliverability infrastructure into a platform for transactional and customer-domain email. - Its deeper opportunity is enabling developers to build products like Resend or Bento without starting from raw email infrastructure. #### 💬 Quote > He's selling that underlying infrastructure. > — Ian Landsman Ian Landsman on PostShiba’s business model #### 📚 Transcript **Ian Landsman:** So Jesse and team are launching a new product called PostShiba, the transactional email platform that powers Bento's amazing delivery rates. Password resets, logins, receipts, emails your users are refreshing their inbox waiting for you to send. And you can even build Bento on PostShiba. So you build a whole email platform on top of PostShiba. It's all the sending power. Jesse's currently gathering interest in customers to test it out. So if you're looking for a replacement to SendGrid, Mailgun, Resend, or SES? Join the waitlist and he will reach out at postshiba.com. New from one Jesse Hanley and old from one Jesse Hanley, the double up, Bento, the email marketing and CRM platform built for the AI era. Send your product marketing and transactional email with Bento. Check out their new transactional email pricing with rates as low as $0.09 per thousand emails. All right, before we go to AI psychosis, we got to talk about this. Jesse's building, well, Jesse's selling, the underlying sending infrastructure that powers Bento. So it's always been in his terms and conditions that you can't build Bento on Bento, for example. Like you can't build a transactional email platform on top of Bento. Right. But now he's doing the hardest thing, which is trying to literally replace SES, SendGrid, Mailgun, etc. Right. Which I think is very interesting because he's like, he's having to secure servers and ASIN blocks and like get in good with all the, you know, the whatever they're called. Not the Postmasters, but whatever. Whatever. Gatekeepers. The gatekeepers. And now you could build your own transactional or marketing email platform on top of PostShiba instead of defaulting to SES, which is interesting. **Aaron Francis:** Yes. Yeah, so it's like, it seems like, we should get some clarification from, because this is interesting, because it's like, No, we'll just make it up here. This is fine. We're just going to make it up here. Let's go. It'll be more interesting if we make it up. But it's like, it seems like you can do both. Is right? Or am I misunderstanding it? Like, you can... What are the boths? That you could just do... It's a transactional email platform in the traditional sense, right? It's postmark. It's resend. Yes. It's right. Okay, I could just use it that way. But if I want to build my own marketing email service on top of this, that I could do that too. Like, I could build Bento on top of this, literally, or I could build ConvertKit on top of this, is the idea. It's not just, yes, we're a transactional emails but we will tell you infrastructure to do email stuff on top of this if you desire. Yes. **Ian Landsman:** So I think you can, I think you can just like, so Altro or Lara Jobs or whatever could just use PostShiba not to like build an email platform on top of. But then if you're, let's roll back, you know, four or five years, however long Resend has been alive. Resend to be, to simplify it unfairly, Resend is an SES wrapper, right? That's an unfair simplification, but it's directionally correct. But what Resend did was built developer experience and UI niceties, and they got super in tight with the React community, and so they added value on top of SES. Now, you could do that exact same thing using Poshiba instead of SES, and my understanding is Jesse has mastered the art of deliverability from running Bento in this way. So Bento, in my understanding, has never been an SES or SendGrid rapper. He's always at least had some part of actually sending emails from boxes, which just sounds like a nightmare which is his do the hardest thing um and so now he's selling that underlying infrastructure so yeah it's really interesting it's bold yeah oh it's super bold i think uh yeah i want to i want to edit his **Aaron Francis:** ad we got to edit his ad jesse we're gonna edit your ad because like i actually really like on the postshiba.com website, the first like under subhead is send from your customers domains. See like that to me, that's an appealing one, right? Like that's always hard. It's super annoying. So when you have a SaaS that like wants to do that, yeah, it's just a complicated thing to embrace so uh yeah it's gonna help me with that that sounds amazing i also don't think it's mentioned at least at the top of the home page but um he's also gonna do inbound which is something bento does not do uh which you know for the complete sort of sass package, I think is nice to have the inbound. So you're just dealing with like all your email stuff in one spot. Most of the time when you're sending email, you might in theory at least want to be able to receive an email. So I think that's nice to just have a single provider there too. You definitely don't want to try to receive emails without a provider. It's like total super nightmare. So yeah, so