WEBVTT 00:00:11.000 --> 00:01:09.000 I'm Dan Gerlanc and welcome to Agents and Engineers, the podcast for engineers building and building with agents. Today I have Greg Sedetsky with us. He is currently the founding engineer at Revise Robotics. Prior to that, he was a fractional CTO with 10X Management for 11 years, where I was also a client. He before that was founder, CTO, and CEO of at least three other startups, including a web mapping startup that was acquired by Apple. During Hurricane Harvey, he created a map that helped rescue 1700 people. And he was awarded by both the US Coast Guard and the Canadian government for that work. Well, welcome Greg. Great to have you on today. 00:01:09.000 --> 00:01:12.000 Welcome. Thank you so much. 00:01:12.000 --> 00:01:20.000 So to start, I guess tell us a little bit about what you're working on today at Revise Robotics. 00:01:20.000 --> 00:01:28.000 Yeah, super fun. Thank you for the intro. Yeah, very kind. I was kinda thinking like, you know what, it's kinda nice to start a day with somebody just listing off and it was like the greatest hits. 00:01:28.000 --> 00:01:31.000 Well, you said impressive background, so 00:01:31.000 --> 00:01:37.000 No, no, no. You you're very kind. yeah. you have to tell everybody about the bourbon company started. Okay. I haven't started this one. That's a dream. 00:01:37.000 --> 00:01:40.000 yeah. 00:01:40.000 --> 00:01:55.000 yeah, so yeah, Revise is super, super fun. startup that I joined quite recently here in New York City. The short of it is that there's a ton of laptops that get kind of discarded after they're used in corporate, you know, the corporate world, right? So huge company, 00:01:55.000 --> 00:01:57.000 Yep. Yep. 00:01:57.000 --> 00:02:12.000 thousands of employees. And those laptops, once they kind of are deemed like not good enough or they're being replaced after a few years, t strangely or I would say shockingly enough, the best thing for them usually ends up in crushing them, like literally just like making little pellets 00:02:12.000 --> 00:02:14.000 So wow. 00:02:14.000 --> 00:02:17.000 and extracting, you know, like a tiny little bit of gold maybe out of them or silver or whatever, copper, 00:02:17.000 --> 00:02:19.000 Yeah. Yep. 00:02:19.000 --> 00:02:22.000 and then like basically like mining them, which is this 00:02:22.000 --> 00:02:55.000 Yes, it is. 00:02:22.000 --> 00:02:55.000 absurd thing because these laptops are perfectly, perfectly still good. So what we do We kind source these laptops and then we use robots to fix them up. I can go into what that entails. basically we make them refurbished. If you've bought a refurbished computer, you know kind of what that is. It's like as good as new. and then and then that's it. And then they're donated right now. So it's very nice. You know, they're donated to people who either can't afford laptops or there's like lots of nonprofits that are like in this space and they're trying to connect people who basically their lives would be change by having a laptop, which you can imagine 00:02:55.000 --> 00:02:58.000 Yeah. 00:02:55.000 --> 00:02:58.000 not having a laptop and having a laptop, that's a really big difference. we're not talking about like up gr sorry. Yeah, right. It's not it's not 00:02:58.000 --> 00:03:02.000 Yeah, no, a hundred percent. Yeah, a hundred percent. Yeah. 00:03:02.000 --> 00:03:12.000 like, like is this the the latest and no it's not it's like no laptop. like imagine, you know, applying to anything and and reading information. So so it's very, very nice. Yeah. The it's very motivating. Mm-hmm. No. It it can that's right. 00:03:12.000 --> 00:03:20.000 Right. It's not like can this run Crysis it's like can I use Chrome to get my email. 00:03:20.000 --> 00:03:55.000 Yeah, can I go on a Google form? You know, I I'm even like, you know, telling my girlfriend it's like you know, she was like, we embed a Google form like on a website we're doing. And I was like, you can't like the mobile like like a mobile Google form form is broken enough. Anyway, laptops are good. and yeah, so it's very, very nice. But one of the sort like one of the many twists that are kind of interesting about what we're doing is that there's a big robotic component because this kind of work right now is done by humans who basically stare at a Windows installation screen. and they just stare at that on like multiple laps, which is a crazy insane thing that we are forcing humans to do. 00:03:55.000 --> 00:03:57.000 like watching the installation process happen? Okay. Yeah. Yes. 00:03:57.000 --> 00:04:54.000 Watching yeah, as interesting as watching grass grow, really. And that's on one side. And then the other side, which is really interesting and kind of like novel and wouldn't have been possible a few years ago, is that we do use AI. But the AI, one of the things that the AI does is that it makes decisions as to what to do next. So it is basically watching the screen with a webcam. And then it's like, okay, like the BIOS menu has this like option that I've never seen, but no problem. Like I know my goal is, you know, to like whatever, wipe the hard drive. And so let me go like arrow down three times, which I've never done. I mean the AI doesn't have these thoughts. But in a sense, we're you know, we're like piggybacking on it's like kind of like the general intelligence of AI, especially with vision models. Which again a few years ago, if you had to write this, you would be like, I can it can only work if this it's the exact same thing, if there's one deviation like Like, you know, obviously you couldn't like let it just figure it out and now you can, so that's a big change. Yep. 00:04:54.000 --> 00:05:01.000 Right. Like how many different models would you need to at each step to get to the next step there? Yeah. Yeah. 00:05:01.000 --> 00:05:31.000 Yeah, and exactly, and how many states a laptop can be in. And you and we're really talking about, you know, laptops that maybe, you know, the battery is like for just faulty enough that it will boot, but then kind of like mid operation shut down, but the AI can still like figure that out or try another strategies, or sometimes it has to press F four, but like more quickly, and it like will direct the robot to like we have like our video on our website is very like you kind of see the robot like just jam F four and that's what it does. so it's it's it's kind of yeah, it's kind of wild. 00:05:31.000 --> 00:05:35.000 So how did had you worked in robotics before this? How did you get into doing this? Okay. 00:05:35.000 --> 00:05:47.000 No, but they knew that. Yeah. They knew that before hiring me. I did not hide my non robotic background. I did have a very old GitHub repo where I listed all the robotic arms that one day I wish to have. And this was my claim to like I've been thinking about it for a while, which is completely true. 00:05:47.000 --> 00:05:52.000 Okay. All right. 00:05:52.000 --> 00:06:16.000 But yeah, all all my like little smart like I have like little art experiments and things that I release on the web. that was definitely always like one day I'll get to and now it's like it's not just one day, it's every day. And it's not just like a little bit like we have four and they can harm us. and of course way less than like others that are like we'll just like rip your body in two if you just pass a negative instead of a positive. Ours are more like yeah, like these arms are 00:06:16.000 --> 00:06:19.000 The robotic arms. Okay. 00:06:19.000 --> 00:06:39.000 just like it's pretty much like a you know, whatever like you see on a construction site, just You control it with numbers and code that you send it and it's absolut it absolutely doesn't care if it moves left or right, it's just a motor. But the forks and the torques and everything is just like it's it's it it's not funny and in a sense you be yeah, whatever, you become aware of the Yeah. Yeah, very careful, yeah. Which is fun 'cause 00:06:39.000 --> 00:06:44.000 Right. You don't you gotta be careful. Yeah. 00:06:44.000 --> 00:06:48.000 it adds a little bit of spice to our the the software development world where my god I 00:06:48.000 --> 00:06:49.000 Right. 00:06:49.000 --> 00:06:54.000 have to, you know, no, I have to reinstall NPM or something. No, this is like you're going to the hospital. 00:06:54.000 --> 00:06:59.000 Right. You have you have to make sure it actually doesn't go back to the old way and smash the laptop. Yeah. 00:06:59.000 --> 00:07:27.000 Yes, yeah, yeah. I I can i e ev even hurting the laptop in the sense I'm like, you know, it's still much better than hurting one of my colleagues. so we try to get away for that, but knock on wood it hasn't happened, but it there's close calls and then we're like, all right, how do we We try to learn we're try learn from the close calls to our yeah, our our exact our post mortems are hopefully not mortems. Haha. So it's but but that's the spirit, yeah, that's the spirit. 00:07:27.000 --> 00:07:36.000 So how mu how much of this has changed in the last year or two years with LLMs, vision models, 00:07:36.000 --> 00:07:38.000 Mm-hmm. 00:07:38.000 --> 00:07:45.000 or how much can be done now that couldn't be done before? What like what does that transition kind of look like? 00:07:45.000 --> 00:09:11.000 Yeah. I'd say yeah, I mean there's two aspects of it. I I think on one hand there's like you know, there's like the company and I junced semi recently, so I I'm also not gonna like opine necessarily what was going on years ago. I think, yeah, from the company's point of view, as I understand kind of its history, specifically vision models, which I have to say, like in a sense that this is a very dumb thing to say, but like I think it's kind of intrigu intriguing that when we got LLMs and we're like, all right, we have this like rock from that you know, flew from another planet and when we plug electrodes into it and it talks back to us and like we love it. I don't know that I would have predicted that they were also gonna like be really good with images. I don't know if everybody was really clear, but in a sense we I kind of feel like we just got another gift out of nowhere. At some point it was like, by the way, they all now work perfectly. It does OCR better than anything. I actually did my mom found like our grandma's like recipe book And they're all like handwritten in Russian. And you know, it was like 50 pages. And I like, I mean, my brain is poisoned now. I go like of c I know exactly what I'm gonna do. I'm not even gonna not do that. I immediately take 15 pictures with my iPhones, write like a six-line script, send it to Clyde, and I'm like, I'm even like it's almost like I'm curious and I'm not curious anymore. I just assume it's gonna work. And it like transcribes Russian handwritten recipes. It does hallucinate a few, like a few fishes, become a different kind of fish. 00:09:11.000 --> 00:09:48.000 Sure. Yeah. 00:09:11.000 --> 00:09:48.000 so we w we do still like validate them one by one, we read them. But who, how, where? Like when does that so yeah, so the vision models, again, back to the company, that made a huge, huge impact because the robot can see the laptop it's working on and make take decisions. the AI LM component that I think we'll talk about a lot in terms of the coding, I mean, in the sense it's just a multiplicator of speed, right? It's like you can have a smaller team, you can go faster, you can make more projects. I of course used Claude to tell me what Claude projects I had worked on. I counted and I was like, get me for the last three to six months. It was like, actually we only have 00:09:48.000 --> 00:09:51.000 Yeah. 00:09:48.000 --> 00:09:51.000 history for 30 days. I was like, no. And it's still t 00:09:51.000 --> 00:09:53.000 In c in Claude or in GitHub? 00:09:53.000 --> 00:09:57.000 No, no, the dot Claude in the dot Claude in my in my like home directory, you know, where it kind of like ho like has like a few like kind of JSON y files. That's right, it's gonna have the session history. 