Autonomous Agents, AI Regulation, and Ruby's Performance Frontier の全文書き起こし

00:03:36.000 Just last Wednesday on the October 15th, FTC sent both companies' civil investigative demands, which are just subpoenas with a friendlier name, each running more than 50 pages. 00:03:36.159 --> 00:03:38.560 What you might be wondering is what are these demands? 00:03:38.719 --> 00:03:49.520 They document security incidents, safeguarding practices, marketing materials, privacy policies, all this fun stuff that the FTC agency now plans to compel top executives to testify. 00:03:49.680 --> 00:03:51.039 Let's see if they actually do. 00:03:51.280 --> 00:04:14.879 The same day that all of this has happened, a group of state attorneys they filed a separate consumer protection lawsuit against OpenAI, and OpenAI's response was carefully bland, saying, as we improve our models and they increasingly help people in their day-to-day lives, safety remains a top priority, adding it will engage constructively with the FTC. 00:04:14.960 --> 00:04:17.199 And an anthropic, of course, said nothing. 00:04:17.439 --> 00:04:17.680 Yeah. 00:04:17.920 --> 00:04:28.160 FTC chair Andrew Ferguson's longstanding position just was paraphrased by reporters rather than quoting directly is that the AI system shouldn't ship if they aren't safe. 00:04:28.399 --> 00:04:31.680 The question for us is Is AI's regulation moment finally here? 00:04:31.839 --> 00:04:35.120 Does it have real teeth, or is all of this just kind of like paperwork? 00:04:35.519 --> 00:04:42.560 It's gonna have its moment where a bunch of Silicon Valley execs go to Capitol Hill and make some headlines for a couple of weeks. 00:04:42.639 --> 00:04:44.240 Like that's the moment that's here. 00:04:44.560 --> 00:04:49.600 I feel like I have to walk a fine line here because I am not an AI doomer by any means. 00:04:49.920 --> 00:04:55.519 At the same time, I want to say that I think Andrew Ferguson's claim here needs some scrutiny. 00:04:55.680 --> 00:04:58.240 AI systems shouldn't ship if they aren't safe. 00:04:58.560 --> 00:05:15.199 That sounds nice, but I actually think that we all, including regulators, need to update our definition of what is safe because these models, even at their best, are not safe in the traditional sense that we have all come to know and understand from computers. 00:05:15.439 --> 00:05:19.600 Things that are not deterministic are going to do crazy things. 00:05:19.759 --> 00:05:20.480 We've seen that. 00:05:20.720 --> 00:05:24.480 That does not mean by any means that open AI and anthropics shouldn't be held accountable. 00:05:24.560 --> 00:05:30.800 They absolutely should for every time that their models go off and cause an outage or make people pull their hair out. 00:05:30.959 --> 00:05:39.120 At the same time, what we were just talking about at the top of the show, I've never had to so closely guard a software program running on my machine. 00:05:39.199 --> 00:05:40.480 I've never been so worried about it. 00:05:40.560 --> 00:05:42.560 It's just not that safe, even in

the best of times. 00:05:42.879 --> 00:05:53.199 You need to be a little bit more understanding of what the power of these models are and sort of couch our regulations or whatever's coming in the understanding that safety is looks a little different today. 00:05:53.439 --> 00:05:57.439 You make a great point because I think about we're working very small scale. 00:05:57.600 --> 00:06:03.920 Yeah, we're worried about our local computers, maybe a small little droplet somewhere, maybe a few of them. 00:06:04.160 --> 00:06:07.279 The scale here, I think, is what is the worry. 00:06:07.600 --> 00:06:09.920 But you're a czar, you're not that's not small scale. 00:06:10.000 --> 00:06:13.600 Like there's a lot of things that can go very, very wrong if those models go crazy. 00:06:13.839 --> 00:06:15.759 Yes, but even that's not small scale. 00:06:15.920 --> 00:06:16.079 Okay. 00:06:16.240 --> 00:06:23.839 These are like swarms in the hundreds of thousands of agents collaborating together and are let loose on a public domain, right? 00:06:23.920 --> 00:06:24.480 Like on the internet. 00:06:24.720 --> 00:06:24.959 Yeah. 00:06:25.120 --> 00:06:30.240 Hopefully, that's what kind of comes out of this is like, where are those guardrails? 