back to index

Talking AI and Agentic Coding with Yury Selivanov


Whisper Transcript | Transcript Only Page

00:00:00.120 | Hello, Juri.
00:00:00.840 | Hi, Armin.
00:00:02.340 | How's it going?
00:00:02.720 | Good.
00:00:04.080 | How about you?
00:00:05.160 | Pretty good.
00:00:07.240 | So we're going to discuss the Chantik coding and how useful AI is for us.
00:00:12.080 | And potentially we're taking opposing sides here, is my guess.
00:00:15.360 | I guess that's the idea, right?
00:00:17.760 | That's the idea, yeah.
00:00:18.920 | So maybe also to set maybe some sort of level here.
00:00:22.320 | Right now I'm basically alone with a co-founder, so I don't have a team more
00:00:26.120 | or less, and so my Chantik coding experience is a lot of just two
00:00:29.800 | people, right?
00:00:30.480 | And so you're probably in a different position too.
00:00:32.320 | I'm a slightly different position.
00:00:33.800 | Sad.
00:00:34.320 | Well, first of all, I'm joining this to play devil's advocate because I kind
00:00:40.560 | of believe in AI coding tools.
00:00:42.520 | I'm using them myself, sometimes productively, sometimes less.
00:00:46.320 | That said, the stuff that I want to discuss, it's real.
00:00:49.480 | I think those problems are real and it's good.
00:00:53.280 | I think ultimately it will help if people talk about those issues.
00:00:58.040 | So I'm not going to be aggressive, but I'm going to be aggressively making
00:01:01.880 | my point that those issues are important.
00:01:04.000 | Even maybe sometimes sort of focusing too much on things that ultimately I
00:01:08.800 | personally don't care much about.
00:01:10.280 | But that's, that's my role in this conversation.
00:01:13.160 | I don't want pro AI people to, to find where I live and like, to a different discussion.
00:01:20.440 | I love AI, but, but, but yeah, there are some, some, some issues with it that we need to discuss.
00:01:25.840 | And, uh, and, uh, I'm excited about this.
00:01:27.240 | So I'm Yuri, Yuri Salivanov.
00:01:29.080 | Uh, I'm a Python core developer for, uh, uh, since 2013.
00:01:33.120 | Get a lot of stuff to Python, like async await syntax, uh, and, uh, um, in famous asynchronous generators.
00:01:40.600 | Maintained async IO, created UV loop.
00:01:42.520 | And, uh, uh, uh, I, uh, I'm a co-founder and CEO of, uh, gel data where we try to fix some things in Postgres.
00:01:51.240 | Um, so that's me.
00:01:52.720 | Uh, I've been, uh, a software engineer for, uh, for a long, long time.
00:01:57.040 | Um, and, uh, I'm taking, uh, full advantage of AI.
00:02:02.160 | At least I wanna, wanna, wanna think of myself this way.
00:02:05.040 | Like I, I'm super excited about this technology.
00:02:07.440 | I think this is, this is something new, uh, and, uh, that's why it's exciting.
00:02:11.160 | Uh, and, uh, sometimes, uh, I, uh, I have some wins with it.
00:02:15.880 | Yeah.
00:02:16.520 | That said, Armin invited me to, uh, be devil's advocate.
00:02:20.760 | So, uh, that's my role.
00:02:21.960 | Yeah.
00:02:22.360 | And so I am basically unemployed since April this year.
00:02:26.240 | I did 10 years of century before.
00:02:28.040 | Four years.
00:02:28.640 | Did a lot of Python trust over the years.
00:02:31.480 | built Flask and a bunch of other things.
00:02:33.640 | And in the last couple of months, fell into a hole of going really,
00:02:37.480 | really deep on authentic things.
00:02:39.640 | And started a company just basically two months ago, uh, where I'm now taking
00:02:46.040 | full advantage of authentic coding, maybe to, to a degree that's slightly unhealthy.
00:02:51.240 | But I was very excited about the technology.
00:02:53.760 | Do you want to play devil's advocate for a change?
00:02:56.520 | So let's maybe start with, I mean, maybe one, one extra piece of context here.
00:03:00.800 | is like, because I did a lot of Python over the years, I've also sort of decided that
00:03:04.960 | this time around, I'm just going to, uh, sub check myself to the choices of the computer.
00:03:09.760 | And I had the AI sort of run evals a couple of months ago, and I was trying to figure out
00:03:13.840 | like, which language should I use if I want to end up with a chat decoding.
00:03:17.200 | And, and the AI had the best result with Go.
00:03:19.520 | So now I'm a Go programmer.
00:03:21.040 | I think your, your chances of finding a co-founder, like dropping the more things.
00:03:26.640 | I already have a co-founder, so I'm done, I'm past that, past that point.
00:03:30.800 | Oh yeah, AI is your co-founder, I think.
00:03:32.400 | Yeah, AI is my co-founder.
00:03:33.200 | Yeah.
00:03:33.680 | Yeah.
00:03:34.000 | Sunnet 4.5.
00:03:36.000 | Wasn't there like a, I think it was like a Y Combinator where it was like the, the goal
00:03:40.320 | is like Y Combinator 1 is like the single person, billion dollar company.
00:03:44.240 | Right?
00:03:44.800 | Yeah.
00:03:44.800 | So that doesn't count if AI is the co-founder.
00:03:47.440 | Well, you have a few, you have JPD5 and you have three co-founders.
00:03:51.760 | Okay.
00:03:52.000 | Let's start with building prototypes with AI.
00:03:54.640 | I think that that's probably the part where there's the most agreement on that there's some
00:03:58.400 | value in it.
00:03:59.120 | I'm guessing because seemingly Lovable exists and they have a bunch of money that they make
00:04:03.440 | and, and bolt and whatever.
00:04:05.200 | Maybe we start with, is it possible to take what seemingly we all agree on, which is like,
00:04:11.520 | you can actually prototype really quickly with this thing and apply that also to prototyping in the
00:04:16.080 | concept context of an existing code base or at the very least using it there.
00:04:21.520 | And then maybe like to which degree does the same apply from your perspective also for just
00:04:27.280 | building in an existing code base.
00:04:28.800 | Yeah.
00:04:29.200 | So this, so this is basically where probably you and I agree.
00:04:32.560 | I think that the coding tools, all of them, not necessarily Lovable, but Cloud Code as well.
00:04:37.040 | And the cursor, obviously like even like vanilla ChatGPT, chat editor, if you want to copy,
00:04:42.880 | paste code, they work amazing for prototyping.
00:04:45.840 | My problem with that is that when you attempt to prototype something or build something within
00:04:51.520 | the existing, within the context of existing code base, suddenly, suddenly it's, it's, it's much harder.
00:04:58.000 | Suddenly you might find yourself spending a lot of time unproductively trying to replicate your recent
00:05:05.440 | success with prototyping something that worked and basically just face, face a wall.
00:05:11.360 | And this, this is kind of the, the start of the, I guess, of the conversation is that I think that
00:05:17.360 | everybody agrees that AI is absolutely amazing for creating prototypes.
00:05:21.680 | Productizing them sometimes is much, much harder.
00:05:25.040 | And, uh, if, uh, and if you have a team of people, this is where, uh, things start slowly sort of, um,
00:05:32.880 | falling out of, of context, uh, for you.
00:05:35.920 | Um, yeah, I had, I had, I had a lot of success with, uh, with building prototypes, but again, uh,
00:05:42.800 | So why, why does the prototype work for you?
00:05:44.240 | And why does it not work for you in, in, in the context of a code base?