they're going to handle that too. So pretty exciting. --- ### [Ian Burned Two Thousand Dollars On Agent Reviews](https://share.snipd.com/snip/d0199208-20bb-48d1-baaa-911b43283bb6) 🎧 28:11 - 34:10 (05:58) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/48d43977-faab-494d-a8e9-6ab800593098" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman used several Tailscale-connected machines to run Solo’s cross-platform tests and review hundreds of commits. - The automated review loop consumed quota across four accounts and led him to spend $2,000 in overflow credits. #### 💬 Quote > The last money in the TryHard Studios corporate account is being spent on tokens. > — Ian Landsman Ian Landsman on Solo’s runaway automated review loop #### 📚 Transcript **Ian Landsman:** Oh my **Aaron Francis:** God, how? They **Ian Landsman:** put you on a plane and you burned it. I know. So I left my laptop in the room at CabanaConf, but the poor guy, you don't want to leave it doing nothing, right? Right, he gets that. It's got all that CPU just sitting there, so I had it churning. So over the past week or two, I've had it just kind of like nighttime churning on getting Linux ready to go. So I now have, I have an Omachi machine that I bought off Amazon, just like IBM ThinkPad, super cheap, crappy computer. And I've gotten it, on there, but then people were like, oh, what about NixOS? What about Ubuntu? I'm like, I don't know what those things are. So instead of buying new computers again, I just got VMs in DigitalOcean, and now I have a Ubuntu machine and a NixOS machine that are now on my Tailscale network. **Aaron Francis:** You're in the Tailscale. I love the Tailscale. Sponsor us. Eric, get Tailscale on sponsoring this show. Come on. **Ian Landsman:** Yeah, sales guy, get Tailscale. So now from my local machine, I just tail scale into i think from this machine into six other computers you know it's crazy i got windows i got a mac studio i got another mac book i got all these linux machines um and so i just had it uh and i also installed like vnc on these uh these linux machines so or i use vnc i don't know if it's installed or not i didn't do any of this um so that i can like pull it up and actually view solo running uh and click around and make sure that it actually works so i had a couple of cabana days where i just left it in the room burning on, Hey, let's run the performance harness on Nix OS, run the performance harness on Ubuntu, which takes a lot. Like the, my, my performance harness takes a lot of work to like figure out what's going on. And then it like, it made a bunch of commits from all these different computers, and then it pushes it all to GitHub. And then I think what's put me down to zero, I'm out of resets. I'm out of quota across four **Aaron Francis:** accounts. I have three resets. Across four accounts. I wish I could sell you my resets. I would buy a reset right now for $200. **Ian Landsman:** So I think what's finally burned through all my quota is after all these machines finished their platform specific fixes, they all pushed up and I pulled it down onto just one machine, one discrete machine. And I said, I need you to review the last seven days commits because it's like all these Linux, you know, all these Linux and Windows fixes. I don't want to make a platform specific fix that breaks Mac OS or whatever. And so I split it up into tranches and had every commit reviewed. And then every defect that it found has to have has to be red-green tested, and then approved by AMP, a reviewer. And we're talking like, I think it came up with 240 things that had edge cases or needed to be fixed. So this has been running since like Friday morning when I, you know, I set it off before I walked down to the cabana and we're now it's now Tuesday and it's still working. Still going? And it's still going. And now I'm out of quota. So it's **Aaron Francis:** like actually doing anything. You think it's like not just stuck? **Ian Landsman:** Oh, no, it's actually doing something. I'm watching it. Yeah. It's got 226 done so there's only like 14 more um but alas I did just I did just load it up with U.S. dollar credits so that should last me four or five minutes um but yeah I mean it is very cool to be able to because it's all on tail scale, it's very cool to tell Codex like, hey, just go run this on the Omarky machine and see what happens. And Codex just like figures out just SSH over there and then I'll do this over there. And then all right, now I can pull that back to this machine. So that part's been really nice. But I got to get this. I got to get this review doom loop under control because it's costing me all my money, man. **Aaron Francis:** How much have you spent in overflow **Ian Landsman:** credits? Just this morning. So I ran out overnight, so I left it running overnight. And I woke up this morning and checked it and it was at zero. And so just this morning, is a business expense. Remember, this is coming out of TryHard Studios Treasury Fund. The last, you know, the last money in the TryHard Studios corporate account is being spent on tokens. I put a thousand into one account and a thousand into another account. Yeah. Okay. So