WEBVTT 1 00:00:05.200 --> 00:00:09.080 Hey, everybody, welcome back to another episode of JavaScript Jabber. 2 00:00:09.320 --> 00:00:11.679 This week, on our panel we have Dan shapiir. 3 00:00:12.119 --> 00:00:15.240 Hey from the still very hot in Monkey tel Aviv. 4 00:00:16.120 --> 00:00:20.039 I'm Charles Maxwood from Top End Devs. Yeah it. Yesterday 5 00:00:20.359 --> 00:00:23.079 we had a high of ninety five, which was nice 6 00:00:23.079 --> 00:00:27.199 and cool compared to what it's been so yeah, I 7 00:00:27.239 --> 00:00:31.559 feel that. We also have Steve Sewell from builder Io. Steve, 8 00:00:31.600 --> 00:00:33.679 do you want to say hello and remind people who 9 00:00:33.679 --> 00:00:34.000 you are? 10 00:00:34.520 --> 00:00:37.679 Yes, everybody, I'm Steve. I'm our co founder and CEO 11 00:00:37.840 --> 00:00:41.520 at Builder. We make cool AI, designed to code, designed 12 00:00:41.520 --> 00:00:44.000 to live. You know, website cm my stuff which I'm 13 00:00:44.000 --> 00:00:46.840 sure going to And it is only fifty eight degrees 14 00:00:47.240 --> 00:00:49.280 farrenheit and San Francisco. It's not hot at all. I 15 00:00:49.320 --> 00:00:50.560 wish it was a little hard to be honest. 16 00:00:51.399 --> 00:00:53.200 Wow, is that why you're sitting in the server room? 17 00:00:53.240 --> 00:00:56.640 Get a little a little warm? Yeah? Exactly? 18 00:00:57.320 --> 00:01:02.880 Yeah, good deal. Well we got you on. I've seen 19 00:01:02.880 --> 00:01:07.480 a whole bunch of videos from you about AI. I 20 00:01:07.519 --> 00:01:12.200 know you, you know, run builder dot io with with 21 00:01:12.319 --> 00:01:16.280 Mishko and a bunch of other folks. And I guess 22 00:01:16.480 --> 00:01:19.519 what I'm wondering and where maybe we should start is okay, 23 00:01:19.560 --> 00:01:23.120 So what is the CEO of this? You know, hey, 24 00:01:23.159 --> 00:01:27.760 build your website with us. You know, why do you 25 00:01:27.799 --> 00:01:31.200 care about AI? Like, how does that fit into the 26 00:01:31.280 --> 00:01:34.359 life that you're running? And you know, how will that 27 00:01:34.400 --> 00:01:37.040 fit into the life of somebody who's going, hey, I'm 28 00:01:37.040 --> 00:01:38.200 a JavaScript developer. 29 00:01:40.519 --> 00:01:44.239 Yeah, no, it's a it's a great question. So, you know, 30 00:01:44.359 --> 00:01:47.280 I actually, as words it sounds, I'm a little bit embarrassed. 31 00:01:47.280 --> 00:01:49.120 In retrospect, I think it's probably still the right idea. 32 00:01:49.879 --> 00:01:53.680 I was excited by l M progress when I was 33 00:01:53.719 --> 00:01:56.280 seeing it. You know, GBT three was interesting but very 34 00:01:56.319 --> 00:01:58.840 hard to get good results from. Three point five was 35 00:01:58.840 --> 00:02:02.200 a big breakthrough, Like, ohater, it was better. It was 36 00:02:02.239 --> 00:02:04.319 noticeably better and easier to work with. You could talk 37 00:02:04.319 --> 00:02:06.280 to like a human, not like you'd weird. You know. 38 00:02:06.280 --> 00:02:08.960 I remember GPT three somewhere I would say something even 39 00:02:09.000 --> 00:02:12.840 as symbol as ending a your prompt to get a 40 00:02:12.840 --> 00:02:14.759 completion off of with like a new line or something, 41 00:02:14.840 --> 00:02:16.400 it would like break the whole thing. So they'd be like, 42 00:02:16.439 --> 00:02:18.719 just don't do that, And like three point five was like, 43 00:02:18.800 --> 00:02:21.759 let's assume humans are humans and we will you know, 44 00:02:21.800 --> 00:02:23.360 make sure it works with whatever you give it. That 45 00:02:23.439 --> 00:02:26.159 was pretty huge breakthrough. But even then you saw like 46 00:02:26.240 --> 00:02:28.759 all these companies rushing to add eight out of the product, 47 00:02:28.879 --> 00:02:31.319 or like we even have vcs who invested in us, 48 00:02:31.360 --> 00:02:34.159 Like every CEO is talking about redoing their whole roadmap 49 00:02:34.199 --> 00:02:36.840 to be focused on AI. People are rebranding their companies AI, 50 00:02:37.520 --> 00:02:39.439 and I was like, you know, I'm thinking of like 51 00:02:39.520 --> 00:02:41.719 crypto and all this stuff, and I'm like, no, we're 52 00:02:41.759 --> 00:02:45.599 we are not doing that at all. We are dabbling, 53 00:02:46.280 --> 00:02:52.080 We're not changing significant plans. As we dabbled further, the potential, 54 00:02:52.240 --> 00:02:54.919 hypothetical potential became really clear. You know, if you work 55 00:02:54.919 --> 00:02:58.039 in that space of like you know, you've got a 56 00:02:58.159 --> 00:03:00.400 design and a program like FIGM on one end, and 57 00:03:00.439 --> 00:03:03.240 you've got a website or app that you probably have 58 00:03:03.439 --> 00:03:06.759 a mix of developers coding on because we focus kind 59 00:03:06.759 --> 00:03:09.520 of larger businesses generally, so you've got developers. You're not 60 00:03:09.599 --> 00:03:11.879 like some small mom and pop shop or something. So 61 00:03:11.919 --> 00:03:14.319 you've got developers writing code, and you have people who 62 00:03:14.319 --> 00:03:16.840 are not developers trying to put out pages or update 63 00:03:16.879 --> 00:03:20.960 pages or something through a CMS. Where can AI help most? 64 00:03:21.080 --> 00:03:22.840 You know, we've seen all these cool demos and this 65 00:03:22.879 --> 00:03:25.159 is where things get confusing too, is there's a lot 66 00:03:25.199 --> 00:03:28.080 of demos that are not representative of the average users experience. 67 00:03:28.159 --> 00:03:31.360 So you could, for instance, go into chat to BT 68 00:03:31.520 --> 00:03:34.080 and have it summarize a long piece of text, and 69 00:03:34.280 --> 00:03:36.680 every demo will do well and every user's experience will 70 00:03:36.680 --> 00:03:38.199 probably be pretty good at that. That's something that the 71 00:03:38.240 --> 00:03:40.039 elms are good at. Take a large amount of information 72 00:03:40.240 --> 00:03:42.479 and condense it down. And if you don't like the 73 00:03:42.680 --> 00:03:44.360 style of how it condensed it, you know, like the 74 00:03:44.439 --> 00:03:47.439 language it used. Like when I paste a huge amount 75 00:03:47.439 --> 00:03:49.120 of stuff and I tell it like turn this into 76 00:03:49.159 --> 00:03:51.759 an email, it assumes that I'm like this corporate person 77 00:03:51.919 --> 00:03:55.159 emailing of a million people. No, no, we're startup for what 78 00:03:55.240 --> 00:03:57.360 are we now? Seventy people? I use a pretty chill 79 00:03:57.439 --> 00:03:59.520 vibe when I talk. Here's some examples of how I 80 00:03:59.599 --> 00:04:02.560 usually talk. Emulate that. GPT is still pretty bad at that, 81 00:04:02.639 --> 00:04:05.520 but Claude as much better in my experience. But anyway, 82 00:04:05.520 --> 00:04:07.680 those are good. But then you see these other demos 83 00:04:07.719 --> 00:04:10.360 of like you know, these one off cherry pick things 84 00:04:10.400 --> 00:04:13.159 of like hey, I built me this whole program. That's awesome. 85 00:04:13.199 --> 00:04:16.720 In fact, Claude with artifacts is great at generating like 86 00:04:16.720 --> 00:04:19.600 a snake game or something. But that gives you a 87 00:04:19.639 --> 00:04:21.920 look into like what people would like to have happened. 88 00:04:21.920 --> 00:04:24.560 You know, people would like to If I'm a developer 89 00:04:24.600 --> 00:04:26.879 that has a Figma design coming from a design system 90 00:04:26.879 --> 00:04:29.920 and it's got behavior implicit in it. This is a 91 00:04:29.959 --> 00:04:33.279 new dashboard mock up with we we've got APIs for 92 00:04:33.319 --> 00:04:35.120 this data, and we've got components for these charts, and 93 00:04:35.160 --> 00:04:37.279 it's you know, it's got a layout and Figma. If 94 00:04:37.319 --> 00:04:40.000 I can just turn that into like almost finished code 95 00:04:40.240 --> 00:04:42.480 to just start me off, there's always nuances I got 96 00:04:42.480 --> 00:04:44.160 to do. But if you can connect it to APIs, 97 00:04:44.279 --> 00:04:47.040 assemble the components, create the layout and tail and whatever 98 00:04:47.040 --> 00:04:49.120 I'm using, and then let me work on it from there, 99 00:04:50.000 --> 00:04:52.199 that's pretty cool, especially when you know those things can 100 00:04:53.160 --> 00:04:57.079 actually work fairly reliably. That becomes like a why not 101 00:04:57.240 --> 00:04:59.199 type of thing? Why am I writing this all by hand? 102 00:04:59.240 --> 00:05:02.360 If the I can do that pretty effectively. On the 103 00:05:02.360 --> 00:05:05.079 flip side, if you're because part of what we are 104 00:05:05.160 --> 00:05:07.920 is headless CMS. So if you're a user trying to 105 00:05:07.959 --> 00:05:11.240 create new pages within a next JS app or whatever, 106 00:05:12.040 --> 00:05:13.839 same thing, you've got this mock up of a page. 107 00:05:13.839 --> 00:05:15.240 If I could just click a button and make that 108 00:05:15.279 --> 00:05:17.439 become real and then maybe use natural language to say, 109 00:05:17.480 --> 00:05:19.839 actually move the button over there, or when I click 110 00:05:19.879 --> 00:05:21.720 the button and should trigger the off flow or whatever, 111 00:05:22.040 --> 00:05:24.240 rather than learn this complicated tool of like, yes we 112 00:05:24.279 --> 00:05:26.839 have the off components registered in the tool, and yes 113 00:05:26.879 --> 00:05:28.959 we have our APIs connected over here, and if I 114 00:05:29.000 --> 00:05:31.319 know how to click one hundred buttons, I could do it. 115 00:05:31.360 --> 00:05:32.920 But if I could just say it and have it happen, 116 00:05:33.639 --> 00:05:35.959 that's the pipe dream. That's the obvious reason to care. 117 00:05:36.240 --> 00:05:38.560 The question obviously is just how well can it do that? 118 00:05:38.720 --> 00:05:42.399 And then more importantly, how can we make a reliable 119 00:05:42.399 --> 00:05:44.439 path to make as many of those dreams come true 120 00:05:44.439 --> 00:05:46.920 as possible without being full of foot guns. And that's 121 00:05:46.959 --> 00:05:49.319 been kind of our focus in research and development of 122 00:05:49.360 --> 00:05:50.240 the last year or two. 123 00:05:51.000 --> 00:05:56.560 So it's basically the domain that the space that you're 124 00:05:56.600 --> 00:05:59.959 working in is essentially Figma to code. 125 00:06:01.120 --> 00:06:02.920 That's a way to think about it. It's not every 126 00:06:02.920 --> 00:06:05.279 customer's use case. You don't as you can imagine if 127 00:06:05.319 --> 00:06:07.920 you could turn prompts into real life stuff using the 128 00:06:08.040 --> 00:06:10.319 React components you have and stuff you don't need Figma, 129 00:06:10.439 --> 00:06:11.800 you can just tell it to make me a thing 130 00:06:11.800 --> 00:06:14.360 and it can make you the thing. But Figma is 131 00:06:14.360 --> 00:06:16.439 probably one of the most common ways you represent in 132 00:06:16.519 --> 00:06:19.759 great detail what you want to happen before it happens. 133 00:06:20.000 --> 00:06:21.680 So you can think of us as like the if 134 00:06:21.680 --> 00:06:23.920 you saw TL draws old make or not old, Like 135 00:06:23.920 --> 00:06:26.680 they're make real demos, you know they diagram you hit 136 00:06:26.759 --> 00:06:28.639 make real becomes real. You can think of us as 137 00:06:28.680 --> 00:06:32.399 like an entire make real application or platform. Most of 138 00:06:32.439 --> 00:06:35.319 the time those are in Figma designs already. Sometimes they're 139 00:06:35.360 --> 00:06:37.560 just in a gyro ticket or a slack conversation. We 140 00:06:37.920 --> 00:06:40.560 we imagine a future world is not too far where 141 00:06:40.560 --> 00:06:42.240 we could have like a slack bot where you tag 142 00:06:42.279 --> 00:06:44.199 the builder slack bot. It looks at a thread, it 143 00:06:44.240 --> 00:06:47.120 summarizes your idea, implements, it sends you back a link. 144 00:06:47.199 --> 00:06:49.439 How does this look? Turn that to code, sink it 145 00:06:49.480 --> 00:06:51.920 to your code base, or just hit publish. Run it 146 00:06:51.920 --> 00:06:54.240 as a five percent test and see how well that does. 147 00:06:54.279 --> 00:06:55.439 Stuff like that, you know. 148 00:06:55.800 --> 00:06:58.480 Can be without the Figma. 149 00:06:58.839 --> 00:07:00.560 Yeah, with or without the Figma, that's one way to 150 00:07:00.560 --> 00:07:01.519 put it totally right. 151 00:07:02.240 --> 00:07:07.199 To be fair, though you're not, it's not really green field. 152 00:07:07.240 --> 00:07:10.680 I mean Figma themselves, as I recall, are looking at 153 00:07:10.720 --> 00:07:16.279 ways of transforming their designs into code m that's correct. 154 00:07:16.800 --> 00:07:19.399 Yeah, I think the green fields of l ms are 155 00:07:20.079 --> 00:07:23.759 what you know. I think the the hype is real. 156 00:07:23.879 --> 00:07:27.600 I was very against making any deviations to our plans 157 00:07:27.680 --> 00:07:29.879 or marketing or anything just because AI looks cool. I 158 00:07:29.879 --> 00:07:32.399 feel like there's a lot of startups who just at school, 159 00:07:32.480 --> 00:07:34.199 we want to be cool. Let's just do that. We 160 00:07:34.319 --> 00:07:36.759 put a lot more thought into it before investing behind it. 161 00:07:37.000 --> 00:07:39.199 And to be honest, in most cases where I saw 162 00:07:39.240 --> 00:07:41.199 a startups just say AI, the school is do AI 163 00:07:41.600 --> 00:07:43.800 for lack of better terms. Some would see spikes and 164 00:07:43.839 --> 00:07:46.160 sign ups, some would do kind of cool things. But 165 00:07:47.240 --> 00:07:51.560 where I saw many fail completely were when they were 166 00:07:51.600 --> 00:07:55.519 adding AI for AI's sake, as opposed to solving major 167 00:07:55.639 --> 00:07:59.639 problems that only large language model school solve. So there 168 00:07:59.639 --> 00:08:01.600 are of ideas of things we could do with AI 169 00:08:01.680 --> 00:08:03.160 for the sake of doing it. But one of the 170 00:08:03.160 --> 00:08:06.879 biggest problems that Pigma has in their code generation that 171 00:08:06.959 --> 00:08:09.639 we have and explorin this too, that site builders have 172 00:08:10.279 --> 00:08:14.720 is that people have a lot more ideas then they 173 00:08:14.720 --> 00:08:17.720 can get into code in a high quality, and any 174 00:08:17.759 --> 00:08:20.360 tool has tried to turn any figma idea into code 175 00:08:20.360 --> 00:08:23.519 in an automated way is bad. It's the simplest way 176 00:08:23.560 --> 00:08:26.079 to put it. They do a very very bad job. 177 00:08:26.199 --> 00:08:28.759 And we found a few techniques that work pretty well 178 00:08:28.839 --> 00:08:32.000 that can only be done with llms. We found that 179 00:08:32.039 --> 00:08:35.840 we at the end of the day, like it's funny 180 00:08:35.879 --> 00:08:38.360 because I don't want to overhype lllms. I think they're 181 00:08:38.399 --> 00:08:41.720 a critical missing piece, but they are not the full solutions. 182 00:08:41.759 --> 00:08:43.399 What I mean by that is, you know, I've got 183 00:08:43.399 --> 00:08:46.000 some videos about like use AI as little as possible, 184 00:08:46.000 --> 00:08:49.240 and I still firmly believe in that most of our 185 00:08:49.320 --> 00:08:52.399 AI solutions that just look like, for instance, a Figma 186 00:08:52.399 --> 00:08:55.159 design is the input or a prompt and the output 187 00:08:55.200 --> 00:08:57.279 is code. It looks like we just fed that into 188 00:08:57.279 --> 00:08:59.360 the LM and we got that out the other side 189 00:08:59.799 --> 00:09:02.200 a more code than that, and that's good because we 190 00:09:02.200 --> 00:09:05.399 have control over that. It makes the product differentiated. There's 191 00:09:05.440 --> 00:09:08.559 not just an enormous amount of code pre processing the 192 00:09:08.559 --> 00:09:10.759 heck out of everything, post processing the heck out everything. 193 00:09:11.080 --> 00:09:14.679 But also we've trained our own models for specific parts too, 194 00:09:15.000 --> 00:09:18.039 and so you can take You could take various approaches here, 195 00:09:18.080 --> 00:09:21.399 but the one that's worked well for us is solve 196 00:09:21.519 --> 00:09:25.279 everything without AI as much as possible, break it down 197 00:09:25.320 --> 00:09:29.519 into the Tinese problems that are just not really solvable 198 00:09:29.559 --> 00:09:32.879 with just typical you know, conditions in code. For us, 199 00:09:32.919 --> 00:09:34.840 that was those were much smaller than you think, especially 200 00:09:34.840 --> 00:09:37.000 if you're really grind out the problem. It's really small. 201 00:09:37.039 --> 00:09:39.200 But you can't do without AI. But there's probably gonna 202 00:09:39.200 --> 00:09:41.279 be some point, like if you're turning designs to code 203 00:09:41.600 --> 00:09:46.000 or prompts to code or whatever. That what you can't 204 00:09:46.360 --> 00:09:51.200 decipher with most models, you could train yourself. You know, 205 00:09:51.279 --> 00:09:54.000 you could do training as simple as things like decision trees, 206 00:09:54.039 --> 00:09:56.399 which can work great if you just don't know what 207 00:09:56.519 --> 00:10:00.320 value to put into a box or set of values decision, 208 00:10:00.320 --> 00:10:02.480 try training with like XG boosts or something. It could 209 00:10:02.480 --> 00:10:05.120 be phenomenal to wipe out a bunch of crappy conditional 210 00:10:05.159 --> 00:10:08.639 code into one. Essentially, the AI takes example datas and 211 00:10:08.679 --> 00:10:10.679 kind of writes conditional code for you in a sense. 