00:09:57.000 --> 00:10:03.000 okay. Hm. Yeah, yeah, the s where it has the session history for Claude. Yeah. 00:10:03.000 --> 00:10:40.000 So I have like and I know it's sure whatever, like it's not it's not a numbers game, but it is, of course. I have like Twenty-seven projects that I worked in the last three days. And then of course when I read I'm like, yeah, that's a huge one. That's a huge one. That's a huge one. And I'm like, that's of course a completely new, like, you know, I th the one time that I remember working on like 20-ish projects was when I did I did my time so to speak and was at Recurse Center, which is a fantastic place. Huge shout out to Recurse. place in New York that you can also do remotely, but mostly in New York, people come for three months, work on their own project. It's amazing. recurse.com There I there it is. I plugged it. I don't get an affiliate, but at Recurse 00:10:40.000 --> 00:10:42.000 I I know lots of folks who've gone there 00:10:42.000 --> 00:10:43.000 It's 00:10:43.000 --> 00:11:09.000 and 00:10:43.000 --> 00:11:09.000 it's it's fantastic. It it's just like yeah, the the best people and the best but when you're there you feel like it's a such sp such a special moment and it kinda gets the best out of you and there I maybe worked on twenty projects at one time. Now I work on twenty projects like all the time. which yeah, it's like that's novel. I don't know what the old number is, but now it's too late, it's out of my mind. anyway, I don't know if that's 00:11:09.000 --> 00:11:20.000 And do you think I mean, I like to think that everyone is now using these, but I mean empirically I can't anecdotally I see it right on social media, but 00:11:20.000 --> 00:11:23.000 Yeah. Hmm. 00:11:23.000 --> 00:11:29.000 also there's still some people who say, Hey, I use this and it's not really doing that much for me. Do you think 00:11:29.000 --> 00:11:30.000 Yeah. 00:11:30.000 --> 00:11:42.000 there's something that separates developers who get a lot out of working with LLMs versus those who don't. 00:11:42.000 --> 00:12:22.000 i yeah, I I it's almost like I can't I almost can't believe that quote unquote people are still saying that because it felt like that's what people were saying six months ago and maybe there was more uncertainty or like they had a better point and now unfortunately their position becomes I think a little bit more untenable. Like we can talk about the harms of AI and I think those are still like in the air and we're probably all like coughing up like bits of plastic because like The AI pollution is gonna like also maybe kill us faster than like whatever. That's still true. I say the following. I I have a little bit of a hard time thinking about this in the sense of like first of all, like who is everyone? Are we talking about developers or like a a a bigger swath of the population? Developers, okay, okay. 00:12:22.000 --> 00:12:24.000 Developers, I'd say. Yeah. 00:12:24.000 --> 00:13:20.000 Well, okay. But even though you said I was hoping you wouldn't. I'll just say one quick thing that was interesting is that I visited that that many names, but they're very nice people and I love them. But I visited two friends and I would say generally they're definitely they've they've never develop you know s you know said that they were developers or anything like that. but they were always interested in I mean anyone's interested in making having my mom loves making eye images just because you know like she types words and I guess something. So these people were like, I want to make whatever, apps, things. I think what was interesting seeing how they work was that They're very lost as to like like they really wouldn't know how to improve the system. They wouldn't understand what was wrong about any of the decisions. And in one specific case, what I witnessed was well, first of all, the person has like obviously like really great ideas. Like obviously the tools have nothing to do with they have a ton of really good ideas and and thoughts and like they should be empowered. So that's absolutely true. 00:13:20.000 --> 00:13:30.000 Yeah, and I think sometimes almost better and sometimes because you're coming at it from what's my problem versus how do I engineer this? 00:13:30.000 --> 00:13:37.000 Yes, exactly. Yeah, they did not care. They they cared for the thing to like exist and then hopefully work and then hopefully kind of like be okay. 00:13:37.000 --> 00:13:56.000 Yeah. 00:13:37.000 --> 00:13:56.000 But they would I think were even very open to the thing like half working because they thought, well if I at least if I get it I can fix it. So the big thing that I saw there that I found really interesting was that they were it really felt like seeing maybe like whatever ourselves or myself like two, three years ago, they were going to claude.ai with Claude like claude.com 00:13:56.000 --> 00:13:57.000 Yeah. 00:13:57.000 --> 00:15:01.000 They were talking to it and then they would just like accept everything it would say, which which is fine. And but every single operation was them copying something or going somewhere and clicking. So in a s in a funny way, they were the AI's tool. like, you know, in this like MCP relationship, they were the tool and the AI was just like telling them what to do. And then it so all the tales that they would tell me was like like they would show me a thing that was deployed on like six different servers. They of course didn't understand why, how, like they would just they follow yeah, one thing is on Netlify there's a Cloudflare worker, there's an a Why d don't ask why. It did work. It it got like them a weird URL and it was always like, this took me six hours. And I was like, I why did it take you six hours? This is like a fifteen line flask script that does one thing with a web UI. And the reason, of course, is that they were actually the reason why it took six hours. So i in one example, which again I kind of find it it's a more pure example because I wasn't dealing with a developer who was like didn't like the goal language or something, like nothing against go, but just like I was just like Enabling a person and so as as somebody said, I really just built a DIY version of lovable, which is also kind of a funny thing. I should have just maybe given them lovable. But 00:15:01.000 --> 00:15:03.000 Yeah. Right. 00:15:03.000 --> 00:15:20.000 I did basically be like, okay, go away. I'll write a markdown file that and this is like I think pretty much like really the only rule that I kind of just am annoyed that it still like doesn't use by default. It's like just use UV, don't do virtual environments, which I love, but we've all moved on. 00:15:20.000 --> 00:15:34.000 I wrote a tool that actually you use it as a Claude hook and it lets you have rewrites where it will block things like Python three and rewrite it as uv run. 00:15:34.000 --> 00:15:38.000 Wait, okay, what's a hook and what's a rerun? I've never heard of those things. 00:15:38.000 --> 00:15:51.000 So in Claude basically you can have hooks at various steps. So one can be any time you're running a bash script or running bat the bash tool. And 00:15:51.000 --> 00:15:52.000 that's interesting. 00:15:52.000 --> 00:15:59.000 so initially I wrote this, it's called MMI or Mother May I. 00:15:59.000 --> 00:16:01.000 I saw that. Yes, yes, yes. Okay. 00:16:01.000 --> 00:16:06.000 And one of so one of the things it does is it lets you configure your allow list. 00:16:06.000 --> 00:16:07.000 Mm-hmm. 00:16:07.000 --> 00:16:11.000 in a way by in a more flexible way than Claude code gives you. 00:16:11.000 --> 00:16:14.000 Mm-hmm. Mm-hmm. Mm-hmm. 00:16:14.000 --> 00:16:18.000 or simpler, like you don't need to understand globbing, things like that. 00:16:18.000 --> 00:16:21.000 Mm. Mm-hmm. 00:16:21.000 --> 00:16:26.000 You can just be like, hey, allow me to run git and any git subcommands. 00:16:26.000 --> 00:16:28.000 Yep, yep, yep, yep, yep. 00:16:28.000 --> 00:16:37.000 And then the other part I added later was to I had the same problem where I was tired of it running Python or Python 3. And 00:16:37.000 --> 00:16:38.000 Mm-hmm. Mm-hmm. 00:16:38.000 --> 00:16:40.000 so basically you can 00:16:40.000 --> 00:16:41.000 Mm. 00:16:41.000 --> 00:16:45.000 it what you what commands you don't want it to run. 00:16:45.000 --> 00:16:46.000 Right. 00:16:46.000 --> 00:16:54.000 And there's a mode in the hooks where it it communicates it back to Claude or other harnesses have this ability where you send 00:16:54.000 --> 00:16:58.000 Mm-hmm. 00:16:54.000 --> 00:16:58.000 it a message and say, hey, don't use Python 3, always use "uv run" 00:16:58.000 --> 00:17:29.000 Hmm. But can it so so so this is like I don't know if this is what part of my opinions. I sort of like, in a sense, because you see, you know, you fill up the context window, it remembers one thing and then it forgets it, it writes a markdown, found and it reads it back. In a sense, I like almost think it's almost like a waste of time to write any markdown instructions. So to me, I'm interested, I mean, I'm interested and generally I I think people should check it out, and I know you do really good work. I'm not doubting that, but I'm less interested if this is more a markdown. If this is a right. It's like it's like like you remove the Python tool. If you can remove the Python tool 00:17:29.000 --> 00:17:34.000 It's not more markdown. Yeah. Yeah. 00:17:34.000 --> 00:17:48.000 from its tools, my God, like please. I absolutely want that. but that's definitely one of the like and I think we should talk about permissions because I have in very recently gone to the dark side. 00:17:48.000 --> 00:17:51.000 So tell tell me more about the dark side. 00:17:51.000 --> 00:17:54.000 I run Claude dangerous 00:17:54.000 --> 00:17:58.000 Okay. Do you run it in a sandbox or on a separate host? 00:17:58.000 --> 00:18:00.000 Nope. Nope. 00:18:00.000 --> 00:18:02.000 Nope, just locally. You're 00:18:02.000 --> 00:18:03.000 Yep. 00:18:03.000 --> 00:18:05.000 living dangerously. 00:18:05.000 --> 00:18:24.000 I have been enabled so the the thing is I got a taste for it in a specific environment. So right. I think the history for me of like my pr my relationship with permissions was that you get the tool and you're like, my god, yes, Claude, to ask me every single question. Yes, you and I were doing 00:18:24.000 --> 00:19:57.000 Yeah. 00:18:24.000 --> 00:19:57.000 this thing. And then you become, of course, you know, the whatever the Homer Simpson scene years ago where he like had a nuclear plant, he gets to like Woodpecker that just presses the button all the time. You just say yes to everything. I'm not gonna catch when it like base sixty four my SSH key and sends it off to like a malware site. Like it will be in like like there's no like i a a good malware would like write itself in the middle of a file that you can't even see because it shows you the first three lines and it's like I'm not gonna catch it. So I just say yes to everything and then of course two things happen. One, I'm annoyed. Two, it's a waste of time because it waits on me and I'm only saying yes. And three, it's it's actually annoying in the sense of like like you said, like it it's it will ask six variations of the same git call because it changed the name of a branch. I'm like, why? Like d the what's c what there's no security here, right? So so I'm kind of having that in the back of my mind. And then I'm like, I'm sure somebody has has a sandbox thing. And then I re read about like, it uses sandbox, but sometimes it can even b like defeat its own sandbox. You should not use its sandbox. I'm like, all right, so now we're like grading sandboxes. I have to get into sandbox line. Then I learned of Nono. I I did see your tool. I I I didn't get to it first. I got to Nono first, which is also like a sandboxing tool. And then they have a dash dash clawed like preset. I'm like, all right, this is it, Greg. Good job. Like you did your d you're not gonna go crazy. Like you're gonna use a good sandboxing tool. And and then I I don't even really know what I want. Sand well, I yes, I I I guess what I liked about Nono potentially was that it was like it should n No no. It's like N O N O, yeah. So 00:19:57.000 --> 00:20:02.000 What is what is this tool? I haven't used it. No no, no. Okay. Okay. 00:20:02.000 --> 00:20:14.000 The my the pitch that I really was like into was it will only read write the directory you're in, which also is a little bit what you kind of assume happens when you start Claude and it goes, do you trust this folder? I'm like, yeah, like 00:20:14.000 --> 00:20:15.000 Yeah. 00:20:15.000 --> 00:20:17.000 okay, like deal Claude, like stay 00:20:17.000 --> 00:20:18.000 Right. 