you got to do what you got to do You got to do what you got to do. You don't make hay while the sun shines. You can't stop. What are you going **Aaron Francis:** to do? It's in the middle of doing stuff. You got to keep going. Yeah. That's what I'm saying. I love this. This is fantastic. You're blowing through the counts. You're using credits. There probably is. I imagine you're doing a lot of weird stuff, right? And there's a lot of files moving around, which I think eats up a lot of stuff. You know, like, there's probably some optimizations you could eventually come up with of, like, probably it's doing dumb stuff in certain spots where, like, it's not token-efficiently doing things, you know? But also kind of who cares, so I don't know. --- ### [Tiny Local Models Can Monitor Agent Terminals](https://share.snipd.com/snip/eabcdbb2-852d-4f3c-85cd-5b33c3914fdd) 🎧 40:02 - 45:28 (05:25) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/d1e0fcf7-619a-4d67-a39d-b34fc9219706" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman is training a sub-megabyte classifier to recognize terminal states such as permission prompts, input requests, and active work. - Real captures from macOS, Ubuntu, and Windows produced a model targeting 98% accuracy with sub-10-millisecond responses. #### 💬 Quote > There are going to be so many tiny on device under a megabyte, super fast, totally offline models. > — Ian Landsman Ian Landsman on the future of specialized local models #### 📚 Transcript **Ian Landsman:** Okay. It's going great. So to catch up the dear listener, I'm training a small classifier model to ingest a terminal grid and output a confidence score on the state of that terminal grid, i.e. is it working? Is it paused on permissions? Is it ready for user input? Is it asking me a question, et cetera? So I now am, this is actually, I think, one of my more clever ideas. I have turned local solo, so behind like an environment variable, so you have to flip it on. I've turned all my local solo instances into auto-capture vectors. So now as I'm just working throughout the day, every 30 seconds, it's capturing the grid and the diagnostics and writing it to a file. So my training data is becoming a lot more diverse because it's not all synthetic. A lot of it is actually real. And so then what I'll do is I go to the Mac Studio inside of Solo and I say, hey, can you pull all of the auto captured grids off the other machines and you know how i'm able to do that ian say it say how say it say the word okay can **Aaron Francis:** i say it say not sponsoring us tail scale tail scale yes by the reprobates who are not sponsoring us come on tail scale get on it so **Ian Landsman:** the codex on the Mac Studio reaches out to all the other machines, which is nice because it's mini platforms. So I'm not just capturing macOS grids. I'm capturing Ubuntu and Windows grids as well. Hold **Aaron Francis:** on. We need an explanation of a grid because like, I mean, you've used the term a few times, but I was thinking that is like the state when it stopped. But are you saying every 30 seconds, just capturing the whole screen, even if it's turning through stuff? **Ian Landsman:** Yeah, every 30 seconds, regardless of what the user is doing, capture the rows and columns of the terminal. That's what I call the grid, the rows and columns, the cells. Even if it's just **Aaron Francis:** spitting out its normal, hey, I'm doing this, I'm doing that, like all that stuff. Okay, so you're getting a ton of data. A ton of data. Every 30 seconds across all your terminals, that's going to be a lot of data. Exactly. And that's **Ian Landsman:** important because I need lots of examples where it's like not waiting on a user permission so that this classifier can understand negative outcomes and positive outcomes both. So then I tell it on the Mac Studio, pull it all, it tail scales, it pulls it all from all the machines and wipes them from those machines. So they don't pile up on, you know, smaller computers like my Mac book. And then, um, uh, Codex and Claude both label all of the grids. So they give it their blessed label of I'm a world-class model. This grid is representing a user permission request. If they both agree, so if Claude and Codex both agree with confidence, this is representing this scenario, then it gets labeled. So then we know what the ground truth is. And then all of those, labeled grids go into this model classifier training pipeline. And out the other side is spit an under one megabyte model that responds in under 10 milliseconds with 98% accuracy. Wow, that's what we thought last week. We weren't sure if it was actually accurate. It was like 60% It accurate. So, **Aaron Francis:** okay. So, like, now to the model training, you're just feeding it only the positives, right? I feed it both. You do feed it both. So you're feeding it, like, tons and tons of data now. **Ian Landsman:** Yeah. We have to have positive examples and negative examples okay so that it can differentiate between the two okay um and then there is a there's like a like basically like an eval set of grids that are not used for training but are fully labeled and then