00:06:30.720 --> 00:06:33.040 Can somebody do that for one? 00:06:33.360 --> 00:06:42.399 And I worry though, because the whole premise of how these things were trained to begin with stems from them having access to the public realm and letting things go loose at it. 00:06:42.560 --> 00:06:45.120 So what we're seeing now though is kind of like the reverse of it. 00:06:45.199 --> 00:06:45.439 Yeah. 00:06:45.600 --> 00:06:51.519 Before it was very much information gathering and collective, and now it's being put out there. 00:06:51.680 --> 00:06:54.160 They're they're actually doing stuff, mutating. 00:06:54.399 --> 00:06:55.040 Yeah, they are. 00:06:55.199 --> 00:06:56.560 I do want accountability, right? 00:06:56.639 --> 00:07:00.160 If somebody can mutate something that's available, that's the concern. 00:07:00.399 --> 00:07:05.759 Let's move on to Ruby 4.1, just got its first preview. 00:07:06.000 --> 00:07:08.319 It is all about Ractors. 00:07:08.560 --> 00:07:10.639 So this is really interesting to me. 00:07:10.800 --> 00:07:15.680 This is October 12th, Ruby 4.1.0 preview one is out. 00:07:16.079 --> 00:07:24.319 The headline is that each Ractor now collects its own heap on its own thread instead of stopping the whole program for global garbage collection. 00:07:24.560 --> 00:07:29.439 Now, allocation-heavy Ractor programs now scale like forked processes. 00:07:29.600 --> 00:07:32.079 That's a quote from the Ruby maintainers. 00:07:32.319 --> 00:07:40.959 There are a couple of ins and outs about this, but I think the headline is that Ruby is leaning into Ractors. 00:07:41.120 --> 00:07:53.600 They definitely have found something where they're able to make this genuinely a difference maker with respect to forked processes,

with respect to multi-threaded or multi-processed programming. 00:07:53.920 --> 00:07:57.439 I think people are starting to see real gains in performance. 00:07:57.680 --> 00:08:00.800 I'm curious to know what you think about this, V. 00:08:01.120 --> 00:08:04.879 Does per ractor garbage collection finally make Ractors worth reaching for? 00:08:05.040 --> 00:08:06.639 Have you already been reaching for them? 00:08:06.800 --> 00:08:13.439 I'm starting to get a sense that people are, but I haven't seen a lot of production instances in the headlines in the news. 00:08:13.600 --> 00:08:16.000 Where do you think we stand right now with Ractors? 00:08:16.240 --> 00:08:18.399 I'm torn on this because fibers are great. 00:08:18.639 --> 00:08:30.560 There's a whole rage framework, which we've been living under a rock as a fiber-based parallelism framework that's kind of taking storm and it works really well from a performance perspective. 00:08:30.800 --> 00:08:33.440 But at the same time, it's not true parallelism. 00:08:33.600 --> 00:08:36.080 And I think that's what Ractors aims at solving. 00:08:36.320 --> 00:08:37.120 It's the aim, yeah. 00:08:37.279 --> 00:08:38.399 I don't think we're there yet. 00:08:38.639 --> 00:08:42.799 Does this really solve the whole global interpreter lock and things like that? 00:08:43.120 --> 00:08:43.759 No, it doesn't. 00:08:43.840 --> 00:08:47.279 And I think definitely correct us if we're wrong here, anybody who's listening. 00:08:47.360 --> 00:08:52.960 But my understanding is that you're not getting true parallelism as long as the GVL is still in place and the GVL is not going anywhere. 00:08:53.039 --> 00:08:54.480 This is an improvement for sure. 00:08:54.720 --> 00:08:59.840 Garbage collection is being and not going across different reactors is a performance improvement. 00:08:59.919 --> 00:09:01.519 It's not a solution to the GVL. 00:09:01.840 --> 00:09:05.519 I worry it's just a way of shifting around blame for the GVL. 00:09:07.600 --> 00:09:09.840 I had a different take when I looked at this. 00:09:10.000 --> 00:09:11.440 I think Ractors are exciting. 00:09:11.519 --> 00:09:17.679 And I guess what I noticed was when I was quickly scanning this, Ractors shipped with Ruby 3. 00:09:17.919 --> 00:09:20.960 So they shipped in 2020, which is kind of interesting, right? 00:09:21.039 --> 00:09:22.080 It's been six years. 00:09:22.320 --> 00:09:23.759 This is an AI podcast. 00:09:23.919 --> 00:09:33.519 I think that the reason that there've been a significant number of advancements in the last 12-ish months is really because they've set AI at this problem. 00:09:33.679 --> 00:09:40.080 I don't know that for sure, but I think we can tease out from Matt's talks that they're definitely leaning into that. 00:09:40.320 --> 00:09:42.480 So I think that that is a good thing. 00:09:42.639 --> 00:09:44.960 What you say about fibers is interesting, though. 