00:05:47.520 | Because I can actually very simply take the opposing view here.
00:05:51.600 | I swear to you, like more than 90% of my, my infrastructure code right now is like a really
00:05:56.480 | big infrastructure component.
00:05:57.680 | Uh, it, it basically is a, is a piece that gives agents mailboxes for emails and it's all.
00:06:02.560 | And it's not slop.
00:06:05.440 | It's good code, I think.
00:06:06.560 | But my point is mostly like, this is no longer a prototype.
00:06:11.600 | Now it's real.
00:06:12.320 | Right.
00:06:12.640 | And sort of like every iterative change that I'm doing to it is, it's like writing to a code base
00:06:17.440 | that has a pretty big size at this point.
00:06:19.440 | I actually, I, I, I, in, in some way, I'm actually impressed how well it still works,
00:06:23.440 | despite it being basically at this point, more 40,000 lines of code code, um, which is maybe not
00:06:28.640 | the largest code base in the world, but it's definitely past the context size.
00:06:31.840 | I kind of wonder like to which, where, where, where do you, like, why does it not work for you?
00:06:37.200 | And why does it seemingly still work for me?
00:06:38.800 | What's the, what's the challenge?
00:06:41.280 | I can basically, uh, give an example.
00:06:43.920 | So I'm building a new tool actually for AI, uh, and to, to, to have it better integrated in my
00:06:50.240 | workflow.
00:06:50.800 | And, uh, to do that, we're, uh, this is like a big spoiler or something that we want to launch,
00:06:55.040 | uh, relatively soon.
00:06:56.080 | But that thing plugs into your terminal, uh, like literally it acts as a transparent proxy.
00:07:02.240 | And I wanted to understand how to build this thing first.
00:07:05.680 | And there are multiple different approaches.
00:07:07.680 | One approach is, uh, to basically take T-Max for example, and, and just instrument, uh, around it.
00:07:12.800 | And, uh, the other approach would be to tap into Pty and whatnot, uh, and then like actually do the
00:07:18.960 | proxy when you sort of like intercept, start in, start out and do all of that.
00:07:22.880 | So my first task was to understand, uh, to what degree I can push T-Max.
00:07:26.720 | And I'm not an active T-Max user.
00:07:28.480 | I've used it a couple of times, but this is where, uh, I think cloud code, uh, shined.
00:07:33.200 | At that time, um, uh, there was SunNet 3.5, I think, and, uh, uh, Opus 4.1.
00:07:40.320 | So SunNet, uh, did not work for me at all.
00:07:42.240 | So I had to use the expensive one, but the expensive one was amazing.
00:07:44.720 | I had one day where I built six prototypes, uh, of this thing.
00:07:48.160 | The first one was using T-Max, the other was like a slightly different thing.
00:07:52.240 | And then I realized, let, let, let me try to tap into those things at like at a more fundamental
00:07:57.520 | sort of IO level, because T-Max, I just couldn't see how I can have decent user experience.
00:08:01.920 | Again, thanks to AI, the prototype worked quickly.
00:08:04.400 | Started using Python and obviously hit the performance wall immediately.
00:08:07.520 | So, uh, just, uh, just realized that, hey, like I'm gonna have milliseconds of
00:08:12.640 | latency on every keystroke.
00:08:14.160 | This will not work.
00:08:15.120 | And Claude and I were managed to improve the performance to be decent, but then, uh, like
00:08:19.840 | in pure shell, but then you run Vim inside that thing.
00:08:22.560 | And like, suddenly, like, you can just see, like, you press a key and it takes some time
00:08:26.800 | for it to appear on the screen.
00:08:27.840 | Okay.
00:08:28.080 | So that didn't work.
00:08:29.040 | So then, well, obviously rust, right.
00:08:31.120 | But I'm not an active, I'm not, I'm not a good rust engineer.
00:08:33.840 | I can read rust code.
00:08:34.880 | I can write some rust code.
00:08:36.080 | You know what's funny about this?
00:08:37.120 | Sorry, I got to interrupt you for a second.
00:08:38.880 | My first vibe coding, like actual proper experiment, I did this sort of night of like,
00:08:43.760 | not sleeping with Mario Zechner and Peter Steinberger is we basically built the thing called
00:08:50.480 | Bipe Tunnel, which was a PTY interceptor that streams the keystrokes to the internet.
00:08:54.880 | So it sounds very familiar.
00:08:56.240 | It's in the air.
00:08:56.720 | And we reviewed it three times.
00:08:58.000 | Once in Node, once in Rust, once in Swift.
00:09:00.880 | Oh my God.
00:09:01.520 | Oh my God.
00:09:02.000 | It's, it's, it's, it's in the air.
00:09:03.600 | So then, then I tried to make it in Rust.
00:09:06.240 | And, uh, there are like a lot of things that I wanted to tap into.
00:09:09.120 | Like for example, I wanted it to react to certain shortcuts.
00:09:12.160 | I wanted it to be able to sort of like multiplex different terminals on one screen, kind of like
00:09:18.400 | T-Max is doing, but do it like on a specific shortcut.
00:09:21.840 | And amazingly, I built all that, all those things in one day.
00:09:26.080 | But then I showed it to Rust engineer, the company who is actually extremely qualified.
00:09:30.560 | And I had two reactions from him.
00:09:32.160 | Uh, like, first of all, like, oh my God, how much did you pay for this?
00:09:35.200 | And I'm like 600 bucks in a day, which is insane, obviously.
00:09:38.880 | But then, uh, his reaction was like, first he was shocked, but then, hmm, this probably sounds
00:09:44.960 | about right because it would take us probably a couple of weeks to build all those prototypes.
00:09:49.680 | And we just saved all this time.
00:09:51.280 | But then we started looking at the code and obviously like it's, it's, it's just, just, just bad code.
00:09:56.000 | So, but for me, for me, it's not possible.
00:09:58.880 | Like I'm an expert Python engineer and I'm like decent, I guess at Rust, but this is where the gap
00:10:05.120 | was basically for me, the Rust looked fine for him.
00:10:07.840 | It was, it looked like a mess.
00:10:09.840 | So then it's easy to extrapolate.
00:10:11.360 | If, if I had, had to do this myself and I would just be, uh, uh, vibe coding this thing further
00:10:17.840 | and further and further building product, I would 100% hit the wall where I would not be able, uh,
00:10:24.720 | to progress much without me building myself deeper understanding of the, uh, of the programming
00:10:31.040 | language and the problem of, of, of, of everything essentially.
00:10:34.880 | And, uh, and we're going to chat about like that specific trap, uh, in, in a little bit,
00:10:39.520 | but that's my complaint about building prototypes is that at some point of time, uh, you have to evolve
00:10:45.440 | that thing.
00:10:46.080 | You have to productize it.
00:10:47.440 | And this is where currently there are limits.
00:10:50.480 | If you are an expert engineer and you know, well, the domain and you know, well, uh, the
00:10:56.000 | programming language, then you can harness the power, I guess, of, of, of the, of the tool.
00:11:00.960 | And this is what you're doing.
00:11:02.560 | This is your success.
00:11:04.000 | I'm an expert programmer now.
00:11:04.720 | Yeah.
00:11:04.960 | You are.
00:11:05.520 | Well, that's interesting actually.
00:11:06.800 | Uh, I mean, for me to chat about that.
00:11:09.280 | For me, the interesting thing is like, I think you're right.
00:11:12.080 | Like you can definitely, um, I call this once, like you can, we can Vibe code a code base to death,
00:11:17.680 | which we've definitely done.