00:20:18.000 --> 00:20:22.000 in this like in the sense I accepted that when I installed Claude Code, like whatever, a year ago or whatever. 00:20:22.000 --> 00:20:27.000 Yeah. 00:20:22.000 --> 00:20:27.000 But already knowing that's not really the case, because it can go, it can also peek in your downloads folder, it can also do this, and then you say yes, and then you're like, So 00:20:27.000 --> 00:20:28.000 Red, yeah. 00:20:28.000 --> 00:20:50.000 the first was kinda bullshit. So So no no is like, no, we will only that fuller. I'm like, that seems like a good deal. And then in terms of networking, which again, like, really what I don't want to happen is two things deleting files on my computer and extracting my SSH keys, right? Which should be encrypted, by the way, which is a whole other conversation. And maybe like installing a keyboard logger, I guess. So, you know 00:20:50.000 --> 00:20:56.000 Yeah. Like deleting writing all zeros to your hard drive. That kind of stuff. Yeah. All all sorts of fun, yeah. 00:20:56.000 --> 00:21:05.000 Yeah, guess yeah, writing zeros, I didn't think of that. Yeah but but like Yeah, but it's it's like will it happen? Anyway, so so then no no promises. 00:21:05.000 --> 00:21:09.000 I mean some it would probably be unlikely to happen unless you were doing some task that 00:21:09.000 --> 00:21:12.000 But but unless you ride the first LM virus. 00:21:12.000 --> 00:21:16.000 yeah, right. I mean if you had something injected into your context that like sure, like you downloaded 00:21:16.000 --> 00:21:19.000 Then no problem. Then no pro 00:21:19.000 --> 00:21:27.000 a website that then tells it to do this, sure. That's I think the risk. But working on your own project probably not. 00:21:27.000 --> 00:21:34.000 like in my right exact. I I I mean it makes mistakes, it does sometimes go like, let me delete this file 'cause it takes up space and then 00:21:34.000 --> 00:21:35.000 Yeah, yeah. 00:21:35.000 --> 00:21:41.000 I guess if I had clicked yes, yes, yes, yes, and then asked me to do and I would well know, I'd be like, Whew it it's good that we're anyway. So 00:21:41.000 --> 00:21:42.000 Yeah. 00:21:42.000 --> 00:21:50.000 so I install Nono basically it stops working. Because of course Claude Code is like releasing probably like fifty thousand updates a day And they just changed 00:21:50.000 --> 00:21:53.000 Yeah. 00:21:50.000 --> 00:21:53.000 one path to one thing. The setting broke, the cookie broke, the login 00:21:53.000 --> 00:21:54.000 Yeah. 00:21:54.000 --> 00:22:08.000 broke. And I'm like, okay, so now I'm annoyed at a sandboxing tool that was supposed to run the other tool that was supposed to do the thing. And it's just like, how many layers up am I gonna go? So I'm like a little bit like, I think I need to go the other direction. I keep that 00:22:08.000 --> 00:22:12.000 Yeah. 00:22:08.000 --> 00:22:12.000 in mind. Then I work on a hardware project. And this is one of the kind of things I also want to say today, 00:22:12.000 --> 00:22:13.000 Mm-hmm. 00:22:13.000 --> 00:22:18.000 is that hardware with Claude is like so fun. the one thing that I did is that 00:22:18.000 --> 00:22:19.000 tell me more. Tell me more about that. 00:22:19.000 --> 00:22:22.000 Well, right. So so I ha I have a few hardware things that I did. 00:22:22.000 --> 00:22:23.000 Okay. 00:22:23.000 --> 00:22:33.000 this one, perfect example. This is like this is school textbook, school book, whatever. So my colleagues w have like a lot of Arduinos, they a lot of things. Basically, there's a lot 00:22:33.000 --> 00:22:43.000 Mm-hmm. 00:22:33.000 --> 00:22:43.000 of like copy pasting, and we have like homegrown libraries. That's all very normal. And I kind of go, Hey, I don't have an Arduino library in the you know, like there's like a the official libraries, whatever, or just like libraries 00:22:43.000 --> 00:22:55.000 Yeah. 00:22:43.000 --> 00:22:55.000 that you're like kind of like NPM install. I go, let's make a library with this code that we reuse all the time, which basically is like RPC. It's just a way to call in the Arduino, ask it to do a thing, and then the Arduino responds with something. Everybody like rewrites that. We use JSON sue us. That's right. That that's 00:22:55.000 --> 00:22:59.000 Yep. The G the G R P C of Arduino. 00:22:59.000 --> 00:23:03.000 right. JSON a little bit of JSON never hurt nobody. 00:23:03.000 --> 00:23:19.000 Yeah. 00:23:03.000 --> 00:23:19.000 It works, it's easy to talk to, serial, nothing crazy. And I'm like, let's let's make a fun library so that it's fun to write on, you know, remove the boilerplate, et cetera. But then I go, Let's actually get a taste. And that was my that was my first hit. That was the free drug. The free opium was that. I get a Raspberry 00:23:19.000 --> 00:23:22.000 Well the 00:23:19.000 --> 00:23:22.000 Pi. So I'm I'm doing it super responsibly. I get a Raspberry Pi. This is two and a half weeks ago maximum. Maybe three. Maybe three. I c I could go see the root. 00:23:22.000 --> 00:23:30.000 So when when was when was this? Okay. Very recent. 00:23:30.000 --> 00:23:33.000 extremely recent. This is why I'm like very high right now on this. So 00:23:33.000 --> 00:23:34.000 Ha ha. 00:23:34.000 --> 00:24:05.000 I get a Raspberry Pi, you know, log into it, install Claude code, connect to my account. There's nothing on the device, so it's great. then I take a real Arduino. I connect it to the Raspberry Pi and then I take some resistors. Then I will admit to it, even though it's a podcast and it will be forever. He teaches me Kirchhoff's laws and then tells me this is an incredible, he's he's amazing. Antonio, I everybody's amazing company, but Antonio, big shout out. He goes, I can't believe you got so far without knowing this. And that was like the funniest thing ever. Because like I've been doing electronics for so long, but I was like still 00:24:05.000 --> 00:24:07.000 Yeah. 00:24:07.000 --> 00:24:15.000 voltage dividing, really? And he was like, No, no, the the math about this is like 19th century. So He very patiently he's an amazing you know, teacher. very patiently walks me through the resistor stuff. 00:24:15.000 --> 00:24:21.000 He what one of the one of the founders of the company or okay. Okay, cool. 00:24:21.000 --> 00:24:30.000 And so we so basically I put the right resistors after like an hour. Again, Simpsons and the Simpsons at one point Bart is explaining something to Homer, and then he at the end he's like using puppets and Homer still doesn't get it. That's how it felt. 00:24:30.000 --> 00:24:31.000 Ha ha ha. 00:24:31.000 --> 00:25:05.000 But I get it, I get it. And so I put the right resistors, and basically what I what it does is that the Arduino forget it LMs, forget anything. If you just ask it what is the voltage at this specific pin, it will always, because of Kirchhoff because of physics, because of everything. It will always say like 2.5 volts. That's just physically what it's gonna do. And then in another pin, it was zero. We put it to ground. So then my prompt on the Raspberry Pi was like, hey, here's a bunch of code. Like I pasted a bunch of like old examples. Here's a bunch of code that we reuse all the time. Write a library. Start by writing a library that like in C C for the Arduino. 00:25:05.000 --> 00:25:06.000 Mm-hmm. 00:25:06.000 --> 00:25:11.000 Burn it on the Arduino, like send that to the firmware, test it, then 00:25:11.000 --> 00:25:31.000 Yeah. 00:25:11.000 --> 00:25:31.000 when you're done. Write a Python library that talks to the Arduino library that you just wrote, and also like make a Python wrapper so that we can call it from Python, test it with the same Arduino, and here's the two pins that you can check your work on. And then and that was of course dash dash then she was get permissions. And then I kind of turn to him and go, like let's go have a break, let's go have a beer. This will take thirty to forty-five 00:25:31.000 --> 00:25:32.000 Yeah. 00:25:32.000 --> 00:26:08.000 minutes. You know, and I'm like, because mentally that's how long it took to just say yes and to switch back. And I think it's like I turn around. I look at him, we maybe talk about one thing. It's five minutes, it's done. And I look back and I was like, what? What what just happened? And then the library is like perfect. It's tested, it added tests, it self-tested itself. And I read it through. I'm like, yeah, this is great. This is great. Obviously, like one line of feedback and it does it immediately. And then I I was really, really hooked. I was really hooked on this, like, huh, like I knew about this, but you know, and of course, 00:26:08.000 --> 00:26:13.000 Mm-hmm. 00:26:08.000 --> 00:26:13.000 this is like a classic thing. It's like The right combination of like a test harness, 00:26:13.000 --> 00:26:17.000 Yeah. 00:26:13.000 --> 00:26:17.000 some hard constraints, obviously a problem that was simple. It wasn't trying 00:26:17.000 --> 00:26:18.000 Yeah. 00:26:18.000 --> 00:26:25.000 to like it it had seen a trillion Arduino examples, a a trillion Python UV rapper serial communication 00:26:25.000 --> 00:26:40.000 Yeah. 00:26:25.000 --> 00:26:40.000 examples out there. Very much in its comfort zone. Pooped out a library, pooped out two library quite easily. And then of course I said, like to make it an official Arduino library, you have to write this weird manifest. I was like, Can you write a weird manifest? Prrr. Because again, thousands of examples, I look it 00:26:40.000 --> 00:26:48.000 Yeah. 00:26:40.000 --> 00:26:48.000 over, sounds good, and now I have an Arduino library in the Arduino. And I hope I don't get banned because I admitted that it was made by AI, but I read it. It we use it. Don't come at me, bro. 00:26:48.000 --> 00:26:54.000 Well, right, I mean you instru you gave it the parameters and 00:26:54.000 --> 00:26:57.000 Yes. 00:26:54.000 --> 00:26:57.000 made sure it did what you want it to do, right? 00:26:57.000 --> 00:27:04.000 We use it. It it's it's short, it's very readable, and it's beautiful C plus plus, I guess. I 00:27:04.000 --> 00:27:05.000 Here. 00:27:05.000 --> 00:27:15.000 admittedly don't have the C plus plus skills to like tell you whether it like use the right C out star star. It works. So that got me hooked. And then I again I have like more example of hardware things. Some gone 00:27:15.000 --> 00:27:23.000 Yeah. And how long do you think it would have taken you to build this by hand? 00:27:23.000 --> 00:27:33.000 I mean, okay, I think it's like a day, right? It's like a day of like I get the C library working, I 00:27:33.000 --> 00:27:42.000 Yeah. 00:27:33.000 --> 00:27:42.000 write I think even like what would have annoyed me, what I would have been like, what's the good C pattern? Because the library like just to get specific, right? On the Arduino side, what I want is for a person 00:27:42.000 --> 00:27:43.000 Yeah. 00:27:43.000 --> 00:27:46.000 who uses this library to not have to think about the library. I want them to just declare, 00:27:46.000 --> 00:27:47.000 Yeah. 00:27:47.000 --> 00:27:56.000 hey, I'm an Arduino that does, I don't know, like we change, you know, the the focus on a camera that it's controlling. So I want to declare set focus and then 00:27:56.000 --> 00:27:57.000 Yeah. 00:27:57.000 --> 00:28:15.000 you library expose that to the serial port to the JSON. You carry the inputs, you transform the types from text, right? To like like you do the string parsing in C correctly. And so I think it's just a lot of like like what's a good pattern for that? In Python, I can definitely write that a lot faster, like, 00:28:15.000 --> 00:28:16.000 Yeah. 00:28:16.000 --> 00:28:25.000 you know, make a class and have like, you know, like these attributes and these functions that are like all dynamic. But yeah, in C and C I haven't done it a thousand times. So 00:28:25.000 --> 00:28:39.000 Right. 00:28:25.000 --> 00:28:39.000 between my like rusty ish, okayish C, but that can still get things wrong, my almost unexistent C, me looking for a pattern, looking around for a library, that would have probably taken like some hours, like three, four hours. Then 00:28:39.000 --> 00:28:48.000 Yeah. 00:28:39.000 --> 00:28:48.000 the Python would have worked, then I would have tested it, I would have found a weird mistake, I would have like, my god And then you know, like between like uploading to the Arduino it always takes like five minutes and then it's like 00:28:48.000 --> 00:28:49.000 Yeah. 00:28:49.000 --> 00:28:52.000 i it so yeah, it's a day. Not five 00:28:52.000 --> 00:28:55.000 Yeah. 00:28:52.000 --> 00:28:55.000 minutes. Not five minutes. There's no world where it would have taken five minutes. 