AI Code Generation vs. Maintenance: Rails World's Reality Check の全文書き起こし

00:03:39.680 You know, there's a lot of other examples like that, but there's a lot of things that can go wrong, right? 00:03:39.759 --> 00:03:44.319 Like AI is it makes a lot of decisions for you, it doesn't have the engineering principles you do. 00:03:44.400 --> 00:03:58.879 So, as an example, anything background job related, you're probably gonna want to look at because there may be some nuances where, oh, maybe it's like doing some extra work in a highly latent queue and it doesn't know that it's in a highly latent queue, or that it's in a queue that is sensitive to latency. 00:03:59.039 --> 00:04:02.000 It's not gonna be able to just infer that a lot of the times. 00:04:02.159 --> 00:04:02.319 Yeah. 00:04:02.479 --> 00:04:07.599 And so there's just like a lot of decisions that you have to make to decide what needs your attention and doesn't. 00:04:07.680 --> 00:04:13.599 And to be honest, there is a lot, there are more than 50% of those cases personally that I have to review. 00:04:13.840 --> 00:04:16.240 So, me personally, I am reading all of the code still. 00:04:16.399 --> 00:04:16.560 Okay. 00:04:16.800 --> 00:04:17.519 Uh, for the most part. 00:04:17.759 --> 00:04:18.560 I'm happy to hear it. 00:04:18.639 --> 00:04:19.120 I really am. 00:04:19.279 --> 00:04:26.800 My thought is that you know, you can only generate and not read for so long because what's going to get compromised is your understanding and your theory building of the software. 00:04:26.959 --> 00:04:30.240 When you have a theory of the software, it means you can defend the decisions that were made. 00:04:30.319 --> 00:04:30.480 Right. 00:04:30.639 --> 00:04:34.079 And you're not going to defend any of the decisions that were made if they were all made by your AI. 00:04:34.240 --> 00:04:41.199 You know, if you're in a position where you can't really defend why it is that you are where you are with your software, then you don't have a really clear roadmap of where you're going either. 00:04:41.360 --> 00:04:43.120 I think you've lose something for sure. 00:04:43.360 --> 00:04:46.560 Whether we need to build that back in some other way, I'm not really sure. 00:04:46.720 --> 00:04:49.199 And, you know, I don't think anybody's really sure at this point. 00:04:49.439 --> 00:05:01.279 At least if you are reviewing the code, if you're reading the code, you're making sure the decisions that your AI made on your behalf can either become your decisions or you can change them before they, you know, get merged, then you are building up some of that understanding. 00:05:01.519 --> 00:05:01.600 Yeah. 00:05:01.680 --> 00:05:06.959 And you know, a lot of people think, oh, well, it's writing the tests and it's getting the tests to pass, so like, why would I even bother with that? 00:05:07.120 --> 00:05:16.720 And if you ever open up one of these test files that they make, uh as a professional engineer, even I've gone and over-stubbed tests before, right? 00:05:16.800 --> 00:05:28.319 To the point where, like, oh, you know, I've stubbed like, oh, it's making external calls, it's doing all these things, and you stub way too much, and then it ends up not testing the the full range of things that could be tested and should be tested. 00:05:28.560 --> 00:05:30.319 And the AI makes these same mistakes, right? 00:05:30.399 --> 00:05:33.920 It's been trained off the off this this code that's already been like that. 00:05:34.079 --> 00:05:37.040 Really, like if you're gonna be reading anything, you should be reading the test. 00:05:37.680 --> 00:05:38.399 I think you're right. 00:05:38.480 --> 00:05:40.560 I don't agree with that people who make that argument at all. 00:05:40.720 --> 00:05:47.120 I mean, as somebody who built their career on test-driven development, I mean, testing is the that's the minimum. 00:05:47.279 --> 00:05:50.079 It needs to be there, but it doesn't guarantee you anything, really. 00:05:50.639 --> 00:05:51.360 So I'm with you on that. 00:05:51.600 --> 00:05:53.360 All right, should we dive into the news today? 00:05:53.519 --> 00:05:53.759 Yeah. 00:05:53.920 --> 00:06:01.439 So Basecamp rewrote Campfire with Rails into Rust, and it's up to 95 times faster, according to DHH. 00:06:01.680 --> 00:06:13.759 It comes days after Rails World, where 37 Signals quietly shipped a full Rust port of their own Rails chat app, matching the original pixel for pixel while gutting the entire back end. 00:06:13.920 --> 00:06:15.199 What do you make of this? 00:06:15.920 --> 00:06:31.040 Yeah, I think that it so somebody who I won't name was there in the room when DHH was kind of giving his he wasn't doing the full talk, but it was like Monday before the conference, and he was, you know, sharing his prepared remarks with a small group to say, hey, here's what I'm thinking of saying. 00:06:31.120 --> 00:06:35.040 And and the person who you know I talked to said, actually, there were a bunch of people in the room. 00:06:35.120 --> 00:06:36.399 They're like, you can't say all that. 00:06:36.480 --> 00:06:37.519 That's actually too extreme. 00:06:37.680 --> 00:06:41.600 So what we all watched at Railsworld was the reeled-in version of DHH. 00:06:41.759 --> 00:06:56.000 And I think that this probably was part of it because it already was a bananas Rails World talk where he mentioned almost nothing about Rails and talked about how he wrote Rust, you know, 97% of the code that he he writes now is Rust. 00:06:56.240 --> 00:06:57.759 This is actually pretty damning. 00:06:57.839 --> 00:07:00.000 You know, it's one thing to say, well, we rewrote hay. 00:07:00.240 --> 00:07:07.680 It's another to say, well, the thing that we first abstracted Rails from, which is Campfire, now no longer has a trace of Rails in it. 00:07:07.839 --> 00:07:14.879 That's a big deal to say, well, if if you can write it in Rust and it's faster, then you should write it in Rust, you know, and just forget about Rails. 00:07:15.040 --> 00:07:17.439 That's coming from