00:11:19.040 | So my co-founder built a really great prototype for what it's worth.
00:11:22.480 | For like, for, for, for a particular thing that we're exploring, which is basically like a machine
00:11:27.680 | that connects to, to your email.
00:11:29.680 | And it does interesting things in the context of email, but, but it was built on top of a combination
00:11:34.800 | of Mailgun and Cloudflare workers and TypeScript and D1.
00:11:39.360 | And it was very, very, very little input of me.
00:11:41.760 | Um, and after three weeks, it created the code base.
00:11:44.960 | It had like multiple tables, duplicates and all kinds of stuff in it.
00:11:48.320 | But you can have many views on this, right?
00:11:50.880 | One of which is, well, that's irresponsible, but it's not really because like the whole point
00:11:55.840 | is like figure out, do we even want to build this in the first place?
00:11:58.160 | And, and so my version of this was then to say, okay, so, so what, what is it that we're
00:12:02.480 | actually doing here?
00:12:03.040 | Like what are the pieces that we want?
00:12:04.560 | And I also used, um, a combination of Codex and Cloud for it.
00:12:10.800 | But I think like in a responsible way where, where like there's certain things in, in software
00:12:15.040 | engineering where it's really, really critical that you get it right.
00:12:17.520 | Where's your data?
00:12:18.320 | How is your data laid out?
00:12:19.600 | Is the problem that you're solving even scaling in the right way?
00:12:22.720 | And not everything is a scaling problem, but like you can definitely build a bunch of
00:12:26.240 | shit in a way that doesn't make it very far.
00:12:28.320 | And then obviously the code, the code can't look like complete trash.
00:12:31.680 | What I found at least is that there's a huge difference in my experience between the kind
00:12:36.160 | of code that you get when you get an AI to program Go as an example, or even Java or PHP versus for
00:12:41.920 | instance, when I get an AI to write when they write JavaScript or when they write to some degree
00:12:47.120 | Python, but definitely Rust.
00:12:48.800 | And I think it has a lot to do with the complexity of the abstractions that are typically in place.
00:12:53.920 | Because one of the reasons I think the TypeScript code base is just in a terrible state is because
00:12:58.240 | it just created the same abstractions over and over in different files and never found them.
00:13:02.560 | And in Go it didn't even get the idea to write the abstraction in first place because Go is just
00:13:06.080 | such a simple language.
00:13:07.200 | So I think that to some degree, I think it, like you definitely have to use your brain.
00:13:11.760 | Like even in the Go code base, I think if I wouldn't have refacted the hell out of it,
00:13:16.080 | it would probably have made a terrible mess too.
00:13:18.720 | But I think for as long as you commit yourself to wanting to uphold a standard and a quality bar,
00:13:24.000 | I think it works.
00:13:25.600 | It doesn't come for free though, but I'm still quicker than if I would have to write
00:13:30.000 | every single line by hand.
00:13:31.040 | I think that's my...
00:13:32.000 | You think that.
00:13:32.720 | That's the main problem is that it's like, I think we should move like to my next talking
00:13:38.960 | point that I have like a little list of things.
00:13:41.440 | Like my next talking point is directly related to this.
00:13:44.880 | This is time and expectation management, which is that you have no idea will it work for you this day
00:13:52.720 | or not.
00:13:53.280 | Sometimes you want to do something complex, you prompt it and it just punch shots it for you.
00:13:58.560 | And it's amazing.
00:13:59.440 | It's an amazing feeling.
00:14:00.480 | Like it's great.
00:14:02.080 | It just worked.
00:14:03.520 | You maybe do a little bit of cleanup, write the commit message to sort of make it not look like
00:14:08.800 | it's AI rhythm.
00:14:09.600 | And that's it.
00:14:11.040 | Like go solve your next task or go walk in the park.
00:14:14.000 | But sometimes something easy can take hours of work.
00:14:19.040 | Like it doesn't work this time, but it's like almost there.
00:14:21.600 | So it's like, it feels like 15 more minutes and you're done.
00:14:25.200 | So you spend the time and suddenly like three hours pass and it's still like where it started
00:14:29.920 | essentially.
00:14:30.800 | And then you can end up in a situation where the whole day is wasted and you haven't done it yourself.
00:14:35.200 | Like maybe you would solve this task in two hours of coding.
00:14:37.600 | And that is the problem.
00:14:40.640 | I sometimes look at AI coding tools like this perfect dopamine cycle.
00:14:47.200 | It gives you a kick when it works and then it doesn't work.
00:14:50.400 | So it gets you slightly depressed, but then it works again.
00:14:52.720 | And I'm not sure that ultimately it saves time.
00:14:54.960 | Like again, like there are clear examples when it does.
00:14:57.840 | I think AI perfectly replaces stock overflow for a lot of times.
00:15:01.920 | Like seven times out of 10, like it will give you the correct result.
00:15:05.120 | It saves you time.
00:15:06.000 | And when it doesn't, it's kind of easy to see and check.
00:15:09.040 | So ultimately 100% time savings.
00:15:11.120 | I can buy that.
00:15:12.560 | With coding though, I'm not so sure anymore.
00:15:15.200 | Perceptually, I have a feeling that I can save a lot of time using AI.
00:15:18.960 | But when I'm objectively looking back at that time, I'm no longer sure that that was the case.
00:15:24.800 | So I think it saves me time and I think objectively so.
00:15:28.960 | But I think that my strong suspicion of how to make it work, and this is sort of like greatly
00:15:34.880 | extrapolated from a data point of me, is I think, so first of all, my opinion has dramatically changed.
00:15:40.880 | And I've told this people before, prior to me, I was like, all right, this is just a bunch of nonsense.
00:15:45.440 | But what I think like the reason why I feel like I can read the machine, like I have an understanding
00:15:50.320 | of what the hell it's doing is because the weird thing is that there is no learning curve.
00:15:57.120 | There's a hill to walk up on, a very, very, very long hill.
00:16:02.400 | And you walk up this hill for like two months and then you feel like there's enlightenment,
00:16:06.000 | because now you know what the machine sort of does and doesn't do.
00:16:09.040 | So it's not about like you, there's a learning curve and you have to learn the machine.
00:16:12.800 | It's like, you have to feel it.
00:16:14.960 | It sounds so ridiculous, but because like the whole point about is that you need to understand
00:16:18.720 | what is a task that it will not be able to do, because otherwise you're going to run into
00:16:22.000 | a situation that you spent three hours on a goddamn thing that will have taken you 15 minutes by hand.
00:16:27.120 | And because of this dopamine thing, you don't really feel like, because there's progress all
00:16:31.200 | the fucking time, it's just a pull request.
00:16:33.600 | And like, if you actually measure the time that you iterate on the damn pull request,
00:16:37.120 | it would actually be really bad.
00:16:38.720 | But the whole like, if you actually spent the time sort of like to onboard yourself
00:16:43.280 | into that properly, the number of cases where you're going to run into this kind of greatly
00:16:48.000 | diminishes because you recognize ahead of time what's going to work or not.
00:16:52.080 | And actually, I made the same experience at one point with Rust, where I felt the only way in which
00:16:56.960 | I could be productive at all in Rust was to figure out what I couldn't do in the language.
00:17:00.560 | And whenever I had a problem, that was like, you have to self-borrow.
00:17:03.120 | You have to, I don't know, do a bunch of stuff that languages can't do.
00:17:06.400 | You just have to recognize very early on that you're doing something that will not lead to
00:17:10.400 | success because otherwise you just, you're grinding out there with no progress.