that is our test set that we then score the model on yeah so cloud and codex will label just hundreds of grids we hold out a certain portion of them train on the rest and then evaluate against our ground truth that it has never seen before um it's working fantastic. Wow. It's great. This is wild. This is the future of like the not leading labs models. There are going to be so many tiny on device under a megabyte, super fast, totally offline models that can just do these little things for us. It's incredible. I'm loving it. I'm having so much fun. Have **Aaron Francis:** you tested running it like on Linux and Windows and all that yet? Do you think it'll work there? It's going to be okay? Yep, **Ian Landsman:** it's going to work there. I know that it works with that data and it is compiled down in such a way that it will be embedded into solo and just called directly from rust and so i've had it you know i've said like hey put the model over there and see if it works and it works because it's just compiled to be non-platform specific man --- ### [Solo Becomes The Harness For Its Own AI](https://share.snipd.com/snip/9191bdc2-ea4b-45b1-8f2b-ab799eaedf56) 🎧 48:26 - 54:46 (06:19) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/b626cc32-6e18-4de3-802c-2f0aee88c7fc" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Solo’s Oracle combines a language model with canonical application operations, letting users manage projects, agents, and tasks through chat. - Documentation search keeps answers current without requiring Solo to be recompiled whenever guidance changes. #### 💬 Quote > I am the harness now. > — Ian Landsman Ian Landsman on giving Solo’s agent control over the application #### 📚 Transcript **Ian Landsman:** right the old self-hoisting um i also on a lark was like what if what if solo had a universal agent where you could just pop open a little like a little UI chat? So not not a terminal. You just hit the little whatever it's going to be a chat bubble or whatever. And then you say, hey, can you add this project to my Solo? Hey, what processes are running? Hey, how do I, what the hell is a scratch pad? How do I do X, Y, Z? And then this universal agent can control solo the application to add projects, inspect processes, answer customer questions, like, because it has access to the documentation. I don't like that one as much. I know, that's fine. **Aaron Francis:** I don't think you're going to be able to pull that off. That seems much harder. That's a much harder thing to pull off than the classifier. The **Ian Landsman:** podcast gods have smiled upon me because **Aaron Francis:** you said that. I **Ian Landsman:** have already pulled it off. This worked out great for me. That segment could not have gone better for me. I got a great idea. You'll never be able to do it. I've done it. Hall of Fame. Hall of Fame. All right. Tell me about this. Now I'm intrigued here. Okay. So I didn't train this model. All right. This is, uh, is an open source, open weights model. Quinn, whatever. But it's **Aaron Francis:** a real model. That's what I'm talking about. Yeah. I was thinking you need a real model for this. That's not just like a **Ian Landsman:** self-trained thing. I never said it wasn't a real model. So it appears you have egg on your face. And it appears, it appears, Ian, that you have egg on **Aaron Francis:** your face. I still have a problem with this anyway. Don't even worry about it because I feel like you could just use all the models that are in there and just use Claude Haiku and be done with it or whatever, Luna. Like, the user has models that are real and you can just use them. Perhaps. **Ian Landsman:** Yeah. Now, let me tell you why you're wrong. Okay. Give it to me. This is my favorite segment of the show. Yes, this is a good one. All right, let's go. All right. So this is primarily a user experience, user onboarding adventure, we'll say. So when a user downloads solo and they just fire it up, we do our best to figure out what agents they already have. So we're doing the most that we possibly can to make the onboarding as fast as possible. However, people still ask me questions that I'm like, this seems obvious to me. Obviously, I built it so everything is obvious to me. And if you want to win, you have to, like, be empathetic and understand, okay, I maybe have more knowledge. The user is not understanding this. That is a me problem, not a them problem. Right? So, now, you can download Solo. And there's a little, I think I've called it the Oracle, which is a cool name. There's a little Oracle where you can pop open this chat and you can ask it, hey, can you add my outro.fm project to Solo? And this agent, it's interesting because this is like the first time I've built a harness, right? else I've done is meta harness which is like let me put some chrome around the lab twoies yeah now I'm just doing raw model calls and I am the one who's like oh you want to use this tool all right well I need to parse that out use the tool and give you the response back yeah so everything is coming together because of the way I've architected this thing. So I've got this layer of like internal operations that's