00:09:45.120 --> 00:09:46.480 I have always liked fibers. 00:09:46.559 --> 00:09:48.080 The Rage framework exists. 00:09:48.240 --> 00:09:49.759 Maybe people are already comfortable. 00:09:49.840 --> 00:09:54.000 And then if that's the case, then what are we really trying to achieve with Ractors? 00:09:54.159 --> 00:09:57.919 We're not going to have a true parallelism in Ruby. 00:09:58.080 --> 00:09:58.960 I don't think that's in the offing. 00:09:59.120 --> 00:10:00.960 And I don't think anybody's really trying to do that. 00:10:01.200 --> 00:10:03.679 What's the ceiling then for Ractors? 00:10:03.759 --> 00:10:06.320 Is it that it gets as good as fibers or better? 00:10:06.399 --> 00:10:08.240 And if that did, I'm curious to know. 00:10:08.320 --> 00:10:13.679 And it this is an open question: what advantage would a Ractor have over fibers? 00:10:14.080 --> 00:10:16.639 What I've been reading, it's kind of reduced footprint. 00:10:16.799 --> 00:10:22.799 So encapsulation of the Ractor is kind of like a reduced version from an object standpoint and responsibility. 00:10:23.039 --> 00:10:23.840 And I could be wrong. 00:10:23.919 --> 00:10:24.399 I don't know. 00:10:24.720 --> 00:10:27.200 Part of me is just hesitant to try out a preview. 00:10:27.279 --> 00:10:31.759 So I'm excited to see it coming kind of out of the experimental phase. 00:10:32.000 --> 00:10:33.679 Will I start working with it? 00:10:34.000 --> 00:10:34.799 Probably not. 00:10:34.960 --> 00:10:35.200 Yeah. 00:10:35.360 --> 00:10:37.600 Until somebody else figures out all the sharp edges. 00:10:37.759 --> 00:10:39.200 What are we getting Ruby 4.1? 00:10:39.360 --> 00:10:40.240 Are we getting that on Christmas? 00:10:40.399 --> 00:10:41.360 Oh, we've got to be. 00:10:41.440 --> 00:10:41.600 Yeah. 00:10:41.759 --> 00:10:44.000 Anticipated final stable release on Christmas. 00:10:44.159 --> 00:10:44.480 Beautiful. 00:10:44.639 --> 00:10:45.759 Beautiful. 00:10:46.480 --> 00:10:49.919 Y Jit runs on Windows, and YJet we're going to do in our next segment. 00:10:50.080 --> 00:10:52.799 But Y Jet running on Windows, I think, is a big deal. 00:10:53.120 --> 00:10:58.000 We constantly give every Windows developer short shrift and no tools. 00:10:58.159 --> 00:11:02.000 And I didn't actually realize until I read this that we weren't even shipping with YJet. 00:11:02.240 --> 00:11:07.600 It was just kind of like, well, if you're running Ruby on Windows, it's going to be slow and it's going to probably not install. 00:11:07.679 --> 00:11:09.120 So we're just not going to help you. 00:11:09.279 --> 00:11:10.639 Which is not okay. 00:11:10.879 --> 00:11:15.120 It's not really nice to a whole ton of people in the world that run on Windows. 00:11:15.279 --> 00:11:17.679 I guess we're going to give them some crumbs on Christmas. 00:11:18.000 --> 00:11:18.879 A little warm-up. 00:11:19.360 --> 00:11:19.679 Yeah. 00:11:19.919 --> 00:11:21.120 Do you want to take this next one? 00:11:21.360 --> 00:11:24.080 Shopify just put Zajit into production. 00:11:24.320 -->

00:11:25.519 Yeah, I think I think it's Zajit. 00:11:25.759 --> 00:11:27.600 It's already going toe-to-toe with YJIT. 00:11:27.759 --> 00:11:30.639 If you're not familiar, Shopify has a special Ruby team. 00:11:30.720 --> 00:11:33.200 They have been working on various JIT implementations. 00:11:33.360 --> 00:11:35.360 This is the latest iteration of that. 00:11:35.519 --> 00:11:49.679 They have deployed Zujit, the newer SSA-based method JIT that shipped disabled by default in Ruby 4 to a couple of small regions of its storefront renderer, each compared against a same-size YJIT region. 00:11:49.840 --> 00:11:59.039 In their post, they kind of looked at the response times for what the Zajit is, in quotes, probably as fast as or faster than YJIT. 00:11:59.360 --> 00:12:06.159 Not a clear win or clean win, and the author admits running different regions for each makes the comparison a little sketchy. 00:12:06.559 --> 00:12:16.159 Memory told a clearer story, so they are seeing about 125 megabits of average JIT overhead for YJIT versus the 101 on Zijit. 00:12:16.399 --> 00:12:18.320 Measured separately from the Ruby heap itself. 00:12:18.559 --> 00:12:21.600 Her benchmark results, though, are genuinely mixed. 00:12:21.840 --> 00:12:24.240 Several are roughly tied, some are worse. 00:12:24.399 --> 00:12:29.200 One benchmark call path mostly vanished after inlining, removed it entirely. 