00:17:14.480 | And so I think like for as long as you steer away from the shit that doesn't work,
00:17:18.160 | it does actually save you time.
00:17:19.520 | My problem is that again, like you walk up that hill and then your module is released.
00:17:24.240 | And then the hill suddenly is slightly different for you.
00:17:26.880 | And sometimes you have to go downhill now.
00:17:29.200 | It's uneven.
00:17:30.240 | Like that's the thing.
00:17:31.120 | Like with any tool that we've had before, be that the programming language or
00:17:35.840 | an IDE or something like that.
00:17:38.000 | All the knowledge that you have accumulates and you build on that knowledge.
00:17:42.240 | And with LLMs, you don't know.
00:17:44.080 | Sometimes, sometimes if you forget to like reset context and then the context overflows
00:17:48.960 | of Claude and then just like starts mumbling and doing stupid things, you have to actively
00:17:54.160 | pull yourself and realize, hmm, I should probably just start a new chat.
00:17:57.120 | Like this, this thing is done.
00:17:58.640 | Like no matter what I say anymore.
00:18:00.880 | And then sometimes weird stuff happens.
00:18:03.280 | Like it just gives up and like, you should do that.
00:18:05.840 | And I'm like, no, you should do that.
00:18:07.680 | Like do this yourself.
00:18:08.880 | You are the machine.
00:18:09.760 | And no, I will not.
00:18:10.960 | You do that.
00:18:11.680 | And it's, it's, it's kind of ridiculous and cool.
00:18:14.480 | And sometimes even refreshing to find yourself in this, but oftentimes it's also just frustrating.
00:18:19.920 | It's like, you know what I mean?
00:18:21.280 | And you can end up in this situation pretty easily.
00:18:22.400 | Again, my main point here is that it's not clearly a hill.
00:18:26.320 | It's something that you, some, some surface that you see like 30 meters in front of you.
00:18:32.160 | And you have no idea if it's uphill or downhill.
00:18:34.480 | What is even going on?
00:18:35.600 | Am I learning anything?
00:18:36.800 | Am I learning how to use it or not?
00:18:38.560 | Or maybe I just forgot to reset the context.
00:18:39.920 | The irony is that I think like, unless I would have actually had the time to, to like not
00:18:44.160 | have pressure, I mean, I had a pressure to start a company, right?
00:18:46.560 | That was an internal pressure, but it wasn't like, it has to happen by day X.
00:18:49.760 | And so I basically had a bunch of time to deal with a bunch of shit and to figure this out.
00:18:55.520 | And what I realized when I talked to a bunch of people in a company, I was like, well,
00:18:59.040 | my leadership makes me want to use AI tools and it doesn't work for me.
00:19:02.800 | I was like, yeah, if, if, if you would basically have to use like 20% time,
00:19:06.000 | whatever they call it at Google to figure out how this shit works.
00:19:08.400 | Like you wouldn't make a progress because it just, it, it's like, it requires trial
00:19:12.880 | and error and experimentation.
00:19:13.920 | Hey, I think you're right in one sense.
00:19:15.600 | That's like, there's, there's no guarantee that your problem tomorrow is going to be like
00:19:19.440 | within sort of your, your feeling of like what the machine can do or not.
00:19:22.640 | And the model might change and they might, I don't know, regress something.
00:19:26.480 | Like there was a time when Antropic had a bunch of server errors where like
00:19:29.840 | the quality of the model went down and you wouldn't know.
00:19:32.320 | It's just like, you got to post more than a month after.
00:19:34.800 | It's like, yeah, your shit didn't work quite as well as it used to.
00:19:37.840 | So I, and, and individually you, you, you feel like gaslit by this thing all the time
00:19:42.560 | because like, I feel like this kind of problem I did successfully before.
00:19:46.000 | And now it doesn't want to do it.
00:19:47.200 | So I think that is correct.
00:19:48.640 | And to some degree it doesn't fully go away.
00:19:51.040 | But, um, but man, I think the reason it's called vibe coding to some degree is because
00:19:56.320 | there are these, these feelings where you get a sense of like, does it work or not?
00:20:00.640 | And it's, it's, it's mind bending in a way because as a programmer, you're used to determinism
00:20:05.360 | or you want to, you chase determinism as much as you can.
00:20:08.080 | And now we have like, yeah, fuck determinism.
00:20:11.440 | It's just whatever.
00:20:12.480 | It just wipes.
00:20:13.280 | So that is really frustrating.
00:20:15.360 | Particularly if you try to use these things for like, also like building an agent or something.
00:20:21.200 | I'm with you on the, on the, like, there's this, this part about it.
00:20:24.960 | But I think the question is like, what's the percentage of it?
00:20:26.880 | For me, the feeling of the percentage of like the shittiness and the weirdness, it's like 20%.
00:20:31.520 | And it's not 90% because it's, it's only that small thing.
00:20:36.400 | I still feel like I get a lot of value out of it.
00:20:38.720 | But I heard a lot of people say like, well, nice.
00:20:40.880 | This is like this 90% of my problem is the shit doesn't work like I want.
00:20:44.240 | And so the 10% improvement is just not worth it.
00:20:46.640 | Like, yeah, if that's how you feel about it, I can see it.
00:20:49.360 | It's just doesn't really, I don't feel like that.
00:20:51.440 | My vibes are different.
00:20:52.400 | By the way, I have a feeling that by the end of this conversation, you might stop using AI
00:20:58.640 | because I'm feeling like I'm convincing you.
00:21:00.160 | So if you want to interrupt the call early.
00:21:02.880 | So I have, I have, I have, the list goes on.
00:21:07.680 | I do have a thing here, which is a time management and expectation management.
00:21:12.240 | And it's like, does it save your time?
00:21:13.280 | And I think like the main way in which it saves me time is concurrency and parallelism.
00:21:17.680 | I feel like I'm solving multiple problems simultaneously in one way or another.
00:21:22.080 | I think a pretty big way in which I do actually save a lot of time is that
00:21:25.360 | there's a bunch of problems that are going on, which are fully solvable
00:21:29.920 | while I'm solving another problem.
00:21:32.000 | And for, for someone who runs a very lean organization right now,
00:21:35.440 | it feels like I can do more, right?
00:21:37.040 | And it's like, even, even Mitchell Hashimoto mentioned this recently.
00:21:42.320 | It's like, you can spend time with your kids and you feel like productive still.
00:21:45.520 | And I leverage that a lot.
00:21:47.200 | Like I really leverage this a lot.
00:21:48.560 | This idea that even when I'm doing something, maybe it's unhealthy too.
00:21:51.920 | Like there's probably some version of this where like, you should really turn off your brain
00:21:55.760 | at a certain time of the day.
00:21:56.960 | But the fact that I can sort of multitask is for me really the thing that gives me the biggest
00:22:05.040 | feeling of like, this is good.
00:22:06.240 | I'm not an expert on this, but I do know.
00:22:08.240 | I've heard multiple times that people who multitask have perceived the perception of
00:22:12.880 | them being productive and in reality, they are not.
00:22:15.120 | So I think this is like one of the traps again, when you can be fully convinced that you're saving
00:22:20.800 | time, but in reality, you are not.
00:22:23.760 | And the problem here is that it's really, really hard to really benchmark it and get
00:22:27.920 | a definitive answer if it works for you or not.
00:22:30.400 | You can't solve the same problem twice.
00:22:33.040 | One week with AI and another week without AI, it's just not going to work, right?
00:22:36.800 | Finding equivalent problems and then sort of benchmarking against that is also like almost
00:22:42.240 | impossible.