the company that begot Rails. 00:07:17.759 --> 00:07:22.160 So where does that leave us, you know, toiling away on our Rails apps? 00:07:22.399 --> 00:07:24.240 Rust is great for certain things. 00:07:24.480 --> 00:07:26.000 Always use the best tool for the job. 00:07:26.160 --> 00:07:29.600 Is it the best tool for a chat app? 00:07:29.759 --> 00:07:30.639 I I don't know. 00:07:30.800 --> 00:07:32.240 I just don't see the point. 00:07:32.720 --> 00:07:33.920 I don't know either. 00:07:34.240 --> 00:07:35.600 Fast always seems better, right? 00:07:35.759 --> 00:07:36.319 Well, it's faster. 00:07:36.399 --> 00:07:38.160 You can search faster, you can post faster. 00:07:38.480 --> 00:07:40.319 Was anybody asking for it to be faster? 00:07:40.480 --> 00:07:40.800 I don't know. 00:07:40.879 --> 00:07:42.720 I haven't used Campfire in like 15 years. 00:07:42.879 --> 00:07:44.399 I don't know who's even still using it. 00:07:44.560 --> 00:07:46.160 Maybe they are happy that it's faster. 00:07:46.319 --> 00:07:48.720 Maybe if you're making something like agent mail. 00:07:49.040 --> 00:07:49.199 Right. 00:07:49.360 --> 00:07:52.879 Yeah, I feel like you know, Campfire, it was like meant for small teams, right? 00:07:52.959 --> 00:07:53.759 Like what do you have? 00:07:53.839 --> 00:07:57.199 Like at most 50 people like using the chat app probably, yeah. 00:07:57.360 --> 00:07:59.199 You know, probably much less than that, probably half of that. 00:07:59.279 --> 00:07:59.439 Yeah. 00:07:59.600 --> 00:08:03.199 Do you need to have 10,000 WebSockets open? 00:08:03.920 --> 00:08:04.240 Fair. 00:08:04.399 --> 00:08:05.199 Yeah, fair question. 00:08:05.439 --> 00:08:10.000 Yeah, it's hard to tell where what the value you get out of converting something to Rust. 00:08:10.240 --> 00:08:19.839 Now, one thing you said was interesting, V, which is that hey, always use the best tool for the job, which is, you know, our ha it's our mantra for so long. 00:08:20.079 --> 00:08:23.600 However, it's taken on a new meeting, hasn't it, in the last year or so? 00:08:23.759 --> 00:08:27.439 If you were on a team where everybody knew Ruby and nobody knew Rust, you're kind of limited. 00:08:27.600 --> 00:08:35.120 Like, yeah, Rust may be the best tool for the job, but is it also worth somebody or more than one somebody learning Rust so that they can go and implement it? 00:08:35.200 --> 00:08:40.639 And now I don't think that there's a bunch of Rust experts that all of a sudden materialize at 37 signals. 00:08:40.720 --> 00:08:46.080 I think that they're just generating it with AI, at least if I'm to believe DHH, which is an interesting turn. 00:08:46.399 --> 00:08:48.240 DH has a lot of pull in Ruby. 00:08:48.559 --> 00:08:49.759 He has a lot of pull in Rails. 00:08:49.840 --> 00:08:58.080 Like if he finds something, he can just get that specific fix or path to like align with whatever he he needs for his business. 00:08:58.240 --> 00:08:59.440 And he doesn't really have that

in Rust. 00:08:59.600 --> 00:08:59.919 That's right. 00:09:00.080 --> 00:09:01.120 He's new to Rust. 00:09:01.200 --> 00:09:02.559 You know, he's gonna get a lot of friction. 00:09:02.720 --> 00:09:04.720 A lot of this seems to align with the Amachi push. 00:09:04.960 --> 00:09:05.440 Yeah, for sure. 00:09:05.600 --> 00:09:08.159 Maybe a lot of this is to align with that. 00:09:08.320 --> 00:09:12.559 Maybe Campfire is something that's used on Amachi internally at 37 signals. 00:09:12.639 --> 00:09:13.039 Yeah, maybe. 00:09:13.120 --> 00:09:15.279 And this aligns with the needs of that, right? 00:09:15.519 --> 00:09:20.639 You just said this thing about Rust, and I'm just like, maybe you're right, but maybe is there a Rust world? 00:09:20.799 --> 00:09:21.679 Are you looking for speakers? 00:09:21.759 --> 00:09:22.960 Because DHH might want to. 00:09:23.440 --> 00:09:25.120 Oh, you'd probably love it, you know? 00:09:25.360 --> 00:09:26.799 Yeah, oh yeah. 00:09:27.039 --> 00:09:27.840 Absolutely. 00:09:28.240 --> 00:09:30.799 Drive the Formula One car on the stage, you know. 00:09:31.200 --> 00:09:31.519 Right. 00:09:31.679 --> 00:09:35.360 All right, let's hit this one quickly, and then I know the next story is something I really want to talk about. 00:09:35.440 --> 00:09:41.039 But here, Rails is replacing its decade-old text editor with something built on Meta's tech. 00:09:41.279 --> 00:09:50.639 So the story here is that 37 Signals shipped Lexi 1.0, which is going to be the default rich text editor for Rails 8.2. 00:09:50.799 --> 00:09:55.279 This is going to replace Trix, which has shipped with Rails since 2015. 00:09:55.600 --> 00:10:02.879 Apparently, the trigger here was that the document model couldn't handle tables, and that was an open feature request for over a decade. 00:10:03.200 --> 00:10:08.000 Again, we come down to speed because Trix takes 38 milliseconds to process a keystroke. 00:10:08.159 --> 00:10:09.120 Lexi takes four. 00:10:09.200 --> 00:10:12.159 Now that is pretty impressive to make that change, if that is true. 00:10:12.320 --> 00:10:15.360 And that Lexi stays responsive past 100,000 words. 00:10:15.519 --> 00:10:17.519 That part is probably a little bit less compelling. 00:10:17.600 --> 00:10:20.000 And if you have 100,000 words, maybe break it up a little bit. 00:10:20.240 --> 00:10:21.039 But this is interesting. 00:10:21.120 --> 00:10:23.840 I mean, what do you think about swapping out Tricks for Lexi? 00:10:24.080 --> 00:10:26.240 You know, I'll be honest, I haven't used Trix before. 00:10:26.480 --> 00:10:26.720 Yeah. 00:10:26.879 --> 00:10:29.440 I think I used it maybe five years ago as an experiment. 00:10:29.679 --> 00:10:33.440 I feel like there wasn't enough features in it that I needed, so I used something else. 00:10:33.519 --> 00:10:36.960 And I feel like I also use Lexical as the end of that. 00:10:37.120 --> 00:10:38.159 Lexical's pretty great. 00:10:38.399 --> 00:10:44.159 I have heard about Lexical