212 00:10:10.960 --> 00:10:13.480 Fantastic We use that for certain things. You can do 213 00:10:13.480 --> 00:10:15.440 fancy stuff like random forest. You know, a bunch of 214 00:10:15.480 --> 00:10:19.399 decision trees helping make a decision awesome. But what nothing 215 00:10:19.440 --> 00:10:23.759 can really do well outside of an LM is understand 216 00:10:23.799 --> 00:10:27.799 meaning of things and that's where again the LLM in 217 00:10:27.799 --> 00:10:30.240 our experience should not be treated as this opaque box 218 00:10:30.360 --> 00:10:33.519 where very raw inputs get out and very finished outputs 219 00:10:33.559 --> 00:10:35.960 get out the other side. But when you start identifying, 220 00:10:36.360 --> 00:10:39.600 here's a basic example. You've got a responsive design and 221 00:10:39.639 --> 00:10:42.440 it's got desktop end mobile, and on desktop you've got 222 00:10:42.480 --> 00:10:44.559 a nav you know, with all the links horizontal, and 223 00:10:44.559 --> 00:10:48.240 on mobile we've got a Hamburger menu. There's no world 224 00:10:48.320 --> 00:10:50.960 where we could solve that without an LLM. But again, 225 00:10:51.000 --> 00:10:53.360 we're not passing the designs in and saying figure it out. 226 00:10:53.399 --> 00:10:55.480 We're not doing screenshot to code, which is kind of 227 00:10:55.519 --> 00:10:59.240 awesome sometimes but always not quite right. People want you 228 00:10:59.279 --> 00:11:01.440 know you designed the position, you use the specific design 229 00:11:01.440 --> 00:11:03.840 tokens of Figma components, so those map to design tokens 230 00:11:03.840 --> 00:11:05.879 compotes in your code. When it's a screenshot, that's the 231 00:11:05.919 --> 00:11:08.840 conduit you lose all that information, and that's try not 232 00:11:08.919 --> 00:11:11.279 to lose that unless you're like, you know, again, you're 233 00:11:11.320 --> 00:11:14.279 just some random you're building some random project on the side. 234 00:11:14.399 --> 00:11:17.000 Maybe you don't care, it's it's just beginning. But if 235 00:11:17.000 --> 00:11:19.320 you're doing Burker Company, you need a lot more than 236 00:11:19.320 --> 00:11:20.600 that for it to be useful. You don't want to 237 00:11:20.600 --> 00:11:23.480 have to get code and rewrite it all, and so 238 00:11:23.600 --> 00:11:26.399 you need to use some type of LLM to understand 239 00:11:26.440 --> 00:11:28.159 the meaning of that and say, oh, this should become 240 00:11:28.159 --> 00:11:30.440 an interactive Hamburger menu. It did not have to figure 241 00:11:30.440 --> 00:11:32.639 out all those other things, how components map, how design 242 00:11:32.639 --> 00:11:34.519 tokens are map, all this stuff. We figured that out 243 00:11:34.559 --> 00:11:37.639 in advance, and we've found a format to pass that 244 00:11:37.720 --> 00:11:41.120 into the LM heavily preprocessed. You know, when all we're 245 00:11:41.159 --> 00:11:45.200 saying is here's baseline code that's almost done. It's just ugly. 246 00:11:45.200 --> 00:11:48.360 It's poorly named, it's poorly structured, and its might have 247 00:11:48.399 --> 00:11:50.960 some semantic misses, like the accessibility might not be quite 248 00:11:51.039 --> 00:11:53.879 right because we programmatically generated this with old school code 249 00:11:53.879 --> 00:11:56.720 and old school models, which are fast and reliable and 250 00:11:56.840 --> 00:11:59.000 best to use when you can. But all you want 251 00:11:59.039 --> 00:12:00.679 to do is do some clean up of the code, 252 00:12:00.759 --> 00:12:03.919 refactor this, name the components better, give it some props, 253 00:12:03.919 --> 00:12:08.240 some better class names, and then anything that's meaningfully wrong, 254 00:12:08.559 --> 00:12:12.159 such as the way we converted the horizontal. Now maybe 255 00:12:12.159 --> 00:12:14.559 our default logic turns it just vertical, because that's kind 256 00:12:14.559 --> 00:12:17.159 of a rule that works pretty well. Responsive design stuff 257 00:12:17.200 --> 00:12:19.639 side by side becomes vertical on the you know, narrow 258 00:12:19.639 --> 00:12:22.919 screen when that clearly didn't work well and you actually 259 00:12:22.960 --> 00:12:25.320 wanted a HAMBURGERM menu. Ho, the LM take a pass 260 00:12:25.360 --> 00:12:28.120 at that. Maybe that's the area that's least accurate, but 261 00:12:28.240 --> 00:12:31.120 still pretty much there, and then you're not having the 262 00:12:31.159 --> 00:12:33.720 developer at the end of the day have to tweak everything. 263 00:12:33.840 --> 00:12:35.399 They may just have to make some small tweaks to 264 00:12:35.399 --> 00:12:37.000 that Hambergerman. That's a bit rambly. 265 00:12:37.039 --> 00:12:37.799 That's maybe an example. 266 00:12:37.879 --> 00:12:41.679 Yeah, I kind of want to back up a little 267 00:12:41.679 --> 00:12:45.120 bit one thing here that I'm just going to throw 268 00:12:45.159 --> 00:12:48.000 out there. We talked to Obie Fernandez on Ruby Rogues 269 00:12:48.639 --> 00:12:51.159 and he's got a book about, you know, building and 270 00:12:51.200 --> 00:12:53.879 working with the LMS, and you know it talks to 271 00:12:53.919 --> 00:12:58.000 APIs and you know it does He's he's building chat assistance, right, 272 00:12:58.039 --> 00:13:00.399 So we're kind of talking about a different problem at 273 00:13:00.480 --> 00:13:02.759 but I think a lot of the ideas are the same. 274 00:13:02.799 --> 00:13:05.600 Where he basically breaks it down and says, yeah, so 275 00:13:05.639 --> 00:13:10.200 you have a like a gethub gethub ai chatbot, right, 276 00:13:10.720 --> 00:13:13.200 and so it knows about it's APIs and it's special 277 00:13:13.200 --> 00:13:15.240 things that it can do. And then you might have 278 00:13:15.320 --> 00:13:18.080 some other bot that knows kind of the next level 279 00:13:18.159 --> 00:13:21.360 up that orchestrates things, right, So he he was advocating too, 280 00:13:21.480 --> 00:13:24.799 to like break things down into really really granular things 281 00:13:25.240 --> 00:13:28.360 and have it come together that way. Of course, his 282 00:13:28.440 --> 00:13:30.720 stack is kind of a stack of AIS and you 283 00:13:30.840 --> 00:13:32.759 kind of figure it out and set up context and 284 00:13:32.759 --> 00:13:36.039 things like that. But I imagine a lot of people 285 00:13:36.399 --> 00:13:39.000 and there are some comments to this effect on here too, 286 00:13:39.159 --> 00:13:41.279 and I think, I think this is kind of where 287 00:13:41.279 --> 00:13:43.840 I want to start with. What you've talked about is 288 00:13:46.080 --> 00:13:48.480 you know, Jack Harrington on Twitter said what are the 289 00:13:48.559 --> 00:13:50.759 wrong ways to integrate AI and your application? And what 290 00:13:50.759 --> 00:13:52.480 are some of the right ways, And you kind of 291 00:13:52.480 --> 00:13:54.440 got into, you know, breaking the problem up and things 292 00:13:54.480 --> 00:14:00.960 like that. But then, yeah, Charles G said, AI is 293 00:14:01.080 --> 00:14:04.720 only reliable for prototyping and search. It feels like vaporware 294 00:14:04.759 --> 00:14:09.919 for most other stuff. It's that use models, you know, 295 00:14:10.000 --> 00:14:13.519 gotten off the hype train, and I think some of 296 00:14:13.519 --> 00:14:16.799 the stuff is going to get better. But yeah, so 297 00:14:16.799 --> 00:14:19.960 so you you talked about these specific instances, but what 298 00:14:19.960 --> 00:14:23.360 what are the problems that it solves? Well, like before 299 00:14:23.360 --> 00:14:27.000 do I look at my stack and go okay, I 300 00:14:27.080 --> 00:14:29.320 want I wanted to hit here, and you know, maybe 301 00:14:29.360 --> 00:14:31.039 I'm going to hit you know, this and this and 302 00:14:31.080 --> 00:14:32.919 this and this and this and this, and I'll have 303 00:14:32.960 --> 00:14:35.440 different models or different l ms that I hit or whatever. 304 00:14:35.840 --> 00:14:38.360 But how do I know that this is a good fit? 305 00:14:39.000 --> 00:14:41.759 And then how do I start putting that in there? 306 00:14:42.240 --> 00:14:45.399 Because it does feel like some things are moving this 307 00:14:45.519 --> 00:14:48.000 way other things. Yeah, it's not there yet. 308 00:14:48.039 --> 00:14:52.960 But so before Steve answers, and I'll definitely let Steve answer, 309 00:14:53.000 --> 00:14:55.519 I just wanted to mention that one of the catalysts 310 00:14:55.519 --> 00:14:59.320 for this entire conversation is an excellent blog post that 311 00:14:59.440 --> 00:15:05.159 Steve ro vote and posted on the Builder blog, which 312 00:15:05.679 --> 00:15:08.600 is titled how to build AI Products that Don't Flop. 313 00:15:09.240 --> 00:15:12.600 So I think that in a lot of ways, what Steve, 314 00:15:13.039 --> 00:15:16.159 what I assume you'll be saying and and kind of 315 00:15:16.200 --> 00:15:18.679 also addresses a lot of the issues that were brought 316 00:15:18.799 --> 00:15:22.679 up the stuff that you actually wrote in that blog post. 317 00:15:23.279 --> 00:15:25.639 So first of all, I I you know, we will 318 00:15:25.679 --> 00:15:30.840 post that blog post here in the chat, so and 319 00:15:31.360 --> 00:15:34.200 I highly recommend for people to go and check that out. 320 00:15:35.080 --> 00:15:38.200 And and it also have to say that it also 321 00:15:38.360 --> 00:15:41.480 resonated a lot with me because of stuff that we 322 00:15:41.559 --> 00:15:44.519 are doing at the size sense, which is the company 323 00:15:44.559 --> 00:15:47.720 that I recently joined, which turns out has a very 324 00:15:47.759 --> 00:15:52.080 similar philosophy and how we're using AI in our own products. Uh, 325 00:15:52.120 --> 00:15:54.960 And maybe I'll touch on that after you kind of 326 00:15:55.000 --> 00:15:56.000 answer the questions. 327 00:15:56.919 --> 00:16:01.919 Yes, no, this is a great question, and so you know, 328 00:16:02.679 --> 00:16:04.720 at the end of the day, maybe it starts by 329 00:16:04.720 --> 00:16:09.000 covering like what are the AIS bad at and why, 330 00:16:09.080 --> 00:16:11.159 and that can help us distill down what they're good 331 00:16:11.200 --> 00:16:14.519 at as an alternative and why. So what they are 332 00:16:14.679 --> 00:16:18.720 bad at is the biggest problem that lms have is 333 00:16:19.039 --> 00:16:21.200 what people call hallucinations. I don't love that term, but 334 00:16:21.200 --> 00:16:23.600 it's the term people use where they just make things up, 335 00:16:23.639 --> 00:16:26.039 and they make it up confidently. And it's probably a 336 00:16:26.080 --> 00:16:28.399 result of how they're trained. They're just trained on lots 337 00:16:28.399 --> 00:16:31.879 of information of people saying you know, they're just saying things. 338 00:16:32.080 --> 00:16:33.559 You know, they're saying things as if they're. 339 00:16:33.399 --> 00:16:36.440 Factually kids say that. It's the same approach that I 340 00:16:36.559 --> 00:16:37.759 take with most everything. 341 00:16:39.600 --> 00:16:41.480 Just say what I think as if it's certain. 342 00:16:41.519 --> 00:16:45.120 I say what they think, and I say it very confidently, exactly. 343 00:16:45.240 --> 00:16:48.720 Yes, I know many people in my life I probably 344 00:16:48.840 --> 00:16:50.759 are in one as well, who will do that as well. 345 00:16:50.799 --> 00:16:53.480 Whatever you think, whatever kind of sounds right. I got 346 00:16:53.480 --> 00:16:55.039 complete confidence in it when's saying it out. 347 00:16:55.360 --> 00:16:57.120 I've never done that everything. 348 00:16:58.399 --> 00:17:00.279 So you think about it, what's the training data? It 349 00:17:01.080 --> 00:17:04.799 is an AI trained on emulating the training data and 350 00:17:04.880 --> 00:17:06.839 training data as humans saying things if they know everything. 351 00:17:06.880 --> 00:17:08.640 So the AI just says things as if it knows everything. 352 00:17:09.680 --> 00:17:14.039 In the however many billion parameters they use, they can't 353 00:17:14.039 --> 00:17:16.720 store all the information of the world. And it seems 354 00:17:16.759 --> 00:17:18.440 like two date, people have still not figured out a 355 00:17:18.480 --> 00:17:20.960 liable way of having LMSA. Sorry, I don't know the 356 00:17:21.039 --> 00:17:23.599 answer to that. They will just make up dates and 357 00:17:23.640 --> 00:17:25.960 times and answers and stuff like that. It's very annoying. 358 00:17:26.480 --> 00:17:29.240 Well, I just want to chime in here because this 359 00:17:29.279 --> 00:17:33.160 has always been a problem with AIS. Right is you 360 00:17:33.240 --> 00:17:36.160 have a certain probability of not getting a right answer. 361 00:17:36.200 --> 00:17:39.839 The difference is is that the answers we're looking for 362 00:17:39.920 --> 00:17:43.799 now are a fully written out email or code or 363 00:17:43.799 --> 00:17:48.279 things like that right where you can be mostly okay 364 00:17:48.960 --> 00:17:51.119 except for these couple of things, where in the past 365 00:17:51.160 --> 00:17:55.079 it was generally something like AI vision or something like that. 366 00:17:55.119 --> 00:17:58.200 And so if it if it didn't always identify the 367 00:17:58.200 --> 00:18:01.160 dog as a dog, people would just kind of you know, 368 00:18:01.319 --> 00:18:04.920 as long as it was generally accurate, it was useful. 369 00:18:05.240 --> 00:18:09.880 And now it's it's problematic because it's generally accurate, but 370 00:18:09.920 --> 00:18:10.680 that's not good. 371 00:18:10.559 --> 00:18:14.759 Enough exactly, or especially if you think about it. Yeah, 372 00:18:14.799 --> 00:18:17.279 a good example is we have like one of those 373 00:18:17.319 --> 00:18:20.039 apps that lets us know when an animal goes in 374 00:18:20.039 --> 00:18:21.839 front of the camera at home, so we think it's 375 00:18:21.839 --> 00:18:23.759 our dog and turns out it was just a shadow. 376 00:18:24.319 --> 00:18:26.200 But it's like, who cares? Like whatever, the AI is wrong? 377 00:18:26.240 --> 00:18:28.640 Who cares? When it's not a who cares is when 378 00:18:28.640 --> 00:18:30.839 it's an essential part of a product. The product has 379 00:18:30.880 --> 00:18:33.319 a flow from start to finish, and the LM is 380 00:18:33.319 --> 00:18:36.279 wrong five percent of the time, that's a huge problem, 381 00:18:36.359 --> 00:18:39.200 especially when you use lms maybe for multiple steps. 382 00:18:39.160 --> 00:18:41.640 Or if it's strong one hundred percent of the time, 383 00:18:41.680 --> 00:18:43.400 but to a five percent extent. 384 00:18:44.200 --> 00:18:47.759 Yeah, exactly exactly, that's a problem. Those add up if 385 00:18:47.759 --> 00:18:50.240 you think about everybody loves the idea of AI agent's 386 00:18:50.240 --> 00:18:52.920 like a complete multi step tasks and to end, well, 387 00:18:52.960 --> 00:18:55.559 those little errors compound to become big problems. And if 388 00:18:55.599 --> 00:18:58.680 you've never used something like autogpt, it derails and it's 389 00:18:58.720 --> 00:19:00.839 a big problem. It's our generator a mess of data 390 00:19:01.160 --> 00:19:02.839 has no clue it's been off the rails for the 391 00:19:02.920 --> 00:19:05.640 last hour. It just makes it worse. And that's a 392 00:19:05.720 --> 00:19:08.039 huge issue with with this kind of future we want. 393 00:19:08.720 --> 00:19:12.519 But there are solutions to this. So actually there's two 394 00:19:12.559 --> 00:19:17.640 solutions that we found to be extremely effective. One has 395 00:19:17.680 --> 00:19:19.440 to do with actually that blog post you mentioned, Dan, 396 00:19:19.480 --> 00:19:22.680 and one has to do with some like microagent techniques. 397 00:19:22.680 --> 00:19:25.519 We've been really investing behind both in an open source 398 00:19:25.559 --> 00:19:29.519 project we've recently open sourced and some work we're doing 399 00:19:29.599 --> 00:19:32.279 internally in the product. So solution one that works fantastic. 400 00:19:32.519 --> 00:19:35.680 Let me give an example. Builder has lots of docks. 401 00:19:36.200 --> 00:19:38.759 People don't know the answers to the questions, and it's 402 00:19:38.799 --> 00:19:40.440 tedious to try and come through all the docks to 403 00:19:40.440 --> 00:19:42.799 find your answers. No matter how we try and restructure 404 00:19:42.839 --> 00:19:44.799 the docks or surface the right information the right time, 405 00:19:45.000 --> 00:19:47.640 it's never good enough. And I know that when I 406 00:19:47.680 --> 00:19:49.440 use other people's products too. 407 00:19:49.759 --> 00:19:51.680 You're not the only ones exactly. 408 00:19:52.640 --> 00:19:56.480 It's everybody's problem. It's it's it's it's very difficult. And 409 00:19:56.519 --> 00:19:59.920 so we're like, okay, let's feed all of the information 410 00:20:00.039 --> 00:20:04.920 of our docks into you know, the context window for 411 00:20:04.960 --> 00:20:07.839 an ll M and that's the big change that's happened recently. 412 00:20:07.960 --> 00:20:10.880 The context one is Windows have gotten freaking massive, and 413 00:20:10.920 --> 00:20:15.519 that's huge. I've even seen papers on using large the 414 00:20:15.599 --> 00:20:19.079 large context windows is more effective than fine tunings. Rather 415 00:20:19.079 --> 00:20:21.880 than fine tuning with hundreds or thousands of examples, just 416 00:20:21.880 --> 00:20:24.680 fit ten examples in the large context and you'll outperform. 