00:22:42.800 | So like, we will never know.
00:22:44.480 | I guess, I guess what will happen is that those tools will keep improving.
00:22:47.680 | And at some point of time, the advantage of using them is going to be so clear.
00:22:50.960 | That's just going to be undeniable.
00:22:52.320 | That's going to be hard for people like me to even make this argument.
00:22:55.200 | But like, at least right now, I find myself as a senior engineer, software engineer,
00:23:00.400 | to be not fully convinced that it saves me time.
00:23:02.960 | I'm still doing it because I'm enjoying part of this process.
00:23:06.080 | And sometimes I want to be this dopamine junkie and play with the stack.
00:23:10.960 | And sometimes I have a feeling that I 100% know that AI will work here.
00:23:14.960 | So it would be silly of me to not use it.
00:23:17.680 | But definitively, I can't say yet that it's 100% improvement in like me writing code myself.
00:23:27.360 | Sometimes, yes, like this example with prototype, 100% AI helped.
00:23:31.440 | Day-to-day in a more complex code base, I'm not so sure.
00:23:36.320 | Like I had plenty examples where I wasted, it felt like I wasted days of
00:23:40.480 | time trying to make AI work.
00:23:42.800 | I got something in return, so it's not completely like a loss.
00:23:46.400 | But this is why I feel like it is because this is a problem that's also hard for an engineer,
00:23:50.960 | or is it a problem that's hard for the AI?
00:23:53.120 | Because I, for instance, like I don't work at Sentry anymore, but I did play a lot.
00:23:58.320 | The last couple of months was like, how good can I do authentic coding on the Sentry code base?
00:24:01.600 | But since it's sort of fair source out there on GitHub, so I can still fuck around with it.
00:24:05.040 | And I found it equally frustrating to work on that code base with the agent as myself.
00:24:10.960 | Because we have created a monster of like a multi-service complex thing.
00:24:15.840 | Who knew collecting errors is hard, right?
00:24:21.360 | If you build a large thing with lots of components at scale, a bunch of stuff is not fun.
00:24:28.320 | But I think it's also not fun.
00:24:29.360 | I'm not joking.
00:24:29.920 | I'm just pointing out that it is actually hard.
00:24:32.160 | Like production services, yeah, it's complicated.
00:24:34.640 | And I think like if I were to take a lens and I said like, well,
00:24:37.360 | the problems that the agent runs into are the problem every engineer runs into.
00:24:41.440 | And one of the ways in which to become productive with a large company is to
00:24:45.600 | figure out which things not to do because there's pain on the other side of it.
00:24:49.680 | But in some ways, the reason we have developer experience teams and
00:24:53.280 | companies is just to reduce the total amount of like, I don't know, like potential code
00:24:59.280 | changes that you have to do where you'll run into a wall and the wall is just terrible developer
00:25:05.680 | experience.
00:25:06.400 | So I wanted to which degree sort of what you run into is also in parts that it's actually
00:25:11.440 | not great for a human either to work on this.
00:25:13.600 | I'm not saying you have a shitty code base, just to be clear.
00:25:15.520 | I just wondered to which degree this is limited.
00:25:19.120 | Yeah, I think it's quite related.
00:25:21.360 | But I want to shift gears a little bit and talk about the implementation of those tools.
00:25:27.120 | And I think it's a little bit related to this problem.
00:25:30.400 | Okay.
00:25:30.640 | We have a super complex code base.
00:25:32.560 | It will not fit in the context.
00:25:34.560 | And even if the context is huge or advertised to be huge of like millions of tokens, in reality,
00:25:40.800 | we know that it's like the performance degrades after like 32,000 tokens or something like that.
00:25:45.520 | So you, and this is the problem of context management.
00:25:47.920 | And this is the weirdest part of me of the, of this whole AI revolution, because we have this
00:25:54.400 | amazing LLMs and it really feels like the future, but under the hood, it's powered by principles from
00:26:01.280 | 1980s, like expert systems and whatnot.
00:26:03.680 | Like you as an engineer pick up what you want to feed that beast.
00:26:07.920 | And that feeding happens on a lot of assumptions, a lot of hard code and luck.
00:26:13.040 | Sometimes it works.
00:26:14.160 | Sometimes it doesn't.
00:26:15.360 | You can trust an LLM to itself form the context.
00:26:19.040 | So you have to feed it with something, with your prompt.
00:26:22.560 | And then your ID adds some context around it.
00:26:25.200 | Maybe your open tabs in your editor.
00:26:27.760 | Maybe some settings in your, I don't know, PyProjectTOML or CargoTOML or something like that.
00:26:33.680 | And then that's it.
00:26:34.720 | Like you just, just, just wait for magic to happen.
00:26:37.360 | But it's an important bit that actually feels like we don't have any progress there at all.
00:26:42.400 | This, this, this notion of forming the context.
00:26:45.280 | And when you are just building a prototype, the context is empty.
00:26:48.640 | All you have is your prompt and maybe a couple of desires, like write it to me and go or whatever.
00:26:54.560 | Or when the project is small, it might all fit there.
00:26:57.200 | In the compass code base, context forming is extremely hard and it might not necessarily be.
00:27:01.760 | Those files that you have in your ID might not be necessarily relevant.
00:27:05.680 | And asking you to just drag and drop tabs meticulously for every prompt when you select
00:27:10.960 | which files it should focus on also feels like a drag, right?
00:27:14.960 | So that's my, that's my problem with it.
00:27:17.120 | I mean, the whole context management thing, it feels completely sort of out of sync with the
00:27:22.560 | whole forward thinking.
00:27:23.600 | So I guess my question to some degree is like, how do you work?
00:27:26.240 | Because, um, I, I'm the kind of person that really only makes progress in anything that I'm
00:27:32.480 | doing when I talk to another person in some bizarre way.
00:27:34.880 | And, and maybe I talk to like another version of me, but I have to talk for a problem.
00:27:39.600 | Like I have to, I have, sometimes there's weird ideas like this shit is going to work.
00:27:43.520 | And then as I'm sort of like working myself for it.
00:27:45.760 | Like I just actually, it's just like, there's a bunch of holes in it.
00:27:47.840 | And so for me, almost a natural part of solving any problem is to work my way through it.
00:27:53.760 | And that lends itself really well to agenda coding because that turns out to feed all the
00:27:57.840 | shitty context into my brain.
00:28:00.000 | And now I just need to get the machine to do it too.
00:28:02.240 | So like, I just really just have, I just talk to my computer a lot and it's a byproduct.
00:28:09.520 | The context problem is almost in quote solved because it's just, I don't know, it, it doesn't,
00:28:14.480 | to me in a way, it feels like I'm almost always anyways, working in a way where I just feel the
00:28:20.000 | context either for myself or for the machine.
00:28:22.160 | So it's not such a huge departure, I guess, to, to be a little bit more descriptive to me
00:28:29.120 | of what I want the machine to do.
00:28:30.400 | And so maybe that also is why I maybe encountered a little bit fear of problems because I don't,
00:28:34.160 | I think you call it a drag of like providing all the files in context.
00:28:37.520 | I don't know, for me that has always been natural.
00:28:39.360 | I used to maintain this, like this, this files where I write down all the steps that I want to
00:28:44.000 | do prior to doing it.
00:28:45.840 | Because I feel like that gives me a better understanding of what my change is or how
00:28:49.600 | I have to structure my change.
00:28:50.720 | Like a lot of it refactoring, it's always like, you got to do this first because otherwise the
00:28:55.360 | second step is much harder.
00:28:56.640 | And so that to me is like constant engineering that you even do without AI.