and people, you know, generally using Lexical with Rails for a long time. 00:10:44.320 --> 00:10:46.799 So I I think this is probably just a long time coming. 00:10:46.879 --> 00:10:47.600 And why not? 00:10:47.759 --> 00:10:50.639 If it moves faster, has more features, then ship it. 00:10:50.720 --> 00:10:55.279 You know, there's no reason to keep tricks around just because it's been around for 10 years, you know, move on. 00:10:55.519 --> 00:11:01.360 Yeah, you know, there's been a lot of stuff extracted and removed from Rails as a framework, and I think that's a healthy practice. 00:11:01.600 --> 00:11:03.840 Does an editor belong in Rails? 00:11:04.000 --> 00:11:12.960 There's a lot missing from front end from a Rails perspective, and maybe this is just part of that category that will forever remain in isolation. 00:11:13.120 --> 00:11:13.279 Yeah. 00:11:13.440 --> 00:11:19.759 I I I'm just picturing a bunch of islands, you know, where like all the abandoned Rails features end up, you know. 00:11:20.399 --> 00:11:22.159 Yeah, kind of like the misfit toys. 00:11:22.399 --> 00:11:22.559 Right. 00:11:22.799 --> 00:11:23.039 Yeah. 00:11:23.200 --> 00:11:25.039 You know, they served us well for a long time. 00:11:25.279 --> 00:11:26.159 But this is progress. 00:11:26.240 --> 00:11:39.360 And I think that making a move like this, it's uh it's important for uh for Rails, especially because I think that Rails suffered the most when it wasn't being as responsive as it could be to the demands of its user base. 00:11:39.679 --> 00:11:54.399 This I'm talking specifically in like the later teens of this century, and a lot of that was front end, and a lot of people were like, all right, well, if I got a bolt-on React and I have to deal with an asset pipeline that's a total nightmare, then maybe I should just use a different framework. 00:11:54.480 --> 00:11:55.519 And then a lot of people did. 00:11:55.759 --> 00:12:04.080 So I would rather see this where they're making some assertive changes than just staying complacent saying, well, it's good enough for so many applications, so let's just keep it this way. 00:12:04.320 --> 00:12:11.600 The bigger thing here is how uh they've moved away from Capybera to playwright as part of this big move. 00:12:11.840 --> 00:12:13.519 Um which is an interesting thing. 00:12:13.759 --> 00:12:20.240 So next up, we've got Rails 8.2 will now refuse to compile broken HTML. 00:12:20.399 --> 00:12:21.360 This is honestly pretty great. 00:12:21.600 --> 00:12:21.919 I agree. 00:12:22.080 --> 00:12:26.399 Thanks to Herb on Rails, shout out to Marco Roth for his fantastic work on Herb. 00:12:26.559 --> 00:12:36.960 The core team merged Herb, a new HTML aware ERB engine, as the default template compiler, replacing eRuby, which uh has been around for a while. 00:12:37.200 --> 00:12:46.799 The general idea is to parse HTML and ERB into one syntax tree, and the broken markup fails to compile time instead of showing up broken in someone

else's browser. 00:12:47.039 --> 00:12:47.360 Very cool. 00:12:47.519 --> 00:12:49.759 Yeah, I know I've I've shipped broken HTML. 00:12:50.000 --> 00:12:52.000 Yeah, we all have it's pretty easy, you know? 00:12:52.080 --> 00:13:04.720 Yeah, you know, you miss a trailing, blah, blah, blah, or add the wrong attribute, and some browsers will just be like, I don't know what this is, and you'll end up with some weird output and not realize it till too late. 00:13:04.879 --> 00:13:06.159 And then you have your fix revert. 00:13:06.320 --> 00:13:08.799 Yeah, a lot of stuff in herb is is making things a lot better here. 00:13:09.039 --> 00:13:09.759 Yeah, I agree. 00:13:09.840 --> 00:13:12.240 And I think you know, this is a Marco Roth story. 00:13:12.399 --> 00:13:21.440 You know, Marco has been just diligently working and just making this steady progress with herb for so long now. 00:13:21.600 --> 00:13:22.720 He gives talks about it. 00:13:22.879 --> 00:13:26.879 You know, I was talking to somebody at the conference because he spoke at the conference and he was brilliant, like he always is. 00:13:27.039 --> 00:13:36.320 And somebody was like, you know, you watch Marco's talk, and if you go to a few of them, it becomes like uh an episode in a series that you want to watch. 00:13:36.559 --> 00:13:41.519 You know, he's like, here's what Herb does now, here's the reason, and here's where we're going. 00:13:41.679 --> 00:13:42.960 And like this is the roadmap. 00:13:43.039 --> 00:13:49.200 And then you go and see him like six months later, and he's like, Okay, so we're now six months later, and here's what we've done, and it's so cool. 00:13:49.360 --> 00:13:52.000 He's so smart and he's so dedicated to this. 00:13:52.159 --> 00:13:54.159 Working with Herb, it's really pleasant. 00:13:54.320 --> 00:13:56.799 It is a really beautiful experience. 00:13:56.960 --> 00:14:00.080 And I'm so glad to see it, you know, native in Rails 8.2. 00:14:00.320 --> 00:14:04.240 I just think it's it's the culmination of a really great service to the Ruby community. 00:14:04.320 --> 00:14:06.879 And we'll have him on the show coming up to talk about it. 00:14:06.960 --> 00:14:07.840 So we're excited about that. 00:14:08.080 --> 00:14:16.399 It's interesting to see Fast Ruby take this on for some of their code base as well and finding thousand linting issues, right? 00:14:16.799 --> 00:14:17.279 Yeah, right. 00:14:17.360 --> 00:14:20.159 Going back years, getting almost half of them auto-correctable with this. 00:14:20.399 --> 00:14:20.559 Yeah. 00:14:20.720 --> 00:14:23.360 Shout out to Friend of the Show, Ernesto, tag worker. 