like create scratch pad, update scratch pad, you know, complete to do. All of these discrete functions live in this middle layer of the application. Then every ingress point, HTTP, MCP, CLI, they all have their own adapters into these canonical operations, right? So again, this comes back to engineering, not just like vibe coding. Right. Now with this Oracle, it basically has access to all of those canonical operations by just writing another thin ingress point that is now the tool call ingress point so i have this quin i've got this quin whatever 3.5 i don't even really know uh model running through uh running through So it's hosted on U.S. inference. It's not hosted in China or anything. So Solo makes this API request to Cloudflare, and it basically passes along, like, here are all the tools that are at your disposal, or if you need to, like, search tools, here's how you can do it. Passes all that to Cloudflare. Quinn gives me the response and inside it embedded our tool calls. Then the Rust code actually executes those tools, gives it back to the model. So now everything that you can do via MCP, you can just do in this one little chat box. So you can be like, hey, how many to-dos do I have open in this project? Well, you've got 40 in that project. Or, hey, can you close all of the agents over there? And so this is something that some people I don't think many people have requested this, but I think many people in their core want this. But some people have requested like I need a better way to just like triage and do stuff without opening a new terminal agent. And so this is my response to that. But this is also my response to onboarding because it has access to all of the web hosted API docs. So it can like, it can call search documentation. And then the rest behind the scenes goes out to the API to read all the documentation. So that means it's alive. It's always living. I don't have to recompile solo because it's searching the web for the docs. And so you can ask the Oracle, hey, how do I spawn a new agent? What's a good orchestration flow? How do I set up XYZ? --- ### [Use Data Structures To Make Bugs Impossible](https://share.snipd.com/snip/37802c07-9964-48a0-a649-18396af42966) 🎧 01:03:49 - 01:06:41 (02:52) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/bad3a45f-d99c-45fb-95a6-709136666e28" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Use a data-structures-and-algorithms audit to replace magic strings, scattered checks, and fragile logic with explicit structures. - Ian Landsman uses the audit before committing feature work so bugs become impossible rather than merely patched. #### 💬 Quote > Make the bug impossible by introducing a named data structure or algorithm. > — Ian Landsman Ian Landsman on his DSA audit prompt #### 📚 Transcript **Ian Landsman:** So the DSA audit is data structures and algorithms. And the whole intent of the audit is look for places where we are basically using data structures, but not. So, you know, if you're like, this should be represented as a state machine, but we're not. Or this should be an enum, but we're just kind of having magic strings passed around. The whole intent of this is break down my application into subsystems, and then every subsystem look for simplifications or bugs that can't not just be fixed, be obviated. Make the bug impossible by introducing a named data structure or algorithm. And it's like, oh, well, if this was just a discriminated union, you wouldn't need all these checks here. And you're like, oh, well, that actually seems pretty nice. And so that's the intent of this giant prompt, which we'll link up in the show notes. But it's very good. Before I ever commit, like if I'm doing a feature spike before I ever commit, I run a minor DSA audit on just that new work. And it always comes back with, yeah, this would be better if this were a struct or an enum or whatever. And it just makes the possibility of bugs, like this specific type of bug, it just makes it impossible, which is fine **Aaron Francis:** yeah yeah no seems like super good it's always nice when these things uh works i gotta put it i just have it like i'm pasting it in right now but i gotta put it maybe i'll make it a sky do you have it just set up as a skill or do you just paste it i have it set up **Ian Landsman:** as a solo prompt template right as a prompt template yeah so so **Aaron Francis:** not a skill though so you're just using it on on demand when it. Yeah. Yeah. Um, which I mean, it is nice that way too, because like for this, like I did, this was in help spot. So it was like, I only wanted to do this on the changes I made for help spot six, which has all the new UI stuff. Like I didn't want to like just go through all help spot and like, I don't want to get into like doing 20 year history. Yeah. Like at some point, maybe we'll do that. But like this one, I just wanted to focus on the new stuff. And yeah, so it would obviously did that and that was fine. And so, yeah, so this way you can like, so I had to tweak your prompt slightly at the beginning of it, right? To like not go through the whole app. So did that. And then, yeah, it was great. So the AI, it's getting there. I