417 00:20:25.000 --> 00:20:26.839 But they didn't even mention, or maybe they did and 418 00:20:26.839 --> 00:20:27.319 I missed it. 419 00:20:27.400 --> 00:20:30.039 The biggest thing is just one of the things talked 420 00:20:30.039 --> 00:20:31.960 about in the beautiful too. 421 00:20:32.680 --> 00:20:34.839 Yeah, and so what's what's practical? Because assuming we're talking 422 00:20:34.920 --> 00:20:37.200 jobs script devs, we're talking practitioners. They want to use 423 00:20:37.240 --> 00:20:39.720 the AI. The practical benefit of that is you don't 424 00:20:39.759 --> 00:20:41.720 need to assemble those thousands informations. You don't have to 425 00:20:41.759 --> 00:20:44.000 make a separate fine tune model instance for every use case. 426 00:20:44.160 --> 00:20:46.640 You can assemble things on the fly or make changes 427 00:20:46.680 --> 00:20:49.839 in experiment at a faster rate. It's it's really opens 428 00:20:49.920 --> 00:20:53.400 up a lot of benefits. And so what didn't work well? 429 00:20:54.119 --> 00:20:56.000 So here's if you go to chat ChiPT and ask it, 430 00:20:56.480 --> 00:20:59.319 ask it a question like how do I add a 431 00:20:59.559 --> 00:21:02.599 user in builder do io via API? It'll tell you 432 00:21:02.640 --> 00:21:05.079 go to build a dio slash API, slash users send 433 00:21:05.079 --> 00:21:07.319 a post that API doesn't exist. There's not an API 434 00:21:07.400 --> 00:21:09.599 to add users to your account and builders through APIs, 435 00:21:09.599 --> 00:21:11.559 though it will tell you that it exists. That's a 436 00:21:11.599 --> 00:21:15.400 huge problem. So then if you then try and augments 437 00:21:15.440 --> 00:21:18.359 this with okay, well we're going to take chatchipt or 438 00:21:18.400 --> 00:21:22.039 anthropicalaud APIs and we're going to supply our API documentation 439 00:21:22.359 --> 00:21:27.039 into it. Maybe we'll do some fancy based on embeddings 440 00:21:27.079 --> 00:21:29.839 and semantic search to find the right docs to include 441 00:21:29.839 --> 00:21:33.680 in that context and send it. It still will it'll 442 00:21:33.680 --> 00:21:35.920 get better, but it still will tell you to add 443 00:21:35.960 --> 00:21:38.720 a user. Go to API v one slash users and 444 00:21:38.759 --> 00:21:40.720 so the thing that works extremely well. So what the 445 00:21:40.759 --> 00:21:43.720 elms are really good at in my experience is condensing 446 00:21:43.759 --> 00:21:46.920 information down. So if you tell it very clearly, here's 447 00:21:46.960 --> 00:21:49.640 a set of information. You are only allowed to answer 448 00:21:49.720 --> 00:21:53.559 questions using this information, and if that information has the answer, 449 00:21:53.559 --> 00:21:55.680 condense it down and provide the answer. If it does not, 450 00:21:55.839 --> 00:21:58.079 say I don't have an answer to this, it probably 451 00:21:58.079 --> 00:22:01.880 doesn't exist. That works wildly well in our experience. Whether 452 00:22:01.880 --> 00:22:04.519 that's here's a lot of code, simplify the code, whether 453 00:22:04.559 --> 00:22:07.160 that's here's a transcript, tell me the key points discussed, 454 00:22:07.519 --> 00:22:11.400 whether that's here's the API documentation. Answer questions about the APIs. 455 00:22:11.640 --> 00:22:14.240 As long as you firmly say you could only use 456 00:22:14.279 --> 00:22:17.599 this information and nothing else, it works. I mean almost 457 00:22:17.599 --> 00:22:20.000 one hundred percent, and I mean almost in terms of 458 00:22:20.039 --> 00:22:21.960 like LM still can surprise you from time to time, 459 00:22:22.440 --> 00:22:24.960 use a good one, So like don't use I hope 460 00:22:24.960 --> 00:22:27.119 peop aren't using GPT three or three point five anymore. 461 00:22:27.119 --> 00:22:29.559 For most things there's better options. But if you're using 462 00:22:29.640 --> 00:22:31.759 something like anthropoclout three point five's on it and you're 463 00:22:31.799 --> 00:22:34.119 giving it lots of information more than a needs, I'm 464 00:22:34.119 --> 00:22:36.680 telling it to reduce and only use that information. It 465 00:22:36.720 --> 00:22:39.720 works great. Another example was so we have this assistant 466 00:22:39.759 --> 00:22:41.759 in our docks. It answers your questions and it works 467 00:22:41.759 --> 00:22:43.599 a lot better. Another thing that started doing, though, is 468 00:22:43.680 --> 00:22:47.119 making up links, like oh I added to the instructions 469 00:22:47.160 --> 00:22:49.319 at one point, like include links as much as possible, 470 00:22:49.599 --> 00:22:51.319 because a lot of what you're doing, or a lot 471 00:22:51.319 --> 00:22:53.200 of what we hope AI can help with the navigating 472 00:22:53.240 --> 00:22:57.200 docs is just knowing what for my use case, what 473 00:22:57.279 --> 00:22:59.799 docs I need to read, and so I told it 474 00:23:00.000 --> 00:23:02.000 will the answer with links. Maybe just a way to 475 00:23:02.039 --> 00:23:04.279 quickly find the right links is a good, good example. 476 00:23:04.599 --> 00:23:06.599 But I started making up links. So what did I do? 477 00:23:06.759 --> 00:23:10.759 I added to the prompt. I took our site map 478 00:23:10.880 --> 00:23:13.640 essentially and some additional contexts on each link, and I 479 00:23:13.680 --> 00:23:17.160 fed into the prompt and said use links only these links, 480 00:23:17.519 --> 00:23:20.440 nothing else. And then fantastic. It always links to things 481 00:23:20.440 --> 00:23:22.519 that are relevant. The more you do that, the more 482 00:23:22.519 --> 00:23:25.039 you tell it to do nothing more than cadets information 483 00:23:25.119 --> 00:23:28.119 and use nothing more than information provided, and just add 484 00:23:28.119 --> 00:23:30.920 a copious amount of information. Because these context windows are massive, 485 00:23:31.319 --> 00:23:34.079 it can work phenomenally well. So that's a use case 486 00:23:34.119 --> 00:23:37.000 that can be great. Another example of that is how 487 00:23:37.000 --> 00:23:40.640 our AI designed to code works is we have old 488 00:23:40.680 --> 00:23:44.799 school code, and old school models generate fairly accurate code 489 00:23:44.799 --> 00:23:47.759 for a design. It's just verbose and ugly. It's just 490 00:23:47.759 --> 00:23:51.279 too much. It's one massive component of div soup and 491 00:23:51.319 --> 00:23:53.119 the classes are named like div one v two to 492 00:23:53.200 --> 00:23:55.599 three to four. If you take that and pass it 493 00:23:55.640 --> 00:23:57.960 to an LM and say, just reorganize this into multiple 494 00:23:58.000 --> 00:24:02.039 files and components, well named renamed classes according et cetera. Again, 495 00:24:02.160 --> 00:24:04.359 give it a model like the latest two point five 496 00:24:04.400 --> 00:24:06.319 son it or GPT four is pretty good, but song's 497 00:24:06.359 --> 00:24:09.440 really good. It's just a fantastic job. And so we've 498 00:24:09.480 --> 00:24:11.880 spent an enormous at the time in the past trying 499 00:24:11.920 --> 00:24:15.200 to take generated code and try and make up class names. 500 00:24:15.279 --> 00:24:16.759 How the heck you're going to do that. That's back 501 00:24:16.759 --> 00:24:20.640 into that bucket of meaning. We don't know meaning in 502 00:24:20.720 --> 00:24:23.359 any type of code or model before at LLM really 503 00:24:23.400 --> 00:24:25.359 for this type of use case. But if you can 504 00:24:25.359 --> 00:24:28.279 distill down that way, it works fantastically well. And then 505 00:24:28.319 --> 00:24:31.079 the last piece, the last piece we've learned here, and 506 00:24:31.119 --> 00:24:33.279 i'd love to touch on the agent piece because there's 507 00:24:33.279 --> 00:24:36.079 one other technique that's great. The last piece we've learned 508 00:24:36.279 --> 00:24:39.160 is the user interface matters a lot too, and there's 509 00:24:39.160 --> 00:24:42.240 a user interface pattern that is just almost always wrong, 510 00:24:42.359 --> 00:24:45.119 yet we see almost everyone jump to all the time, 511 00:24:45.160 --> 00:24:48.279 including us, which is you want to assume that the 512 00:24:48.279 --> 00:24:51.119 AI is never perfect, and especially because when we're talking 513 00:24:51.119 --> 00:24:55.599 about meaning and summarization and stuff like that, it will 514 00:24:55.640 --> 00:24:58.599 not get everything right the first time. Even if it's accurate, 515 00:24:58.599 --> 00:25:00.640 it may not be quite exactly what you want. So 516 00:25:00.680 --> 00:25:02.519 what you don't want is a UI like we used 517 00:25:02.519 --> 00:25:04.599 to have, which is like if I wanted to get 518 00:25:04.680 --> 00:25:07.119 builder to update my content in some way. It used 519 00:25:07.119 --> 00:25:08.880 to be like click a button, then you get a box, 520 00:25:08.960 --> 00:25:10.240 then you type in it and hit some of it, 521 00:25:10.279 --> 00:25:12.079 and the box goes away and it makes your update 522 00:25:12.359 --> 00:25:14.960 and it acts like you're done. Most likely you're not done. 523 00:25:15.039 --> 00:25:17.000 Most likely it got you closer, but you're not quite 524 00:25:17.039 --> 00:25:20.200 there at least some portion of the time we've actually 525 00:25:20.240 --> 00:25:23.119 been landing on. A chat interface is almost always the 526 00:25:23.200 --> 00:25:25.240 right interface for an LM, even if your use case 527 00:25:25.319 --> 00:25:28.799 is not chat. So if your use case is imperfect 528 00:25:28.839 --> 00:25:31.160 but designed to code, and here's code, you should have 529 00:25:31.200 --> 00:25:33.000 a chat interface next to it saying what would you 530 00:25:33.000 --> 00:25:35.000 like to change about the code? Oh, I forgot to mention, 531 00:25:35.039 --> 00:25:38.839 I'm using tailwinds update it. There's too many components or 532 00:25:38.839 --> 00:25:41.759 there's not enough components. You should be able to constantly iterate, 533 00:25:41.839 --> 00:25:44.839 and the chat format lets you maintain that context and 534 00:25:44.920 --> 00:25:46.799 never assume it has to be done the first time 535 00:25:46.920 --> 00:25:50.839 or what's coming soon for us is importing Figma designs 536 00:25:50.960 --> 00:25:53.799 and then like a Pigma prototype where you have like 537 00:25:54.319 --> 00:25:55.920 you know, you have this mockup of like clicking this 538 00:25:56.000 --> 00:25:58.680 launches this modal and does this thing and updates this data. Well, 539 00:25:58.680 --> 00:26:01.039 we're going to suck that in and make real but 540 00:26:01.400 --> 00:26:05.319 not every You know, sigma is not a spec it 541 00:26:05.480 --> 00:26:08.319 is a suggestion, right, and so we need to treat 542 00:26:08.359 --> 00:26:11.599 it that way. And it's a description, So we're going 543 00:26:11.680 --> 00:26:13.200 to treat it that way. We're going to make some 544 00:26:13.240 --> 00:26:16.359 assumptions and then when actually you want clicking this to 545 00:26:16.359 --> 00:26:17.759 do that, you should be able to say it and 546 00:26:17.839 --> 00:26:21.519 see that happen. So that's one big bucket of learnings. 547 00:26:22.079 --> 00:26:25.000 The other big bucket of learnings is if you want 548 00:26:25.039 --> 00:26:27.599 to use lllms on a loop, you know, you could 549 00:26:27.680 --> 00:26:30.079 describe an AI agents just an LLLM on a loop. 550 00:26:30.519 --> 00:26:33.200 Do a thing, analyze a thing, do the next thing, 551 00:26:33.400 --> 00:26:34.839 decide when you need to stop. 552 00:26:35.480 --> 00:26:37.880 Right. The only thing I would add is that whatever 553 00:26:37.920 --> 00:26:40.119 you're telling it gets added to the context window and 554 00:26:40.160 --> 00:26:42.880 things like that, so that it knows what you've already 555 00:26:42.920 --> 00:26:45.279 done and what you've already told it correct. 556 00:26:45.319 --> 00:26:47.680 And you could also add that it's taking actions at 557 00:26:47.680 --> 00:26:51.480 each step, so the prior thing feeds into the next thing, 558 00:26:51.519 --> 00:26:54.359 the next context, and it takes another action and it continuous. 559 00:26:55.279 --> 00:26:58.400 Those are I think people firmly understand how we will 560 00:26:58.400 --> 00:27:00.759 be able to do even more magn wild things of 561 00:27:00.799 --> 00:27:02.960 AI at some point, you know, like I mentioned, we 562 00:27:03.079 --> 00:27:06.359 envision a world where another example is like assign a 563 00:27:06.519 --> 00:27:09.720 Gerra ticket to Builder and have it just implement that thing, 564 00:27:09.799 --> 00:27:12.359 and then you take a look at that or the 565 00:27:12.400 --> 00:27:15.680 Slack example too. Here's our idea. Implement it because Builder 566 00:27:15.720 --> 00:27:17.319 has an API and could hook right up to your 567 00:27:17.359 --> 00:27:19.279 live app. You could have it. If it powers your 568 00:27:19.279 --> 00:27:22.279 homepage like it does for lots of customers like j 569 00:27:22.440 --> 00:27:25.200 Crew or Zapy or whatever, you could just tell it 570 00:27:25.240 --> 00:27:26.720 like hey, up to the homepage with this and they'll 571 00:27:26.759 --> 00:27:28.839 just go do it. Maybe it to be safe. Ruends 572 00:27:28.839 --> 00:27:30.599 it as a one percent test shows you in a 573 00:27:30.680 --> 00:27:32.400 day that the data is good, the ab test is 574 00:27:32.440 --> 00:27:34.799 winning some you know metric, and then you great scale 575 00:27:34.839 --> 00:27:36.599 it up. You know stuff like that. Awesome, Just be 576 00:27:37.039 --> 00:27:38.960 be my homide that helps me do TV stuff that 577 00:27:39.000 --> 00:27:42.920 I didn't feel like doing. The problem though with agents 578 00:27:42.960 --> 00:27:46.240 that that have to take a sequence of actions is 579 00:27:47.119 --> 00:27:49.720 how they derail. Those small errors compound, so you need 580 00:27:49.759 --> 00:27:53.480 some type of mechanism to bring them back and analyze 581 00:27:53.519 --> 00:27:56.160 are they off track or not. And so this is 582 00:27:56.400 --> 00:27:58.200 a technique we've been using. We call it like a 583 00:27:58.240 --> 00:28:01.960 microagent technique because it's about specialized and I think this 584 00:28:02.119 --> 00:28:04.519 is a topic that has to do with product development 585 00:28:04.519 --> 00:28:06.680 in general, which is, don't try and boil the ocean, 586 00:28:06.720 --> 00:28:10.240 like don't try and build Devin the world's do everything 587 00:28:10.319 --> 00:28:14.480 software engineer. Then suddenly you're solving everyone's problems simultaneously. It's 588 00:28:14.480 --> 00:28:17.880 an impossible task. It's not a good idea. Rather, if 589 00:28:17.880 --> 00:28:20.359 you're an agent for one specific type of thing, starting 590 00:28:20.400 --> 00:28:23.519 there and then building up over time through feedback, through iteration, 591 00:28:23.640 --> 00:28:26.319 through all that stuff. That's how I always believe in 592 00:28:26.359 --> 00:28:30.880 building products. It's work better for me. So we have 593 00:28:30.960 --> 00:28:34.720 this open source project called Microagent, where the technique we've 594 00:28:34.920 --> 00:28:38.119 we realized works really well is if on each step 595 00:28:38.720 --> 00:28:42.160 you can have something that's not AI. We're exploring if 596 00:28:42.200 --> 00:28:43.720 AI can be the step, but I haven't had good 597 00:28:43.720 --> 00:28:46.640 results yet. But something that's not AI can essentially test 598 00:28:46.759 --> 00:28:48.960 if that step was successful and if it was not, 599 00:28:49.119 --> 00:28:51.839 feed feedback in and then let the LM run again, 600 00:28:51.920 --> 00:28:53.440 so it has the feedback from the last iteration. 601 00:28:53.559 --> 00:28:59.759 So for example, if you said that you had the 602 00:29:01.079 --> 00:29:08.039 LM output explanations that were link heavy, and you said 603 00:29:08.039 --> 00:29:11.599 that you verified the links by limiting it to your 604 00:29:11.680 --> 00:29:14.880 site map, but Another thing you could do is obviously 605 00:29:15.000 --> 00:29:17.319 just look at the links at the response, test them 606 00:29:17.319 --> 00:29:19.640 out and if you get I don't know, five hundred 607 00:29:19.759 --> 00:29:23.640 or something, then you can feed that back and say okay, 608 00:29:23.720 --> 00:29:27.079 that link is broken. Don't provide that as an answer 609 00:29:27.160 --> 00:29:27.960 or something like that. 610 00:29:28.519 --> 00:29:31.000 Spot on. No, that's a perfect example, and that's what 611 00:29:31.000 --> 00:29:32.480 we've seen. So I'll give you a couple more use 612 00:29:32.480 --> 00:29:35.519 cases for this. So one how the microagent project works 613 00:29:35.559 --> 00:29:37.680 as a CLI and instead of going to like chat 614 00:29:37.759 --> 00:29:40.559 ept and saying giving me a give me some code 615 00:29:40.559 --> 00:29:45.000 that converts marked down to HTML, you instead well run 616 00:29:45.000 --> 00:29:48.279 the cli microagent and describe that convert marked on cod 617 00:29:48.319 --> 00:29:51.000 to htmail instead of just giving you code and it's 618 00:29:51.039 --> 00:29:53.359 your problem if it doesn't work. Instead, it generates a 619 00:29:53.359 --> 00:29:56.599 test first, and it'll have all these input output examples 620 00:29:56.599 --> 00:29:59.000 of marked down to htmail and then it'll say it 621 00:29:59.000 --> 00:30:00.759 doesn't look good. You can give feedback on the test, 622 00:30:00.839 --> 00:30:02.960 like this is wrong, or add more or less or whatever, 623 00:30:03.359 --> 00:30:07.599 or just say okay. Then it'll it'll right the test 624 00:30:08.039 --> 00:30:11.319 and then it'll write code, and every time it writes code, 625 00:30:11.359 --> 00:30:15.079 it'll run the test and any testing id feedback. It's 626 00:30:15.119 --> 00:30:18.480 a I TDD and it works really really well, especially 627 00:30:18.480 --> 00:30:21.720 for certain use cases that AI generally wasn't good at before, 628 00:30:22.440 --> 00:30:24.640 And so you get this sort of guarantee that with 629 00:30:24.799 --> 00:30:28.720 AI spitting out code, you don't have to any longer 630 00:30:29.839 --> 00:30:34.359 hope that that code works because it'll look good and 631 00:30:34.400 --> 00:30:37.000 it might work on one example but not others. The 632 00:30:37.079 --> 00:30:38.920 test will ensure it works on all the examples. And 633 00:30:38.920 --> 00:30:40.960 when I run microagent and have it generate big complex 634 00:30:41.000 --> 00:30:43.319 things for me, I rest easy that if it passed 635 00:30:43.319 --> 00:30:44.880 all the tests and the tests are pretty thorough that 636 00:30:44.880 --> 00:30:47.599 it generates it works, and then if we find any 637 00:30:47.640 --> 00:30:50.599 issue subsequently, we'll update the tests accordingly. So it actually 638 00:30:50.720 --> 00:30:53.799 is a much more like high confidence solution. One other 639 00:30:53.839 --> 00:30:58.160 example is Figma has this concept called component sets, where 640 00:30:58.359 --> 00:31:01.279 it's it's like how it represents a button with all 641 00:31:01.359 --> 00:31:06.279 kinds of variations, you know, like color primary. It's a 642 00:31:06.279 --> 00:31:09.440 design system feature exactly. It's how you have something similar 643 00:31:09.440 --> 00:31:11.359 to React opponents and Figma, so you can place the 644 00:31:11.359 --> 00:31:14.640 button and change the color from primary secondary, you know, 645 00:31:14.720 --> 00:31:17.839 aero state, true, you know whatever, it's cool. It's weird 646 00:31:17.880 --> 00:31:19.799 as hell how it works in Figma and Figma you 647 00:31:19.799 --> 00:31:22.519 actually design so if you have like three different props 648 00:31:22.519 --> 00:31:24.920 with three different options each. You have to make nine designs. 