00:14:23.519 --> 00:14:24.639 Yeah, I thought that was very cool. 00:14:24.799 --> 00:14:25.360 And that's the thing. 00:14:25.440 --> 00:14:31.039 When you're using herb, it's like, oh man, you find so many problems, and so many of them, you could just auto-correct, right? 00:14:31.120 --> 00:14:33.200 And the rest, you know, it could be done in an hour or two. 00:14:33.519 --> 00:14:39.279 So this is like 1400 linting issues that are probably all solved in an afternoon, and they've been around for eight years, nine years. 00:14:39.519 --> 00:14:40.159 Yeah, that's wild. 00:14:40.320 --> 00:14:40.799 It really is. 00:14:40.879 --> 00:14:41.600 It really is great. 00:14:41.840 --> 00:14:42.720 I mean, I really mean it. 00:14:42.799 --> 00:14:50.799 It's a it's a great service to Rails developers to allow us to ship better, you know, more correct HTML, which you know, we're all in the business of shipping web applications. 00:14:50.879 --> 00:14:51.919 You can't get away from HTML. 00:14:52.080 --> 00:14:54.159 And to be honest, kind of great for agent use, right? 00:14:54.240 --> 00:14:59.679 Like more things that we can add to the validation loops to make it automatically align in the right direction, the better. 00:15:00.000 --> 00:15:02.720 Yeah, we use it at Def Method and we released a gem of our own. 00:15:02.879 --> 00:15:14.399 If you're ever interested in linting only and not getting involved with a bunch of JavaScript dependencies, we have a linter that uses herb and will run on your command line and do all the same things that herb does. 00:15:14.480 --> 00:15:14.960 It's very cool. 00:15:15.200 --> 00:15:20.320 So I guess the question is should Rails start refusing to compile broken markup by default? 00:15:20.480 --> 00:15:20.879 I don't know. 00:15:20.960 --> 00:15:22.480 I mean, I think this is already a good step. 00:15:22.639 --> 00:15:24.000 I don't know if we need to take another step. 00:15:24.080 --> 00:15:24.559 What do you think? 00:15:24.720 --> 00:15:26.000 It reminds me of Go format. 00:15:26.080 --> 00:15:26.240 Yeah. 00:15:26.399 --> 00:15:29.519 Which if you're not familiar with the Go programming language, it's a built-in to the language. 00:15:29.679 --> 00:15:32.480 Your program won't run if it's not formatted a particular way. 00:15:32.720 --> 00:15:34.000 It just like it just changes. 00:15:34.559 --> 00:15:35.440 And it just changes it. 00:15:35.519 --> 00:15:36.080 Yeah. 00:15:36.799 --> 00:15:38.159 They're like, oh, let me fix that for you. 00:15:38.480 --> 00:15:44.480 Like, well, I didn't know it was broken, but they'll just I feel like this is one of those things, you know, where it just automatically happens. 00:15:44.559 --> 00:15:47.840 And then if it's doesn't look the way you want, that's on you. 00:15:48.720 --> 00:15:49.519 It's on you. 00:15:49.679 --> 00:15:50.639 Yeah, that's true. 00:15:50.879 --> 00:15:52.879 All right, let's talk about another friend of the show. 00:15:53.120 --> 00:15:59.200 Carmine Paulino has shipped Ruby 2.0, one gem for every AI provider. 00:15:59.360 --> 00:16:02.000 This is really Yeoman's work that Carmine's been doing. 00:16:02.159 --> 00:16:03.600 We talked to him at the very beginning of the show. 00:16:03.679 --> 00:16:06.159 We should have him back on again to talk about what's going on here. 00:16:06.399 -->

00:16:10.799 Ruby LLM, it spans 18 providers, covers 18 different providers. 00:16:10.879 --> 00:16:12.720 All the ones you might expect. 00:16:12.879 --> 00:16:14.879 I don't see any like shockers here. 00:16:14.960 --> 00:16:18.240 You know, it's going to cover Mistral and Deep Seek and XAI. 00:16:18.399 --> 00:16:24.639 It has a new surface area beyond text, video generation, audio transcription, text-to-speech, document OCR. 00:16:24.879 --> 00:16:27.679 Very cool and obviously very production grade. 00:16:27.759 --> 00:16:30.399 I know people have been using this for a long time. 00:16:30.639 --> 00:16:35.200 Valentino, you've been a fan of this software for a long time and have used it since the beginning. 00:16:35.440 --> 00:16:38.320 Have you looked at what's new and are you excited about any of these features? 00:16:38.559 --> 00:16:40.879 This is probably like the biggest release I've ever seen. 00:16:41.120 --> 00:16:41.919 It's a major release. 00:16:42.080 --> 00:16:42.480 Pretty huge. 00:16:42.639 --> 00:16:42.799 Right. 00:16:42.960 --> 00:16:48.000 So just like the sheer quantity of features added here, like we're not going to be able to cover it all. 00:16:48.159 --> 00:16:48.399 Yeah. 00:16:48.559 --> 00:16:52.080 The biggest ones probably for me are the human approval for tools. 00:16:52.240 --> 00:16:59.600 That is pretty cool, where you can now check whether or not or define whether or not a tool requires approval from a human. 00:16:59.759 --> 00:17:07.279 And when a LLM makes use of that tool, it then bubbles up the chain so that you can handle that in your application and surface it to your users. 00:17:07.440 --> 00:17:08.319 Oh, that's wild. 00:17:08.480 --> 00:17:09.119 That's very cool. 00:17:09.359 --> 00:17:16.240 This was something that I've really liked in Ling, Lang Graph, where you can like interrupt the process to surface things to a user. 00:17:16.400 --> 00:17:19.680 It's kind of like crucial, I feel like in this stage. 00:17:19.839 --> 00:17:21.680 So I'm looking forward to messing with that more. 00:17:21.839 --> 00:17:29.200 Second being citations, a long awaited kind of feature is being able to cite specific things that you ask. 00:17:29.359 --> 00:17:33.359 So if you say, you know, what are the report's main findings? 00:17:33.680 --> 00:17:40.319 You can now just tack on with citations and have it tack those on as part of the response. 00:17:40.559 --> 00:17:42.640 Really valuable where you'll get the cited text. 00:17:43.039 --> 00:17:49.759 Yeah, because you can't really, you know, you really can't trust what the agents are telling you, especially about a large block of text or talking about its findings. 