Transcript EN 0:00 But you can't be loud. 0:01 Be quiet with the sword. 0:03 Be quiet with the sword. 0:07 Code is less important than it used to be, obviously. 0:11 But the question is, how good can your agent deal with a shit ton of code, 0:17 given its limitations with respect to context window? 0:21 So the agent is ultimately hurting itself by generating too much code. 0:26 Is the code as important as it used to be? 0:28 No, definitely not. 0:30 But I would say that it is still important in Astra is cool and all, 0:35 just like Fable, but there's still slot machines. 0:38 And I would say there is a lot of code where you don't need to know what it does and how it works. 0:43 But there's also a lot of very important infrastructure code where you still want to know what it does. 0:50 I think ultimately it comes down to control. 0:53 I personally like controlling interfaces and types. 0:56 And as long as I have control of that, the LLM is kind of limited in how much damage it can do. 1:02 Yes, the models can shit out immense amounts of code now, but by doing so, they're hurting themselves because ultimately they still have a limited context window. 1:09 And if it doesn't fit, you get more slop. 1:13 What I found interesting with the Astra release is that I, 1:18 for the first time, felt like less and less desire to actually look at the code that it produces because it's so hard to see. 1:25 And not, I mean, it's hard to look at. 1:27 It is incredibly compressed in situations in where it shouldn't be, 1:32 even when it actually uses code generation for writing code. 1:36 But it's particularly unreadable when it uses code to call tools. 1:40 So like actually following along what it's doing is also getting harder and harder just by how it writes code. 1:47 And I think like to Mario's point, like it actually does get, like during the training process, 1:52 it probably gets rewarded for token efficiency. 1:55 But as a result of this token efficiency, it is hard to read. 2:00 And there's a tension, I think, sort of forming where I think even if you get the 2:06 interfaces right and everything else, like if I look at an Astra produced code, 2:11 I feel an urge to do code formatting. 2:13 Which I didn't feel for a very, very long time. 2:16 I basically haven't used code formats at all because I felt like the LLM is going to follow along the style of the code that's already there. 2:22 And now it doesn't. 2:23 Yeah, and it is kind of odd. 2:24 It is really odd. 2:25 And I don't know where it's going. 2:27 But I'm a little bit feeling like this might become more of a thing now. 2:32 And I don't know how I feel about this. 2:35 You said something earlier, which I thought was interesting. 2:38 As the models are trained to not only be coding agents, 2:41 but also like personal open claw-like assistants, it's that's a reason that the training might 2:48 contain traces or things like just create whatever the 2:54 fucking code you want to create, as long as it does the thing the user asked for. 2:58 And the user is a non-technical user and the user doesn't write production systems, 3:01 but just tiny little things like build me a daily news brief website or something. 3:08 Or Minecraft clone or anything. 3:09 So if the models are now being trained to produce slop code just to make their users happy, 3:15 which is totally fine, then we'll have a detrimental effect on using these models for production software. 3:20 So what I found quite interesting was, 3:25 I noticed as you sort of look into the code that it produces, 3:31 like in tests, in shaders, in embedded JavaScript, 3:34 you can clearly notice that it's way worse than in the regular code that it writes. 3:38 So I wrote in a blog post, I think it's like one step removed from regular code. 3:41 It sort of starts falling into this pattern of really highly compressed things. 3:45 But then I was like, okay, maybe this just happens to me. 3:47 So I asked a bunch of people on Twitter, like, hey, if you build like an Astro game over the weekend and it works for you, 3:53 just show me the code. 3:54 And there were two people who replied like, oh yeah, it wasn't so bad. 3:58 And I just opened the code and it was really, really bad. 4:01 And then it was, I think both of them replied, yeah, it's probably not nice. 4:06 The thing is, like, we don't really expect it to actually write code this bad, I think. 