649 00:31:24.920 --> 00:31:28.599 You have to manually code up or design every possible combination. 650 00:31:28.680 --> 00:31:31.799 It does us weird, it's funky, but it's it's effective. 651 00:31:32.720 --> 00:31:34.920 So how you translate that to code. What we do 652 00:31:35.200 --> 00:31:39.720 is we generate baseline code for every single variant and 653 00:31:39.759 --> 00:31:43.000 then we tell the LM consolidate down this all down 654 00:31:43.000 --> 00:31:46.079 into one piece of code with props you know, and react. 655 00:31:46.359 --> 00:31:49.240 And what we do is to verify the LM did 656 00:31:49.240 --> 00:31:52.799 that correctly, we will run it through tests. So because 657 00:31:52.839 --> 00:31:55.440 we know what the end state of every combination of 658 00:31:55.480 --> 00:31:57.720 prop should look like, we then take what the LM 659 00:31:57.759 --> 00:32:00.200 provide it and test it against every end state, give 660 00:32:00.200 --> 00:32:02.279 it every combination of props and make sure that the 661 00:32:02.880 --> 00:32:06.240 essentially the snapshot is correct to the original spec and 662 00:32:06.279 --> 00:32:09.119 if it's not, we feed the feedback in and if 663 00:32:09.160 --> 00:32:11.079 we've actually found you can play with it. You can 664 00:32:11.119 --> 00:32:13.480 either feed this to a slow high quality model like 665 00:32:13.519 --> 00:32:17.000 before three point five sonnet with anthropic there was claud Opus, 666 00:32:17.559 --> 00:32:19.640 which it was not obvious. I've had way a better 667 00:32:19.640 --> 00:32:22.079 experience with anthropic models in general than open A models. 668 00:32:22.559 --> 00:32:25.319 As becoming more popular now and more discussed but it 669 00:32:25.359 --> 00:32:26.799 was a little bit of a hot take in the 670 00:32:26.839 --> 00:32:30.640 past or less common knowledge. But anyway, you can play 671 00:32:30.640 --> 00:32:32.599 with the knobs. You can either say we're going to 672 00:32:32.680 --> 00:32:35.119 do the big, expensive model and probably takes less iterations, 673 00:32:35.480 --> 00:32:38.200 or you can scale down to like we actually had 674 00:32:38.200 --> 00:32:41.240 good results with Anthropic Haiku their smallest model, and it 675 00:32:41.240 --> 00:32:43.119 would take four or five iterations, but that would run 676 00:32:43.200 --> 00:32:45.279 faster than the two and three on opus and you're 677 00:32:45.319 --> 00:32:47.920 still guaranteed an accurate result at the end. So if 678 00:32:47.960 --> 00:32:51.640 you have that automated check that can feed into the AI, 679 00:32:52.200 --> 00:32:53.960 it could work really well. And that's kind of my 680 00:32:54.000 --> 00:32:58.079 point again about hopefully this illustrates good examples of a 681 00:32:58.240 --> 00:33:00.720 I can't be the entirety the brain of your product, 682 00:33:01.200 --> 00:33:03.880 but you can isolate these different techniques and use it 683 00:33:03.920 --> 00:33:05.920 to accomplish things that just would have been kind of 684 00:33:06.000 --> 00:33:06.920 impossible previously. 685 00:33:06.920 --> 00:33:10.000 So to speak, related to what you just said, one 686 00:33:10.039 --> 00:33:13.759 of the most salient points that you made in the 687 00:33:13.759 --> 00:33:16.039 blog post, and it's kind of related to what you 688 00:33:16.160 --> 00:33:19.720 just said right now, is that in all cases, it's 689 00:33:19.799 --> 00:33:24.839 not about building your product to be Holy Eye centric 690 00:33:24.920 --> 00:33:29.480 with a thin layer around it. It's taking some existing 691 00:33:30.319 --> 00:33:35.480 service product that solves a real problem and then sprinkling 692 00:33:35.599 --> 00:33:41.359 in AI to make it better. And so it's the 693 00:33:41.440 --> 00:33:49.480 reverse of what the VCS might have liked, but it's 694 00:33:49.720 --> 00:33:53.319 much more grounded and down to earth. And like I said, 695 00:33:53.359 --> 00:33:58.480 where I'm seeing very similar thing. At Sizense, we kind 696 00:33:58.480 --> 00:34:05.359 of build dashboards and stuff for BI and so in 697 00:34:05.400 --> 00:34:09.880 the past it was like, you know, you basically need 698 00:34:09.960 --> 00:34:13.719 to either best case scenario or dragon drop do you 699 00:34:13.840 --> 00:34:17.400 build your dashboard or you might you know, do some 700 00:34:17.440 --> 00:34:21.719 coding to build your dashboard. Well it's now we're working on, 701 00:34:21.880 --> 00:34:23.920 like you said, the sort of a chat thing where 702 00:34:23.960 --> 00:34:27.639 you can describe the dashboard that you that you like 703 00:34:28.239 --> 00:34:30.360 and you get it. But at the end of the day, 704 00:34:30.400 --> 00:34:34.199 it uses all the components and know how in capabilities 705 00:34:34.599 --> 00:34:40.599 that we've already built for constructing dashboards. So it's so 706 00:34:40.880 --> 00:34:46.880 it's it's adding AI to existing infrastructure to make it 707 00:34:47.079 --> 00:34:50.800 better and to make it more approachable, I might say, 708 00:34:51.440 --> 00:34:57.119 and easier to configure and without you know, specialized knowledge. 709 00:34:58.000 --> 00:35:00.880 It what you're talking about, Dan, And it sounds like 710 00:35:01.800 --> 00:35:04.559 it's a different level of interactivity. 711 00:35:04.239 --> 00:35:06.760 Right Look at the end of the day, I you know, 712 00:35:06.920 --> 00:35:11.599 even like then you when Google came out, I say, 713 00:35:12.039 --> 00:35:15.679 what was it? It's twenty something years ago. I said, 714 00:35:15.800 --> 00:35:19.559 why aren't are you? Are you eyes like that? Instead 715 00:35:19.599 --> 00:35:22.559 of pulled down menus and stuff, just give me a 716 00:35:22.559 --> 00:35:25.760 text box where I can say what I want and 717 00:35:25.800 --> 00:35:28.840 the software does it. Well, we're finally kind of getting 718 00:35:28.880 --> 00:35:30.360 that in a way. 719 00:35:32.480 --> 00:35:35.760 I think I have a question related to this because 720 00:35:37.000 --> 00:35:39.800 you know, Dan says, we're finally starting to get to this, 721 00:35:40.440 --> 00:35:42.800 and it seems like, you know, we've talked about, Hey, 722 00:35:42.840 --> 00:35:47.079 we had GPT three, which was hmm, we had GPT 723 00:35:47.199 --> 00:35:53.440 three point five, which was okay, GPT four is pretty good. 724 00:35:56.360 --> 00:35:57.840 I mean, are we going to continue to see this 725 00:35:57.960 --> 00:36:00.559 kind of thing too, where you know, we we get 726 00:36:00.800 --> 00:36:04.440 more of the text boxer? I mean the demos and 727 00:36:04.800 --> 00:36:07.599 the demos kind of made me upset because you know, 728 00:36:07.800 --> 00:36:09.840 then you'd see people trying to do the demo where 729 00:36:09.840 --> 00:36:12.480 they were talking to GPT for and it wasn't working 730 00:36:12.519 --> 00:36:16.320 for them like the demo work. But you know, we're 731 00:36:16.360 --> 00:36:20.280 a lot closer. You know, how does this continue to advance? 732 00:36:21.840 --> 00:36:26.320 Yeah? I mean the assumption we're making is that it's 733 00:36:26.320 --> 00:36:28.960 hard to perfectly quantify this, but the malls will get 734 00:36:29.000 --> 00:36:32.400 something like ten better year reyear they'll get you know, 735 00:36:32.719 --> 00:36:37.639 some percent less hallucinations or weird hiccups or weird problems. 736 00:36:37.800 --> 00:36:39.880 If you have a ninety five percent success right, maybe 737 00:36:39.880 --> 00:36:41.679 next year you have a ninety six. Maybe next year 738 00:36:41.679 --> 00:36:44.800 of a ninety seven. And these things tend to slow down, 739 00:36:44.840 --> 00:36:47.960 though they also tend to be little s curves as well, 740 00:36:48.000 --> 00:36:53.039 Like I think one one not counter argument, but counterpoints 741 00:36:53.079 --> 00:36:56.760 to the idea of it slowing down completely is things 742 00:36:56.840 --> 00:36:59.360 like I don't know if you've used rock g RRO 743 00:37:00.000 --> 00:37:01.760 and I'm forgetting what spelling one is Elion mustling, but 744 00:37:01.800 --> 00:37:04.039 one is like a hardware company. A hardware company, I 745 00:37:04.039 --> 00:37:08.760 think it's jer Okay is super interesting because they can 746 00:37:08.840 --> 00:37:12.199 run Lama three at insane speeds. You type in their box, 747 00:37:12.239 --> 00:37:14.880 you just get tons of text quickly. And they were 748 00:37:14.920 --> 00:37:17.960 one of the first to realize at a full ability 749 00:37:17.960 --> 00:37:21.519 to commercialized level, you can make hardware optimized for LMS 750 00:37:21.960 --> 00:37:24.199 and it's not just like a small percent better, it's 751 00:37:24.280 --> 00:37:29.000 like ten times faster and cheaper, which is crazy. And 752 00:37:29.079 --> 00:37:31.559 so we are not currently using Rock because we need 753 00:37:31.679 --> 00:37:34.199 larger context windows and LAMA three and all the others 754 00:37:34.199 --> 00:37:36.719 don't support the huge context windows that like Opening Eye 755 00:37:36.719 --> 00:37:40.000 and anthropic have I'm hoping, I know, medicinal training LAMAI 756 00:37:40.159 --> 00:37:42.599 three four hundred B. I think it is. And so that, 757 00:37:42.760 --> 00:37:44.679 you know, maybe a magic there is a larger context window. 758 00:37:44.719 --> 00:37:48.079 I don't know that could be huge, but breakthrough innovations 759 00:37:48.119 --> 00:37:52.079 like that could accelerate things. One example of that is, like, hey, 760 00:37:52.079 --> 00:37:54.719 people love the idea of you know, as an example, 761 00:37:54.760 --> 00:37:58.599 because build has an API and SDK, it can dynamically 762 00:37:58.679 --> 00:38:01.639 render your reaction app and components. When you drag and 763 00:38:01.719 --> 00:38:03.559 drop or use AI to modify it, you can see 764 00:38:03.559 --> 00:38:05.159 it all in real time, so it feels like Figma, 765 00:38:05.280 --> 00:38:08.280 but it's actually your react or quick or view or 766 00:38:08.280 --> 00:38:11.519 whatever app. So that's cool. And then changes can turn 767 00:38:11.559 --> 00:38:13.760 the code or published in their live to your users, 768 00:38:13.760 --> 00:38:15.760 so you can get people who are not even developers 769 00:38:15.800 --> 00:38:18.920 making changes, pushing updates, marketing pages or whatever you want. 770 00:38:20.360 --> 00:38:23.039 But what people really love the idea of is well, 771 00:38:23.440 --> 00:38:25.599 why are we hard coding apps at all? Can't we 772 00:38:25.639 --> 00:38:28.519 make components and understand things about our users? And when 773 00:38:28.559 --> 00:38:31.719 I jump into you know, JCrew dot com, can it 774 00:38:31.800 --> 00:38:35.519 just dynamically produce an experience that's one to one, personalized 775 00:38:35.519 --> 00:38:38.599 and fitted for me. Now that's like Wow, that's cool. 776 00:38:38.920 --> 00:38:41.159 That's no, we're not even close to that. I've watched 777 00:38:41.159 --> 00:38:43.920 the code generate slowly enough times to say we're not 778 00:38:43.960 --> 00:38:46.440 even remotely there. But then when things like Groc come out, 779 00:38:46.440 --> 00:38:51.239 which is probably the first commercially successful commerciuccessfulse and like 780 00:38:51.280 --> 00:38:53.880 I see people using it, I'm using it myself from 781 00:38:53.920 --> 00:38:58.360 time to time. Breakthrough an LM performance that was a 782 00:38:58.440 --> 00:39:01.079 big step towards that. You know, you see that paper 783 00:39:01.119 --> 00:39:03.880 of like what is it one point five bit YadA 784 00:39:03.960 --> 00:39:06.159 YadA for LMS, and it's like, hey, you might take 785 00:39:06.159 --> 00:39:08.039 another le forward as well. I don't think Rock is 786 00:39:08.079 --> 00:39:09.840 using an architecture think used to something differently, so that 787 00:39:09.920 --> 00:39:13.360 might be a subsequent innovation to add on. Now it's like, okay, 788 00:39:13.400 --> 00:39:16.199 if we can do these in real time, maybe that 789 00:39:16.320 --> 00:39:18.519 is possible only for certain parts of your app. It's 790 00:39:18.519 --> 00:39:20.800 not you know, it's not like it's not like you 791 00:39:20.840 --> 00:39:24.079 deploy an empty repo and suddenly applications build themselves in 792 00:39:24.079 --> 00:39:26.480 front of users in real time. But like you can 793 00:39:26.519 --> 00:39:28.360 start small, Like we were working with this very large 794 00:39:28.360 --> 00:39:30.960 company on the use case where they want to just 795 00:39:30.960 --> 00:39:33.320 be able to type in a very common use case. 796 00:39:33.360 --> 00:39:36.519 They want to type in a query and fetch the 797 00:39:36.639 --> 00:39:38.960 data associated to whatever that is. So like, I just 798 00:39:39.000 --> 00:39:40.480 want to see this data, and it can pull the 799 00:39:40.519 --> 00:39:42.800 data and visualize it for you. And the way they're 800 00:39:42.800 --> 00:39:46.239 doing visualization is code generation. But just like any tool, 801 00:39:46.280 --> 00:39:51.800 this is very common, or they're using anthropic claude artifacts 802 00:39:52.039 --> 00:39:57.280 or tail draws make real, you always get just like funky, 803 00:39:57.280 --> 00:40:00.000 plain old raw HTML REACT. It's not like using your 804 00:40:00.000 --> 00:40:03.000 APIs your components, You're you're using tail whatever it's it's 805 00:40:03.039 --> 00:40:05.400 not using your it's not code. It's gonna go to production. 806 00:40:05.440 --> 00:40:09.320 It's all throw away code. But if you could instead 807 00:40:09.519 --> 00:40:12.360 be able to in real time, either offline or online, 808 00:40:12.719 --> 00:40:16.559 assemble what you have your pieces. Offline use case would 809 00:40:16.599 --> 00:40:19.800 be like you know, I'm just gonna import a page 810 00:40:19.840 --> 00:40:21.599 to Figma and then I'll hit publish and then I'll 811 00:40:21.639 --> 00:40:25.400 go online when I'm done editing. So that's ultimately the 812 00:40:25.440 --> 00:40:27.760 offline use case. The online use case would be showing 813 00:40:27.760 --> 00:40:29.920 me these things on demand. So this large company that 814 00:40:29.920 --> 00:40:33.480 wants to surface your data, wants to generate code effectively 815 00:40:33.559 --> 00:40:35.840 with your components, Well, what can do that? We already 816 00:40:35.840 --> 00:40:38.519 have these SDKs where you if they're aware of your components. 817 00:40:38.559 --> 00:40:41.519 They can dynamically render out in real time your stuff 818 00:40:42.440 --> 00:40:44.880 we're exploring, like, hey, yeah, you queer your stuff and 819 00:40:44.920 --> 00:40:48.000 you've it's aware of your just the components you want 820 00:40:48.039 --> 00:40:50.239 it to be for this use case charts, diagrams, pipe 821 00:40:50.320 --> 00:40:53.239 chart table, et cetera. I can just dynamically produce u 822 00:40:53.280 --> 00:40:56.360 eys based on your specific query in pretty real time 823 00:40:57.039 --> 00:41:00.400 using LAMA three on rock. That actually works pretty well. 824 00:41:00.760 --> 00:41:04.639 So that's sort of like your first like online on 825 00:41:04.760 --> 00:41:08.280 demands generation and that makes sense for loged users maybe 826 00:41:08.320 --> 00:41:10.119 paying a certain amount per month, you know, that could 827 00:41:10.159 --> 00:41:13.159 be justified by the cost of serving that. But if 828 00:41:13.159 --> 00:41:16.119 we keep making these innovations on cost and speed, et cetera, 829 00:41:16.360 --> 00:41:19.480 you could get to a world where online generation of 830 00:41:19.519 --> 00:41:21.800 parts or larger parts. I mean, let's take if you're 831 00:41:21.840 --> 00:41:24.519 Jay Crew, maybe you want to manually merchandise that hero. 832 00:41:24.639 --> 00:41:26.119 You want to promote this new product lines as the 833 00:41:26.159 --> 00:41:29.280 hero is this thing, but maybe down below whatever Amazon 834 00:41:29.400 --> 00:41:31.960 highly personalizes the products you see. What if you could 835 00:41:31.960 --> 00:41:34.079 just throw that at the LM and the whole UI 836 00:41:34.559 --> 00:41:37.440 is based on what you'd be interested in seeing. Those 837 00:41:37.440 --> 00:41:38.960 are kind of interesting. I can't remember if there was 838 00:41:38.960 --> 00:41:40.840 a question here or what I was answering. But I 839 00:41:40.880 --> 00:41:42.920 want to throw that in because I think it's interesting 840 00:41:42.920 --> 00:41:43.960 direction that we might get to. 841 00:41:44.519 --> 00:41:46.920 Maybe we'll also get to a world where it's not 842 00:41:48.119 --> 00:41:54.800 just what's their name? The only company actually making money 843 00:41:54.840 --> 00:42:01.719 off of AI is Uh, I'm blanking out Microsoft. 844 00:42:03.199 --> 00:42:05.159 I was gonna say Microsoft. 845 00:42:04.920 --> 00:42:10.599 No, who makes the hardware? Oh yeah, and VideA and Vidia, 846 00:42:10.639 --> 00:42:13.360 And VideA is the only company actually making money off 847 00:42:13.400 --> 00:42:16.000 of the off of the AI revolution. 848 00:42:16.079 --> 00:42:18.280 Their stock price definitely yeah. 849 00:42:17.920 --> 00:42:21.079 Yeah, yeah, because you raise money as a startup and 850 00:42:21.119 --> 00:42:24.880 then spend all that money paying in VideA. 