00:28:55.000 --> 00:28:58.000 Yeah, s so is what is that like a hundred 00:28:58.000 --> 00:28:59.000 Right. 00:28:59.000 --> 00:29:03.000 X speed up relative to doing it by hand? 00:29:03.000 --> 00:29:08.000 Right. I mean that's that's like like how can you not call that a drug? Like it's literally like 00:29:08.000 --> 00:29:09.000 Yeah. 00:29:09.000 --> 00:29:13.000 you've walked all your life and someone was like, You want to try a bicycle? And it's like, Yeah, that's especially like 00:29:13.000 --> 00:29:14.000 Yeah. 00:29:14.000 --> 00:29:16.000 a a bicycle going downhill. Which is also maybe that that that 00:29:16.000 --> 00:29:17.000 Right. 00:29:17.000 --> 00:29:24.000 could be like a funny thing of like sometimes it's still a bike. It's just the battery's out, it's a s it's a heavy city bike. You're going uphill 00:29:24.000 --> 00:29:25.000 Yeah. 00:29:25.000 --> 00:29:34.000 and there's you know, and there's like potholes. You still like the bike. You still like the idea of pushing this bike, but maybe like walking would have been less effort. Blah blah blah. Maybe this is a bad analogy. But 00:29:34.000 --> 00:30:00.000 So well so that well that that sounds like another scenario. Is like is that like where does that case come up that you found where yeah, that's still easier than doing it by hand, but there's some potholes or it's a city bike that maybe fell in the East River. Yes. Yes. 00:30:00.000 --> 00:30:17.000 sunk cost fallacy. The one example where it still did things right, but it was more like when it got things wrong, it like kind of like put us into trouble and stuff, is I've vibe set up the networking at our 00:30:17.000 --> 00:30:30.000 Mm. 00:30:17.000 --> 00:30:30.000 workplace. And we basically went from like having like a home router that like could have like a few ports. Twell's like, all right, like let's have a switch, let's have a router, let's have, you know, a few access points. You want like DHCP. And I like maybe 00:30:30.000 --> 00:30:34.000 Yeah. 00:30:30.000 --> 00:30:34.000 this is again kind of like a pattern a little bit like I know just enough to get in trouble and I've done it manually 00:30:34.000 --> 00:30:35.000 Yeah. 00:30:35.000 --> 00:30:39.000 enough, but wouldn't call myself like a like a real networking expert. 00:30:39.000 --> 00:30:42.000 So is this like for everyone's laptops at the office or is this also the hardware? Okay. It's like also all 00:30:42.000 --> 00:30:45.000 Yes. For everything. Yeah, for everything. 00:30:45.000 --> 00:30:48.000 the machine all all the robots, everything too. Okay. 00:30:48.000 --> 00:30:56.000 Everything, right? And it's like when we got the space, there were like Ethernet jacks, but basically the previous tenant had like just cut all the cables. So one thing was hardware, which was 00:30:56.000 --> 00:31:03.000 Okay. 00:30:56.000 --> 00:31:03.000 really fun, and I've done it before. Some of it was a little bit new, but I got like a punch down tool if anybody has ever done that. Basically you get a bunch of cables. You know it's like those panels with like all the Ethernet jacks. So I punched 00:31:03.000 --> 00:31:06.000 Yeah. Yeah. 00:31:06.000 --> 00:31:08.000 them one by one. I got the right tools. I tested that was fun. That was like the very Yes. 00:31:08.000 --> 00:31:17.000 So you've got a ma so you've got things that are hardwired, you've got Wi Fi and you've got things that need a good network connection. hard wire connection. 00:31:17.000 --> 00:31:32.000 Well, right. It's like yeah, and yeah, and it's like exactly, yes, that's that's correct. And we have FIOS and but but it's more like, you know, at the end of the day, you really to speak in networking terms, like you just want like that basic firewall and you kind 00:31:32.000 --> 00:31:40.000 Yeah. 00:31:32.000 --> 00:31:40.000 of want those wireless access points to kind of work. And there's like some kind of magic when access points like use the same network name, computers and phones will mostly magically 00:31:40.000 --> 00:31:41.000 Right. 00:31:41.000 --> 00:31:43.000 kind of like transfer blah blah. But I also 00:31:43.000 --> 00:31:50.000 Right, so you so you can switch between different access points based on which one is stronger or things like that. 00:31:50.000 --> 00:31:52.000 closer, but then like also Macs and other devices will like make their own decisions 00:31:52.000 --> 00:31:59.000 Right. 00:31:52.000 --> 00:31:59.000 as to when to swap 'cause there's like whatever. So I've used and kinda like MikroTik equipment and maybe I don't 00:31:59.000 --> 00:32:07.000 Mm-hmm. 00:31:59.000 --> 00:32:07.000 think it's like masochistic, but it is a little bit like more complicated than it should be. And it but it's kind of cheaper and it's it's actually like quite powerful. 00:32:07.000 --> 00:32:16.000 Yeah. 00:32:07.000 --> 00:32:16.000 It's been very stable. I've I've kind of used it for like many, many years now, like in my home apartments and stuff. And I've I think I've used it at one or two kind of like these like s semi-corporate gigs where I kind of like help out 00:32:16.000 --> 00:32:17.000 Yeah. 00:32:17.000 --> 00:32:20.000 somebody where I'm like, your Wi Fi's terrible, let me put like a few antennas, blah, blah, blah. But a lot 00:32:20.000 --> 00:32:21.000 Mm-hmm. 00:32:21.000 --> 00:32:24.000 of it was always like, woof, like if you click enough options, 00:32:24.000 --> 00:32:25.000 Ha ha. 00:32:25.000 --> 00:32:27.000 like it gets like like I don't know. Like I don't know between 00:32:27.000 --> 00:32:28.000 Yeah. 00:32:28.000 --> 00:32:33.000 like the bridge on the thing with the like it's it's too much networking for me, even though I think I understand something, but not everything. 00:32:33.000 --> 00:32:34.000 Right. 00:32:34.000 --> 00:32:39.000 So now, New World, and this is like a even pre-dangerly skip permissions, but th that 00:32:39.000 --> 00:32:46.000 Yeah. 00:32:39.000 --> 00:32:46.000 doesn't really matter. I am like, wait, like, does those do those routers have APIs? And and they sure 00:32:46.000 --> 00:32:47.000 Yeah. 00:32:47.000 --> 00:32:58.000 do. And they have like REST APIs and they have like other kinds of APIs. You can like SSH into them. So I kind of go, hey Claude, you you must have seen a bunch of MikroTik setups in your day. Do you think you can 00:32:58.000 --> 00:32:59.000 Yeah. 00:32:59.000 --> 00:33:04.000 help me with this? Like I have a switch and I have a router. And it basically like it's very text-based. It's the config of a 00:33:04.000 --> 00:33:10.000 Yeah. 00:33:04.000 --> 00:33:10.000 router are like it's literally just a list of text with like slash DHCP, nah da. One very quick example. 00:33:10.000 --> 00:33:11.000 Makes sense. Yeah. 00:33:11.000 --> 00:33:12.000 Yeah, right. It's is right. It w what would it be? An image, right? maybe 00:33:12.000 --> 00:33:15.000 Sounds all right. Yeah. 00:33:15.000 --> 00:33:30.000 one thing that will also kind of like cement like kind of what I was doing, or like just another example, is that when a device joins using DHCP, it's given like an IP address, but that IP address like it's not baked. It's just like it's kinda temporary, you might lose it, like there's 00:33:30.000 --> 00:33:35.000 Yeah. 00:33:30.000 --> 00:33:35.000 no guarantee. But you can tell it, hey, this device with this MAC address forever it should be this IP address, right? So we had a list like that before, and I was kind of bring Yes. 00:33:35.000 --> 00:33:42.000 Right. But in ish yeah. Like if you like your printer, right, you probably always want at the same IP address. Yeah. Things like that. Yeah. Yeah. 00:33:42.000 --> 00:33:49.000 The exactly. And especially like for us, like our robots. Like we run our robot so you know, and other devices because we want to address them with a stable IP. 00:33:49.000 --> 00:33:52.000 Yeah. 00:33:49.000 --> 00:33:52.000 So and then of course I can hear even from this record 00:33:52.000 --> 00:33:57.000 Right. I mean you can do fancy things like give it a name and do DNS resolution, r right, but that's more work. Yeah, yeah. 00:33:57.000 --> 00:34:06.000 Yes, you can do tail scale. Yeah, you can do telescale. There there's like it's not it's not like this is the only way, but it's got it's almost like also a like an easy, quick solution to just like bake those IPs. So you know, I was also like migrating that Yes, yes, yes, exactly. Which and and then it could 00:34:06.000 --> 00:34:14.000 Yeah. Yeah. But first adding a lot of extra complexity of running separate DNS, you know. 00:34:14.000 --> 00:34:22.000 Yes, yes, yes, exactly. And then you everybody has to use a DNS, but yeah. So I was kind of bringing all of those configs, you know, into the micro tick, but I let Claude 00:34:22.000 --> 00:34:28.000 Yeah. 00:34:22.000 --> 00:34:28.000 like talk to the router directly. And Roughly speaking, I would say 00:34:28.000 --> 00:34:49.000 Mm-hmm. 00:34:28.000 --> 00:34:49.000 it's still a success. I'm still happy that I did that. I do think that there is a bit of like it's true that I don't fully know whether it might have opened the, you know, firewall incorrectly of wanting to like optimize some weird thing. This is also what's kinda scary with networking is that no one will tell you that you have like a huge, you know, like huge 00:34:49.000 --> 00:34:50.000 Yeah. 00:34:50.000 --> 00:35:01.000 vulnerability in the sense it's just there. and there I maybe like I'll also say like my in my mind I have my like Well I I have the French expression, it's like the the elders, the the what do you call like the committee of elders is when 00:35:01.000 --> 00:35:04.000 Mm-hmm. Yeah. 00:35:04.000 --> 00:35:28.000 when I'm in like grave doubt about what I'm doing, I actually bring it to the three elders, which is Claude Gemini and Chat GPT, and I sort of get them to agree. And this is when I'm like I'm making a bit like for instance, like if I'm doing tax stuff with AI or even just like I give it a bunch of documents, it just even for it to like give me a hint as to what to ask my accountant. I'm not I'm not yet using it purely to just make my taxes and not look at it. Maybe next 00:35:28.000 --> 00:35:52.000 Right. 00:35:28.000 --> 00:35:52.000 year. But even just that, I will go to like the three elders and be like, like, if you don't agree, I'm not even like I'm not even touching that. So so definitely one thing that I've done is that I've copied paste this micro tick config into like different, you know, different AI systems and be like, please like find a flaw. Like you are a red hat. Like, how would you hack this? And once they all go, yeah, no, this is this is how you do it. There's no hole. I'm kind of like, There's probably no hole. 00:35:52.000 --> 00:35:57.000 So so you're using basically like a different agent adversarial verification. 00:35:57.000 --> 00:36:08.000 Yes. Yeah, at the material stuff is like I think I think there's a th it's like I almost want to say maybe this is one of the like almost most crucial thing. It's this like 00:36:08.000 --> 00:36:15.000 Yeah. 00:36:08.000 --> 00:36:15.000 and you know, I can see it in myself. It's like it's something about, you know, the difference between just having these thoughts and then when you write down and then you start kind of reading 00:36:15.000 --> 00:36:22.000 Yeah. 00:36:15.000 --> 00:36:22.000 yourself or like even the rubber duck phenomenon where like even hearing yourself say say a thing, you can like react to it, which is sometimes different than if you just have those 00:36:22.000 --> 00:36:23.000 Yeah. 00:36:23.000 --> 00:36:44.000 thoughts bouncing around. And in a certain sense, like More agents all more looking at one problem. So yeah, so my colleague and friend, Antoine, that I also told you like, you know, like worked a lot w works a lot right now with like AI agents, rewriting like a giant library. Even though his last email was less encouraging than usual, he was like, I'm not sure about the quality. That was like an interesting I was like, my god, 00:36:44.000 --> 00:36:50.000 Mm. 00:36:44.000 --> 00:36:50.000 what do you mean? What do mean? But he's has he's had this like gaggle of AI agents almost 00:36:50.000 --> 00:36:51.000 Yeah. 00:36:51.000 --> 00:36:54.000 like all having different like maybe like kind of like schizophrenic personalities. 