4:11 Like, I got fully trusted that the models are at least certain level of, 4:16 like, passing bar. 4:18 And it feels a little bit like it violated this now. 4:21 I don't know. 4:22 It definitely feels like a shift. 4:25 Yeah, just yesterday I did some work on an old C, C++ code base in Spine. 4:30 I switched back to 5.6 Sol 4:35 because Astro just couldn't write any sensible kind of code. 4:40 It's very weird. 4:41 It's very, very weird. 4:43 I blame personality fans. 4:46 One thought that I had was that if you have the idea, 4:53 at least in theory, that you don't have to at all consider the code because, 4:57 like, the test coverage is going to guarantee you that the code works perfectly. 5:00 Then, even if you have, like, 100% branch coverage, 5:05 which I don't think is likely, do your reader cover all the possible inputs to that function? 5:10 And one of the things that I noticed, like, with this is that because it 5:16 compresses the code so much, I, 5:20 when I just look at the diff that it does, I don't have any confidence that, 5:25 like, I could even without asking the LLM tell you if the tests are sufficient. 5:30 I'm really struggling with this whole thing because, 5:34 like, if the code is too unreadable, then, and I'm not saying, 5:38 like, it's completely unreadable, but, like, it becomes so that you're one step further away from it. 5:43 And then you rely more on, like, okay, but there are tests and the tests pass and I can still navigate in my slop game. 5:49 And so, surely the whole thing must be working. 5:52 But it's, like, it becomes so much harder to understand if you feel repulsed looking at the code. 6:00 So, at that point, the tests also just become performative, right? 6:04 It just gives you a good feeling of there are tests. 6:08 I don't fucking know what they do and if they actually cover what we need to cover in the tests, but they exist. 6:13 And that's why I feel good. 6:16 It's very emotion-based, all of it. 6:19 But if they produce code that is unreadable, then you have even less incentive of checking whether the code they produced makes any kind of fucking sense. 6:26 So, it feels like we are getting pushed to being removed from the code. 6:31 Okay. 6:34 What if you use something like Astra, which is clearly, 6:37 like, trained to succeed, 6:41 but you use it only for the tool calls. 6:45 Like, you don't use it for the writing. 6:47 So, like, whenever it tries to write some code, you're like, oh, you stop. 6:50 Some agent, I don't know, something else you implement now. 6:57 Because I really feel like there's a tension in the training process. 7:00 They're like, either this is really good and we should like this because, 7:05 I don't know, software engineering is solved and it's just a bug fix that we have to deal with now. 7:10 And so, we don't look at the code anymore. 7:11 So, it doesn't matter how it looks. 7:13 And it's, like, really purely optimized for the LLM. 7:15 Or, these models are actually only behaving this way because they're being trained to do other things, 7:20 particularly really good tool calling using code. 7:23 And so, there's a tension in the training process and you should actually use a different model that's being trained to be an 7:29 expert software engineer. 7:31 And I really don't know what the answer is here because, 7:34 like, it's not like you go to OpenAI page and it's like, yeah, do that. 7:40 No, but maybe that's like a model capacity issue now. 7:44 Like, as they try to accommodate more and more use cases or vertical slices of 7:50 specific industries, obviously, the model needs to know about how to operate within these environments. 7:57 But that will ultimately mean that there are less parameters available to the things they used to be good 8:03 for, like software engineering. 8:06 So, maybe we are hitting that now. 8:08 Maybe it's the, you're trying to accommodate as many industries as possible. 8:13 So, your industry, specifically the software industry, is now suffering because there's not enough parameters left for your specific industry. 8:20 I don't know. 8:20 The token spent on engineering went 8:26 up in a certain way. 8:27 So, you could see, like, month by month what that was. 8:30 But none of the top line numbers moved up at all. 8:35 In a way, that was correlatable to this. 8:38 But obviously, the number of commits went up. 8:39 The number of, like, other things that you could measure went up. 8:42 And so, it went from, like, an insignificant part of, like, the expenditure of the company to a really meaningful part of the expenditure of the company. 