851 00:42:24.960 --> 00:42:27.559 You know this this kind of gets into one of 852 00:42:27.599 --> 00:42:30.079 the other questions I had, and you mentioned this. The 853 00:42:30.320 --> 00:42:34.119 previous question that you were answering was you know, you know, 854 00:42:34.159 --> 00:42:36.119 how does this continue to advance? And I think you 855 00:42:36.159 --> 00:42:40.400 pretty well answered that. Beyond that, you kind of got 856 00:42:40.440 --> 00:42:46.159 into more of the arena of cost and speed right 857 00:42:46.159 --> 00:42:51.239 as opposed to capability, And of course, costs and speed 858 00:42:51.360 --> 00:42:54.360 kind of play into capability, right because if it has 859 00:42:54.400 --> 00:42:56.679 infinite costs, then I can't provide it to my customer 860 00:42:56.760 --> 00:43:01.719 unless they have an infinite bank account. And similarly, you know, 861 00:43:01.880 --> 00:43:05.159 if it's not fast enough, then again, you know, it 862 00:43:05.559 --> 00:43:09.280 lowers the utility. So one of the things that I've 863 00:43:09.280 --> 00:43:13.239 talked to some people who are beginning to adopt AI 864 00:43:14.119 --> 00:43:18.639 features into their stuff is yeah, up to a certain point, 865 00:43:18.880 --> 00:43:22.719 it you know, it's it's great, and then it gets expensive. Right, 866 00:43:23.239 --> 00:43:27.000 So so how do you start to how do you 867 00:43:27.000 --> 00:43:29.360 start to manage some of that. 868 00:43:29.360 --> 00:43:31.800 That's a great question. That's something we've looked a lot 869 00:43:31.840 --> 00:43:34.800 at in a few different kind of findings of ours. 870 00:43:36.119 --> 00:43:37.880 First one, I could tell you, just from a consumer 871 00:43:37.920 --> 00:43:40.360 point of view, I hate the idea of all these 872 00:43:40.360 --> 00:43:42.679 different applications trying to charge me another twenty dollars per 873 00:43:42.719 --> 00:43:45.239 month per user for their AI features. Hey, right, you 874 00:43:45.280 --> 00:43:46.119 know part of me. 875 00:43:46.039 --> 00:43:47.840 Is like you've solved my problem, thank you. 876 00:43:48.679 --> 00:43:52.280 Yeah, I don't want to do for it exactly. And 877 00:43:52.320 --> 00:43:55.920 it's like the part that kills me is and I 878 00:43:55.960 --> 00:43:58.000 know no product should ever do this. It just doesn't 879 00:43:58.000 --> 00:44:00.280 make sense. But I'm like, I know you charge me 880 00:44:00.320 --> 00:44:03.159 that amount because the LM is expensive. Can I just 881 00:44:03.199 --> 00:44:06.079 supply everyone my open AI or anthropic key and they 882 00:44:06.119 --> 00:44:08.400 can just build me based on usage and you can 883 00:44:08.480 --> 00:44:10.719 just make sure the feet that the way it works 884 00:44:10.760 --> 00:44:13.519 with your products really well. I don't imagine anybody doing that. Obviously, 885 00:44:13.519 --> 00:44:16.360 when we make open source AR projects, we have several Yes, 886 00:44:16.440 --> 00:44:18.559 you just supply your key and then it works that way. 887 00:44:19.960 --> 00:44:21.719 But yeah, those things really add up, and in a 888 00:44:21.719 --> 00:44:23.639 lot of cases, I think the companies are just trying 889 00:44:23.679 --> 00:44:26.360 to make sure that their bottom line is taken care of. 890 00:44:26.440 --> 00:44:28.119 They know if they don't charge twenty dollars a month, 891 00:44:28.119 --> 00:44:30.280 if you use this heavily, they might be underwater on you, 892 00:44:30.320 --> 00:44:32.760 and that's a big problem. And a SaaS company usually 893 00:44:32.800 --> 00:44:36.119 wants to have eighty percent margins, So if you're going 894 00:44:36.199 --> 00:44:36.679 to cost. 895 00:44:36.480 --> 00:44:38.920 Them and we're back to the we're back to the 896 00:44:39.000 --> 00:44:41.559 uber model, like raise a whole lot of money and 897 00:44:41.599 --> 00:44:44.880 then basically subsidize your users. 898 00:44:45.440 --> 00:44:47.639 That's how a co pilot started out. They were losing 899 00:44:47.679 --> 00:44:50.400 money on a per user basis, but aren't anymore, at 900 00:44:50.480 --> 00:44:54.119 least that's what was reported, And so yeah, I think 901 00:44:54.119 --> 00:44:57.039 we're definitely seeing that across the board. Also a lot 902 00:44:57.079 --> 00:44:59.519 of startups losing money on the training costs and all 903 00:44:59.519 --> 00:45:02.199 the other as well. You know from VC money, but 904 00:45:03.519 --> 00:45:06.920 you know, you could probably put a value on it 905 00:45:07.079 --> 00:45:10.079 doesn't even take that much math or mental exercise to 906 00:45:10.079 --> 00:45:13.639 put a value on various things. For instance, we just 907 00:45:13.760 --> 00:45:17.760 have our AI in our docs for free for anyone 908 00:45:17.800 --> 00:45:21.079 anyway can prompt and use it. In a lot of cases, 909 00:45:21.119 --> 00:45:22.840 you can one just ship it and see what it costs. 910 00:45:22.880 --> 00:45:25.639 Doesn't cost a ton, and two we actually know you know. 911 00:45:25.679 --> 00:45:29.320 Our platform is probably in the bucket of it's not 912 00:45:29.599 --> 00:45:32.360 trying to be like a fort seal, which is like 913 00:45:33.039 --> 00:45:35.719 it's one button and you never do anything. You never 914 00:45:35.760 --> 00:45:38.920 have to know anything. It's definitely there are things to 915 00:45:39.039 --> 00:45:41.400 learn and more power to get as you learn. 916 00:45:41.800 --> 00:45:44.480 Also, your model is different. I mean, at the end 917 00:45:44.519 --> 00:45:51.079 of the day, your model is is a customer every 918 00:45:51.559 --> 00:45:56.679 once in a while needs a I in order to 919 00:45:56.719 --> 00:46:01.199 translate their Figma designs into code. I assume they don't 920 00:46:01.239 --> 00:46:06.079 do that every day. Whereas firstyle is you're not even 921 00:46:06.119 --> 00:46:10.079 a paying customer yet. Here we'll show you how we 922 00:46:10.199 --> 00:46:14.039 take your whatever and use AI to turn it into whatever. 923 00:46:15.519 --> 00:46:20.320 It's a totally different model in terms in terms of 924 00:46:20.760 --> 00:46:23.440 the finances exactly. 925 00:46:23.559 --> 00:46:25.719 And that's where you need to make sure. You just 926 00:46:25.719 --> 00:46:29.559 have to work make sure that the value you're providing 927 00:46:29.599 --> 00:46:31.880 in the rate that the user needs AI to get 928 00:46:31.880 --> 00:46:35.119 that value is less than you're charging them and an 929 00:46:35.159 --> 00:46:37.440 ideal world. You know what we've done in a lot 930 00:46:37.440 --> 00:46:40.519 of cases is it's kind of back to that point 931 00:46:40.599 --> 00:46:42.440 of making sure you do as much as possible without 932 00:46:42.480 --> 00:46:45.360 the LM. It's going to be faster, cheaper, better for 933 00:46:45.639 --> 00:46:48.960 ninety nine percent of the flow between whatever input whatever 934 00:46:49.000 --> 00:46:51.360 output is needed. And so that's how we cut down 935 00:46:51.360 --> 00:46:54.199 on cost traumatically as well use our own trained models 936 00:46:54.199 --> 00:46:57.599 when we can. If you are in a world where 937 00:46:57.800 --> 00:47:00.480 you know, Charles, you were mentioning the example of like 938 00:47:01.280 --> 00:47:03.320 you might use multiple llms, and we've done that in 939 00:47:03.360 --> 00:47:07.360 some cases. Like there's various ways you can do this, 940 00:47:07.559 --> 00:47:10.159 but there's certain use cases where you might want an 941 00:47:10.360 --> 00:47:13.280 LM to plan the work first and then another LM 942 00:47:13.320 --> 00:47:16.480 to execute on the plan. That can be a really 943 00:47:16.480 --> 00:47:19.639 good one. It can be not expensive if you don't 944 00:47:19.679 --> 00:47:21.760 need to feed too much context in multiple times. The 945 00:47:21.800 --> 00:47:23.719 output of a plan is usually not a lot of tokens, 946 00:47:24.800 --> 00:47:26.480 but it can be. Let's say you needed a whole 947 00:47:26.519 --> 00:47:29.159 chain of lllms, or let's say you had a situation 948 00:47:29.239 --> 00:47:32.559 where LM one takes lots of context and provides a plan, 949 00:47:32.960 --> 00:47:35.159 and then you have different LMS executly each step with 950 00:47:35.199 --> 00:47:37.320 a plan. But they each need the large amount of context. 951 00:47:37.760 --> 00:47:39.760 Then it can start adding up and it's like, ah, 952 00:47:39.840 --> 00:47:42.519 you know, can we do that without LMS or can 953 00:47:42.559 --> 00:47:44.960 we do other things? In the design to code flow, 954 00:47:45.000 --> 00:47:49.239 there's a lot of different steps and for us, some 955 00:47:49.280 --> 00:47:50.719 of the steps that can't be done with code I 956 00:47:50.760 --> 00:47:53.880 mentioned we train our own models. Those are it's hard 957 00:47:53.880 --> 00:47:57.800 to underscore how drastically faster, cheap, and more reliable training 958 00:47:57.800 --> 00:48:00.079 your own models than an LLM if it's a fit 959 00:48:00.280 --> 00:48:02.480 use case, and the fitting use cases might be more 960 00:48:02.480 --> 00:48:04.679 out outside of the box than you might think. Like, Yes, 961 00:48:04.760 --> 00:48:07.360 one of our use cases is image detection. When you 962 00:48:07.360 --> 00:48:09.679 have a Pigma design, certain things that are like one 963 00:48:09.760 --> 00:48:11.800 hundred vectors should actually be one image when it gets 964 00:48:11.840 --> 00:48:13.719 to a web or app, and then the text around 965 00:48:13.760 --> 00:48:17.119 it should stay text and UIs and so that is 966 00:48:17.159 --> 00:48:20.000 a good sufficient use case for like an object detection 967 00:48:20.079 --> 00:48:23.559 model which uses a convolutional neural net very common. It's 968 00:48:23.719 --> 00:48:25.679 hot dog, not hot dog. It's identifying dog, but it's 969 00:48:25.719 --> 00:48:28.639 identifying an image from training data, which in our case, 970 00:48:28.679 --> 00:48:30.480 you can generate that training data from the web. Scrape 971 00:48:30.480 --> 00:48:34.039 web pages, see what are images? Screenshot it, give it 972 00:48:34.079 --> 00:48:36.519 the screenshot plus the bounding boxes where the images work to. 973 00:48:36.559 --> 00:48:39.039 Now you've got infinite training data. You can train your models. 974 00:48:39.519 --> 00:48:42.400 But when it comes to things like decision trees, I 975 00:48:42.400 --> 00:48:45.039 think are way more interesting than people realize. You can 976 00:48:45.039 --> 00:48:47.400 break a lot of problems down to decision trees, where 977 00:48:47.400 --> 00:48:53.559 you basically provide for a very specific problem. In our case, 978 00:48:53.559 --> 00:48:55.840 it can be things like here's a whole mess of 979 00:48:56.320 --> 00:48:59.679 layers and figma, which of these should be considered groups together, 980 00:49:00.119 --> 00:49:02.559 like you know, into a flex row or column. Things 981 00:49:02.599 --> 00:49:04.760 like that. You can actually generate decision trees and be 982 00:49:04.800 --> 00:49:06.760 able to figure out or rain a forest, or you 983 00:49:06.760 --> 00:49:10.000 can get you can start simple like conditions and code, 984 00:49:10.400 --> 00:49:13.159 upgrade to desisionery if you need to upgrade to a 985 00:49:13.199 --> 00:49:14.679 rain of forest, if you need to upgrade to a 986 00:49:14.679 --> 00:49:17.119 neural network, if you need to upgrade to an LM 987 00:49:17.199 --> 00:49:19.960 at that point, and you can play the cost knob 988 00:49:20.039 --> 00:49:22.840 as much as you want. And there's definitely cases of 989 00:49:22.880 --> 00:49:25.880 products I see sometimes online where I'm like, the economics 990 00:49:25.920 --> 00:49:28.079 are just not gonna work on that product. I know 991 00:49:28.119 --> 00:49:30.159 you're charging twenty month now, but it's not gonna last. 992 00:49:30.440 --> 00:49:31.719 You have to find one that will. 993 00:49:32.119 --> 00:49:34.880 So I have to ask, I mean, I assume that 994 00:49:35.960 --> 00:49:38.599 what is it? When did you found the builder io? 995 00:49:39.519 --> 00:49:42.280 Oh goodness, twenty nineteen. I think end of your twenty 996 00:49:42.360 --> 00:49:43.079 nineteen something like that. 997 00:49:43.119 --> 00:49:45.960 Oh, five years ago? Yeah, well four or five years ago. 998 00:49:46.000 --> 00:49:48.239 I assume that four or five years ago, a lot 999 00:49:48.320 --> 00:49:51.519 of the stuff that you're talking about today you didn't know. 1000 00:49:52.239 --> 00:49:52.559 Correct. 1001 00:49:53.000 --> 00:49:56.199 Yes, so a lot of what we're talking about today 1002 00:49:56.239 --> 00:50:00.480 are things that you learned during the past four years, 1003 00:50:00.480 --> 00:50:05.400 maybe even one or two years. How did you go 1004 00:50:05.480 --> 00:50:06.760 about learning all this stuff? 1005 00:50:07.440 --> 00:50:08.559 A great question? 1006 00:50:08.840 --> 00:50:11.199 I do have an end the Internet into his brain model. 1007 00:50:12.400 --> 00:50:17.159 I'm sorry, no, it's it's more of an agentic approach. 1008 00:50:17.239 --> 00:50:19.400 I hate that word. By the way people keep saying agentic. 1009 00:50:19.559 --> 00:50:21.440 It's like, I don't know if that's a real word. 1010 00:50:21.559 --> 00:50:25.079 I hate it so much. It means agents, like it's 1011 00:50:25.079 --> 00:50:27.559 like referring to like an approach, like an agent agentic. 1012 00:50:28.320 --> 00:50:31.039 It just feels like a VC talk term. But whatever, 1013 00:50:31.920 --> 00:50:33.079 it's the morning agents on. 1014 00:50:33.079 --> 00:50:34.559 Some shades and use code words. 1015 00:50:34.679 --> 00:50:38.000 I mean, because he AI was this academic thing a 1016 00:50:38.039 --> 00:50:41.280 couple of just a few years ago, and all of 1017 00:50:41.280 --> 00:50:43.800 a sudden, it feels like you need to be an 1018 00:50:43.840 --> 00:50:48.239 expert and otherwise you're kind of potentially left behind. So 1019 00:50:48.320 --> 00:50:50.639 it's a whole lot of stuff like that you need 1020 00:50:50.679 --> 00:50:52.360 to learn very quickly. 1021 00:50:53.159 --> 00:50:56.400 Well, so, in my opinion, you know a lot of 1022 00:50:56.400 --> 00:51:00.800 people naturally say I want to get good with AI, 1023 00:51:01.119 --> 00:51:03.000 so I need to learn how AI works. I need 1024 00:51:03.000 --> 00:51:05.400 to get a book on machine learning. I need to 1025 00:51:05.440 --> 00:51:07.360 take a course on machine learning, like how the neural 1026 00:51:07.440 --> 00:51:09.760 networks are trained, or how transformers work under the hood, 1027 00:51:09.760 --> 00:51:10.199 et cetera. 1028 00:51:10.840 --> 00:51:14.000 I've done. It's hard. 1029 00:51:14.760 --> 00:51:18.960 Hard. In my opinion, it's like trying to learn physics 1030 00:51:18.960 --> 00:51:22.719 to answer questions about biology or psychology. It's not usually 1031 00:51:22.800 --> 00:51:24.280 a good idea. It's not going to give you that 1032 00:51:24.400 --> 00:51:28.719 high of ROI and so, and I'll give you an example. 1033 00:51:28.719 --> 00:51:31.360 I majored in cognitive science in college, which is like 1034 00:51:31.400 --> 00:51:33.800 the closest thing to an AI specific major that existed 1035 00:51:33.840 --> 00:51:37.280 at Berkeley, and I dropped out up to two years 1036 00:51:37.280 --> 00:51:41.039 because it was actually horribly impractical. It was a cool information, 1037 00:51:41.119 --> 00:51:42.719 but it wasn't as practical as I just want to 1038 00:51:42.719 --> 00:51:46.159 get hands on and build stuff. And still today, in 1039 00:51:46.159 --> 00:51:48.199 my opinion, the way you'll learn more about how to 1040 00:51:48.400 --> 00:51:53.119 build effective AI products with LLLMS, et cetera, is to 1041 00:51:53.159 --> 00:51:55.280 build air practic with LMS and run into all the issues. 1042 00:51:55.320 --> 00:51:57.920 I don't think the academics could tell me all the 1043 00:51:57.960 --> 00:52:00.440 strengths and weaknesses for our type of customer, you case 1044 00:52:00.519 --> 00:52:03.119 that an LLM. Willer once have if they haven't actually 1045 00:52:03.119 --> 00:52:05.440 gone through it all. Now, if you look at like 1046 00:52:05.519 --> 00:52:09.239 the research papers coming out, there actually are different categories 1047 00:52:09.280 --> 00:52:12.440 of research papers. Some are actually useful, some are literally 1048 00:52:12.519 --> 00:52:15.519 sort of like the analysis of the usage of these things, 1049 00:52:15.559 --> 00:52:18.800 like the one I mentioned comparing fine tuning to context models. 1050 00:52:18.840 --> 00:52:21.639 That's actually useful, but that's not necessarily big brain stuff. 