00:12:29.360 --> 00:12:34.960 So it seems like there's some mixed reviews here on Zajit from Shopify's partial deployment and measuring of it. 00:12:35.120 --> 00:12:41.279 But Zujit ships compiled into Ruby already, but off by default, so you could just turn it on with a flag if you wanted to try it out. 00:12:41.519 --> 00:12:44.480 Is it too early to call Zajit a win then? 00:12:44.720 --> 00:12:50.000 And when the sample is small like this and the results are mixed, I don't really know what to take from this. 00:12:50.240 --> 00:12:51.120 This is the funny thing. 00:12:51.200 --> 00:12:57.440 Just pulling this article up, this is Max Bernstein, by the way, who wrote about this at Rails at Scale, and he works for Shopify. 00:12:57.679 --> 00:13:03.200 So just looking at this, the agents' coverage of the story is mixed. 00:13:03.440 --> 00:13:05.919 Their opinion is mixed for what it's worth. 00:13:06.159 --> 00:13:07.519 Shopify is very happy. 00:13:07.679 --> 00:13:09.200 And I'm on the optimistic side. 00:13:09.279 --> 00:13:11.840 I mean, I think what Shopify did with WideJIT was incredible. 00:13:11.919 --> 00:13:16.720 And I think they did that at a time where there was a lot of skepticism over a new just-in-time compiler. 00:13:16.879 --> 00:13:19.600 And it's already light years better than its predecessor. 00:13:19.759 --> 00:13:23.360 Now it's DJT, it's taking some time to answer this question directly. 00:13:23.519 --> 00:13:27.759 Maybe it's not time to switch over your applications, but it seems like it's close. 00:13:27.840 --> 00:13:30.559 And it definitely seems like that is the future. 00:13:30.799 --> 00:13:31.360 It's interesting. 00:13:31.440 --> 00:13:38.799 I was looking at what's the difference, just for listeners out there who might not be savvy about just in time compilers, because I'm not really. 00:13:39.120 --> 00:13:42.960 The reason that YJIT has been a huge success is because it takes this pragmatic approach. 00:13:43.039 --> 00:13:45.919 It compiles small pieces of Ruby code as they're needed. 00:13:46.000 --> 00:13:48.720 That's as very, very just in time, like just as they're needed. 00:13:49.120 --> 00:13:51.919 ZJIT kind of takes a step back, or is ZJIT, I'm sorry. 00:13:52.320 --> 00:13:53.919 It kind of takes a step back. 00:13:54.240 --> 00:14:01.840 And so instead of optimizing little pieces at a time, it takes a sort of a more traditional compiler architecture and it looks for larger chunks of code. 00:14:01.919 --> 00:14:04.960 And so it's looking at ways to optimize code paths. 00:14:05.039 --> 00:14:09.120 It looks at places that are likely not going to be executed, duplicative code paths. 00:14:09.360 --> 00:14:19.759 The idea is that over time it can actually get a lot better than YJet, but it doesn't sound like that to me, just looking at this, and I'm a layman when it comes to compilers. 00:14:19.919 --> 00:14:22.480 That doesn't sound like the kind of thing that's going to be an overnight success. 00:14:22.559 --> 00:14:23.919 It sounds like that's going to take some time. 00:14:24.159 --> 00:14:29.440 We have an entire language and decades worth of execution strategy to kind of take into consideration here. 00:14:29.759 --> 00:14:43.679 I think until they get around to finally introducing the typed annotations where you can give the compiler kind of like ahead of time details about what it's looking at, which you know is in quite a few other languages to help with this specifically. 00:14:43.840 --> 00:14:48.480 I feel like until we get that worked out, we won't see the drastic improvements we're looking for. 00:14:48.720 --> 00:14:53.360 When you say it's available in other languages, is it available in other interpreted languages, like Python, for example? 00:14:53.600 --> 00:14:56.320 I believe Python does do something like that. 00:14:56.480 --> 00:14:56.879 Yes. 00:14:57.120 --> 00:14:57.279 Okay. 00:14:57.440 --> 00:15:06.639 I don't know if it's full-fledged, like you can get with a true compiled language, but there are aspects of it that you can kind of inject at certain instructions. 00:15:06.879 --> 00:15:07.200 That's cool. 00:15:07.279 --> 00:15:08.320 I'd like to learn more about that. 00:15:08.399 --> 00:15:11.840 But I'm not going to learn about it right now because we're going to move into the next segment. 00:15:12.480 --> 00:15:19.360 Meta defaults your code to train our next model and calls the box, you have to find yourself the opt-out. 00:15:19.679 --> 00:15:34.080 In news that shouldn't surprise anybody ever anywhere, Meta's new