00:29:00.480 | And so that is always how I worked in a way.
00:29:02.880 | So I don't feel it's a drag, I think.
00:29:04.880 | And I don't know if I even want a machine that figures the shit out itself because then like,
00:29:08.480 | what is my role in it?
00:29:09.680 | I feel like my role is to provide the context.
00:29:12.560 | Right.
00:29:13.920 | So I don't know, is it not, is it not related to you in a way that I feel like that's the
00:29:18.000 | part of engineering that's actually like.
00:29:20.080 | No, no, no.
00:29:20.480 | It's how you piece it together.
00:29:21.520 | It's true.
00:29:22.880 | Like what you're saying, like for example, I have ultra wide monitor.
00:29:25.680 | Why?
00:29:25.920 | Because like when I'm writing software, I want to have multiple columns and I want
00:29:29.680 | the context for myself.
00:29:30.960 | I have sometimes the same file open like side by side at different locations because I need to
00:29:36.080 | stare at them to make progress.
00:29:37.840 | So just like you, I also have to have context before I'm writing code.
00:29:42.560 | I have to understand what files it will touch and what files are related in the project and
00:29:47.520 | all of that.
00:29:48.400 | Like, like naturally, I think any engineer does that.
00:29:51.760 | Maybe like one of the failure points for me is that I'm not talking to the computer.
00:29:56.240 | I just recently started experimenting with that.
00:29:58.400 | Just like bound, like double tap on a fan key to capture audio and transcribe it.
00:30:04.000 | And it worked fine.
00:30:04.720 | But like, it's not a habit for me yet.
00:30:06.160 | So I don't yet have definitive answer if it works or not.
00:30:09.360 | But still, sometimes you might have a file that has 5,000 lines of code in it, which for what
00:30:15.280 | it's worth is questionable.
00:30:16.400 | Like, why do you have files that long?
00:30:18.080 | It's, it's, it's, it's, it's, it's also weird, but I can see Claude losing the threat often
00:30:25.120 | in a complex code base.
00:30:26.320 | When you have too much of that stuff.
00:30:28.080 | Sometimes the problem is just like that you have to touch 20 files to solve it.
00:30:32.400 | And it's hard for you to make the.
00:30:34.160 | Context smaller, like to subdivide this problem into smaller problems.
00:30:38.320 | Sometimes I realize, hmm, it will take me like three times longer to write this prompt,
00:30:43.200 | to explain it, everything in nitty gritty details, like what it's supposed to be doing.
00:30:47.600 | But if I do all of that work, maybe just easier for me to just go and do it myself
00:30:52.080 | at this point when I'm explaining everything.
00:30:54.480 | So it's, it's like really, really hard, at least for me.
00:30:57.760 | Maybe, maybe our project is specific.
00:30:59.520 | Did you, did you see a difference between, I know that you also work on CPython, right?
00:31:02.800 | So did you see a difference from working on CPython versus like on GEL, for instance,
00:31:06.400 | in your effectiveness of AI?
00:31:08.320 | So I'm not actively working on CPython, but recently I rewrote UUID implementation
00:31:14.800 | from Python to C. It definitely helped. Like, I have a feeling that it helped,
00:31:19.200 | but also it was a clean slate problem. And also LLMs are amazing at translating
00:31:27.200 | a thing from one language to another language. So this is where it helped.
00:31:31.040 | It still made like a lot of non-idiomatic to CPython things. So ultimately not, I doubt that
00:31:37.520 | there is like a single line of code that survived ultimately. I rewrote the whole thing, but it did
00:31:41.920 | save time because a lot of boilerplate was, was, was, was reduced for me. I didn't have to copy,
00:31:47.040 | paste and change things as I usually do. LLM can do that for me and then I can edit. So that's my
00:31:52.160 | recent contribution. And again, like it's a bad example because it felt like a new thing. I should
00:31:56.880 | try to use it more often.
00:31:58.320 | I found, I found one kind of interesting failure case. And I talked with Lukas at Europython this,
00:32:04.880 | which is I actually found it to be much harder to work on code, which is in the training set,
00:32:11.680 | then on code that's not in the training set. Interesting.
00:32:14.080 | Um, and it's, it, I think it's sort of counterintuitive, but it, it might also make
00:32:20.400 | sense because one of the problems is that you basically work on something that's overrepresented
00:32:24.320 | in the training sets. Like the CPython code base, probably there's, there's a lot of it in the training
00:32:28.480 | set, but, but the one that you're working on is actually not the one that has seen. It's the,
00:32:33.200 | it's, it's like the nine months newer version of it. And I, I've also encountered this with
00:32:37.520 | the century code basis. Like it, it, there are millions of forks of century out there,
00:32:41.280 | some of which are really, really old. So it, it just hallucinates code together,
00:32:45.200 | which has not been in the code base for a long time because it thinks it's still there. So that I
00:32:51.280 | found really interesting because that's an entire class of problems that where, where, where, like, I don't
00:32:57.200 | know if it's going to be a representative going forward, but for instance, like you get this idea
00:33:00.800 | that like, if you have this huge context and like all of your shit is in the training data, then it
00:33:05.840 | should be so much better. But the reality for me at least is that unless it's sort of day-to-day,
00:33:10.320 | up-to-date, which unlike to be is, it's actually not helpful. So now I actually remember there was one
00:33:15.760 | problem with my UUID work at the Python core sprint, writing it in C. The problem is that Python C API
00:33:22.000 | has like an old style way of declaring a module, Python module in C, and there is a new way. New way is
00:33:30.640 | like two phases in civilization. There is like a notion of module context where you sort of put your global
00:33:37.680 | variables and whatnot. There is something required for you to do if you want to use free-threaded Python
00:33:42.880 | in the future or sub-interviews, stuff like that. So it's required. And obviously if you are working on
00:33:50.160 | Python code base, anything new that you do must, must follow that new thing. And it's quite different API,
00:33:56.400 | like internally, just the different arrangement of the code and type declaration. And there are some
00:34:00.400 | gotchas there. And this is where it felt miserably. So essentially it was insisting on doing things the
00:34:05.040 | older way. And yeah, I had to actively fight it. And ultimately, I think I give up and just focus
00:34:11.200 | on like little things, like write me the body of this function. That's the part where I feel like
00:34:14.960 | I kind of want to talk about open source a little bit and the impact that this whole thing has on it.
00:34:18.640 | Because that's the part where I might actually take the opposing view and say, it's going to be terrible.
00:34:23.120 | Yeah, yeah, yeah. I think we might align. And again, you might stop using AI for this conversation.
00:34:28.800 | I have a lot of thoughts on that. I'm not sure. But I actually, I worry about this quite a bit because
00:34:32.880 | I actually feel like, I actually could take the view here where it's like, unless we're really
00:34:39.280 | careful, we're going to make a huge mess of it. The whole thing of open source for me has never been
00:34:44.560 | that we need more open source. I always felt like what you need is like open source libraries are sort
00:34:51.120 | of common problems that a lot of people are sort of banding together and they're delivering the best
00:34:55.280 | quality code that can be so that we also overall can build better companies. So that to me was like
00:35:01.120 | the idea is like, if there's a really hard problem, you get the best people together, you cooperate
00:35:05.280 | together, you build this thing, and then everybody is hopefully going to leverage this. With the cost of
00:35:10.000 | code generation going down and seemingly everybody loving the idea that GitHub stores is all that
00:35:14.560 | matters. And we're going to have millions of modules on NPM and like having thousands of dependencies
00:35:19.280 | of an application is actually the way to build shit. Actually, I think like because it is such a
00:35:24.880 | predominant view, this, this can only end terribly with so much more actual real slop going around.