1051 00:52:21.639 --> 00:52:24.760 Even if they have some weird calculus equations. You can 1052 00:52:24.800 --> 00:52:26.360 feed it into. You could upload a chat to BT 1053 00:52:26.519 --> 00:52:28.239 and ask what you want to ask and it'll tell 1054 00:52:28.239 --> 00:52:31.079 you what you're looking for from the paper, which is nice. 1055 00:52:31.320 --> 00:52:36.199 But all the ways I have learned anything is just 1056 00:52:36.480 --> 00:52:39.880 rapid iteration. So that applies to learning AI. That applies 1057 00:52:39.880 --> 00:52:42.599 to building a company in general, building a product. That 1058 00:52:42.639 --> 00:52:44.840 applies to learning how to market. I didn't know how 1059 00:52:44.880 --> 00:52:46.480 to market anything. We made a ton of mistakes in 1060 00:52:46.480 --> 00:52:50.400 the past. We still make mistakes now selling It's just 1061 00:52:50.679 --> 00:52:53.000 I'll take it build and needed its first customer. What 1062 00:52:53.039 --> 00:52:56.400 did I do? I asked around like how what is 1063 00:52:56.639 --> 00:52:58.239 sales like? And people like I don't know. You go 1064 00:52:58.280 --> 00:53:00.000 talk to people and see if they want to buy. 1065 00:53:00.039 --> 00:53:01.719 So I just start talking to people seeing what they 1066 00:53:01.760 --> 00:53:03.280 want to buy. And I made a lot of mistakes, 1067 00:53:03.320 --> 00:53:05.000 and now I could tell you a thousand things not 1068 00:53:05.039 --> 00:53:07.280 to do, a thousand things to do instead, just from 1069 00:53:07.320 --> 00:53:09.559 trial and error. And that's where I think a lot 1070 00:53:09.559 --> 00:53:13.760 of people, whether they're learning to program, build more complex software, build, 1071 00:53:14.320 --> 00:53:16.639 you know, get users for their software, whatever it is. 1072 00:53:18.159 --> 00:53:20.239 Two frequently people fail to just say I'm going to 1073 00:53:20.320 --> 00:53:22.320 be bad at this, I'm gonna do it anyway, and 1074 00:53:22.360 --> 00:53:23.960 I'm gonna do so many times that I'm going to 1075 00:53:24.039 --> 00:53:26.920 learn infinite things not to do, which leaves a smaller 1076 00:53:26.960 --> 00:53:29.039 and smaller window of what to do. And once I 1077 00:53:29.079 --> 00:53:31.119 have that small window of what works and that long 1078 00:53:32.000 --> 00:53:34.559 knowledge of what doesn't, at that point you could say 1079 00:53:34.559 --> 00:53:36.320 I'm actually kind of good at that thing. And all 1080 00:53:36.320 --> 00:53:39.960 it takes is doing And just like the agent's example, 1081 00:53:40.320 --> 00:53:42.440 it just means lots of feedback. So doing something and 1082 00:53:42.440 --> 00:53:44.920 getting no feedback, you're not gonna learn anything. But I 1083 00:53:44.960 --> 00:53:47.559 think like a big learning of mine is like, hey, 1084 00:53:47.719 --> 00:53:49.360 you know, we'll be like one hundred person company by 1085 00:53:49.400 --> 00:53:51.599 the end of the year. I've never run one hundred 1086 00:53:51.599 --> 00:53:53.280 percent company. I think the biggest team I managed previously 1087 00:53:53.280 --> 00:53:55.079 this company is eight people. So it's like, okay, how 1088 00:53:55.079 --> 00:53:55.960 do I know what to do? 1089 00:53:56.400 --> 00:53:56.800 What I do? 1090 00:53:56.840 --> 00:53:58.760 I get as much feedback as possible. People tell me 1091 00:53:58.800 --> 00:54:00.960 what's wrong, Tell me's wrong? Fix it? You know that 1092 00:54:01.199 --> 00:54:03.280 that loop works the product too. Get the product in 1093 00:54:03.280 --> 00:54:06.880 people's hands, Let them tell you what's wrong. Fix try stuff. 1094 00:54:07.159 --> 00:54:09.360 I think that abstract really well to so many things. 1095 00:54:09.679 --> 00:54:13.039 Do you have dedicated AI people in your company? 1096 00:54:14.519 --> 00:54:18.239 Yes, ish we're not large enough, but we really do 1097 00:54:18.280 --> 00:54:20.719 have an AI team. Yeah, we do about three four 1098 00:54:20.840 --> 00:54:22.599 five people. Some people dip in and out of it. 1099 00:54:24.000 --> 00:54:27.679 But even then that team probably wouldn't self identify as 1100 00:54:27.840 --> 00:54:30.679 AI people. They would just be like people who are 1101 00:54:30.679 --> 00:54:32.599 working on AI here and getting really good at it 1102 00:54:32.679 --> 00:54:34.679 quickly by working so hard at it. 1103 00:54:35.440 --> 00:54:37.519 Yeah, I want to chime in and just back up 1104 00:54:38.800 --> 00:54:40.719 half a step. And that is the way that you 1105 00:54:40.760 --> 00:54:43.440 describe the way you learn this stuff. What I find 1106 00:54:43.559 --> 00:54:46.039 is that's the way that the people who really note 1107 00:54:46.199 --> 00:54:50.880 something have learned it, right. Yeah, And so if you're 1108 00:54:51.039 --> 00:54:53.559 feeling like, oh, well, I don't want I don't know 1109 00:54:53.599 --> 00:54:56.000 if I want to learn AI because it looks hard. 1110 00:54:56.760 --> 00:54:59.280 It's just like, look, you know, I mean I learned 1111 00:54:59.280 --> 00:55:01.679 web development by hitting my head on the wall a 1112 00:55:01.800 --> 00:55:05.400 zillion times. Right, Oh it's broken. Oh it's broken again. 1113 00:55:05.519 --> 00:55:08.639 It's broken again. It doesn't look right now, but it works. Okay, 1114 00:55:08.679 --> 00:55:11.039 now it doesn't work, but it's you know, you know, 1115 00:55:11.119 --> 00:55:13.119 and you just you just do it. And so I 1116 00:55:13.159 --> 00:55:15.000 just want to encourage people, like, if you're looking at 1117 00:55:15.000 --> 00:55:18.119 this and you're going, boy, you know it sounds like 1118 00:55:18.199 --> 00:55:21.039 kind of a slog and kind of a journey, that's 1119 00:55:21.079 --> 00:55:23.599 just the way it is, right. I will point out that, 1120 00:55:26.480 --> 00:55:29.119 you know, I have a computer engineering degree, and so 1121 00:55:29.800 --> 00:55:32.760 you know, I got into you know, writing code, and 1122 00:55:33.039 --> 00:55:34.599 I still bang my head against the wall a whole 1123 00:55:34.599 --> 00:55:36.840 bunch of times. I just had a little bit deeper 1124 00:55:36.880 --> 00:55:39.719 foundation than maybe somebody else. But if I wanted to 1125 00:55:39.719 --> 00:55:42.239 build the castle, all the all the all the head 1126 00:55:42.239 --> 00:55:43.800 start I had was that I had a little bit 1127 00:55:43.800 --> 00:55:45.960 deeper foundation. I didn't have any of the walls up, 1128 00:55:46.039 --> 00:55:48.400 I didn't have the boat dug none of that stuff, right, 1129 00:55:48.880 --> 00:55:52.440 And so you know, you as you're looking at how 1130 00:55:52.480 --> 00:55:54.000 you break it down, I just I really want to 1131 00:55:54.079 --> 00:55:57.119 encourage people to just just look at this. I'm working 1132 00:55:57.159 --> 00:56:00.320 on putting together a boot camp that you know, through 1133 00:56:00.320 --> 00:56:02.960 a lot of this stuff, do it in October, right, 1134 00:56:03.039 --> 00:56:05.960 And that's the same thing, right is you know, I 1135 00:56:06.239 --> 00:56:07.639 kind of want to be there. So when you run 1136 00:56:07.719 --> 00:56:09.920 into the wall, you know, you're not stuck trying to 1137 00:56:09.920 --> 00:56:11.840 figure out how to get off the wall and go 1138 00:56:11.920 --> 00:56:14.360 do something. You know, do the next step. Right, you 1139 00:56:14.440 --> 00:56:16.760 kind of get there faster, but you're still going to 1140 00:56:16.840 --> 00:56:19.039 run into this even if you have somebody holding your hand. 1141 00:56:19.119 --> 00:56:21.519 So yeah, just just be aware that that that's kind 1142 00:56:21.559 --> 00:56:24.519 of the way that a lot of this goes. I 1143 00:56:24.599 --> 00:56:26.840 kind of want to pivot this. You were talking about, 1144 00:56:26.880 --> 00:56:30.239 you know, getting left behind right on some of this stuff, 1145 00:56:30.360 --> 00:56:34.079 or you know, it moves ahead quickly. And yeah, one 1146 00:56:34.119 --> 00:56:36.840 of the questions that I get from people is, Okay, 1147 00:56:36.880 --> 00:56:39.280 well do I have to learn it? And then the 1148 00:56:39.320 --> 00:56:43.119 other question I get is because you're talking about specifically. 1149 00:56:43.199 --> 00:56:46.320 So when we talked to Obi Fernandez on Ruby Rogues, 1150 00:56:46.559 --> 00:56:51.719 he was talking about the AI systems that replace you know, 1151 00:56:51.800 --> 00:56:56.280 services like copywriting and add optimization and things like that, right, 1152 00:56:57.159 --> 00:57:00.199 you know, aspects of running your business, right, and so 1153 00:57:01.079 --> 00:57:03.559 you know, programmers, you know, would listen to that and 1154 00:57:03.559 --> 00:57:06.960 not feel super threatened, right because they're gonna be in 1155 00:57:07.000 --> 00:57:10.639 a place where they're going You know, that's not something 1156 00:57:10.679 --> 00:57:13.840 I do anyway. But you're talking specifically about, Hey, I'm 1157 00:57:13.840 --> 00:57:15.480 going to take this figma and I'm going to have 1158 00:57:15.639 --> 00:57:18.760 working code at the end, right, And yes, you've kind 1159 00:57:18.800 --> 00:57:22.039 of let us know that there are various stages of 1160 00:57:22.079 --> 00:57:26.119 effectiveness to this. But is it eventually going to get 1161 00:57:26.119 --> 00:57:30.360 there where if I'm writing react or quick or something 1162 00:57:30.360 --> 00:57:32.880 else that you know, my job is going to be 1163 00:57:32.960 --> 00:57:36.320 you better write a really good prompt for the system 1164 00:57:36.440 --> 00:57:39.639 so that it'll give you the right code. And is 1165 00:57:39.639 --> 00:57:41.480 it going to make it harder for people to get in? 1166 00:57:43.159 --> 00:57:45.400 You know, is it going to cut salaries because the 1167 00:57:45.519 --> 00:57:48.239 AI does a bunch of my work? I mean, how 1168 00:57:48.639 --> 00:57:50.559 how vulnerable are we to this stuff? 1169 00:57:51.599 --> 00:57:54.559 Great question? Let me give you an example. So the 1170 00:57:54.639 --> 00:57:58.719 short answer is, I don't think developers have anything to 1171 00:57:58.760 --> 00:58:01.199 worry about. I think they only things to be excited about. 1172 00:58:01.239 --> 00:58:02.760 And I mean that genuinely one of them. 1173 00:58:02.920 --> 00:58:04.559 I like that. That's a really good way to put it, 1174 00:58:04.559 --> 00:58:06.039 because that's what I keep trying to tell people. 1175 00:58:06.639 --> 00:58:10.159 Yeah, it's it's leg about it this way too. I'll 1176 00:58:10.159 --> 00:58:13.360 make a statement and give an example. These tools are 1177 00:58:13.360 --> 00:58:18.239 gonna give you amazing, amazing superpowers. Companies want more people 1178 00:58:18.239 --> 00:58:24.159 with superpowers than less. They don't want another example, and 1179 00:58:24.199 --> 00:58:26.119 this is this is a misleading example. I'll explain why 1180 00:58:26.119 --> 00:58:26.599 its misleading. 1181 00:58:26.639 --> 00:58:27.119 Come back to it. 1182 00:58:27.280 --> 00:58:30.079 Let's say AI only work for front end tools. Would 1183 00:58:30.119 --> 00:58:33.960 you rather have more superpowered front end developers because of 1184 00:58:34.000 --> 00:58:36.280 these tools that make them superpower or more I'll just 1185 00:58:36.320 --> 00:58:38.280 say back end developers that didn't have superpowers. I want 1186 00:58:38.320 --> 00:58:41.719 more the superpower people. Now, realistically, there's different tools for 1187 00:58:41.760 --> 00:58:44.159 the front end superpowers in the back end superpowers. You're 1188 00:58:44.159 --> 00:58:47.199 gonna want both. If you're a developer and you're not 1189 00:58:47.480 --> 00:58:49.280 playing with AI tools to see how it can help 1190 00:58:49.280 --> 00:58:51.360 your workflow, I don't mean you have to be exhaustive. 1191 00:58:51.400 --> 00:58:53.079 I mean, like you know, if you want to start 1192 00:58:53.079 --> 00:58:56.320 with the basics, use chat Gypto. I personally recommend Claude instead. 1193 00:58:56.440 --> 00:58:59.159 Cost the same and it does better, especially with code. 1194 00:58:59.519 --> 00:59:02.199 And if you're not using get a copilot or whatever 1195 00:59:02.760 --> 00:59:05.559 most synonymous thing you have for your ide I would 1196 00:59:05.920 --> 00:59:08.480 highly suggest you do those things because most importantly what 1197 00:59:08.519 --> 00:59:10.519 they do is they build an intuition of what's AI 1198 00:59:10.559 --> 00:59:13.320 good at and not and when it's out of suggesting. 1199 00:59:13.920 --> 00:59:15.599 There are certain things that I just know that AIS 1200 00:59:15.599 --> 00:59:18.320 can do a good job on every time, and I 1201 00:59:18.360 --> 00:59:21.119 have that intuition through repetition of just seeing it happen 1202 00:59:21.159 --> 00:59:23.000 in real time with no effort for me, and I 1203 00:59:23.039 --> 00:59:25.000 know things it will not do good at, and those 1204 00:59:25.239 --> 00:59:28.000 things build a mental model. It's like training your brain, 1205 00:59:28.079 --> 00:59:30.159 your neural network to know what the AI is good 1206 00:59:30.159 --> 00:59:32.000 at and not. So when you're building products with it 1207 00:59:32.119 --> 00:59:34.440 or when you're trying to find productivity games for yourself, 1208 00:59:34.800 --> 00:59:36.400 you have that intuition and where it's going to work 1209 00:59:36.400 --> 00:59:39.119 and not. I see people every day with the wrong intuition, like, oh, 1210 00:59:39.159 --> 00:59:42.360 can AI just do everything for me? It's like you 1211 00:59:42.440 --> 00:59:44.639 have you used AI tools? No, No, it's not going 1212 00:59:44.679 --> 00:59:48.360 to do that. That's going to go badly. So if 1213 00:59:48.400 --> 00:59:51.159 you think about like how it affects jobs, let me 1214 00:59:51.199 --> 00:59:54.039 give you another example. Builder as a tool has always 1215 00:59:54.079 --> 00:59:57.119 been a tool that takes the code and components you 1216 00:59:57.199 --> 01:00:00.800 have and brings and takes work that you have as 1217 01:00:00.880 --> 01:00:03.320 developer that generally you'll find tedious. And I think if 1218 01:00:03.320 --> 01:00:06.119 you're thinking calmly and dispassionately. It's the tedious work you 1219 01:00:06.119 --> 01:00:08.079 don't want to do, like marketing, wanting to move buttons 1220 01:00:08.079 --> 01:00:09.800 around the homepage or change the color of this to 1221 01:00:09.840 --> 01:00:11.840 see you know, and realize the test failed to they 1222 01:00:11.880 --> 01:00:14.360 undo it. It doesn't feel good to be a developer 1223 01:00:14.400 --> 01:00:15.880 doing that. You don't want to be like a middle 1224 01:00:15.920 --> 01:00:20.239 person in between, like say marketing and your homepage. That sucks. Objectively, 1225 01:00:20.280 --> 01:00:24.840 it sucks, and so with builders. Sometimes people before adopting 1226 01:00:24.880 --> 01:00:26.719 the product, they have a concern like, wait, is this 1227 01:00:26.719 --> 01:00:28.440 going to take my job away? Is this? Do I 1228 01:00:28.480 --> 01:00:30.000 have to be worried about this? I would say only 1229 01:00:31.440 --> 01:00:33.559 one to five percent of people have that concern that 1230 01:00:33.599 --> 01:00:36.360 CanCERN happens. I've heard it zero percent from real customers, 1231 01:00:36.400 --> 01:00:38.320 and I keep really close to our customers. Nobody's ever 1232 01:00:38.360 --> 01:00:41.760 adopted the product that makes their development more efficient and 1233 01:00:42.039 --> 01:00:44.880 actually worried about their job security Afterwards. When businesses can 1234 01:00:44.920 --> 01:00:48.360 do more, they just want more. Their appetite grows faster. 1235 01:00:48.719 --> 01:00:52.440 It kind of reminds me of you know, when people said, 1236 01:00:52.480 --> 01:00:55.880 like I'm trying to remember who's this stand up said 1237 01:00:55.920 --> 01:00:59.960 it that if you're worried about illegal immigrants taking your job, 1238 01:01:00.039 --> 01:01:05.159 job and you've got the wrong job. Uh, it's it's 1239 01:01:05.199 --> 01:01:07.800 it's you don't want to work in the job, in 1240 01:01:07.800 --> 01:01:11.440 in those kind of jobs anyway. But is what I'm saying. 1241 01:01:11.599 --> 01:01:18.320 So what I'm understanding from you is that it's that 1242 01:01:18.599 --> 01:01:22.440 superpower to uh, to do away with a lot of 1243 01:01:22.559 --> 01:01:31.719 the repetitive and let's let's call it less thoughtful parts 1244 01:01:31.719 --> 01:01:37.159 of the job. Uh it's is that kind of what 1245 01:01:37.199 --> 01:01:38.079 you're saying. 1246 01:01:38.400 --> 01:01:40.480 Yeah, I can, I can ahad. 1247 01:01:41.360 --> 01:01:43.559 Yeah, I was just gonna say, you know, I did 1248 01:01:43.559 --> 01:01:46.039 a couple of job interviews, and I've been using get 1249 01:01:46.079 --> 01:01:49.159 hoo co pilot for a while and and one thing 1250 01:01:49.199 --> 01:01:53.159 that it does for me is right, occasionally it'll try 1251 01:01:53.199 --> 01:01:55.679 and fill in the whole class in Ruby or the 1252 01:01:55.719 --> 01:01:59.519 whole you know, component in JavaScript, but most of the 1253 01:01:59.559 --> 01:02:02.480 time it's just filling in the part that's sort of 1254 01:02:02.519 --> 01:02:06.880 an atomic piece of the thing. And what it does 1255 01:02:06.920 --> 01:02:10.599 for me that kind of makes my development better is 1256 01:02:10.639 --> 01:02:14.719 that it gives me enough of the pieces that you know, 1257 01:02:14.760 --> 01:02:16.599 from my experience, I can look at it and say 1258 01:02:16.639 --> 01:02:20.039 that's what I want or that's close, right, and so 1259 01:02:20.199 --> 01:02:23.039 from there then you know, maybe I hit tab and 1260 01:02:23.079 --> 01:02:24.639 I have it drop it in and then I go 1261 01:02:24.679 --> 01:02:26.960 and I modify it to be what I need, and 1262 01:02:27.320 --> 01:02:29.920 you know, sometimes it's seventy five percent, sometimes it's one 1263 01:02:30.000 --> 01:02:32.800 hundred percent, and sometimes it's no, that's just not what 1264 01:02:32.840 --> 01:02:37.679 I want. Right, you don't understand what I'm doing, But yeah, 1265 01:02:37.880 --> 01:02:40.400 it totally opens those doors. And the other thing that 1266 01:02:40.440 --> 01:02:42.880 I want to jump on here is you know you 1267 01:02:42.960 --> 01:02:47.719 mentioned that, Yeah, if you have superpowered engineers, people are 1268 01:02:47.760 --> 01:02:49.840 thinking that these companies are just going to cut their 1269 01:02:49.880 --> 01:02:52.880 costs down to the bone and just have bare bones 1270 01:02:53.000 --> 01:02:59.440 development AI generated code. That's generally not how companies work. Generally. 1271 01:02:59.440 --> 01:03:01.000 The way that they is they're going to spend what 1272 01:03:01.039 --> 01:03:04.239 they spend to move forward with all the things that 1273 01:03:04.280 --> 01:03:06.480 they think they want to provide to their customers. And 1274 01:03:06.519 --> 01:03:08.119 so you're just going to be able to provide more 1275 01:03:08.159 --> 01:03:08.639 of that. 1276 01:03:09.679 --> 01:03:13.400 Correct, and you'd be I recommend people don't underestimate how 1277 01:03:13.480 --> 01:03:19.519 much when capacity explodes, ideas and needs and desires explode. 1278 01:03:19.840 --> 01:03:23.400 And I think it's not just proportionate, it's above proportionate. 