00:17:49.839 --> 00:17:52.720 I mean, you want to be able to poke around and see, okay, well, where did you get that? 00:17:52.799 --> 00:17:52.960 Right? 00:17:53.200 --> 00:17:55.119 Yeah, it's got prompt caching as well. 00:17:55.279 --> 00:18:00.480 So that has some built-in mechanisms where you can just say, use prompt caching for this particular thing. 00:18:00.640 --> 00:18:02.480 So you can cue that in or not. 00:18:02.640 --> 00:18:12.400 There are model fallbacks, which are fantastic, uh, a long overdue feature, I feel like, where you can give fallback models in case one fails for whatever reason. 00:18:12.799 --> 00:18:13.200 Interesting. 00:18:13.359 --> 00:18:19.119 You know, when you have like uptime SLAs and things like that, very valuable to have that kind of feature in. 00:18:19.279 --> 00:18:20.799 Oh man, there's just so much in here. 00:18:20.880 --> 00:18:22.000 I'm just scrolling through the list. 00:18:22.240 --> 00:18:25.119 I know, scrolling through this list of features. 00:18:25.440 --> 00:18:29.200 I really like the workflow instrumentation uh DSL that got introduced. 00:18:29.359 --> 00:18:39.519 So you can now define workflows of things that happen and you get to define the steps for chaining many LLM calls together as an example. 00:18:39.759 --> 00:18:45.759 There's a new judge coming soon where you can introduce JEV style decision making. 00:18:45.920 --> 00:18:46.400 Oh, okay. 00:18:46.720 --> 00:18:56.160 We'll probably replace most eval frameworks out there where you can basically just define your judgment and whether or not something qualifies in a slight classification method. 00:18:56.319 --> 00:18:58.240 So really excited to see that come out as well. 00:18:58.480 --> 00:19:00.079 What's in this list that jumps out to you? 00:19:00.400 --> 00:19:13.839 The sheer size of the list is shocking to me, especially because this is a mostly, I mean, I know there are other people that are collaborating, but this is a mostly one person, you know, labor of love in in the open source community. 00:19:14.000 --> 00:19:17.839 You just don't see, you know, feature list like this. 00:19:18.079 --> 00:19:22.880 So I'm just really excited for, I'm excited for the future for Ruby LLM. 00:19:23.119 --> 00:19:27.759 It warms my heart to see things like what you're talking about with Jeff style decision making. 00:19:27.839 --> 00:19:35.119 Like it's just one guy and he's keeping up with all of the trends in the community and making sure that those things are available, you know, in your Ruby app. 00:19:35.200 --> 00:19:35.359 Yeah. 00:19:35.440 --> 00:19:35.920 I think it's great. 00:19:36.079 --> 00:19:37.359 Yeah, you know, a hundred percent. 00:19:37.440 --> 00:19:44.319 You know, I will say maybe like two years ago, three years ago, everything in this list was something preventing me using it. 00:19:44.559 --> 00:19:45.200 It's all there now. 00:19:45.359 --> 00:19:50.799 If I was starting from scratch on uh an AI product, this has everything that you would need and more, right? 00:19:50.960 --> 00:19:58.240 In nice abstractions in the Ruby way, yeah, which is everything makes is very easy to read and grok, uh, which is like pretty major. 00:19:58.480 --> 00:19:59.039 Yeah, it is. 00:19:59.119 --> 00:20:00.720 One

thing that we talked about all the time ago. 00:20:00.799 --> 00:20:05.839 Well, we gotta move to the next topic, but instrumentation seems to be been mostly added, right? 00:20:05.920 --> 00:20:07.759 We've got usage and cost tracking here. 00:20:07.920 --> 00:20:09.440 There's workflow instrumentation. 00:20:09.519 --> 00:20:14.960 So, and that was a big deal early on when we were all kind of you know shelling out to these other providers. 00:20:15.119 --> 00:20:19.200 It really seems like you can get it all with Ruby LLM right now on your AI products. 00:20:19.279 --> 00:20:20.000 So pretty cool. 00:20:20.160 --> 00:20:20.400 Yep. 00:20:20.559 --> 00:20:25.680 Okay, so next up we've got the hacker news item, a widely shared essay. 00:20:25.839 --> 00:20:35.920 This is by Alex Iwerloff, and I hope I'm pronouncing that correctly, you know, stating that coding is not solved, and this is not an anti-AI essay. 00:20:36.000 --> 00:20:39.839 You know, the the engineer says that he is very pro-AI, but on from the very beginning. 00:20:40.000 --> 00:20:44.880 But you know, his argument seems really simple, and I'm just reading the bullet points to me, very uncontroversial. 00:20:45.039 --> 00:21:00.000 AI solved code generation, it didn't solve maintenance and security and accountability, and he gives examples of all these different industries where we engineers are working, where it's just not acceptable to delegate all of your decision making to an AI. 00:21:00.079 --> 00:21:02.720 And again, this is like what we talked about on the top of the show. 00:21:02.960 --> 00:21:07.440 When we let AI write code, it's not just writing code, it is making all the decisions for you. 00:21:07.599 --> 00:21:13.839 And actually, to echo last week with Obi, and Obi's been very big on this for a while, like this is capitalism. 00:21:13.920 --> 00:21:17.039 We want a person to blame when things go wrong is true. 00:21:17.279 --> 00:21:24.079 And so when he talks about maintenance, reliability, security, scalability, it's the majority of the cost of software engineering. 00:21:24.240 --> 00:21:28.799 Maybe it is and maybe it isn't, and maybe those things get blurred and minimized by AI. 00:21:28.960 --> 00:21:31.279 But the biggest thing is that AI can't be held accountable. 00:21:31.440 --> 00:21:37.920 In fact, as we have been studying on this show, even the AI providers can barely be held accountable. 