also feel like this is one that feels like improved significantly with the better model. Like I feel like Astrid is probably a lot better at this than, I don't know, just something about the way it's structured. Like I've done these things before, somewhat similar. I don't know if I've done this exact, maybe I even did this prompt on Sol once, but I don't know. It's structure, it returned. It's like confidence levels, how it described each one just felt like really tight and good. You know, it's like, oh yeah, like it's, it's, it's got it. --- ### [Higher Reasoning Found The Issues Medium Missed](https://share.snipd.com/snip/71fa3b9e-1cd7-4b3c-a158-5fbec2fe608d) 🎧 01:06:44 - 01:07:33 (00:48) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/6d4329d9-640e-4cbf-a27c-9055717f2561" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Aaron Francis found Astra dramatically more capable after moving from medium to high reasoning. - The higher setting connected impacts across the codebase instead of merely identifying isolated issues. #### 💬 Quote > With the high it does that, and I'm like, yes, I want that. > — Aaron Francis Aaron Francis on increasing Astra’s reasoning level #### 📚 Transcript **Aaron Francis:** I did tweet this out that I moved from the astro medium to high and that seems like a market improvement in everything. Like I was trying to do the medium cause the SIBO guy said do the medium and the medium is much better. I don't know. I feel like there's a big difference between medium and high and high is just like significantly better. **Ian Landsman:** I agree. I agree. **Aaron Francis:** Is high your baseline? **Ian Landsman:** Yeah. I just want the best, man. Yeah. **Aaron Francis:** Like I wanted to catch things. Like I wanted to find issues. Like I don't know. With medium, it was never being like oh and then like I saw over here that relates to this like we should fix that too or we should do this thing or whatever and like with the high it does that and I'm like yes I want that I want it to know that like to look around and be like oh this impacts these other areas we should do that over there like yeah that's what I want like I want you in there looking around so make it better --- ### [A Zoox Ride Felt Like The Future](https://share.snipd.com/snip/707f492e-13c6-4f23-968c-cba45030f920) 🎧 01:07:40 - 01:09:58 (02:18) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/fbbbdf94-741d-4f87-adc8-77cc54d6f521" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ian Landsman took a driverless Zoox from the Las Vegas airport and worked alone inside its four-seat, train-like cabin. - The seamless ride convinced him autonomous vehicles could carry most driving miles within a decade. #### 💬 Quote > I walked out of the Vegas airport and [...] I'm not taking an Uber. > — Ian Landsman Ian Landsman on choosing a Zoox over taxis and Uber #### 📚 Transcript **Ian Landsman:** I took a Zoox for the first time. Oh, you did take a **Aaron Francis:** Zoox. That's right. **Ian Landsman:** I took a Zoox from the airport to the resort, the casino. The Zoox is a four-passenger vehicle with no steering wheel no driver front seat back seat it's just like a miniature train car but it's better than trains because i still have my personal autonomy as an american which is nice um so i took a zooks from the airport and it's awesome was **Aaron Francis:** it good it **Ian Landsman:** was great so it's two seats on either side facing each other i was in there by myself which is the ideal um and you have you know you've got your little controls for like air conditioning music whatever um and i just got my laptop out i'm just working in this little train car all by myself uh and then you pull up to the resort and the double doors just kind of like slide open and you just step out and you never say anything. It was very interesting because I was like, I walked out of the Vegas airport and you know, there's the line of taxis. And I look at the taxis and I kind of scoffed myself. I'm not taking a taxi. Come on. This isn't 1999. And then I walked downstairs and I see the Ubers and I'm like, come on, I'm not taking an Uber. This is a 2025. And so I follow the signs to a Zoox and then I get in the car that is not a car and has no driver and has no steering wheel. And I'm like, obviously this is what I'm supposed to be doing. And I do, I think it is. I think it is the future. I think in 10 years, I think the majority of miles driven will be driverless. And I think many of them will be these like tightly packed trains of Zoox-like things following each other on the highway and peeling off and coordinating with each other and driving in **Aaron Francis:** tunnels. No way, Elon. You're way off. You're way off on that, Elon. A hundred percent. Way off on that. A hundred percent. I think there'll be more than now, of course. But no. We're building coal plants and stuff, man. We're going the opposite direction. I don't care. The electricity **Ian Landsman:** has to come from somewhere. I don't have a problem with coal. Coal's fine. Coal's terrible. But in terms of self-driving, I give it 10 years before the majority of miles are self-driven. ---