1279 01:03:23.400 --> 01:03:25.239 And suddenly, when you think you're introduced can do magic, 1280 01:03:25.480 --> 01:03:27.960 you want to do more than magic, and you're excited 1281 01:03:28.000 --> 01:03:30.639 because remember, companies live in a competitive environment. So I 1282 01:03:30.639 --> 01:03:32.159 guess I could put the CEO head on for a 1283 01:03:32.159 --> 01:03:35.159 minute and say, what am I paranoid against competitors? I 1284 01:03:35.239 --> 01:03:38.400 paranoid that we're doing more faster than them, which means 1285 01:03:38.519 --> 01:03:43.159 I want more superpowered aidevs who are more accelerated by 1286 01:03:43.159 --> 01:03:45.480 the AI tools than my competitors, so we leave them 1287 01:03:45.519 --> 01:03:48.119 in the dust. I'm not thinking, oh, we can do more, 1288 01:03:48.159 --> 01:03:50.800 cut people. I'm saying I'm looking at we can do more, 1289 01:03:51.079 --> 01:03:54.280 get ahead of the competition, maximize the resources we have, 1290 01:03:54.639 --> 01:03:57.400 even hire more because now the ROI makes more sense. 1291 01:03:57.719 --> 01:03:59.800 I can hire another developer and pay them at developers 1292 01:03:59.800 --> 01:04:01.599 out and get way more ROI than I did for 1293 01:04:01.599 --> 01:04:04.000 that money. So it's an obvious investment that kind of 1294 01:04:04.000 --> 01:04:05.960 has happened. As we got more productive, we started hiring 1295 01:04:06.000 --> 01:04:08.480 more of these developers. And the way I would suggest 1296 01:04:08.480 --> 01:04:11.960 thinking about it is here's an example. Has anybody ever 1297 01:04:12.079 --> 01:04:15.480 hand coded a PDF? Heck no, You go to Chrome 1298 01:04:15.519 --> 01:04:17.719 and you say exports pdf and it generates the PDF 1299 01:04:17.760 --> 01:04:21.440 for you. In my opinion, figma designs are similar. If 1300 01:04:21.480 --> 01:04:23.280 you get to a point where you can generate it 1301 01:04:23.320 --> 01:04:25.440 and then you can work with it from there, you're 1302 01:04:25.440 --> 01:04:27.719 not gonna go back to hand doing it, it's just 1303 01:04:27.719 --> 01:04:29.400 a waste of time. So the way you can think 1304 01:04:29.400 --> 01:04:31.599 about your work is and I won't solve everything. You 1305 01:04:31.599 --> 01:04:33.880 have a massive, multimillion line of code base, you have 1306 01:04:33.920 --> 01:04:36.400 a very esoteric bug. You're gonna have to look into 1307 01:04:36.440 --> 01:04:38.039 that and you're gonna have to actually dive into it. 1308 01:04:38.079 --> 01:04:41.880 A agents are so far from handling things like that. 1309 01:04:41.960 --> 01:04:45.159 But when you're producing and getting ideas out and stuff 1310 01:04:45.199 --> 01:04:48.000 like that, the drafting iteration phages can move so much 1311 01:04:48.039 --> 01:04:50.639 faster if you can just suck in this dashboard and 1312 01:04:50.679 --> 01:04:52.119 you see it. But you're like, I want to move this, 1313 01:04:52.159 --> 01:04:53.840 I want to reconnect this. I'm gonna get on the 1314 01:04:53.840 --> 01:04:55.400 code for this. I'm gonna get on the prompt for this. 1315 01:04:55.920 --> 01:04:58.760 You become like this orchestrator. And I think, you know, 1316 01:04:58.800 --> 01:05:02.800 a lot of engineer don't like to step into a 1317 01:05:02.800 --> 01:05:05.480 lot of engineers might like the idea of management because 1318 01:05:05.480 --> 01:05:09.039 they like this idea of like parallel execution. You know, 1319 01:05:09.119 --> 01:05:13.079 like you have people working under you, and so by 1320 01:05:13.159 --> 01:05:14.960 you can be in charge of a project and you 1321 01:05:14.960 --> 01:05:17.599 have ten engineers working on it, you can do theoretically 1322 01:05:17.880 --> 01:05:19.960 ten times as much as if you were just the 1323 01:05:20.000 --> 01:05:23.159 only one working on it. Cool but you realize people 1324 01:05:23.480 --> 01:05:26.320 are difficult. You spend your work not solving cool problems. 1325 01:05:26.360 --> 01:05:28.639 You spend some the personnel issues. This person's like that person, 1326 01:05:28.679 --> 01:05:31.440 This person is not convinced of the roadmap. And all 1327 01:05:31.480 --> 01:05:34.440 this stuff happens and people realize, wait, I like building stuff. 1328 01:05:34.440 --> 01:05:36.159 This is building stuff, and I get out of management. 1329 01:05:36.599 --> 01:05:39.639 The best part of I could describe using AI effectively. 1330 01:05:39.880 --> 01:05:41.880 It's like all the best parts of management. You have 1331 01:05:41.920 --> 01:05:44.239 these minions or these tools generating things for you, but 1332 01:05:44.320 --> 01:05:47.960 without the downsides. They don't have any types of emotions 1333 01:05:47.960 --> 01:05:48.639 about anything. 1334 01:05:48.920 --> 01:05:52.639 You can work so far, so far, I don't think 1335 01:05:52.639 --> 01:05:55.519 that's going to change, because why would we program that 1336 01:05:56.679 --> 01:05:58.320 or at least go back to the older AI that 1337 01:05:58.320 --> 01:06:01.239 doesn't have that problem go personality. 1338 01:06:02.559 --> 01:06:05.000 Yeah, well, you know. 1339 01:06:05.239 --> 01:06:07.719 The other thing is you're talking about these iteration cycles 1340 01:06:07.760 --> 01:06:10.880 that you know that it speeds up. And I think 1341 01:06:11.039 --> 01:06:12.840 one thing that a lot of people don't understand is 1342 01:06:12.880 --> 01:06:17.320 that those iteration cycles happen now right, you know, And 1343 01:06:17.440 --> 01:06:19.519 and if you've been on a project long enough, what 1344 01:06:19.679 --> 01:06:23.079 typically happens is you'll get to a point where either 1345 01:06:23.159 --> 01:06:25.639 you've accrued so much technical debt that it's impossible to 1346 01:06:25.679 --> 01:06:30.039 work on or what what will happen is you'll you'll 1347 01:06:30.119 --> 01:06:32.400 kind of get a couple of pieces in place, and 1348 01:06:32.400 --> 01:06:34.519 then all of a sudden, everything else gets easier, right, 1349 01:06:35.039 --> 01:06:37.360 and so then the sky's the limit and and what 1350 01:06:37.400 --> 01:06:39.840 the what you're saying is is a lot of the 1351 01:06:40.239 --> 01:06:43.639 foundational pieces and a lot of the foundational thinking. You 1352 01:06:43.679 --> 01:06:46.079 know that that's mostly just you just have to grind 1353 01:06:46.119 --> 01:06:49.000 through it until you're done with it. That stuff goes away. 1354 01:06:49.079 --> 01:06:51.159 And so then all of a sudden, it's we have 1355 01:06:51.280 --> 01:06:56.079 these greater capabilities, these greater opportunities, and so we can 1356 01:06:56.199 --> 01:06:59.320 instead of the grind taking two months, the grind takes 1357 01:06:59.719 --> 01:07:02.920 two weeks because we get that much further ahead, and 1358 01:07:02.960 --> 01:07:05.280 then we can modify what we got from the AI 1359 01:07:05.840 --> 01:07:07.679 or work with the AI for the two weeks to 1360 01:07:07.719 --> 01:07:09.800 get where we needed to go, and then we can 1361 01:07:09.840 --> 01:07:14.000 turn around, and that iteration cycle happens faster exactly. 1362 01:07:14.119 --> 01:07:16.000 You can think of it like, also, you're no longer 1363 01:07:16.039 --> 01:07:18.320 the fluidist in the band, you're the entire orchestra. You're 1364 01:07:18.320 --> 01:07:20.559 the conductor. Yeah, you can make these happen. 1365 01:07:20.760 --> 01:07:23.360 But here's the thing that I wanted. I know where 1366 01:07:23.400 --> 01:07:26.400 we're running along on time and we'll probably finish soon. 1367 01:07:26.440 --> 01:07:30.760 But there's this one question that I really need to 1368 01:07:31.280 --> 01:07:34.000 get off my chest, as it were. So you've been 1369 01:07:34.039 --> 01:07:38.119 describing and I totally agree with the fact that you know, 1370 01:07:38.360 --> 01:07:40.719 we need to learn how to we need to learn 1371 01:07:40.760 --> 01:07:44.440 these to use these AI tools. Obviously we learned need 1372 01:07:44.480 --> 01:07:48.000 to learn how to best use these AI tools, and 1373 01:07:48.119 --> 01:07:52.480 these AI tools will make us more productive, and I 1374 01:07:52.519 --> 01:07:56.480 totally agree with that, But it's still it. There's still 1375 01:07:56.480 --> 01:07:59.639 the question of how knowledgeable do I need to be 1376 01:07:59.719 --> 01:08:04.199 at AI? Because there's a difference between being able to, 1377 01:08:04.400 --> 01:08:08.039 I don't know, drive a car and be the person 1378 01:08:08.119 --> 01:08:13.519 who can pop the hood and start fiddling with the 1379 01:08:13.559 --> 01:08:17.560 stuff there. So how knowledgeable do I need to be 1380 01:08:17.920 --> 01:08:23.039 about AI in order to be effective as a developer, 1381 01:08:23.199 --> 01:08:29.119 to be desirable as a developer in the upcoming one, two, five, 1382 01:08:29.239 --> 01:08:29.960 ten years. 1383 01:08:30.039 --> 01:08:35.119 Do you think, Yeah, it's a great question. The A 1384 01:08:35.119 --> 01:08:37.960 couple quick thoughts on this. The first one is I 1385 01:08:38.039 --> 01:08:39.960 personally are like The high level one is I personally 1386 01:08:40.000 --> 01:08:42.880 think people just need to know how to use AI 1387 01:08:43.000 --> 01:08:46.800 to be more productive for their own work. In doing that, 1388 01:08:46.960 --> 01:08:51.720 you'll actually learn a lot about the underlying pieces, so 1389 01:08:51.840 --> 01:08:54.880 to speak, and how the technology is suitable for certain 1390 01:08:54.960 --> 01:08:58.000 use cases versus not. A simple example there is like 1391 01:08:58.039 --> 01:09:00.680 you might be writing docs in markdown and you see 1392 01:09:00.680 --> 01:09:03.000 how well at auto complete certain docks. Let's say you 1393 01:09:03.039 --> 01:09:05.880 make a writing products well by learning how copi has 1394 01:09:05.920 --> 01:09:08.079 worked well or not, or Tatchubet's work well or not. 1395 01:09:08.520 --> 01:09:10.760 And now suddenly you work on some type of writing product, 1396 01:09:10.800 --> 01:09:13.319 a Google Docs thing, and the company wants you to 1397 01:09:13.319 --> 01:09:15.800 build some type of AI feature. Well, you already have 1398 01:09:15.880 --> 01:09:19.239 some baseline knowledge and intuition built around this. So the 1399 01:09:19.239 --> 01:09:22.199 two categories probably are one. Using AI tools to be 1400 01:09:22.239 --> 01:09:24.399 a more effective developer. You should always try and be 1401 01:09:24.439 --> 01:09:27.199 more effective because if you're not, your peers are, And 1402 01:09:27.600 --> 01:09:30.000 in a certain sense, you don't want to be just 1403 01:09:30.119 --> 01:09:32.720 so behind the train. Everybody's at a whole new level 1404 01:09:32.760 --> 01:09:34.960 than you because you had just hopped on the train. 1405 01:09:35.439 --> 01:09:37.199 You don't want that to happen. You want to be 1406 01:09:38.239 --> 01:09:40.560 using the best tools, the best ability to be as 1407 01:09:40.600 --> 01:09:43.239 good as your peers, be one of the best of 1408 01:09:43.279 --> 01:09:46.359 your peers. On the other side, there's building AI products 1409 01:09:46.600 --> 01:09:49.920 that is shockingly similar. And if you really want to 1410 01:09:49.960 --> 01:09:52.479 be if you think in five years every product will 1411 01:09:52.479 --> 01:09:54.439 be an AI product, I think that's an extreme thinking, 1412 01:09:54.479 --> 01:09:57.359 but you could think of it differently, as more products 1413 01:09:57.399 --> 01:09:59.279 will have more AI features, and if my company might 1414 01:09:59.319 --> 01:10:01.880 want me to work on it, it's nice to know 1415 01:10:01.920 --> 01:10:03.600 how to work on it. Again, I don't think the 1416 01:10:03.640 --> 01:10:05.199 way to do that is to go read a bunch 1417 01:10:05.199 --> 01:10:07.840 of white papers, build your neural network from scratch, understand 1418 01:10:07.840 --> 01:10:09.800 how transformer works into the hood. Again, I think that's 1419 01:10:09.840 --> 01:10:12.039 like trying to solve psychology problems with physics and a 1420 01:10:12.119 --> 01:10:14.239 knowledge of F equals M. It's not going to get 1421 01:10:14.279 --> 01:10:18.920 you that far. Solve psychology problems by learning about psychology 1422 01:10:18.920 --> 01:10:23.039 and practicing psychology in some form, studying it directly and 1423 01:10:23.119 --> 01:10:26.680 learning that way. And I think that the beautiful thing 1424 01:10:26.680 --> 01:10:28.720 about AI is it can solve your own problems. You know, 1425 01:10:28.840 --> 01:10:31.239 we talk about like the types of work dev don't 1426 01:10:31.239 --> 01:10:33.439 want to do. You have this really exciting opportunity. Let's 1427 01:10:33.439 --> 01:10:36.000 say I think we've talked about code debt, like, could 1428 01:10:36.079 --> 01:10:39.239 you make an AI workflow to help you with managing 1429 01:10:39.279 --> 01:10:42.159 code debt, refactoring, cleaning, as cetther code. That's a problem 1430 01:10:42.159 --> 01:10:43.840 that you right now, as an engineer, might say I 1431 01:10:43.880 --> 01:10:46.159 hate dealing with this. I can play around with AI 1432 01:10:46.239 --> 01:10:49.000 to try and solve my own problem. Engineers love over 1433 01:10:49.039 --> 01:10:51.640 automating things. You know, let's spend a week automating something 1434 01:10:51.680 --> 01:10:54.720 that could have taken a few hours of front work. Everybody, 1435 01:10:54.760 --> 01:10:56.840 everybody's guilty of it. If you're probably any good, you 1436 01:10:56.840 --> 01:10:57.960 probably this is how you think. 1437 01:10:58.359 --> 01:11:01.439 You know, you could probably use an AI tool to 1438 01:11:01.520 --> 01:11:04.159 help you build your AI tool exactly. 1439 01:11:04.760 --> 01:11:09.800 Make your own projects. Get up Copilot, microagent, CHATTPT, whatever 1440 01:11:09.840 --> 01:11:11.920 can help you build it. And that's a good way 1441 01:11:11.920 --> 01:11:13.720 to learn the stuff too. So I always think that 1442 01:11:13.800 --> 01:11:16.119 side projects are awesome. You can learn your day to 1443 01:11:16.199 --> 01:11:18.680 day using AI tools, try new ones, and again, I 1444 01:11:18.720 --> 01:11:20.720 don't think you have to make an exhaustive list. Everybody 1445 01:11:20.800 --> 01:11:23.479 keeps saying Cursor AI is great. I don't care that much. 1446 01:11:23.520 --> 01:11:25.279 I love to the features again today it's the same 1447 01:11:25.279 --> 01:11:27.760 features give Copilot with some nuances. So no, I don't 1448 01:11:27.800 --> 01:11:30.359 think you need to waste tons of time adopting every tool. 1449 01:11:30.359 --> 01:11:32.800 There is a learning curve, but just adopt the basics 1450 01:11:32.880 --> 01:11:35.760 and use your own time to do your own side 1451 01:11:35.760 --> 01:11:38.159 products sometimes to use AI to solve your own problems. 1452 01:11:38.319 --> 01:11:39.920 I think that'll make you equiped to be a very 1453 01:11:39.960 --> 01:11:43.359 future proof, very capable developer. One years, five years, ten 1454 01:11:43.439 --> 01:11:44.039 years from now. 1455 01:11:44.720 --> 01:11:48.439 Very interesting in this context is I think it will 1456 01:11:48.479 --> 01:11:51.600 see how far it goes. But the fact that Google, 1457 01:11:51.720 --> 01:11:56.439 for example, is actually building their own micro model into 1458 01:11:57.079 --> 01:12:01.000 Chrome itself, that will give an interesting opportunity for people 1459 01:12:01.039 --> 01:12:02.159 to play with the technology. 1460 01:12:03.239 --> 01:12:05.680 It is exciting, it's fast too, and that opens up 1461 01:12:05.720 --> 01:12:08.319 new worlds too. This is also why you should probably 1462 01:12:08.359 --> 01:12:09.680 get your hands on the stuff and try and solve 1463 01:12:09.720 --> 01:12:12.760 real problems. Is you start realizing why that matters for us. 1464 01:12:12.760 --> 01:12:14.439 One of the most common things that comes up is 1465 01:12:15.279 --> 01:12:18.319 giant Enterprise bank wants to use our AI features, but 1466 01:12:18.359 --> 01:12:20.880 they can't send their coat over the wire to back end. 1467 01:12:21.119 --> 01:12:23.279 So that's why you start exploring local models, like let's 1468 01:12:23.319 --> 01:12:25.880 make the AI work with Olama, run it all locally 1469 01:12:25.920 --> 01:12:28.600 on your machine. Honestly, if weren't building a products, I 1470 01:12:28.600 --> 01:12:31.199 wouldn't know why people care that much about that. Sure, 1471 01:12:31.279 --> 01:12:33.439 cost is a thing, and so we start exploring it, 1472 01:12:33.479 --> 01:12:36.199 and we tend to find that the local models aren't 1473 01:12:36.600 --> 01:12:39.800 powerful enough yet to solve these use cases without computer. 1474 01:12:39.680 --> 01:12:44.920 So they can be just too big to download exactly. 1475 01:12:45.000 --> 01:12:46.840 That's what I found that the models you need to 1476 01:12:46.840 --> 01:12:49.319 be effective with our product right now they're too big 1477 01:12:49.359 --> 01:12:51.800 to download and too big to run on a local computer. 1478 01:12:51.920 --> 01:12:53.439 The running is the hard pore. You just run out 1479 01:12:53.439 --> 01:12:57.399 a RAM. But again, when you're kind of immersing the ecosystem, 1480 01:12:57.560 --> 01:13:00.560 the other stuff that's happening makes more sense. People care 1481 01:13:00.600 --> 01:13:04.039 about this. Does it work yet? Window dot ai will 1482 01:13:04.039 --> 01:13:06.560 not replace chat GBT. It's gonna be a much smaller model, 1483 01:13:06.560 --> 01:13:09.279 which are a lot dumber, But for certain use cases 1484 01:13:09.279 --> 01:13:11.359 it's good enough. And that's where getting your hands on 1485 01:13:11.439 --> 01:13:13.600 is the best way to know those types of things. 1486 01:13:14.119 --> 01:13:17.640 Yep, right, good deal. Well, I'm gonna kind of wrap 1487 01:13:17.720 --> 01:13:22.880 us up. I recommend you go check out the article 1488 01:13:23.239 --> 01:13:27.000 that Steve wrote. There's also a video I saw that 1489 01:13:27.359 --> 01:13:28.720 has the same name, and I don't know if you 1490 01:13:28.760 --> 01:13:30.079 put that out or if it was somebody else. 1491 01:13:31.359 --> 01:13:33.439 The same thing, the same thing that's in the blog 1492 01:13:33.520 --> 01:13:35.039 is basically covering the video and vice versa. 