00:35:33.200 | So I don't know, do I have a counter to that? Because my view on that is actually, unless.
00:35:37.040 | I have a lot of thoughts on that. First, let's just, just to get that elephant out of the room.
00:35:42.080 | I think there could be a more fundamental problem with all of this is because it's proven now that if AI
00:35:49.600 | is trained on content generated by AI, the quality degrades significantly and the open source is just
00:35:56.560 | this perfect mixer of things. Because now you have pull requests partially written by AI. It's like
00:36:01.680 | really hard to separate which part of it is written by AI, which part of it is written for human, even for
00:36:06.320 | humans. That's part of my few complaints in a minute. But it's also hard for AI and for all those training
00:36:13.120 | pipelines or whatever they have to create those modules. So I'm not an AI researcher, but like I
00:36:19.920 | have enough in my context to know that this is a problem. So I'm a little bit worried about this.
00:36:24.560 | We're going to see a lot of new code written, potentially 10x, maybe 100x new code written. There's
00:36:32.160 | going to be a sea of new stuff and like who knows which part of it is AI, which part of it is human brilliance.
00:36:37.440 | So with that out of the way, problems that I see, like let's get back for example to this comfortable
00:36:43.440 | example of UUID. It's sort of required by CPython contribution policy that you sort of notify people
00:36:49.920 | if you use the AI or not. And here I am. I don't know. Like I really don't know how to answer this
00:36:54.720 | question. So it's 20, 100 lines of C code, which is significant. Most of this code is kind of mundane.
00:37:02.000 | This is not like, I don't know, dict object C where we have a complicated pointer, arithmetic
00:37:07.280 | magic and whatnot. It's mundane C, but it's still C. So it's sharp. You can die there easily and take
00:37:14.560 | the interpreter with you easily. So it has to be reviewed. And responsibly, I wrote that code
00:37:19.840 | responsibly. So I'm not, I don't have a single line where it's just AI generated and I haven't
00:37:24.560 | touched it, rewrote it or reviewed it really carefully. But then the instant I'm saying I use the AI for this,
00:37:31.200 | the whole thing is like dismissed because people, nothing that's 20, 200 lines of C. He's not an
00:37:36.080 | insane person. He probably generated half of it. So it dismisses my work now, but also I myself in a
00:37:42.480 | similar situation, even like I know Lukasz, for example, he submitted that and he said I use the AI.
00:37:47.280 | I would also be dismissive in this case. So how can I even trust people now with this kind of stuff?
00:37:52.960 | Like what is the social dynamic here? The social dynamic is like really, really hard. And I was, I think like,
00:37:59.680 | this is one of the reasons why I feel like I don't even want to take a side on anything when
00:38:05.120 | it comes to the social dynamic, because like, this is just going to have to play out somehow.
00:38:08.560 | But one of the areas where I definitely noticed, like, this is, is all kind of wonky. It's like,
00:38:12.320 | if you, I definitely have released source code out there, which is a hundred percent AI generated.
00:38:17.600 | And one of which was this Vite plugin. It's a very simple Vite plugin. It just forwards the console
00:38:22.080 | log to the Vite logger output so that you can see the console log in the browser in the terminal.
00:38:28.160 | A hundred percent AI generated. And I, and then I was like, okay, so I'm going to publish this now.
00:38:31.840 | And then I wrote a license under Apache 2. And I also said like, if that applies, because quite frankly,
00:38:37.760 | laws in the US, courts in the US have already said like, that's not human generated output. So there's also the
00:38:43.520 | question of like, if you actually have a significant amount of, of code being traded by, by AI, like,
00:38:49.840 | does it still cross the threshold of what we say, like, that is actually genuine human creation worthy
00:38:54.880 | of, of, of copyright. So like, even on that level, I'm now starting to like, look at a lot of source
00:38:59.840 | code out there. Like I've, I ran into a company where I, they have code on GitHub. I think it might be,
00:39:06.720 | I don't know if it's open source license or if it's just happens to be in GitHub, but,
00:39:09.600 | but I found like implementations, uppercase implementation_summary.md in the full of emojis
00:39:16.240 | somewhere in the code base. And I looked at the code was like, this is like, someone is vibing hard
00:39:20.960 | here. And, and, and if you, if you hang around and sort of the startup kind of ecosystem right now,
00:39:27.280 | there's so much of people throwing shit on GitHub, which is probably just a full AI output.
00:39:32.800 | Yeah.
00:39:33.440 | And how are we going to respond to that? Like it's one thing of seeing this against like a pull request on
00:39:38.240 | an like established open source project. And maybe there's at least something in place,
00:39:41.280 | but there's a, there's going to be a whole range of people, which is like creating this amalgamation of,
00:39:46.240 | of, of, of, of different kinds of things there, which are just regurgitated in some way, human output.
00:39:52.880 | Yeah. It, it, it erodes trust. That's the problem.
00:39:56.240 | For sure.
00:39:56.880 | Because I might know that, Hey, like, well, I'm not going to be using code written by you anymore,
00:40:01.920 | but like, for example, let's say Lukas, like I, I know Lukas well, he's, he's a brilliant engineer.
00:40:06.880 | So here's the library written by him. And I'm pretty sure that if this let's, let's assume
00:40:12.240 | it's something like to do with like high performance IO. And I know that Lukas would like look at
00:40:16.640 | everything, everything that touches like the core logic and the, it will be solid, but there is a lot
00:40:21.200 | of code on the outskirts of it. So if there is like disclaimer, I use the AI here, can I trust that
00:40:26.400 | that code is fine? Maybe he didn't even review it. I don't know. Maybe it just appears to be working.
00:40:31.840 | So this whole, this whole problem of like me trusting someone personally or some organization
00:40:36.560 | personally. And, and, and just, just because of it's that organization, I trust that the code
00:40:41.680 | is good. I feel that no longer applies and it's huge.
00:40:44.640 | Yeah.
00:40:44.640 | No, I, I shared a concern a lot actually, because I don't, that's the thing where like, I, I don't actually,
00:40:50.800 | there's one thing where I know like what the machine does to me. Right. But it's a data point of one,
00:40:55.920 | you know, in a rather narrow set of things. Right. And so I cannot even argue, like I, I trust the AI,
00:41:01.680 | I don't trust the AI. It's like the only view I can take is like whatever I create together with the AI,
00:41:06.720 | I'm responsible for. Yeah.
00:41:08.320 | But I don't think that's the view that most people take. And I don't, and there's like no social standard
00:41:12.800 | for it, nor anything. And it would be irresponsible for me to say like, well, because my experience is
00:41:17.360 | this, that's sort of like what everybody's experience is going to be. And, and we're also very early in
00:41:22.720 | this because in some ways we're, we're doing like this version of open source agent decoding for like
00:41:27.280 | six months, give or take, even less, I think in some ways. Because for sure, what I see is that
00:41:32.560 | after cloud code and after new codecs, there's a lot more, like we're way past this sort of little bit
00:41:38.480 | of autocomplete, where you were at least like very actively paying attention to every single thing that
00:41:43.040 | sort of autocompletes out. Now we're like, well, let it run for like 15 minutes and then let's do some
00:41:48.160 | code review here. So I, I'm with you on the, on the trust part. And I also think that one of the
00:41:52.640 | problems with this is now it has created, look, there are also positive things, right? I think
00:41:56.160 | like creating repo cases for bug reports, great time saver in the context of open source. Because
00:42:01.680 | you used to get this, this bug reports of like, well, shit doesn't work. And then the only thing
00:42:06.000 | I could do is like works on my machine, give me some more details. And now, now I get this to like
00:42:10.160 | copy paste this, try to make a repo case and usually gets one. So, so those are positive things,
00:42:15.440 | but I'm, I'm definitely suspicious of pull requests now. And it has made a lot of things much cheaper
00:42:20.160 | than actually were pretty good that they were not cheap before. Because some of, some of the issue
00:42:24.240 | reports are clearly not even human generated. They're also like AI generated. And security reports
00:42:29.040 | against libraries, they're like partially completely, Claude, invent some security issues so that can
00:42:34.160 | annoy some people to get a CVE for. Like all of it's just insane now. It's like objectively insane.