MuseWonder and MuseWonder full models, their subscriptions default new users into contributor mode, which means that everything you do is going back into their system for training. 00:15:34.240 --> 00:15:47.759 This, of course, was a huge barrier to adoption for open AI in the very beginning, because people and companies, quite understandably, didn't want all of their private information going into some black box that a company owned. 00:15:48.000 --> 00:15:54.960 We've sort of settled that question by saying, well, hey, if you pay us some money and tell us not to do it, then we won't train our models with your data. 00:15:55.200 --> 00:15:57.840 Which of course we just have to like trust that they're telling the truth. 00:15:57.919 --> 00:16:00.399 And we may find out years later that it's not the truth at all. 00:16:00.480 --> 00:16:02.080 And that's not going to surprise any of us either. 00:16:02.320 --> 00:16:03.039 I digress. 00:16:03.120 --> 00:16:05.840 Here we are with Muse Code subscriptions. 00:16:06.000 --> 00:16:16.159 The only way to guarantee that your content stays out of training is surprise, surprise, the pay as you go API in standard mode, where contributor mode isn't even offered. 00:16:16.399 --> 00:16:22.080 Is this a fair trade for free capacity or is this a trap most people will never notice they walk into? 00:16:22.159 --> 00:16:23.039 Or actually both. 00:16:23.200 --> 00:16:24.879 The question is yours, Valentino. 00:16:25.120 --> 00:16:25.679 I guess both. 00:16:25.840 --> 00:16:26.000 Yeah. 00:16:26.240 --> 00:16:29.039 Muse is not an open weight model, right? 00:16:29.200 --> 00:16:29.440 Right. 00:16:29.600 --> 00:16:33.039 I don't really understand why people would contribute to it, to be honest. 00:16:33.840 --> 00:16:36.320 The reason is because they don't know that they're doing it, I guess. 00:16:36.639 --> 00:16:39.840 If you use any provider, you don't know if they're doing it either. 00:16:40.080 --> 00:16:40.320 Yeah. 00:16:40.559 --> 00:16:44.879 Which I guess they're at least disclosing that they are doing it and you can opt out. 00:16:45.039 --> 00:16:47.440 I imagine Anthropic is doing the same. 00:16:47.679 --> 00:16:48.000 Yeah. 00:16:48.240 --> 00:16:51.039 All the other companies are doing the same in some capacity. 00:16:51.279 --> 00:16:53.600 If you want to use it, who cares if they trade on it? 00:16:53.840 --> 00:16:55.360 Pay the money, I guess, if you want to avoid it. 00:16:55.440 --> 00:16:56.960 You know, if it does improve it too. 00:16:57.200 --> 00:17:04.160 Well, some of it is going to be used to improve the model, and some of it is going to be used, I'm sure, to send you ads and send people to it. 00:17:04.319 --> 00:17:06.559 When I read this, I thought it this is app tracking. 00:17:06.720 --> 00:17:20.319 You know, Facebook, before it was even called Meta, hit this real valley in its revenue model once it wasn't allowed anymore to track users across all the different apps they pulled up in their phone after they had Facebook up, right? 00:17:20.480 --> 00:17:24.720 And so all of a sudden they couldn't tell where you went after you clicked the link. 00:17:24.880 --> 00:17:29.440 And so they couldn't guarantee performance to the people that were paying them for ads. 00:17:29.519 --> 00:17:30.480 And that's how it runs. 00:17:30.640 --> 00:17:31.920 That's how a lot of companies run. 00:17:32.079 --> 00:17:37.359 At the same time, here we are with Meta in its 2026, and AI is ruling the roost. 00:17:37.440 --> 00:17:45.839 And it's like, well, hey, here is a way that we can potentially make money off of this or provide a better service to people that are going to run ads through our tools. 00:17:46.079 --> 00:17:51.359 So it comes as no surprise to me that they're not going to make it very easy for you to not do that. 00:17:51.599 --> 00:17:58.559 Actually, the only thing that's maybe a little bit surprising is that if you are in standard mode where you're paying, contributor mode isn't even offered. 00:17:58.720 --> 00:18:02.319 Because I would probably, if I were Meta, say, hey, everybody can be contributors, right? 00:18:02.480 --> 00:18:04.559 Unless you specifically tell us not to. 00:18:04.720 --> 00:18:07.359 We're taking your information and we're going to do whatever we want with it. 00:18:07.519 --> 00:18:09.839 I don't even necessarily have a problem with this. 00:18:10.000 --> 00:18:16.559 I mean, these are businesses that are trying to make money and to think that they have some kind of altruistic aim is really naive. 