00:36:54.000 --> 00:36:55.000 Yeah. 00:36:55.000 --> 00:37:06.000 all like just like arguing against each other until there is sort of a consensus. And of course, that to me that still feels like well, if you got a bunch of six year olds arguing, 00:37:06.000 --> 00:37:07.000 Yeah. 00:37:07.000 --> 00:37:17.000 the consensus will still be the consensus of a six year old mind. No one will be like, you know what, we should be d reasonable and bury the hatchet and you know, go have a coffee with the spirit. It will still be like we should put glue on their seat and call them a bad 00:37:17.000 --> 00:37:22.000 Sure. Yes. Yeah. 00:37:22.000 --> 00:37:53.000 you know, a bad whatever, and then we should still like pull their hair or something. So I think that's still that still is the case, right? Like these agents still they're kind of like some limit it it's it's kind of like a dice, like it doesn't matter if you throw it a hundred times or but the probabilities are the same of them getting to any number. And in the same sense, you know, if it's wrong twenty percent of time, I'm not a stats person, don't come at me, but there's something that feels like more eighty percent just gives you eighty percent or less than eighty percent. It doesn't increase the probability of it not make Right, or worse. Yeah, the Right. I know it should, but it I also don't want to be like yeah, I it's like 00:37:53.000 --> 00:38:02.000 Probably much worse than eighty percent. So it's as a compound because those compound at each step, I guess, right? 00:38:02.000 --> 00:38:10.000 but yes, yes. It's like yeah, get a room of, you know, reasonable people is there gonna be a breakthrough? I I don't know. It's it's hard. 00:38:10.000 --> 00:38:28.000 But how'd you feel where it is now? Is it 80% or like 95%? Because I felt like pre-Opus 4.5 and into November that we were kind of at 80% a lot of times. And 00:38:28.000 --> 00:38:30.000 Uhhuh, uhhuh. Uhhuh. 00:38:30.000 --> 00:38:36.000 it was just very hard to I didn't really trust Claude Code to be editing my code because 00:38:36.000 --> 00:38:37.000 Yeah, yeah, yeah, yeah, yeah. 00:38:37.000 --> 00:38:49.000 it was too 80%. but now I feel like maybe we're I mean a lot of times getting closer to like ninety five percent, maybe even 00:38:49.000 --> 00:38:51.000 Yeah. 00:38:51.000 --> 00:38:53.000 with and with repeated iterations, because code 00:38:53.000 --> 00:38:54.000 Yes, yes. 00:38:54.000 --> 00:38:57.000 is something you make up, right? It's a system that you've designed so you can kind 00:38:57.000 --> 00:38:59.000 Mm-hmm. Mm-hmm. 00:38:59.000 --> 00:39:01.000 of endlessly verify that it works how you think it does. you 00:39:01.000 --> 00:39:04.000 Mm-hmm. Mm-hmm. 00:39:04.000 --> 00:39:08.000 can at least iterate at 95 or greater percent towards a better 00:39:08.000 --> 00:39:09.000 Mm-hmm. 00:39:09.000 --> 00:39:11.000 percentage of being right. 00:39:11.000 --> 00:39:25.000 Mm-hmm. Mm-hmm. I I think I think that the number you know, I like I'm I I'm able to make averages, but I almost kind of like almost, you know, like like like air bars or whatever. Like I want 00:39:25.000 --> 00:39:26.000 Yeah. 00:39:26.000 --> 00:39:33.000 to look at the you know, the shape of these errors. I think that I have a lot of one hundred percent. A ton, right? 00:39:33.000 --> 00:39:55.000 Yeah. 00:39:33.000 --> 00:39:55.000 I look at my list of projects again, these like twenty projects, like many of them were like entire full on And and it's like i and are we defining it that did I have to correct it as a like to me that doesn't feel like me giving it more little bit instructions felt more like, well I drove the car, the car wasn't wrong because it went to the left, like I was driving it. Like it's fine, right? Like 00:39:55.000 --> 00:39:56.000 Right. 00:39:56.000 --> 00:40:26.000 I got to the thing, the car didn't explode, the car is one hundred percent successfully a car or useful for like transport. It's really when it gets things wrong, and of course, like we we I mean, we're not just like learning to learn We know these patterns and I guess we're like just getting deeper into them to try to see whether they're solvable or why or how, but between it being convinced it's okay and telling me, my god, you're so right, and then getting in the loop and like what is the nature of the loops when it is both wrong but thinks it's right, or it thinks it's on the right 00:40:26.000 --> 00:40:43.000 Yeah. 00:40:26.000 --> 00:40:43.000 path, but will never be I think that's what's interesting. We're kind of digging into like this is why I kind of I want to zone in on the like I think there's a lot of a hundred percents and the not a hundred percents to me are Like and obviously like all labs are working on this, so it's like, my god, 00:40:43.000 --> 00:40:54.000 Right. 00:40:43.000 --> 00:40:54.000 Greg was the first one to figure But I yeah, I even want to brought up th th this I don't know if it's relevant, but I I remembered this book about you know, the field guide to human error investigations. And it 00:40:54.000 --> 00:41:07.000 Uh-huh. 00:40:54.000 --> 00:41:07.000 it the the the the book the book the book the one of the points of the book is that with hindsight, it always feels like it was a human error. Because it always feels like, well, yeah, the you press the button and the plane, you know, whatever, crash. v versus like that's that's 00:41:07.000 --> 00:41:10.000 Yeah, you guessed. Sure, you're right. 00:41:10.000 --> 00:41:13.000 Yeah, but that's a that's the worst way to look at it. I mean, that's the the book argues that that's 00:41:13.000 --> 00:41:21.000 Mm-hmm. 00:41:13.000 --> 00:41:21.000 a bad way to look at at a thing because y you have to put yourself in the the head and the circumstances of the person at every step before that, leading 00:41:21.000 --> 00:41:22.000 Yeah. 00:41:22.000 --> 00:41:27.000 them that path. Why did they take all those decisions? And also why did the system let them make those decisions? And in a sense, 00:41:27.000 --> 00:41:37.000 Yes. 00:41:27.000 --> 00:41:37.000 that's exactly maybe like what we're talking about, right? It's like now if you told me if you told me, hey, I have a more reliable claw, it's a better percentage, it's a new 00:41:37.000 --> 00:41:43.000 Yeah. 00:41:37.000 --> 00:41:43.000 model. But you know, for some technical reason you can't question, it can't call tests. It can't 00:41:43.000 --> 00:41:44.000 Yeah. 00:41:44.000 --> 00:41:49.000 like it can't call tools. I will not I will be like, get get it away from me. I am way 00:41:49.000 --> 00:41:50.000 Yeah. 00:41:50.000 --> 00:41:55.000 more into new, interesting kind of tooling. Which also, by the way, 00:41:55.000 --> 00:42:01.000 Yeah. 00:41:55.000 --> 00:42:01.000 I will now scream. No one effing told us that testing was gonna solve this. Like, I don't think that a year 00:42:01.000 --> 00:42:02.000 Ha ha 00:42:02.000 --> 00:42:09.000 ago, while we were AI coding Like sure, like some people like, testing could have like its own like kind of a second life, but in a sense it was like, 00:42:09.000 --> 00:42:19.000 Yeah. 00:42:09.000 --> 00:42:19.000 wait, it actually is like a f an incredible, fundamentally now structural, like like it is just how to AI code. AI coding has 00:42:19.000 --> 00:42:20.000 Right. 00:42:20.000 --> 00:42:42.000 a giant self checking and this is also what I know. Is it just like making a lot of mistakes but like self correcting it a little bit? Or is it just giving it reassurance? Or is it this is the biggest gap that I'm gonna jump, or is it Realizing the mistake that it's doing because it sees the tests fail, and then it can reflect, my god, it's true, those six tests failed. That's because 00:42:42.000 --> 00:42:48.000 Yeah. 00:42:42.000 --> 00:42:48.000 I'm going the wrong way. Let me change the strategy. And again, at the speed that it can do that, maybe that's almost 00:42:48.000 --> 00:42:49.000 Yeah. 00:42:49.000 --> 00:42:52.000 as close as how it fell to code as a human. I also 00:42:52.000 --> 00:42:53.000 Yeah. 00:42:53.000 --> 00:42:55.000 write bad C, kind of get 00:42:55.000 --> 00:42:56.000 Yeah. 00:42:56.000 --> 00:43:02.000 it to work and go, my god, like what's different? That's exactly how I used to code more slowly. I also never wrote code and one r ran it and like I I know 00:43:02.000 --> 00:43:07.000 Yeah. Mm-hmm. Right. Yeah. 00:43:07.000 --> 00:43:10.000 when that happens. I call them they're they're the special days. It's I write more than a hundred lines, 00:43:10.000 --> 00:43:13.000 Yeah. Right. Yeah. 00:43:13.000 --> 00:43:17.000 I never run them, I run them once and it immediately works. Come on, that's not every day 00:43:17.000 --> 00:43:19.000 Yeah. 00:43:17.000 --> 00:43:19.000 as a developer. That's not true. That's bullshit. Like that's not how we develop. You you d 00:43:19.000 --> 00:43:24.000 No, or right. Or it might run but not do the thing you want it to do. Yeah. 00:43:24.000 --> 00:43:29.000 What exactly, of course. And it's like we were never like, my god, like I feel so bad, I'm not No, y it's 00:43:29.000 --> 00:43:32.000 Yeah. 00:43:29.000 --> 00:43:32.000 not like Linus Torvalds when he wrote Linus's code, 00:43:32.000 --> 00:43:33.000 Yeah. 00:43:33.000 --> 00:43:46.000 run no. It's like we already have this like I mean the iteration is one thing, the testing is one thing. And of course there's still an ethereal human whatever this is, still kind of like an excited eleven year old who just learned to code, 00:43:46.000 --> 00:43:47.000 Yeah. 00:43:47.000 --> 00:43:49.000 kinda, and reaches for Python three even though we beg it not to. 00:43:49.000 --> 00:43:53.000 Right. Yeah, no, for sure. 00:43:53.000 --> 00:43:59.000 But anyway, so so so so I don't know about the percentage anymore. I'm really, really confused about the percentage. 00:43:59.000 --> 00:44:07.000 Yeah, it's hard to say what is the percentage and I'm I'm sure it varies by how well defined is your problem. Like, 00:44:07.000 --> 00:44:08.000 Yes. 00:44:08.000 --> 00:44:21.000 do you have tests that actually test the core idea of the problem and not just a specific that doesn't generalize, things like that. 00:44:21.000 --> 00:44:31.000 That's right. Yeah, yeah, it's really the that kind of generalization. I'll I'll give two very specific examples where it totally failed me. never got out of trouble, was never 00:44:31.000 --> 00:44:35.000 Mm-hmm. 00:44:31.000 --> 00:44:35.000 helpful, and brought and this is truly the like AI psychosis feeling 00:44:35.000 --> 00:44:36.000 Yeah. 00:44:36.000 --> 00:44:41.000 of like you're going crazy with it, it keeps convincing you it's gonna get it right, or it almost got it 00:44:41.000 --> 00:44:47.000 Yeah. 00:44:41.000 --> 00:44:47.000 right, or next time, and it never does. And it can also, even more fundamentally, never actually 00:44:47.000 --> 00:44:48.000 Mm-hmm. 00:44:48.000 --> 00:44:58.000 Not because like I it's like limited, but it seems like it's just like stuck at a level where it just needs one more meta level, maybe, or even the meta-level needs whatever. So the two 00:44:58.000 --> 00:44:59.000 Yeah. 00:44:59.000 --> 00:45:06.000 examples are very, very specific. So one is I there's a thing you've all heard Claude Shannon. We're all here because of him, and Claude is called because of that, which I find annoying. So Claude Shannon 00:45:06.000 --> 00:45:09.000 Yeah. Yeah. 00:45:09.000 --> 00:45:18.000 in the in the 60s, after you know inventing computers and everything and figuring out relays and everything. Basically builds this tool that is kind of meant for people to like work with relays and understand like computing 00:45:18.000 --> 00:45:20.000 Mm-hmm. 00:45:18.000 --> 00:45:20.000 and things. It's called a minivac 601. It's like 00:45:20.000 --> 00:45:21.000 Mm-hmm. 00:45:21.000 --> 00:45:32.000 a device that has six relays and a few buttons. You would think, my god, that's so silly. You can't do anything with six relays. Turns