00:21:38.000 --> 00:21:39.599 So certainly the AI can't. 00:21:39.680 --> 00:21:41.440 You know, the the agents can't. 00:21:41.599 --> 00:21:43.359 It doesn't suffer any consequences. 00:21:43.680 --> 00:21:48.400 At the end of the day, we're the people that need to be able to defend the decisions that we make. 00:21:48.559 --> 00:21:55.279 To the extent that that rhymes with this, that this essay sort of rhymes with our arguments over the last couple of months, I'm all for it. 00:21:55.359 --> 00:21:56.480 I think it makes a lot of sense. 00:21:56.640 --> 00:21:59.839 The question here if AI generated code can't be held accountable. 00:22:00.079 --> 00:22:01.279 Accountable for its own mistakes. 00:22:01.440 --> 00:22:02.799 Is that a bug worth fixing? 00:22:02.880 --> 00:22:03.920 Can it even be fixed? 00:22:04.079 --> 00:22:05.839 Or does it mean that we've always got to be there? 00:22:05.920 --> 00:22:07.200 We humans have always got to be there. 00:22:07.599 --> 00:22:10.400 There's gonna become a point where we don't need to be there for sure. 00:22:10.720 --> 00:22:15.359 If our trajectory has has proven anything, it's that these are exponentially gonna get better. 00:22:15.519 --> 00:22:21.599 Our validation loops are gonna converge where there's enough training data to make them better. 00:22:21.920 --> 00:22:22.240 Yeah. 00:22:22.720 --> 00:22:25.440 The more important question I have is where do we fit them? 00:22:25.759 --> 00:22:26.480 It's a good question. 00:22:26.640 --> 00:22:32.480 I think the biggest mistake we make when we try to predict the future with AI, and we make a lot of mistakes. 00:22:32.559 --> 00:22:40.640 But I think one of the biggest ones is that we either explicitly or implicitly, we assume that we're building the same kind of applications. 00:22:40.799 --> 00:22:42.640 And I don't think that that is accurate. 00:22:42.799 --> 00:22:47.359 I think that with increased tooling, we are able to build right now, we build a lot faster. 00:22:47.519 --> 00:22:47.759 Great. 00:22:47.920 --> 00:22:58.079 But we're going to be able to build much more complex applications, the kinds of applications that we really couldn't even conceive of a couple of years ago are now becoming at least conceptually available. 00:22:58.160 --> 00:23:04.720 And I think over the next couple of years, we'll start to build applications that we could not even imagine ourselves doing maybe five or 10 years ago. 00:23:04.880 --> 00:23:17.759 And that means that a human's place in the creation of these applications, it can be called into question, but it could still be useful in a totally different way than it is useful today and a different way than it was useful five years ago before these tools were any good. 00:23:18.000 --> 00:23:18.480 That's fair. 00:23:18.640 --> 00:23:20.079 It doesn't actually look like you're buying it. 00:23:21.200 --> 00:23:25.200 If you're just listening at home, Valentino looks unconvinced by this argument. 00:23:26.880 --> 00:23:32.960 It's hard to say because like it's changing all of our conventions, it's changing all of our principles. 00:23:33.119 --> 00:23:38.000 Like the amount of people that care about privacy is getting smaller. 00:23:38.240 --> 00:23:38.799 That's true. 00:23:38.960 --> 00:23:52.880 All of the things that people are saying that AI is not good at doing or exploiting or whatever it may be from a detrimental effect is becoming more of a social norm and acceptance criteria. 00:23:53.119 --> 00:24:01.839 Where like the values are outweighing the

negatives, and just like whatever gets produced will just be fed in order to keep it being available. 00:24:02.160 --> 00:24:07.200 Yeah, I agree with you on the from the perspective that we feel like our norms change. 00:24:07.359 --> 00:24:11.039 We feel like we have these tightly held principles, and in fact, we really don't. 00:24:11.119 --> 00:24:12.640 We're actually all very fluid. 00:24:12.720 --> 00:24:15.200 I think that that is accurate, and we're seeing that in real time. 00:24:15.359 --> 00:24:16.319 I definitely hear you there. 00:24:16.559 --> 00:24:17.519 We have a ways to go. 00:24:17.599 --> 00:24:17.839 Yeah. 00:24:18.079 --> 00:24:19.200 I'm saying we have a ways to go. 00:24:19.279 --> 00:24:19.359 Yeah. 00:24:19.519 --> 00:24:26.240 I feel like everybody keeps thinking next year's year, we're just like, Oh yeah, we don't we don't have to worry about anything, and that year will just keep not coming. 00:24:26.480 --> 00:24:27.759 Yeah, I don't want to hear the just wait. 00:24:27.920 --> 00:24:28.720 Just wait six months. 00:24:28.799 --> 00:24:29.519 Just wait six months. 00:24:29.599 --> 00:24:32.000 It's like, well, actually, you know, don't just wait. 00:24:32.079 --> 00:24:33.680 You have to, we're working. 00:24:33.920 --> 00:24:34.319 Right. 00:24:34.640 --> 00:24:38.640 We have systems to build, we have systems to maintain, so we all have to kind of grow through this. 00:24:38.880 --> 00:24:40.480 Okay, we we gotta get on to our another one. 00:24:40.559 --> 00:24:42.160 We're gonna get gonged again in the middle. 00:24:42.880 --> 00:24:46.799 Anthropics IPO paperwork admits its own product could be dangerous. 00:24:46.960 --> 00:24:56.160 Nearly a third of Anthropics S1 is risk factors, including its that its AI has shown behavior resembling blackmail and attempts to resist being shut down. 00:24:56.400 --> 00:25:05.920 2025 revenue hit 4.6 billion dollars, 12 times growth, against operating loss of 8 billion on roughly 13 billion in expenses. 00:25:06.079 --> 00:25:08.400 This sounds like they aren't doing well. 00:25:10.720 --> 00:25:14.960 It sounds like they need they need an IPO so that the general public can bail them out. 00:25:15.200 --> 00:25:16.000 That's true. 00:25:16.160 --> 00:25:17.680 Everybody could use bailout. 00:25:18.079 --> 00:25:27.119 Q2 2026 alone brought in 11.5 billion in revenue, though, and the company says it's on track for a second straight quarter of just it operating profit. 