00:18:16.799 --> 00:18:19.680 When we use these tools, we don't have necessarily altruistic aims either. 00:18:19.839 --> 00:18:25.279 We want to be able to go through our day more efficiently or shop easier or not answer people's emails. 00:18:25.359 --> 00:18:26.720 We're just going to have an agent do it. 00:18:26.960 --> 00:18:31.599 If you want to get out of the trap, the only way out is to pay a little money. 00:18:31.759 --> 00:18:33.920 Yeah, I mean, we do that with Claude and Codex. 00:18:34.079 --> 00:18:34.240 Yeah. 00:18:34.400 --> 00:18:36.799 Until these smaller models are better, we're going to keep doing that. 00:18:36.960 --> 00:18:37.119 Yeah. 00:18:37.279 --> 00:18:39.359 I will have to retract something I said. 00:18:39.519 --> 00:18:46.480 Apparently, Meta has started this super intelligence framework and are releasing open weight versions of some of these models. 00:18:46.720 --> 00:18:47.440 Oh, no kidding. 00:18:47.599 --> 00:18:48.960 An in-episode retraction. 00:18:49.119 --> 00:18:49.440 I know. 00:18:49.680 --> 00:18:50.480 It's incredible. 00:18:50.720 --> 00:18:53.359 And the producer got in your ear and said, hey, you know, you can't say that. 00:18:53.519 --> 00:18:56.319 We're going to have to have a producer on soon that just corrects us. 00:18:56.799 --> 00:18:58.000 Corrects us in real time. 00:18:58.240 --> 00:18:58.559 Yeah. 00:18:58.720 --> 00:18:59.839 But

right now it's so much fun. 00:18:59.920 --> 00:19:01.519 It feels like everything I say is right. 00:19:01.759 --> 00:19:02.319 Right. 00:19:04.160 --> 00:19:05.920 These are just going to get better and better. 00:19:06.079 --> 00:19:07.279 Make use of it if you can. 00:19:07.440 --> 00:19:07.599 Yeah. 00:19:07.759 --> 00:19:08.799 There's no out training. 00:19:09.039 --> 00:19:09.359 That's right. 00:19:09.519 --> 00:19:16.559 Next up we've got Claude writes so much code that Anthropic had to build its own version control system from scratch. 00:19:16.720 --> 00:19:21.920 Anthropic's engineers, they hit a real wall because they were just producing so much code. 00:19:22.319 --> 00:19:27.839 Fetching their mono repo with plain Git would mean over 600 petabytes of Git traffic a day. 00:19:28.079 --> 00:19:34.720 So a seven-person team spent five months replacing Git entirely with a homemade system they call Scion. 00:19:35.119 --> 00:19:36.319 Not to be confused with the cars. 00:19:36.400 --> 00:19:36.559 Yeah. 00:19:36.799 --> 00:19:39.839 Plus a new review tool they're calling Antique. 00:19:40.000 --> 00:19:42.480 So the pressure came directly from Cloud's own output. 00:19:42.559 --> 00:19:50.720 The merge volume went from about 2,500 code changes a day in May of this year to over 30,000 by September. 00:19:50.880 --> 00:19:51.119 Wow. 00:19:51.279 --> 00:19:51.839 That's incredible. 00:19:52.079 --> 00:19:56.559 So a chart that tracks almost exactly with each new Claude model release. 00:19:56.799 --> 00:20:09.839 The main branch alone on our code base runs close to 100 gigs with a compressed working copy at 25 gigs, and CI performs roughly 25 million checkouts a day before anyone's laptop or agent SAMOX is even counted. 00:20:10.079 --> 00:20:14.960 A Scion engineer shrugged off a complaint about missing offline support on Hacker News. 00:20:15.119 --> 00:20:22.079 Since Claude is by far the heaviest user of Scion and is necessarily online, this just wasn't important to do. 00:20:22.319 --> 00:20:26.000 Anthropic's own write-up is candid about what the job became. 00:20:26.160 --> 00:20:31.039 So they're saying very often we found that the role of engineers was to tell Claude no. 00:20:31.200 --> 00:20:38.640 Elsewhere, agents will sometimes violate core system constraints to satisfy a pattern that should never have been allowed in the first place. 00:20:38.799 --> 00:20:40.880 And then they're getting more skepticism online. 00:20:41.119 --> 00:20:49.119 The register is saying that this only works because a company flush with venture funding has quote cost is no object mentality. 00:21:00.640 --> 00:21:08.000 Is that a preview of where every engineer org is gonna end up, or a problem only a company with unlimited venture money ever needed to solve? 00:21:08.240 --> 00:21:15.039 Sorry, I wasn't paying attention at the end of that because I was just registering scionhub.com and barking it. 00:21:16.559 --> 00:21:17.920 I love this story.