1493 01:13:35.920 --> 01:13:38.600 Yeah, but I highly recommend that. I mentioned the Obi 1494 01:13:38.640 --> 01:13:41.680 Fernandez episode on Ruby Rogues. He actually has a book 1495 01:13:42.119 --> 01:13:44.239 that he wrote on how to use models like this. 1496 01:13:45.079 --> 01:13:47.199 I think some of the code samples are in Ruby, 1497 01:13:47.239 --> 01:13:50.720 but for the most part, it's kind of language agnostic. 1498 01:13:50.760 --> 01:13:53.319 It's just hey, you know, if you understand these APIs 1499 01:13:53.359 --> 01:13:57.640 and capabilities, then here's what you can build. And yeah, 1500 01:13:57.680 --> 01:14:01.000 I mean I think, I think think this really is 1501 01:14:01.039 --> 01:14:04.720 going to continue to change the way that we're working. 1502 01:14:04.800 --> 01:14:07.920 And so you know, the more you can get into 1503 01:14:08.000 --> 01:14:10.520 it ahead of the curve, I think, the better off 1504 01:14:10.560 --> 01:14:14.760 you're going to be. So thanks for coming, Steve, Thanks 1505 01:14:14.760 --> 01:14:17.439 for having Yeah, we're going to do our picks. Then 1506 01:14:17.479 --> 01:14:19.760 we're going to wrap up. Dan, do you have some 1507 01:14:19.760 --> 01:14:20.600 picks for us? 1508 01:14:21.119 --> 01:14:25.039 Yeah, it's not exactly a pick. So I've not been 1509 01:14:25.079 --> 01:14:29.039 speaking about the current situation in Israel and in Gaza 1510 01:14:29.199 --> 01:14:32.279 in the recent months, maybe because it was just too 1511 01:14:32.279 --> 01:14:34.760 painful for me to just think about all the time. 1512 01:14:35.560 --> 01:14:39.600 But this week, actually a person is going to come 1513 01:14:39.760 --> 01:14:42.279 to our company to talk about the fact that his 1514 01:14:42.439 --> 01:14:47.039 son is kidnapped in Gaza. People tend to forget that 1515 01:14:47.039 --> 01:14:50.439 there are still one hundred and twenty Israelis kidnapped in Gaza, 1516 01:14:50.520 --> 01:14:57.960 including one child and one baby. And it's a situation 1517 01:14:58.119 --> 01:15:01.520 that I hoped would have been longer and still hasn't been, 1518 01:15:02.680 --> 01:15:06.359 and hopefully it will be. What can I say, It's 1519 01:15:06.359 --> 01:15:10.239 going to be very, very difficult to even look this 1520 01:15:10.319 --> 01:15:12.479 person in the eye let's put it this way. It's 1521 01:15:12.560 --> 01:15:17.439 it's I'm it's it's it's just such a statues situation. Anyway. 1522 01:15:18.479 --> 01:15:21.880 Sorry for bringing in such a bummer, but that's the 1523 01:15:21.920 --> 01:15:24.920 only thing that I really had that I wanted to mention. 1524 01:15:25.760 --> 01:15:29.199 Uh So, so yeah, hopefully you can pick up the mood. 1525 01:15:31.359 --> 01:15:35.079 Yeah, hopefully I can. But yeah, it is sad, and 1526 01:15:35.600 --> 01:15:38.600 you know, I think sometimes we lose track of these 1527 01:15:38.600 --> 01:15:41.359 things after a certain amount of time when they're not 1528 01:15:41.640 --> 01:15:44.399 in the forefront of you know, what we're listening to 1529 01:15:44.600 --> 01:15:47.600 or watching or things like that. So yeah, keep in 1530 01:15:47.640 --> 01:15:51.840 mind that these are people that we should be thinking about, 1531 01:15:51.920 --> 01:15:55.520 praying for, and and looking for solutions here, and encourage 1532 01:15:55.520 --> 01:15:59.159 our leaders, whether it's in you know, Congress or the 1533 01:15:59.239 --> 01:16:08.239 president or who ever, to help figure some of this out. Yeah, 1534 01:16:08.319 --> 01:16:10.560 We've had some stuff going on in this country too, 1535 01:16:10.560 --> 01:16:11.960 but I'm not going to go into it. 1536 01:16:12.039 --> 01:16:15.319 I you know, spoke a little bit about it before 1537 01:16:15.359 --> 01:16:15.720 the show. 1538 01:16:16.039 --> 01:16:19.119 Yeah, but be kind to your neighbors, right, I think 1539 01:16:19.199 --> 01:16:21.039 a lot of this just comes down to the way 1540 01:16:21.079 --> 01:16:23.239 we demonize each other and we don't need to do that. 1541 01:16:24.399 --> 01:16:27.239 I'm going to jump in and do a game pick. 1542 01:16:28.680 --> 01:16:30.520 I'm going to pick a game that I've picked in 1543 01:16:30.560 --> 01:16:34.119 the past. I just didn't get together with my buddies 1544 01:16:34.119 --> 01:16:38.600 this week to play a board game. And this is 1545 01:16:38.680 --> 01:16:45.560 one my So my nephew's here from Illinois and his 1546 01:16:45.560 --> 01:16:48.960 his parents are going and doing job interviews in Wyoming, 1547 01:16:49.000 --> 01:16:51.479 and so they dropped him off here in Utah instead 1548 01:16:51.479 --> 01:16:53.760 of dragging him around Wyoming and having him be bored 1549 01:16:54.479 --> 01:16:57.800 with them. He can play with my kids. I say, 1550 01:16:57.800 --> 01:17:01.960 play with my kids like he's five, he's seventeen. But 1551 01:17:02.399 --> 01:17:06.319 the game I'm gonna pick is Mysterium. I always do 1552 01:17:06.319 --> 01:17:09.479 a board game kicks pick Steve at the beginning of 1553 01:17:09.520 --> 01:17:14.319 my picks. And so Mysterium is a game where you 1554 01:17:14.399 --> 01:17:16.159 have one person giving clues and then you have other 1555 01:17:16.159 --> 01:17:20.119 people trying to figure out what the clues mean. And 1556 01:17:20.239 --> 01:17:23.439 so you have all the players but one are psychics. 1557 01:17:24.159 --> 01:17:26.840 And then the person giving the clues is a ghost. 1558 01:17:26.960 --> 01:17:29.520 And there have been murders in this house, and so 1559 01:17:29.560 --> 01:17:33.000 there they hand cards to the psychics, and the psychics 1560 01:17:33.000 --> 01:17:37.159 try and determine, you know, who the person, the place, 1561 01:17:37.359 --> 01:17:42.359 or the murder weapon is. And then there are expansions 1562 01:17:42.399 --> 01:17:45.920 where instead of doing a murder weapon, you do a motive. 1563 01:17:47.199 --> 01:17:49.880 But it's it's a pretty fun game. Takes an hour 1564 01:17:49.960 --> 01:17:55.960 or so to play. Sometimes it's really hard because you know, 1565 01:17:56.000 --> 01:17:58.960 you've got all kinds of interesting things on this card. 1566 01:17:59.039 --> 01:18:05.520 It's just this, you know, this picture, and so anyway, Yeah, 1567 01:18:05.600 --> 01:18:09.199 you go through the rounds. You pick the first person first, 1568 01:18:09.279 --> 01:18:12.760 you guess the place next, and then the object, and 1569 01:18:12.800 --> 01:18:15.039 if you get all three and everybody gets if all 1570 01:18:15.039 --> 01:18:18.560 the psychics solve their murder, then the ghost has one 1571 01:18:18.600 --> 01:18:21.000 final round where they give three cards and depending on 1572 01:18:21.039 --> 01:18:23.640 how well you were doing it guessing whether or not 1573 01:18:23.680 --> 01:18:26.960 the other psychics were correct or not, and how early 1574 01:18:27.000 --> 01:18:29.279 you got yours done, you get to see one, two, 1575 01:18:29.319 --> 01:18:33.680 or three of those cards, and those are hints toward 1576 01:18:34.279 --> 01:18:37.520 one of the murders that the psychics solved, and that's 1577 01:18:36.880 --> 01:18:41.439 the ghost, the person who's beginning to concluse that's their murder. 1578 01:18:41.960 --> 01:18:44.760 And you know, it's kind of a majority wins thing 1579 01:18:44.800 --> 01:18:48.479 on that one. It has a board game geek weight 1580 01:18:48.600 --> 01:18:52.600 of one point nine. I keep telling people that kind 1581 01:18:52.600 --> 01:18:57.319 of the average you know, friendly board game that's approachable 1582 01:18:57.359 --> 01:19:00.439 to people who don't play board games is about a two, right, 1583 01:19:01.079 --> 01:19:04.159 and so this is right in there in that area. 1584 01:19:04.640 --> 01:19:09.880 It's kind of a fun social game, and I really 1585 01:19:10.000 --> 01:19:13.239 enjoy it. The kids were playing it. They helped my 1586 01:19:13.279 --> 01:19:15.359 eight year old play it. I don't know if an 1587 01:19:15.399 --> 01:19:18.279 eight year old could pick up on all the nuances 1588 01:19:18.279 --> 01:19:22.399 of it without help, but they just helped to play. 1589 01:19:22.479 --> 01:19:26.800 And so anyway, fun game. Like I said, there are 1590 01:19:27.279 --> 01:19:29.920 expansions for it. It came out twenty fifteen, so I'm 1591 01:19:29.920 --> 01:19:37.560 gonna pick mysterium. I was talking to Dan before Steve 1592 01:19:37.600 --> 01:19:39.680 got on. He asked me how I was, and I 1593 01:19:39.720 --> 01:19:41.399 was like, tired because I had just gone for a run. 1594 01:19:41.439 --> 01:19:45.840 I'm training for another marathon, and so I'm going to 1595 01:19:45.960 --> 01:19:49.479 just do a couple of shout outs on that. I 1596 01:19:49.560 --> 01:19:53.760 have been doing the trainings off of a Training Peaks 1597 01:19:53.800 --> 01:19:56.800 training that I bought. So Training Peaks is free, and 1598 01:19:56.840 --> 01:19:59.239 you can buy training that you can stick on your 1599 01:19:59.239 --> 01:20:01.079 calendar and those are just fixed costs. So I think 1600 01:20:01.119 --> 01:20:04.279 I bought this one for twenty or thirty bucks. And 1601 01:20:04.319 --> 01:20:07.239 so you just tell it when your race is and it, 1602 01:20:08.359 --> 01:20:12.199 you know, Mind's what a sixteen week program, so you know, 1603 01:20:12.279 --> 01:20:13.800 I had to figure out when it started and just 1604 01:20:14.000 --> 01:20:17.319 stick it in and it puts the workouts on my 1605 01:20:17.399 --> 01:20:19.600 phone on my garment app, which sinks to my watch. 1606 01:20:20.479 --> 01:20:23.159 I have a Garment four Runner two thirty five, which 1607 01:20:23.199 --> 01:20:26.039 is not anywhere near the nearest newest model, but it 1608 01:20:26.079 --> 01:20:28.880 works great. So anyway, that's what I'm doing for running. 1609 01:20:29.199 --> 01:20:32.640 And then finally this week, I'm working on getting AI 1610 01:20:32.880 --> 01:20:36.720 for JavaScript dot com up. I'm also doing AI for 1611 01:20:36.880 --> 01:20:39.920 Ruby dot com. And so you can go and you 1612 01:20:39.920 --> 01:20:41.600 can get on the email list, and I'm going to 1613 01:20:41.640 --> 01:20:43.600 be emailing out, hey, this is what I'm doing with 1614 01:20:43.640 --> 01:20:48.119 AI this week. These are the APIs I used. I 1615 01:20:48.199 --> 01:20:50.720 plan on doing all my examples in both languages, so 1616 01:20:50.760 --> 01:20:52.920 you'll get the JavaScript ones on the JavaScript list and 1617 01:20:52.960 --> 01:20:56.119 the Ruby ones on the Ruby list. And I'm also 1618 01:20:56.159 --> 01:21:00.000 looking at putting together summit at the end of September 1619 01:21:00.159 --> 01:21:03.039 or the end of August. Beginning of September. We'll probably 1620 01:21:03.119 --> 01:21:05.319 have people like Steve. I've been talking to a bunch 1621 01:21:05.319 --> 01:21:08.239 of the other folks in the Ruby community that I 1622 01:21:08.279 --> 01:21:11.479 know that are doing AI, and really I want to 1623 01:21:11.479 --> 01:21:15.359 be hitting it at this level right where it's it's 1624 01:21:15.399 --> 01:21:20.000 not hey, here's how you math your way into models 1625 01:21:20.039 --> 01:21:22.760 that work. This is hey, you've got a model that works, 1626 01:21:22.920 --> 01:21:25.319 or you know, here's how maybe you modify a model 1627 01:21:25.359 --> 01:21:28.279 that you have to train it a little bit, but 1628 01:21:28.399 --> 01:21:30.319 I want it to be approachable for people who want 1629 01:21:30.319 --> 01:21:34.760 to add AI features too applications, not how to solve 1630 01:21:35.399 --> 01:21:38.840 whatever thing by you know, managing a data lake and 1631 01:21:38.880 --> 01:21:42.479 then feeding it into a system that generates the model 1632 01:21:42.520 --> 01:21:44.520 and then how to test the model and all that stuff. 1633 01:21:44.600 --> 01:21:47.079 Or we'll get into some of that at a high level, 1634 01:21:47.199 --> 01:21:49.079 but mostly we're going to be talking about, Hey, here 1635 01:21:49.079 --> 01:21:54.720 are the APIs for Claude or GPT four or mid 1636 01:21:54.800 --> 01:21:59.000 Journey or you know whatever. You know, I do podcasts, 1637 01:21:59.039 --> 01:22:02.159 so Whisper, you know, So here's how you use these 1638 01:22:02.199 --> 01:22:04.720 to get what you need. And then you know, here's 1639 01:22:04.760 --> 01:22:06.680 maybe how you tie a few of them together to 1640 01:22:06.720 --> 01:22:09.439 get a more complicated result than you want. Then you 1641 01:22:09.479 --> 01:22:12.119 can get from any one of them. And so that's 1642 01:22:12.119 --> 01:22:16.560 what I'm looking at. And so anyway, I'll be emailing 1643 01:22:16.600 --> 01:22:20.039 people on the email list right now. I'm just finalizing 1644 01:22:20.119 --> 01:22:23.600 my system because I've kind of had to rework bits 1645 01:22:23.720 --> 01:22:26.800 of the email stuff that I've been doing. So if 1646 01:22:26.840 --> 01:22:30.079 you're interested, go to AI for JavaScript dot com. I'm 1647 01:22:30.119 --> 01:22:32.680 also going to be doing weekly calls for JavaScript geniuses. 1648 01:22:32.720 --> 01:22:35.119 You can find that at JavaScript Geniuses dot com and 1649 01:22:35.119 --> 01:22:37.720 that's going to be sort of a back and forth, 1650 01:22:38.000 --> 01:22:41.439 ask questions, get answers. If you have feedback or ideas 1651 01:22:41.439 --> 01:22:44.159 for people, you can help out, you know, and help 1652 01:22:44.199 --> 01:22:46.399 people get where they want to go with that as well. 1653 01:22:46.920 --> 01:22:50.720 But kind of more of a mastermind or group coaching 1654 01:22:51.119 --> 01:22:53.760 sort of setup is kind of a blend. But we'll 1655 01:22:53.800 --> 01:22:55.640 be doing calls for that every week, and you'll be 1656 01:22:55.680 --> 01:23:00.199 getting some AI stuff in there too. So anyway, are 1657 01:23:00.239 --> 01:23:03.800 my picks and myself promos? Steve, what are your picks? 1658 01:23:04.199 --> 01:23:09.039 Yeah? I think probably just the obvious ones. Check out Builder, 1659 01:23:09.119 --> 01:23:11.720 check out visual co pilot, and I mentioned briefly, but 1660 01:23:11.920 --> 01:23:15.399 Microagent is a pretty interesting CLI tool that if you 1661 01:23:15.479 --> 01:23:19.920 want to categorize AI tools into sort of like commonly 1662 01:23:20.039 --> 01:23:23.159 used today, you know, Chatgypt, get a co pilot Claude, 1663 01:23:23.680 --> 01:23:26.600 and then where things are going agents. You know. I 1664 01:23:26.880 --> 01:23:28.520 do very much believe in agents in the medium to 1665 01:23:28.560 --> 01:23:31.600 long term. I don't see many practical use cases for 1666 01:23:31.640 --> 01:23:34.600 them today, but micro agent is the only one that 1667 01:23:34.720 --> 01:23:38.760 I've used. Obviously, I'm biased here obviously, but it kind 1668 01:23:38.760 --> 01:23:40.479 of solves some of the fundamental challenges. I've got a 1669 01:23:40.479 --> 01:23:42.960 blog post talking about this in better detail, but some 1670 01:23:43.000 --> 01:23:46.479 of the challenges of agents, and I think that I may. 1671 01:23:47.880 --> 01:23:48.840 Perfect micro agent. 1672 01:23:50.039 --> 01:23:53.640 Like many things the you know, I believe in technology 1673 01:23:53.680 --> 01:23:56.760 growing in layers incrementally, as opposed to like you know, 1674 01:23:56.960 --> 01:23:59.079 the V one is supposed to solve everything for everyone, 1675 01:23:59.199 --> 01:24:02.399 like Devin or whatever, a super micro or super AI agent, 1676 01:24:02.439 --> 01:24:05.560 this is everything. And so what's interesting about microagent and 1677 01:24:05.600 --> 01:24:07.479 why I suggest people try it out is it is 1678 01:24:07.520 --> 01:24:10.840 a very I think, new technique and it can lead 1679 01:24:10.840 --> 01:24:12.479 to a lot of interesting things. So people trying it 1680 01:24:12.479 --> 01:24:14.319 and giving feedback I think is really interesting because I 1681 01:24:14.359 --> 01:24:16.680 think that's how we solve these problems over time, is 1682 01:24:16.680 --> 01:24:19.239 how to make agents work. Make it work really well 1683 01:24:19.319 --> 01:24:22.000 for something, And I think this tool works really well 1684 01:24:22.039 --> 01:24:24.800 for a certain type of problem. And I think I 1685 01:24:25.000 --> 01:24:27.960 have ideas and experiences that suggest what's really good at 1686 01:24:28.239 --> 01:24:29.960 But I'd love to get more feedback from people what 1687 01:24:30.000 --> 01:24:32.479 they've found success and not success with, and just more 1688 01:24:32.520 --> 01:24:35.319 input gehab issues, poll requests. People have added some cool 1689 01:24:35.359 --> 01:24:37.439 new features to it. So I think it's an interesting 1690 01:24:37.479 --> 01:24:41.840 area of both a practical tool and like an avenue 1691 01:24:41.880 --> 01:24:43.600 for research and development that you could be part of 1692 01:24:43.880 --> 01:24:46.079 in AI and AI agents and stuff like that that 1693 01:24:46.680 --> 01:24:48.399 I would value. I want the project to be as 1694 01:24:48.399 --> 01:24:51.119 community driven as possible, so check it out, give feedback 1695 01:24:51.119 --> 01:24:54.920 great issues, send pull requests for improvements or fixes or whatever, 1696 01:24:55.640 --> 01:24:56.479 and let me know what you think. 1697 01:24:57.439 --> 01:25:00.479 Awesome people want to find you on the Internet. Where 1698 01:25:00.479 --> 01:25:01.239 do they find you, Steve? 1699 01:25:02.000 --> 01:25:04.880 I am Steve at seven oh Steve eight seven o eight, 1700 01:25:04.920 --> 01:25:09.199 on Twitter, YouTube, TikTok, LinkedIn. I don't know all the places, 1701 01:25:09.720 --> 01:25:12.199 but those are the ones I'm probably most commonly active on. 1702 01:25:13.479 --> 01:25:15.880 Awesome. Well, thanks again for coming. This has been awesome. 1703 01:25:16.479 --> 01:25:18.239 Thanks for having me. This is really fun. 1704 01:25:18.560 --> 01:25:20.840 All right, folks, We're gonna wrap it here till next time. 1705 01:25:20.960 --> 01:25:24.720 Max out