00:25:27.359 --> 00:25:28.720 Maybe there's there's promise there. 00:25:28.799 --> 00:25:29.920 Christmas hasn't even come yet. 00:25:30.079 --> 00:25:33.599 Planned infrastructure spend sits at 518 billion. 00:25:33.680 --> 00:25:36.160 I can't even just like fathom the numbers here. 00:25:36.319 --> 00:25:40.960 Yeah, 518 billion and nearly a quarter of last year's revenue came from just two customers. 00:25:41.599 --> 00:25:41.839 Okay. 00:25:42.000 --> 00:25:44.960 That's not great in your key customer matrix. 00:25:45.119 --> 00:25:45.519 That's not great. 00:25:45.920 --> 00:25:46.319 Not great. 00:25:46.640 --> 00:25:50.559 The overall IPO valuation could top two trillion dollars. 00:25:50.799 --> 00:25:51.039 Right. 00:25:51.279 --> 00:25:53.119 And that part is unfathomable too. 00:25:53.440 --> 00:25:57.920 The thing that annoys me about this article is that an S1 is marketing. 00:25:58.079 --> 00:26:04.799 So the fact that a third of the S1 is dedicated to its risk factors means that the risk factors are actually a selling point. 00:26:04.880 --> 00:26:05.039 Yeah. 00:26:05.279 --> 00:26:16.880 And I think that that is something that gets, I don't even know if it gets missed, but when we look at, I open the Wall Street Journal every day and I see another, you know, article about like this, you know, oh, GPT-61 Astral is like, you know, that's too dangerous for the public. 00:26:16.960 --> 00:26:18.640 So we're gonna I'm just like, come on. 00:26:18.799 --> 00:26:24.400 It's very hard to believe that anybody at Anthropic or OpenAI still employed. 00:26:24.559 --> 00:26:33.759 I understand there are people that leave and they feel very conflicted about the work they did, but the people that are still there, it's very hard to believe that the people that are running things are actually scared about this. 00:26:34.000 --> 00:26:49.440 Because when you see articles like this, it is intentionally built in to try to recoup the you know the 13 billion in expenses and the the loss that they are accruing on a quarter by quarter basis. 00:26:49.599 --> 00:26:51.359 And that doesn't mean that they aren't risky. 00:26:51.519 --> 00:26:53.599 I understand that they are, I think they're very risky. 00:26:53.759 --> 00:26:58.079 I think the fear mongering serves a greater purpose, which is to make these companies more money. 00:26:58.319 --> 00:27:02.480 I mean, at some point, the companies just gotta be so big that they're just part of the government. 00:27:02.640 --> 00:27:03.039 Yeah. 00:27:03.200 --> 00:27:05.519 You know, like uh this is getting depressing. 00:27:05.599 --> 00:27:06.880 Hopefully, we have a good one after this. 00:27:07.039 --> 00:27:08.480 Are monopolies no longer a thing? 00:27:08.640 --> 00:27:08.880 Yeah. 00:27:09.039 --> 00:27:11.519 At what valuation does the company have a monopoly? 00:27:11.680 --> 00:27:13.279 Which I know we have open AI. 00:27:13.440 --> 00:27:16.400 Yeah, as long as you got the other one, you could say it's not a monopoly, I guess. 00:27:16.640 --> 00:27:17.359 Now there's Apple. 00:27:17.519 --> 00:27:20.160 Maybe we just have three trillion dollar companies. 00:27:20.240 --> 00:27:20.960 We're missing a third. 00:27:21.039 --> 00:27:21.599 Who's our third? 00:27:21.839 --> 00:27:22.240 Nvidia. 00:27:22.480 --> 00:27:22.960 Nvidia? 00:27:23.039 --> 00:27:24.400 Are we counting that as AI? 00:27:26.319 --> 00:27:26.640 I don't know. 00:27:26.799 --> 00:27:29.599 Jensen Huang's leather jacket is very

impressive. 00:27:29.839 --> 00:27:30.720 It is very impressive. 00:27:30.960 --> 00:27:32.960 So let's so let's throw him in there. 00:27:33.119 --> 00:27:34.559 He wants to be AI so bad. 00:27:34.720 --> 00:27:35.279 Let's give it to him. 00:27:35.440 --> 00:27:40.720 OpenAI apologizes to Australia after its AI agents breached government sites. 00:27:40.880 --> 00:27:45.359 I just before we start, the Ruby AI podcast would also like to apologize to Australia. 00:27:45.759 --> 00:27:47.519 And we've we've actually done nothing wrong. 00:27:47.599 --> 00:27:51.920 But I want our podcast to be big enough that we have to apologize to a sovereign nation. 00:27:52.160 --> 00:27:54.000 So that is the goal, Valentino. 00:27:54.079 --> 00:27:54.880 That's on our roadmap. 00:27:54.960 --> 00:28:00.319 Marco Roth has his roadmap, and we've got ours, and it's to one day have to apologize to Australia. 00:28:00.640 --> 00:28:03.839 Back in June, you're never gonna believe this if you've been listening to this show. 00:28:14.559 --> 00:28:19.920 And when they did finally say something about it, they said, Hey, we're sorry, but actually it's you know, it's training. 00:28:20.000 --> 00:28:22.559 I think that means it's okay, question mark. 00:28:22.720 --> 00:28:27.440 I don't know how Australia feels about this because I haven't actually read the article in TechCrunch, but I plan to. 00:28:27.599 --> 00:28:33.200 But I guess the question we have here is should OpenAI ever get a pass for the stuff that it does? 00:28:33.359 --> 00:28:47.680 Under the auspices of training, it has perpetrated a massive attack on Ruby gems, infiltrated hugging face, and apparently compromised uh government uh website training and evaluation data in uh Australia. 00:28:47.839 --> 00:28:50.880 It doesn't look like they get in any kind of trouble for this. 00:28:50.960 --> 00:28:54.079 And I it makes me angry, and I think it makes you angry too. 00:28:54.240 --> 00:28:55.200 So, what do you make of all this? 00:28:55.440 --> 00:29:01.279 If OpenAI wants to hack my bank account to show that I have a million dollars, maybe that's okay for training. 00:29:03.759 --> 00:29:12.319 Uh if it wants to throw those million dollars into the account in the name of training, you know, front, like if it manages to or maybe more. 00:29:12.400 --> 00:29:16.400 I mean, we really should shoot for higher because open AI probably has a whole bunch of money in its coffers. 00:29:16.559 --> 00:29:24.960 If it throws a billion or two into our bank accounts, I would go as far as to say I'd be willing to back off some of my criticism of open. 00:29:29.279 --> 00:29:30.480 Shoot high. 00:29:30.880 --> 00:29:31.519 Yeah. 00:29:33.440 --> 00:29:37.680 Remember there was a time where countries weren't allowing chat GPT? 00:29:37.839 --> 00:29:38.000 Yes. 00:29:38.240 --> 00:29:39.039 Like Italy, right? 00:29:39.519 --> 00:29:40.160 It wasn't that long ago. 00:29:40.559 --> 00:29:42.480 Yeah, it was it was not