8:49 And it might just be that we haven't figured out yet how to make that work. 8:55 Or, like, what is being produced actually isn't that much 9:01 more meaningful overall. 9:04 And so, because now everybody, like, as an example, like, the baseline expectation that we 9:10 have now for a piece of software is that it does more. 9:14 So, you need now to compete with everybody else who's in your space as well. 9:17 So, like, of course, your thing now should have a chatbot in it. 9:20 Of course, there should be, I don't know, like, the UI should behave in a certain way that was previously not necessary. 9:26 Like, more people now build mobile apps, I think, than they did before just because it became easier. 9:31 So, maybe this is sort of, like, a really weird form of, like, we are required to build more. 9:37 But, like, the amount of consumers is still the same. 9:40 So, it just has become more expensive to the software engineering. 9:44 I don't know if that's the answer. 9:45 I mean, I guess it's hard to measure at the moment just because 9:51 of the economic turmoil overall with oil being up and so on and so forth. 9:57 So, just taking GDP is probably a bad, bad idea. 10:01 But if you look at individual companies, I don't see them producing more meaningful features for the users if you're looking at software companies. 10:08 I just don't see it. 10:09 Like, can you name a single software or SaaS product that has meaningfully changed in the past 12 months? 10:17 I mean, I think in the AI space, you see that a little bit. 10:20 Yeah, but then the question is, who is using all of that stuff from the AI space? 10:24 Yeah. 10:24 I found, like, iOS 27 to be an interesting proof point here because, 10:29 like, clearly there should be more people in Apple using AI tools. 10:35 Like, I heard from people at Apple that there is, like, plenty amount of, like, AI supports of engineering going on. 10:41 But, like, from iOS 26 to 27, I felt like it was the most 10:46 boring iOS release. 10:50 Like, if anything, it's sort of the push against this in a way. 10:55 I mean, I don't think that's bad, actually. 10:57 I don't want a Cambrian explosion of new features and the home of Simpsons car, right? 11:03 But then I must wonder, does the additional resources that you have through agentic coding just ensure that 11:10 whatever you push out has higher quality? 11:12 I would like to think so. 11:14 I personally would like to use agents in a way where I can say, 11:17 okay, my velocity is kind of the same, but everything I release is actually of higher quality, 11:22 even if the quantity stays kind of the same. 11:25 That would be the goal for me when using agentic engineering tools. 11:29 But I'm not sure if you're seeing that either, if you look at companies with their 999s going into the 11:35 9100 territory. 11:40 I think where you see it a little bit in terms of, 11:43 like, quality, presumably, like, depending on how you look at this, is that software should become more secure because we are having 11:49 to work on all of the security reports. 11:52 Right? 11:52 And I think, like... 11:54 Well, that's not happening, I can tell you from first-hand experience. 11:57 It's definitely not more secure. 12:01 But it's also riskier to be on the internet now with software. 12:07 One of our colleagues, like, he's a little bit more now worried about, 12:11 like, smart contracts and crypto wallets just because they have become such attractive targets for 12:17 automated security research. 12:21 But, 12:25 I mean, there are definitely certain areas where, like, on average, 12:30 like, there should be more fixes going into, like, curl is a good example, 12:34 or, like, other software. 12:35 This is really, really old. 12:36 So, they got a lot more valid security reports, so they fixed them. 12:40 So, like, in general, curl is probably safer now than it ever was. 12:47 And it doesn't slop a ton of new features into it. 12:50 But also, the risk profile just went up simultaneously because of the same thing that... 12:55 And that is the problem, because the security patches still take humans to verify and apply. 13:01 But finding the security vulnerabilities is now, like, 100x easier. 13:06 I literally just had my clanker kind of try to break the DRM of one of my old 13:12 products, and it just did it. 13:14 It just, like, without intervention from my end. 13:17 So, yeah, interesting times.
Download captions .vtt ↓