out because he's a genius and the whole thing is like really smartly done and the manuals are beautiful. You 00:45:32.000 --> 00:45:34.000 Like and what is what is a relay for 00:45:34.000 --> 00:45:48.000 Yes, of course. Relay and and it is kind of very fundamentally at a very core, kind of what computers are doing like trillions of times a second. But basically a relay is a is a is a device, it's a physical device. if you drive a car and you ever 00:45:48.000 --> 00:45:51.000 Yeah. 00:45:48.000 --> 00:45:51.000 put a blinker and you hear click, click, click, click, that's a 00:45:51.000 --> 00:45:54.000 Yeah. 00:45:51.000 --> 00:45:54.000 relay turning something on, your blinker, and then turning it off. 00:45:54.000 --> 00:45:55.000 Yeah. 00:45:55.000 --> 00:46:23.000 So a relay is a thing that you can go, hey, can you turn on a thing and can you turn it off? It's kind of like a light switch. But the thing is, once you if you have one relay, that's kind of boring. But if you have more than one relay, there are Very kind of strange, but you can learn them ways of having some relays control other relays. So as one thing turns on, it can turn on another thing that can turn the other thing off. And once you kind of build upon that, you get logic and then you kind of get computers 00:46:23.000 --> 00:46:38.000 Yeah. 00:46:23.000 --> 00:46:38.000 and you get programming and you get everything that we're doing. So so short long story short, he builds this device and you can kind of like experiment with these circuits. And again, six relays, that sounds like very simple. It's like what can you do with six light switches? Just put on a really, you know, bad light show. 00:46:38.000 --> 00:46:46.000 Right. It's like But wait. Yeah. Mm-hmm. 00:46:38.000 --> 00:46:46.000 But turns out he builds like a tic-tac-toe machine. Yeah, he builds like a thing that actually does recognizes characters like OCR and can like like pretend to be an elevator. So it's really surprising. Anyway, 00:46:46.000 --> 00:46:49.000 Mm-hmm. Right. 00:46:49.000 --> 00:47:12.000 so I don't have one. It's expensive. There's not a lot of them left. I want to build a JavaScript version of that. And it's kind of like a long project. I work on it sort of like on and off. Also very like pre-AI, post-AI, kind of like over over two years in the last two years. And the thing that I don't get, which is really my fault, is that I'm actually not I'm actually not simulating relays, which seem like they can only be on and off. 00:47:12.000 --> 00:47:13.000 Mm-hmm. 00:47:13.000 --> 00:47:15.000 That should be really easy to write in code. I'm actually 00:47:15.000 --> 00:47:16.000 Yeah. 00:47:16.000 --> 00:47:19.000 need to simulate electricity, which is harder. And it's 00:47:19.000 --> 00:47:21.000 Okay. Mm-hmm. 00:47:21.000 --> 00:47:24.000 not like we don't know how to do it. We have, you know, people do it every day, but 00:47:24.000 --> 00:47:25.000 Right. 00:47:25.000 --> 00:47:27.000 you you need to know the Kirchhoff law, throwback. 00:47:27.000 --> 00:47:28.000 Yeah. 00:47:28.000 --> 00:47:36.000 You need to understand physics, you need to know this, and you can't pretend that physics don't exist in this world because just of how relays work. So I for a long time 00:47:36.000 --> 00:47:39.000 This like what languages like Verilog and things like that have 00:47:39.000 --> 00:47:48.000 Yeah, yeah, the very exactly. Or y or or basically like if I gave you if I just like if you know somebody hands you like a circuit diagram, like 00:47:48.000 --> 00:47:59.000 Yeah. 00:47:48.000 --> 00:47:59.000 yes, you can read it, you can be like, this is this component. But if if you were to go how does this circuit work and you were like stepping it step by step, you would have to do the math that again, nineteenth century Kirchhoff, 00:47:59.000 --> 00:48:00.000 Right. 00:48:00.000 --> 00:48:04.000 you would have to go, Well, this voltage will drop by this much with this 00:48:04.000 --> 00:48:13.000 Yeah. 00:48:04.000 --> 00:48:13.000 amp, and then you multiply some numbers, this resistance, and you have to do that for every single and it that's called like circuit or node analysis. So again, we know how to do this. This is taught in all schools. I haven't 00:48:13.000 --> 00:48:22.000 Right. 00:48:13.000 --> 00:48:22.000 gone to that school. So have I gone even to any school that I have. But but it it I I'm like I'm a very beginner in this. So I spend a lot of time 00:48:22.000 --> 00:48:24.000 Something you'd need a sim essentially you need a simulator that does this. 00:48:24.000 --> 00:48:36.000 Yes. Yes. And a valid one that's actually mathematically the valid enough. So I for a long time reject that idea and kind of write my own really bad version of that with like 00:48:36.000 --> 00:48:37.000 Yeah. 00:48:37.000 --> 00:48:44.000 graphs and nodes. And you know, it works until it like it's kinda like, you know, Newton laws work until you go really fast and then it they don't work at 00:48:44.000 --> 00:48:56.000 Yeah. 00:48:44.000 --> 00:48:56.000 all. So not that I was reinventing and I I just compared myself to Newton, I really didn't, but like my stuff would work and then immediately if you put like two components in a different way, it would break and I'd be like, And you know, a little bit like an AI, like, let me just fix this problem. But that 00:48:56.000 --> 00:48:58.000 Yeah. Yeah. 00:48:58.000 --> 00:49:04.000 wasn't a problem. So I keep a lot of time going, going, going. I have one chat with Gemini where it admits we're both 00:49:04.000 --> 00:49:18.000 Yeah. 00:49:04.000 --> 00:49:18.000 going crazy, we should stop. Like, and I'm like, yeah, we're we're both going crazy and we will never succeed. And at some point, I z I zone out, I take a big walk, and I remember that I whatever took an online MIT class about like circuits. 00:49:18.000 --> 00:49:19.000 Mm-hmm. 00:49:19.000 --> 00:49:31.000 And that they had a simulator in a little JavaScript thing that worked perfectly. And you could like drag little resistors and connect them and it would tell you the voltage and everything. And I think, huh, like that wasn't AI coded at all. And that worked really well. What if I gave that 00:49:31.000 --> 00:49:33.000 Right. 00:49:33.000 --> 00:49:35.000 as the like layer that then 00:49:35.000 --> 00:49:38.000 Mm-hmm. 00:49:35.000 --> 00:49:38.000 I can build stuff on top of? So I'm just gonna tell the AI, like, we're doing 00:49:38.000 --> 00:49:39.000 Yeah. 00:49:39.000 --> 00:50:04.000 this Minivac simulator, but when you like make the circuit, don't simulate it. Give it to this other tool that was written by like a people person who had like, you know, a physics background. And then it immediately worked. And it like I had resolved the like don't solve physics in your head using AI and don't try to reinvent physics laws. And that was successful. But had I not done that, 00:50:04.000 --> 00:50:08.000 So you use that as the the kind of kind of the core engine to run this and then built on top of that. 00:50:08.000 --> 00:50:40.000 Yes. That's exactly right. Yes. And yeah, and on top of that, no problem. It can figure out components, it can place them, it can call the engine. But it still felt like I don't know, maybe it's a silly gap and everyone is laughing at me right now and pointing at fingers and taking screenshots and putting it on TikTok. This guy thought Kershop's laws could be vibe coded. And I didn't think that, but I anyway, there was like a disconnect between what I thought was possible and of course what the AI was convincing me was almost about to get right. So that's like one example. The other example that I won't go in as much, but it was yeah. Right. 00:50:40.000 --> 00:50:48.000 Well I guess it didn't have the specification. Like if you had the once you had that, you essentially had a specification of the system. 00:50:48.000 --> 00:51:00.000 I had a spec, but I also just like, okay, imagine if you didn't have UV and you didn't have virtual environments. Imagine if you get Claude raw Python with system packages as is, 00:51:00.000 --> 00:51:01.000 Yeah. 00:51:01.000 --> 00:51:05.000 and you asked it to figure out Python dependencies, it wouldn't build 00:51:05.000 --> 00:51:06.000 Right. 00:51:06.000 --> 00:51:08.000 UV for you. It would make a mess. It would 00:51:08.000 --> 00:51:09.000 Yeah. 00:51:09.000 --> 00:51:14.000 it would write files all over, it would like l syspath them all over. Th this is the 00:51:14.000 --> 00:51:15.000 Yeah. 00:51:15.000 --> 00:51:19.000 kind of thing again. It's it's up to you to stop yourself and to stop it and be 00:51:19.000 --> 00:51:40.000 Right. 00:51:19.000 --> 00:51:40.000 like, buddy, like we're not gonna b we're not gonna build a house with like sticks and no glue. It there's just something, it's like it's not the right tools, it's not the right structure, it's not the right approach. It's like we have just a bunch of hammers and not that everything looks like a screw, we're just we we're bringing a bunch of hammers to like I I was gonna say something really tight, but look, I'm gonna say to a kindergarten a kindergarten. Like something feels like why do you 00:51:40.000 --> 00:51:43.000 Yeah, right. 00:51:43.000 --> 00:51:57.000 have these like it's just it's not at all it's just like d doesn't fit. The the the the model doesn't fit. And then the question is, okay, well who has to figure that out? And I think maybe in a year they'll just figure it out themselves and this will podcast serve as a historical capsule. And when we used to go crazy because it couldn't 00:51:57.000 --> 00:52:00.000 Right. Yeah. 00:52:00.000 --> 00:52:33.000 figure out stuff that is just like not think more but think more abstractly or like figure out if that's even the right problem to solve or or i you know b because also maybe I'll kind of add These tools are so they seem so eager to please that I think that there is a knob where if you ask it to code, but it stopped you and it was like, we can't code here, buddy. We gotta we gotta think about what we're doing first. You wanna just write a bunch of a bunch of code that's not gonna run? It doesn't do that. It's like, my God, yes, sir. I'm gonna just 00:52:33.000 --> 00:52:42.000 Yeah. 00:52:33.000 --> 00:52:42.000 here's all the tokens. And it it's it's so so so that's like kind of like then you have to resist, you have to stop it from you have to kind of use your leash to pull it back. Anyway, the other example. 00:52:42.000 --> 00:52:45.000 Yeah. Yeah, I agree that if you're not in 00:52:45.000 --> 00:52:52.000 okay, sorry. 00:52:45.000 --> 00:52:52.000 at the correct level of abstraction, like if you knew you needed a database, say, right? Like similar to 00:52:52.000 --> 00:52:55.000 Yes, yes, yes, yes, yes. Correct. 00:52:55.000 --> 00:52:57.000 if you need a package manager or virtual environment, 00:52:57.000 --> 00:52:59.000 Yep. Yep. 00:52:59.000 --> 00:53:04.000 like maybe Claude will start trying to build the database, but 00:53:04.000 --> 00:53:09.000 Yeah. 00:53:04.000 --> 00:53:09.000 Like historically it would take yeah team like ten years to write a database and that's with knowing here's the proper like, what are 00:53:09.000 --> 00:53:13.000 That's right. That's right. That's right. 