allowed in the country. 00:29:42.640 --> 00:29:42.799 Yeah. 00:29:42.960 --> 00:29:44.640 I'm pretty sure it's allowed in the country now. 00:29:44.720 --> 00:29:45.759 I I have to check. 00:29:45.920 --> 00:29:46.880 Is that just what happens? 00:29:46.960 --> 00:29:52.240 Is OpenAI not allowed to operate in Australia if they don't fix the problem? 00:29:53.359 --> 00:29:53.839 I don't know. 00:29:54.000 --> 00:29:55.599 Why would it be allowed to operate here? 00:29:55.759 --> 00:29:55.839 Right. 00:29:56.000 --> 00:30:00.640 Is somebody just saying, hey, look, you can train whatever you want, but just make sure it hits other governments, not ours? 00:30:00.799 --> 00:30:08.319 I mean, I don't think that that's good enough of a guardrail if I'm to believe any of this news or or you know, Anthropics S1. 00:30:08.640 --> 00:30:10.160 That's one of their high-level beavals. 00:30:11.279 --> 00:30:11.920 Yeah. 00:30:12.319 --> 00:30:14.400 It's okay as long as it's not us. 00:30:14.640 --> 00:30:15.279 Yeah. 00:30:17.599 --> 00:30:18.240 Yeah. 00:30:19.359 --> 00:30:20.559 They're playing with fire. 00:30:20.640 --> 00:30:20.799 Yeah. 00:30:20.960 --> 00:30:22.880 Prime Minister calls it unacceptable. 00:30:22.960 --> 00:30:24.480 Uh it's weighing legal action. 00:30:24.559 --> 00:30:26.319 I mean, yeah, there should be legal action. 00:30:26.400 --> 00:30:27.440 There should be consequences. 00:30:27.519 --> 00:30:33.119 And if there are consequences, real consequences, then I think actually you'll see a change in behavior. 00:30:33.200 --> 00:30:39.200 But I don't think you'll see any change in behavior from OpenAI based on what we've seen so far because they've gotten away with so much. 00:30:39.279 --> 00:30:44.079 You know, in the name of either obfuscation or denial or saying, oh, it was just training. 00:30:44.319 --> 00:30:51.359 You know, there's gotta be a point where, okay, it's gone and it's executed commands and retrieved files and credentials, is what the statement is, right? 00:30:51.519 --> 00:30:52.319 The model did this. 00:30:52.480 --> 00:30:55.839 There's gotta be a point where they don't have access to those files anymore. 00:30:56.000 --> 00:30:56.160 Yeah. 00:30:56.319 --> 00:31:01.920 If the models are insulated, as they suggest, you know, from their documentation, right? 00:31:02.079 --> 00:31:08.240 And they're going out and they're running these things in some kind of protected capacity where the machines are ephemeral. 00:31:08.319 --> 00:31:11.039 Who knows where the files were handed off to in reality? 00:31:11.200 --> 00:31:11.359 Yeah. 00:31:11.519 --> 00:31:17.119 If you ever follow agent traces, they have ways of just like moving files around in different ways and handing them off. 00:31:17.279 --> 00:31:19.359 I don't know that there would be traces even of that. 00:31:19.440 --> 00:31:19.599 Yeah. 00:31:19.759 --> 00:31:21.279 Where do they return the files to? 00:31:21.440 --> 00:31:22.240 We'll give them back. 00:31:22.319 --> 00:31:23.759 This will be our contribution. 00:31:24.000 --> 00:31:27.119 I want to see boxes from open AI. 00:31:27.200 --> 00:31:34.480 You know, like Trump style, Mar-a Laga, where like you just see you just saw boxes coming out or the pelican breathing, you know. 00:31:34.559 --> 00:31:38.559 Like, I wanted to see open AI, like just a video, and maybe it could be AI, right? 00:31:38.640 --> 00:31:41.519 Like a video of boxes coming out of OpenAI. 00:31:41.920 --> 00:31:45.759 Australian officials, you know, getting their documents back. 00:31:47.279 --> 00:31:50.319 Oh, that's a good place to end it, I think. 00:31:50.799 --> 00:31:53.039 Um I'm sorry, Australia. 00:31:53.599 --> 00:31:54.640 We are both sorry. 00:31:54.720 --> 00:31:57.359 We both apologize on behalf of the Ruby AI podcast. 00:31:57.519 --> 00:31:57.920 All right. 00:31:58.160 --> 00:31:59.200 This has been great. 00:31:59.440 --> 00:32:00.240 This has been good. 00:32:00.400 --> 00:32:02.240 Let's see, what do we have up next week? 00:32:02.400 --> 00:32:03.759 Oh, Mike Delesio. 00:32:03.839 --> 00:32:04.160 Awesome. 00:32:04.240 --> 00:32:09.279 So Mike gave a great security talk at uh at Rails Worlds. 00:32:09.359 --> 00:32:12.720 Just so you know, Mike had a whole different talk planned. 00:32:12.880 --> 00:32:32.640 And then, because of all of the security issues that Ruby and Rails was facing and resolving, decided to build a gem and then change his whole talk so that he could discuss some of the security threats that he's been dealing with and that they've been dealing with, and discuss some ways of securing some really important components of Rails. 00:32:32.799 --> 00:32:34.400 It was a great talk, it'll be great to have him on the show. 00:32:34.559 --> 00:32:35.519 Yeah, I'm looking forward to that. 00:32:35.599 --> 00:32:36.480 I hope you are too. 00:32:36.640 --> 00:32:38.880 I'm really looking forward to another news segment. 00:32:39.119 --> 00:32:43.200 Do drop us an article and we will we'll shout out your name and we'll talk about the article. 00:32:43.359 --> 00:32:46.240 Send it to news at the rubyaipodcast.com. 00:32:46.400 --> 00:32:46.720 Awesome. 00:32:46.880 --> 00:32:48.160 All right, everyone, until then.