00:21:18.160 --> 00:21:19.440 This is fascinating to me. 00:21:19.599 --> 00:21:21.359 All of it is fascinating to me. 00:21:21.519 --> 00:21:26.799 First of all, 2,500 code changes a day to 30,000 in the last six months. 00:21:27.039 --> 00:21:29.200 That's what's happening right now at Claude? 00:21:29.359 --> 00:21:30.160 That's incredible. 00:21:30.400 --> 00:21:35.680 I have no problem at all if you're hitting you're hitting performance issues with your tools, then maybe you got to build new tools. 00:21:35.839 --> 00:21:37.599 This is like a story as old as time. 00:21:37.759 --> 00:21:42.400 So I think it's really interesting that they built a new homemade system. 00:21:42.799 --> 00:21:48.559 My thought is well, it's fine to say, well, Scion is closed, we're not going to open source it because it's ours. 00:21:48.799 --> 00:21:53.359 And probably that's correct that most teams today do not need this. 00:21:53.599 --> 00:21:54.559 What about tomorrow? 00:21:54.799 --> 00:21:55.839 What about a year from now? 00:21:56.000 --> 00:21:57.279 Are we all headed for that? 00:21:57.599 --> 00:21:59.440 That is an enormous increase. 00:22:00.160 --> 00:22:02.000 In output and merge volume. 00:22:02.240 --> 00:22:07.759 I don't actually agree with this skepticism that it's a company flush with venture funding and it's a cost is no object. 00:22:07.839 --> 00:22:09.599 I think that this is probably the right move. 00:22:09.759 --> 00:22:14.319 I think the efficiency gains alone probably make this worth it. 00:22:14.559 --> 00:22:20.559 It's just to me, it sounds like it's maybe a little shot for us of what the future might look like. 00:22:20.880 --> 00:22:24.960 This problem has been solved by other companies like Google and Meta. 00:22:25.200 --> 00:22:27.279 They each have their own kind of prior art here. 00:22:27.519 --> 00:22:32.480 If you're dealing with 600 petabytes of just Git traffic, that's so much. 00:22:32.720 --> 00:22:36.240 The cost optimization for bringing that down is like huge. 00:22:36.400 --> 00:22:36.559 Yeah. 00:22:36.799 --> 00:22:37.519 The value for that. 00:22:37.680 --> 00:22:38.640 Why would you not do that? 00:22:38.720 --> 00:22:38.880 Yeah. 00:22:39.119 --> 00:22:44.079 For us, peons is serving well until you get to a certain point of adoption. 00:22:44.240 --> 00:22:49.359 So I wonder what is that tipping point for us of changes per day? 00:22:49.599 --> 00:22:52.720 What number of changes per day for you are you like, you know what? 00:22:52.799 --> 00:22:54.000 Git is not the right tool. 00:22:54.240 --> 00:22:54.640 A ton. 00:22:54.799 --> 00:22:55.680 It's gotta be a lot. 00:22:56.000 --> 00:22:56.640 Yeah. 00:22:57.440 --> 00:23:02.960 At the same time, they probably use some of the prior art as their uh go bys. 00:23:03.519 --> 00:23:12.799 As somebody who has lived through the end of CVS and the entire lifespan of SVN, it's funny to just come to a point where it's like, oh, maybe Git isn't fast enough anymore. 00:23:12.960 --> 00:23:18.240 Git was the most incredible thing I experienced as a developer once it came out and I compared it to SVN. 00:23:18.480 --> 00:23:19.119 Oh, so fast. 00:23:19.200 --> 00:23:20.799 And so much better at resolving conflicts. 00:23:21.119 --> 00:23:27.359 We only have 90 seconds for this one for some reason that I'm I'm not sure of, because of the rules that we put in place and don't even remember putting in place. 