00:53:13.000 --> 00:53:15.000 the properties you need in a database? Like it's probably not gonna get to that. 00:53:15.000 --> 00:53:33.000 Right. Yeah, that's a great ex yeah. Yeah yes, exactly. Yeah, it wouldn't event sequel. Yeah, and and that's like I think I think we are in that state right now. It's like imag that's that's perfect. I I was I thought at first I was thinking you were kind of saying like you you didn't specify like no, but if it didn't know what a database is, which is exactly the case for a lot of things that it doesn't know, 00:53:33.000 --> 00:53:35.000 Yeah. 00:53:35.000 --> 00:53:46.000 it's not a good like it's just not there yet. That the six the percentage three. Like one. You know, it's like it's bad. And it adds the layer of 00:53:46.000 --> 00:53:50.000 Yeah. 00:53:46.000 --> 00:53:50.000 being convinced. Terrible. That's just like a headache. If it stopped itself, yeah, no. It yeah, it would be just like 00:53:50.000 --> 00:53:58.000 Yeah, you don't want it to build the database as part of the project. As part of building like your circuit designer, it's like, we need a database. Let's build a database. No, that's not gonna work. Right, yeah. Yeah. 00:53:58.000 --> 00:54:08.000 Yes. Yeah, no. No, it should just stop itself. It should be like, Well, I hey you're you know, it's like this is of course like you you see to happen. It's like it's kind 00:54:08.000 --> 00:54:09.000 Yeah. 00:54:09.000 --> 00:54:13.000 of if you're like, Hey, look let's invent let's invent teleportation and it goes, No problem, let's get started and 00:54:13.000 --> 00:54:14.000 Yeah. Yeah. 00:54:14.000 --> 00:54:19.000 it's like and you know, because it knows that teleportation is hard, it will not do it. 00:54:19.000 --> 00:54:23.000 Yeah. 00:54:19.000 --> 00:54:23.000 But the things that it thinks is like, Okay, the other example, I'll just mention it just in case it can 00:54:23.000 --> 00:54:30.000 Yeah. 00:54:23.000 --> 00:54:30.000 scare somebody or like i i you know, it get somebody excited about solving it. 3D stuff has been sort of a disaster with it. I've worked, I've like, we're doing a lot of CAD. 00:54:30.000 --> 00:54:37.000 Mm-hmm. That's cool. That's cool because that it's not something I have any experience working with. So cool to hear about. Mm-hmm. 00:54:37.000 --> 00:54:44.000 Yeah. Right. And and again, I of course am also coming at it very, very new. I'm like new to cat, I'm new to this. We're 3D printing. We're doing a lot of robotic, like 00:54:44.000 --> 00:54:56.000 Yeah. 00:54:44.000 --> 00:54:56.000 optic like recognition. It is really fascinatingly like also convinced, writing a lot of code, running it in simulators, not seeing what it does. And you know, as someone very smart, like actually my opposing 00:54:56.000 --> 00:54:57.000 Right. 00:54:57.000 --> 00:55:03.000 right now, very good friend, he was like, Well, it doesn't it doesn't have like a cube in its head when it thinks about a cube, 00:55:03.000 --> 00:55:17.000 Rare. 00:55:03.000 --> 00:55:17.000 right? And so it's like Yeah, that but like times a million that it doesn't have in its head about three D. And so it's like it's been very comical. Very, very comical to kind of try to do a three D. However, there's a world that will be unlocked once it does. But just closing that parenthesis, if anybody Yeah. 00:55:17.000 --> 00:55:27.000 Yeah, I mean that's interesting in the sense of do you need to how do you present that world to it? Like do you need to show it s images of the rotation, things like that, right? 'Cause otherwise I mean it 00:55:27.000 --> 00:55:32.000 Screenshot Yeah. But is that enough? 00:55:32.000 --> 00:55:34.000 makes sense, it can't iterate. Like I th yeah. 00:55:34.000 --> 00:55:37.000 No, it can't iterate. Exactly. And and it's mm-hmm. 00:55:37.000 --> 00:55:40.000 I mean I think that's part of why it's successful in code and a lot of things is because it can iterate and t like 00:55:40.000 --> 00:55:44.000 Yes. Yep. Yep. 00:55:44.000 --> 00:55:47.000 the tests tell it when it's wrong. So then it can iterate. yeah. 00:55:47.000 --> 00:55:54.000 Yep. Yeah, it's a it's a very closed loop. Loops are good. Loops are good, you know, dangerously skip dangerously do more loops is good. 00:55:54.000 --> 00:56:05.000 Yeah. Which I still I I still don't run dangerously skip permissions locally. I know there is a new met there is a new kind of somewhere in between method where Claude actually 00:56:05.000 --> 00:56:06.000 Mm. 00:56:06.000 --> 00:56:08.000 will evaluate. You can have Claude 00:56:08.000 --> 00:56:10.000 Mm. 00:56:10.000 --> 00:56:12.000 they added it as an option and it will actually run each 00:56:12.000 --> 00:56:13.000 Interesting. 00:56:13.000 --> 00:56:20.000 command through a prompt to see if it's safe. 00:56:20.000 --> 00:56:26.000 I feel even more validated by that. I'm like everybody every the kids dangerously skip permissions, 00:56:26.000 --> 00:56:33.000 Yeah. 00:56:26.000 --> 00:56:33.000 you know, and I'm I think I don't know if it came from React. React has dangerously set inner HTML, like has one permission, 00:56:33.000 --> 00:56:36.000 No. 00:56:33.000 --> 00:56:36.000 not one per one attribute. It was kind of the joke, they invented this really 00:56:36.000 --> 00:56:37.000 Right. 00:56:37.000 --> 00:56:42.000 funny name where they were like, Look, you're just breaking React, you're not even using it properly, things are gonna be terrible. Sure, if you want to. 00:56:42.000 --> 00:56:44.000 Yeah. 00:56:44.000 --> 00:56:55.000 And it was always 00:56:44.000 --> 00:56:55.000 Right. 00:56:44.000 --> 00:56:55.000 funny, it was like, that's a nice escape hatch. This is not an escape hatch. This is the future. The future, that's why open claw was so popular. People want this thing autonomous, while loop true. And I understand 00:56:55.000 --> 00:56:56.000 Yeah. 00:56:56.000 --> 00:57:07.000 this will lead to horrific consequences. There's absolutely no doubt in my mind. 100% guaranteed this is like a chainsaw. it's like a baby with a chainsaw 00:57:07.000 --> 00:57:11.000 Yeah. 00:57:07.000 --> 00:57:11.000 on some sort of like a hundred percent. It's just too good. It's just too good. 00:57:11.000 --> 00:57:18.000 Yeah, I mean I've been testing with doing things like using fly.io has a service, sprites.dev where you can run 00:57:18.000 --> 00:57:20.000 Yes, that's right. That's right. 00:57:20.000 --> 00:57:23.000 VMs that have this. And like you could say here's only right, yeah, 00:57:23.000 --> 00:57:26.000 Yeah, exe.dev. Yep. 00:57:26.000 --> 00:57:28.000 here's only like a git token for this repo, right? So limit what can go wrong, but it takes more time. Yeah. 00:57:28.000 --> 00:57:48.000 A VM. Yep. Yes. I I yeah. That's right. Takes more time. And I fully this is the thing. If I'm responsible or somebody was like, looked me in the eye and said, What do you really want? You just want to run a piece of code on a piece of compute. You're not desperate to give it your SSH keys. I'm I don't want it to 00:57:48.000 --> 00:57:50.000 Yeah. 00:57:48.000 --> 00:57:50.000 delete my like family pictures. It's just, yeah, 00:57:50.000 --> 00:57:51.000 Rip. 00:57:51.000 --> 00:57:56.000 convenience wins and blah blah blah. But a hundred percent. I I think of course if we can only do that. 00:57:56.000 --> 00:57:57.000 Yeah. 00:57:57.000 --> 00:58:03.000 these autonomously. I think we really get like we get our lunch and to eat it too or a cake or whatever. I th I think it's that. 00:58:03.000 --> 00:58:04.000 Yeah. 00:58:04.000 --> 00:58:07.000 But clock code is a thing you install on your development computer. So are 00:58:07.000 --> 00:58:08.000 Right. 00:58:08.000 --> 00:58:11.000 they gonna give us compute in the Claude for free? Is this part of the next plan? 00:58:11.000 --> 00:58:21.000 Well no, yeah, it is I mean, it adds another cost to it because you have another computer running and I mean you need enough power on these to be able to actually run the code and the tests. So it's not like a small computer, yeah, too. So it is I think that 00:58:21.000 --> 00:58:30.000 That's right. The thing. That's right. Yeah. And my code talks to my robot. Yep. That's right. So I'm 00:58:30.000 --> 00:58:35.000 adds to the challenge of it is you need a real machine that can do this. Yeah. 00:58:35.000 --> 00:58:50.000 Yeah, maybe I should just, you know, move off my pictures and my SSH keys and just kinda consider my comp like I should have another computer. It's a whole computer. I mean again this is so absurd. but I think when you asked me how long ago that happened, that's a scary question. It's like, yeah, this just just happened. Just just just 00:58:50.000 --> 00:58:59.000 Yeah. 00:58:50.000 --> 00:58:59.000 just just happened. So the the like living whatever living through this day by day has a very, very special feeling where 00:58:59.000 --> 00:59:00.000 Yeah. 00:59:00.000 --> 00:59:06.000 we're all discovering this kind of thing that we're That's bringing us to space. And we didn't know we were 00:59:06.000 --> 00:59:09.000 Yeah. 00:59:06.000 --> 00:59:09.000 in it anyway. So but this yeah, this is extremely fun. This 00:59:09.000 --> 00:59:10.000 Yeah, for sure. 00:59:10.000 --> 00:59:12.000 is extremely, extremely fun. 00:59:12.000 --> 00:59:15.000 Yeah, one if you have to go. But yeah, 00:59:15.000 --> 00:59:17.000 No, no, no, it's all good. No, no, please, please, please. Yeah, yeah. 00:59:17.000 --> 00:59:26.000 one th so this one part that I actu have some talked to some people about, which I do think is I mean I know I've felt this as 00:59:26.000 --> 00:59:28.000 Mm-hmm. 00:59:28.000 --> 00:59:32.000 just there's more context switching if I'm working 00:59:32.000 --> 00:59:33.000 Mm-hmm. 00:59:33.000 --> 00:59:36.000 on like twenty different projects you could work on at 00:59:36.000 --> 00:59:47.000 Yeah. 00:59:36.000 --> 00:59:47.000 once, right? Before it just wasn't feasible. Like are you more tired at the end of the day working with this less in a different way? 00:59:47.000 --> 00:59:52.000 Yeah. W it's true. I was like asking my girlfriend, why can't I sleep well in why why am I waking up at seven AM in fright? 00:59:52.000 --> 00:59:57.000 Yeah. 00:59:57.000 --> 01:00:24.000 Yeah, it's it's totally What is it? It yeah, it feels so different and so happening right now that I don't have the perspective. Is it more tiring? I think ri yeah, the the part that I really, really th so I I think when I would feel that my emotions are kinda like you know, like I bring work to home sort of deal, like 01:00:24.000 --> 01:00:27.000 Yeah. Yeah, yeah. 01:00:27.000 --> 01:00:39.000 I f I I remember it as well, if I was still trying to solve a thing. And I didn't solve it, or it really felt like I was stuck, and I come home, I kinda have a bad evening and I'm a little bit grumpy. 01:00:39.000 --> 01:00:40.000 Yeah. 01:00:40.000 --> 01:00:46.000 And and then if I like, my god, it works, and I stop right there, which then I 01:00:46.000 --> 01:00:53.000 Yeah. 01:00:46.000 --> 01:00:53.000 started doing more, like stop earlier. If you kind of stop at a good point, your whole evening is like, I'm a genius, this worked, I knew it. 01:00:53.000 --> 01:00:55.000 Yeah. 01:00:55.000 --> 01:01:19.000 the world should celebrate me. So so that that used to be the case. Now I think being frustrated That it didn't work now also includes a like, why am I using this? Because it's like if I wasn't able to solve the problem, sure, like I'm a bad developer, I'm not smart enough, I should have gone to like a math school, blah blah blah. You can feel bad about yourself, but then you can be like, no, you know, I have also good days and I've solved other problems, 01:01:19.000 --> 01:01:20.000 Right. 01:01:20.000 --> 01:01:32.000 and it's just this is a hard problem. You can talk to your friends and they can be like supportive. But I think using a tool by choice and then that tool disappointing you and then being like angry at the screwdriver. Or like, yeah, 01:01:32.000 --> 01:01:41.000 Yeah. 01:01:32.000 --> 01:01:41.000 angry at the drill for right. I'm angry I know what I'm angry at. I'm I I once went as Home Depot and I'm installing something in my plaster, whatever, home, and 01:01:41.000 --> 01:01:42.000 Yeah. 01:01:42.000 --> 01:01:48.000 I this guy looked like more of a construction guy than I do. And I go, hey, should I get these like it's I'm trying to get everybody 01:01:48.000 --> 01:01:49.000 Yeah. 01:01:49.000 --> 01:01:55.000 to never use them. It's like it's like a really huge plastic screw, and the idea is that you screw it and then you screw it anchors. Right, because there's always a 01:01:55.000 --> 01:01:58.000 like those anch those anchors. Yeah, yeah, yeah, yeah. 01:01:58.000 --> 01:02:00.000 million anchors. It's like which one is the is the the butterfly one, the other 01:02:00.000 --> 01:02:01.000 Red. 01:02:01.000 --> 01:02:06.000 one? And so and I was like, I've never used them because I don't like the idea, it's gonna make a really big hole. And the guy's like, 01:02:06.000 --> 01:02:09.000 Yeah. 01:02:06.000 --> 01:02:09.000 I use them all the time on the construction side, you should get them like trust me. I'm like, I actually love an expert. And so 01:02:09.000 --> 01:02:11.000 Right. 01:02:11.000 --> 01:02:16.000 I trust the guy, I get them home, they make a mess. It's like they crumble, they leave huge holes, 01:02:16.000 --> 01:02:19.000 Yeah. Yeah. 01:02:19.000 --> 01:02:24.000 and I'm swearing at it, but I am swearing at a tool, I'm swearing 01:02:24.000 --> 01:02:25.000 Yeah. 01:02:25.000 --> 01:02:38.000 at somebody who convinced me who I both trusted and didn't trust. And I'm also angry at myself for like being incompetent, not having the ability to know how to solve this, now being in a treble that now I created a new problem. I have holes in my walls. That feels more like 01:02:38.000 --> 01:02:41.000 Yeah. Right. 01:02:41.000 --> 01:02:50.000 an AI tool bringing you to a bad place, convincing you it was gonna good be a good time, and then you were like, I I was with my friends twenty minutes ago doing 01:02:50.000 --> 01:02:51.000 Yeah. 01:02:51.000 --> 01:03:03.000 small things and you brought me over because it was gonna be like a big thing, but it's terrible over here. And so that's that's a that's a frustration I haven't really learned to live with. because 01:03:03.000 --> 01:03:12.000 Do you think it's more frustrating when Claude gets something wrong than if you got it wrong personally? Yeah. 01:03:12.000 --> 01:03:23.000 Yeah, because you trusted it and then I I I I do think, yeah, I do think it's like it's less pure. It's more like you ate a bunch of nachos, your tummy hurts, and then you're like, Why am I doing this to myself? Like I know that 01:03:23.000 --> 01:03:27.000 Yeah. 01:03:23.000 --> 01:03:27.000 I should be eating veggies and I I had veggies in the fridge, but 01:03:27.000 --> 01:03:28.000 Yeah. 01:03:28.000 --> 01:03:54.000 I I felt a little like nachos are kind of it's like there's a poisonous comfort and like laziness and a little bit of a fast thing that is like addictive. And you kind of accept the cost of it because when the cost is zero because it gives you things super quickly, you're like, this is it. But yeah, I think it's a little bit like having a third coffee in your day. 01:03:54.000 --> 01:04:12.000 Yeah. 01:03:54.000 --> 01:04:12.000 Like, nah, you're just gonna get really ill after some time. And it it it it it it But I I have a hard time telling you it has dawned on me and I've reflected on this and I've cut back. No. Not at all. I'm 01:04:12.000 --> 01:04:14.000 what that's what we're doing, right? Like that's we're not stopping. Yeah. Right. Not not that 01:04:14.000 --> 01:04:19.000 Right now? Is this no. No. No. We're 01:04:19.000 --> 01:04:22.000 I'm running claw I'm not running Claude right now. Right. Yeah. 01:04:22.000 --> 01:05:14.000 Yeah, yeah. No, we're just waiting for the next model to solve all of our problems. I'm just waiting for the next update. I'm just trying to out trick it. And I'm trying to also think that I'm getting better. And this is like almost like somebody who has like an addiction I'm not to compare, of course, but like there's something about the lying to yourself, which is like, is this what I'm doing? I'm just being like, No. It's like kind of yes, back to the people who are like they have a hard time with it, which I I'm sure is true. And then other people are like, I you're not using it right. I have those thoughts sometimes, but those are crazy thoughts. Like not using it right. Like, what is it? What is right? And again, it did I go from seventy percent to seventy three percent, and I feel superior to the people who feel that it is seventy percent, and they're right anyway. So so it's like not not not great, not the best. I think when Tailscale came out, 01:05:14.000 --> 01:05:15.000 Yeah. 01:05:15.000 --> 01:05:19.000 it was a net positive for the entire world. And I don't think anyone want my mental health is 01:05:19.000 --> 01:05:20.000 Yeah. 01:05:20.000 --> 01:05:25.000 damaged. This has repercussions. And 01:05:25.000 --> 01:05:29.000 What does that mean for us going forward though? Like what if we're doing this, I g I d I don't know. But 01:05:29.000 --> 01:05:38.000 But how could we know? How could we How could we Hindsight is twenty twenty, but we we're we're in the middle of a crash and it's gonna be human error 01:05:38.000 --> 01:05:54.000 Yeah. 01:05:38.000 --> 01:05:54.000 at the end, and it was like I have no idea. I th I th I think it's obvious that it's gonna lead to some really, really bad stuff. I think huge leaks, huge systems that go haywire in a way that we're not like people have always there's always been bad code around, but this is gonna like be 01:05:54.000 --> 01:05:55.000 Red. 01:05:55.000 --> 01:06:08.000 like dramatically terribly fast deployed vast amounts of like very vulnerable code, et cetera. I mean the people running r writing malware with LMs right now must have so much fun. It must be an incredible time for them. we are unprepared. and then you know 01:06:08.000 --> 01:06:21.000 Right. Yeah. Yeah. Yeah. Well, I guess that was the the latest lovable breach where 01:06:21.000 --> 01:06:26.000 was it? 01:06:21.000 --> 01:06:26.000 I think they were saying that well I don't I just I think when 01:06:26.000 --> 01:06:30.000 Yeah, yeah, yeah. 01:06:26.000 --> 01:06:30.000 I was reading the reports on it, the speed with 01:06:30.000 --> 01:06:32.000 Yeah. 01:06:30.000 --> 01:06:32.000 which it happened, they were 01:06:32.000 --> 01:06:35.000 huh. Mm-hmm. 01:06:35.000 --> 01:06:41.000 In reporting on it they seemed impressed by the speed with which with things progressed, which to me s made it seem like this is LLM accelerated, otherwise 01:06:41.000 --> 01:06:46.000 Mm. Yep. Yep. 01:06:46.000 --> 01:06:49.000 finding these exploits or these combinations of exploits is just 01:06:49.000 --> 01:06:52.000 Right. Takes time. 01:06:52.000 --> 01:06:57.000 super ti I mean it's just more time intensive, right? Like it's an accelerator. 01:06:57.000 --> 01:07:17.000 Yeah, yeah. I have I have I have to read that. Yeah, it's an accelerator. Exactly. Yeah, it's an accelerator. It's like it's an accelerator. Yeah. So so I don't think we're all gonna be like this was a net positive for the entire world. It was so f like this is this is this is definitely like the invention of gasoline or some some other 01:07:17.000 --> 01:07:18.000 Yeah. 01:07:18.000 --> 01:07:24.000 like substance that we're like we're kinda stuck with and it does a lot for us, but 01:07:24.000 --> 01:07:27.000 Yeah, I think pro automobiles is probably a good a analogue, right? Like there's how 01:07:27.000 --> 01:07:31.000 Yeah, yeah. Yeah. 01:07:31.000 --> 01:07:33.000 many people are killed or injured in automobiles every year, right? Yeah. 01:07:33.000 --> 01:07:46.000 Yes. Yeah. All transit systems being destroyed because now everyone will have a car. Of course we don't need those like, you know, those rail cars, those to be rail cars. Let's get rid of all of them. let's build highways for everybody. Of course 01:07:46.000 --> 01:07:48.000 Yeah. 01:07:46.000 --> 01:07:48.000 everybody then like just the other pollution. Yeah, it's like 01:07:48.000 --> 01:07:49.000 Yeah. 01:07:49.000 --> 01:08:03.000 I I suppose the world is like whatever, worlding better and I don't know, you know, hearts can be transplanted faster because we can transport them not on on the back of a donkey. Like I suppose. Like there's some argument for cars that is like 01:08:03.000 --> 01:08:07.000 Right. Would we have any of like would we even have the internet, any of these other things without that? Probably not. 01:08:07.000 --> 01:08:20.000 Yeah. Yeah, no, if we're still y you know, but but yeah, but being part of a technological revolution doesn't feel like yeah, like e let's go even faster. It just feels like, 01:08:20.000 --> 01:08:21.000 Yeah. 01:08:21.000 --> 01:08:25.000 yeah, this this will have like terrible harms, intended and unintended. Because intended, 01:08:25.000 --> 01:08:39.000 Yeah. 01:08:25.000 --> 01:08:39.000 of course, like people who are like really looking forward to like firing a bunch of people are like so excited about this, of course. and replacing them with AI. I think it's also like obviously the like How is this helping anyone eat if they can't afford to? I 01:08:39.000 --> 01:08:42.000 Yeah. Yeah. 01:08:42.000 --> 01:08:47.000 I I think obviously it's like, well, that's a done pro that's unrelated. And and 01:08:47.000 --> 01:08:57.000 Well. 01:08:47.000 --> 01:08:57.000 and that's of course like one of the the the horrors of all these things is that the trillions invested to make this acceleration accelerate even more have kind of nothing to do with any other problem that I obviously think we should be Yeah, yeah. 01:08:57.000 --> 01:09:03.000 Right, it's invested here versus other I mean, there's always choices in how 01:09:03.000 --> 01:09:04.000 T then Yeah. 01:09:04.000 --> 01:09:07.000 we as a society invest our resources, right? 01:09:07.000 --> 01:09:12.000 Yeah, and Tom and and you know, it's like all the nuclear plants coming back online and like, you know, e 01:09:12.000 --> 01:09:18.000 Yeah. 01:09:12.000 --> 01:09:18.000 every company being like, we had a climate pledge, but that that was before we needed like all the GPUs to compute 01:09:18.000 --> 01:09:27.000 Yeah. 01:09:18.000 --> 01:09:27.000 more tokens. It's like I I don't find it funny. I find it like it's like a cynically disgusting kind of thing of like, yeah, the other words were words and now we 01:09:27.000 --> 01:09:37.000 Yeah. 01:09:27.000 --> 01:09:37.000 we'll have new words. It doesn't matter, we know what we're really after. and yeah, and it's not the betterment of like it's not about endangered species, not you know, breathing 01:09:37.000 --> 01:09:40.000 Yeah. 01:09:37.000 --> 01:09:40.000 in GPU exhaust. so so yeah, so that 01:09:40.000 --> 01:09:53.000 Yeah. Well may maybe this the GPUs will push us push us to full solar power 'cause that's I think is the cheapest installation to do these days now. 01:09:53.000 --> 01:09:56.000 Hey, look at that. Finally. If it's yeah I d mm-hmm. 01:09:56.000 --> 01:10:02.000 Yeah, I th I I'm pretty sure that it's if you're adding a marginal unit of power today, the cheapest 01:10:02.000 --> 01:10:05.000 Hmm. Right, right, right, right, right. 01:10:05.000 --> 01:10:07.000 way to do it is solar. So 01:10:07.000 --> 01:10:36.000 Solar. Well that's it. Then, you know, I don't know. Something something good to look to to to look to look up to to look forward to. but yeah, I I have a hard time saying that I'm an optimist and at the same time I'm such a complete devoted user of it that it's like it's it's it's a it's a it's a strange and I know that my guilt will not do anything to solve anything, so I guess I'm stuck. 01:10:36.000 --> 01:10:41.000 Well, I'm I'm hoping for I mean the open source models and even the local models I think continue to improve. 01:10:41.000 --> 01:10:51.000 Yes, yes, yes, yes. Yep, yep, yep. Yep. Yeah, that's huge. Of course. Yep. We'll we'll get there. We'll get we'll get somewhere. We'll get somewhere. 01:10:51.000 --> 01:10:53.000 I think that increasingly we'll be 01:10:53.000 --> 01:10:54.000 Mm-hmm. 01:10:54.000 --> 01:11:02.000 having control of these and and eventually I do think they will run more locally. 01:11:02.000 --> 01:11:06.000 Yeah, yeah, yeah. We'll see what the world is by then. By that time by that time. Yeah, once we have our little 01:11:06.000 --> 01:11:19.000 By the by then when when that is, yeah. When I get my I'm waiting to get my next I still have a twenty twenty Mac Mini, so I can't really run any serious LLMs on that, but 01:11:19.000 --> 01:11:20.000 Yeah. 01:11:20.000 --> 01:11:27.000 the next Mac I get, hoping can run a decently good local model. 01:11:27.000 --> 01:11:30.000 What's a good model do you recommend? Just curious, just to like 01:11:30.000 --> 01:11:33.000 I think the latest Qwen 3.6 models have been rated pretty well. 01:11:33.000 --> 01:11:38.000 The Qwen, yeah. People like Qwen. That's pretty impressive, yeah. 01:11:38.000 --> 01:11:44.000 those I I think I've heard as in terms of four local models seem 01:11:44.000 --> 01:11:45.000 Yeah, yeah, yeah. 01:11:45.000 --> 01:11:47.000 to be the strongest. 01:11:47.000 --> 01:11:55.000 Yeah, that's interesting. Yeah. Yeah. Yeah, what I I'm I'm yeah, I'm just trying to like yeah, sort of follow along by reading like one thread a every few 01:11:55.000 --> 01:12:27.000 Yeah. 01:11:55.000 --> 01:12:27.000 weeks on H N. But yeah, yeah, I I I've seen that that's that that's interesting. Something to consider for sure. Yeah, yeah, yeah.