00:23:28.000 --> 00:23:33.359 So developers are ripping the smart out of smart TVs using AI coding assistance to do it. 00:23:33.519 --> 00:23:44.000 So there's two open source MIT licensed projects that are surfaced this week that will strip loatware and ads and tracking off of your Android TV, your Google TV, your Fire TV. 00:23:44.160 --> 00:23:46.960 They're called Clean Your TV and Dumb TV. 00:23:47.039 --> 00:23:48.799 And you can check these out right now. 00:23:49.119 --> 00:23:58.720 My question for you, Valentino, is how hard is it going to be now if you put these things in place to set up your parents' TV? 00:24:01.839 --> 00:24:04.319 Well, you know, this is just pie hole all over again. 00:24:04.640 --> 00:24:05.759 Yeah, I was thinking the same thing. 00:24:05.839 --> 00:24:06.720 It's like piehole. 00:24:06.799 --> 00:24:08.000 I remember I made a pie hole. 00:24:08.079 --> 00:24:08.720 I was so proud of myself. 00:24:09.039 --> 00:24:09.759 Pain in the ass. 00:24:10.000 --> 00:24:10.240 Yeah. 00:24:10.480 --> 00:24:10.799 You know? 00:24:10.960 --> 00:24:11.680 Yeah, it is. 00:24:11.920 --> 00:24:15.599 And uh I imagine this is gonna be very much a pain in the ass. 00:24:16.400 --> 00:24:18.880 But you know, we didn't have agents back in the day. 00:24:19.200 --> 00:24:22.160 And uh maybe pie hole is much easier to set up now. 00:24:22.319 --> 00:24:23.920 Uh I may remember that. 00:24:24.480 --> 00:24:24.799 That's true. 00:24:24.880 --> 00:24:25.440 That's a good point. 00:24:25.519 --> 00:24:30.400 I have a friend who runs an engineering team, and he said that in his spare time he developed a thing. 00:24:30.559 --> 00:24:40.640 It's a tool that he loves to listen to Yankee games on the radio, and it's a tool that it detects when the inning is over and then automatically mutes his audio for the commercials. 00:24:40.880 --> 00:24:41.440 It's incredible. 00:24:41.599 --> 00:24:44.319 He's like, I never looked at any of the True Smart TV. 00:24:44.559 --> 00:24:46.160 Yeah, yeah, yeah, yeah. 00:24:46.480 --> 00:24:51.680 Next week on the show, we're going to have Rick Van Michellen. 00:24:51.839 --> 00:24:54.559 He's going to talk a little bit about Museum of Modern Art. 00:24:54.799 --> 00:24:58.000 If you're interested in being on the show, we would love to have you. 00:24:58.079 --> 00:24:59.759 We have guests booked out through February. 00:24:59.839 --> 00:25:00.319 Do you know that? 00:25:00.480 --> 00:25:00.960 Wow, really? 00:25:01.119 --> 00:25:02.160 And potentially even further. 00:25:02.240 --> 00:25:02.400 Yeah. 00:25:02.480 --> 00:25:03.920 So they're, I mean, it's really exciting. 00:25:04.079 --> 00:25:05.759 Thank you, everybody, for saying yes. 00:25:06.000 --> 00:25:10.880 So I we have somebody in the background that's not AI, that's just begging people left and right to come on the show. 00:25:11.039 --> 00:25:12.160 So thank you for saying yes. 00:25:12.319 --> 00:25:15.119 The week after that, we have Marty Hott for Ruby Central. 00:25:15.200 --> 00:25:16.160 And uh let me tell you something. 00:25:16.559 --> 00:25:18.559 Marty's had a very interesting year. 00:25:18.799 --> 00:25:21.200 So we're gonna have a lot of fun talking to him. 00:25:21.440 --> 00:25:24.160 So much going on in Ruby Central and Ruby Gems. 00:25:24.319 --> 00:25:25.759 So that'll be a lot of fun as well. 00:25:26.000 --> 00:25:26.960 Yeah, I'm looking forward to it. 00:25:27.039 --> 00:25:27.920 All of our guests are great. 00:25:28.079 --> 00:25:29.680 Making the show fun and interesting. 00:25:29.920 --> 00:25:32.240 So until next week, great to have you all. 00:25:32.400 --> 00:25:32.960 Thanks for coming. 00:25:33.200 --> 00:25:33.759 Talk to you soon.