00:42:40.560 | It's insane that the well-written issue is a red flag. Or code with comments like, no, like, I'm not
00:42:48.240 | going to touch it. I used to like, if I, if I got a long, like a long issue against one of my project,
00:42:54.960 | I was like, oh, this is like going to spend some time reading for this. Like, this is great. And I
00:42:58.240 | get a long issue. It's like, oh my God, someone went hard on Claude here. But it's just like, this is,
00:43:04.160 | this is the trust eroding part. I really, really hate. Because like, that's, that's not so much about like,
00:43:09.280 | what the machine does with you. It's like what, what we as a group of engineers find. Like, I find
00:43:14.400 | it irresponsible when people sort of shit some AI stuff into an issue tracker and don't declare it.
00:43:19.920 | Or into my mailbox for what it's worth. I get so much email now that looks like slop. And I, I don't
00:43:26.400 | even want to engage with that person. Well, the idea is that you don't. You ask AI to do that and just
00:43:31.840 | never check your email again. No, but that's really trust eroding and I hate it. I really, really hate it.
00:43:37.920 | Yes. And I, it's like, it makes people lazy.
00:43:41.520 | We're going to talk about it in a second. That's, that's, that's the thing. Like,
00:43:46.640 | trust, trust erodes, not just like between me and somebody who I know or, or know about.
00:43:54.880 | Trust erodes within the team. That's the thing. Like, it's hard to understand. It's hard for people to
00:44:00.480 | sort of know what works and what doesn't work. And sometimes you can see a pull request. I know that,
00:44:06.320 | that the code in that pull request is written by human and I know that humans, so that is fine.
00:44:10.640 | I can review it. But then tests. And I see that tests are generated by AI and it's such a common
00:44:16.800 | sentiment online that, hey, just make it run tests. But you know that probably that having a big test
00:44:22.400 | suite might not necessarily be a good thing. Like if you have, if you have duplicate tests or if you have
00:44:28.080 | too many mock tests, it's, it's, it's, it's, it's, it's worse than not having tests at all.
00:44:32.400 | Sometimes tests are expensive, extremely expensive to maintain, to evolve, to do anything. And if you
00:44:37.600 | heavily rely on mocking, for example, you might not have a test suite at all. You might have an illusion
00:44:42.400 | of having a test suite. And AI is really good at creating this illusion of being able to write tests.
00:44:48.080 | And because tests are always treated like a second class citizen, but a lot of engineers say, like,
00:44:52.240 | I'm not going to be investing much time into reviewing that part. And let's just hope that it
00:44:56.000 | gets it right. I can see myself just like really being scared of, of, of the situation. Like we're
00:45:01.200 | just not going to write high quality software because high quality software demands having nice tests.
00:45:05.920 | So I think like, this is sort of a meta point for me in general. I've written about some Twitter
00:45:11.600 | autonomous. The, the, the quality of AI generated tests is like, it's so bad. That's, that's the part,
00:45:17.920 | that's real slow. But I also think that this is, is to a large degree because
00:45:22.000 | like we as engineers also suck at writing tests, like really, really bad because I, I've seen many
00:45:26.880 | more bad tests, mocks, for instance, I hate them. There's so many situations where it mocks out the
00:45:32.160 | part that's actually the one that's most likely to fail and then it fails. Right. Or, or like so many
00:45:38.160 | tests where people just write integration tests that, that rely on bizarre sort of side effects in the
00:45:43.840 | system somewhere. Like let's put some sleep here. Let's do some, like all of this. Right. And, and you're
00:45:49.120 | going to get a lot of these tests generated by AI now, but they were already pretty bad test patterns
00:45:54.640 | before. And yeah, we treat them as second-class citizens. That's, that's. I think like as engineers,
00:45:58.640 | we don't know at scale, we don't know how to write good tests and we're even rewarding people for writing
00:46:03.760 | bad tests. That's the problem with tests. Is that like the more you sort of progress in your software
00:46:10.720 | engineer career, the more you understand that tests are hard and yes, mock tests are horrible. That,
00:46:15.680 | that knowledge internalizes in time. Usually companies go down because of that lack of that
00:46:20.480 | knowledge, but tests are also extremely demanding, as you said, and to not have mocks means building
00:46:25.920 | infrastructure for tests. And sometimes you might have, you might end up building more infrastructure
00:46:31.200 | for your tests than infrastructure for other parts of like the actual production code sometimes. And then
00:46:36.480 | once you build that infrastructure, you need tests for that infrastructure. It's like, it's, it's a
00:46:40.240 | fractal of a problem to have good, reliable test suite. Running it, paralyzing it, make sure that like
00:46:46.000 | it doesn't run hours on GitHub CI and slows everything down. This is really hard. And yes, AIs are not trained
00:46:53.520 | on that. There are not too many good examples of good test suites. And even if they are, usually those
00:46:58.960 | test suites are so highly specialized to that specific project, they're an integral part of that
00:47:04.160 | code base. You can just separate it. So it needs deep insight to get any, like any, takes deep knowledge
00:47:10.880 | to get like any good insights that you can just apply to another project after that. It's really hard.
00:47:15.680 | And I have a feeling that it will take a couple of years of like LLM progress before, like they can sort of
00:47:20.400 | extract that information and reapply it to your domain.
00:47:23.600 | I think like at one point you sort of, the only way I think that is going to get better is like,
00:47:28.080 | if, if we maybe also solve the other problem, which is sort of like maybe the quality of AI goes down
00:47:34.560 | because they're trained on an LLM output. Right. So at one point I think like someone has to find a way
00:47:40.320 | to judge the quality of a code base and be more selective on, on the, on the learning part. And I think
00:47:45.520 | for tests, it will be necessary to some degree because there's just so many bad tests out there. And
00:47:49.360 | they're like their entire test frameworks, which encourage shitty tests. And there has been a
00:47:53.520 | generation of programmers that really believe that those shitty tests are exactly the gold standard
00:47:57.040 | of a test that you should write. And, and I think like that's, that's actually a problem for open
00:48:00.800 | source to large degree. Books written, books, whole books written about like writing tests in a horrible
00:48:05.200 | way by prominent publishers. I mean, this is, this is the weird thing is like, that's actually why I feel
00:48:09.680 | like my, my job is going to be secure for, for generations to come is because really like eventually you realize
00:48:16.000 | that you're getting really good, but being very counter cultured because the culture is, is sort
00:48:21.200 | of, is sort of going to the median of software engineering, which is where not good quality is
00:48:26.240 | being created. Yeah. But yeah. So I have a feeling that the problem here is like even more fundamental
00:48:33.600 | and it probably has to do with the current like technology of LLMs. Again, I'm not an AI researcher,
00:48:37.520 | but I'm hearing from here and there from prominent AI researchers that LLMs are like either at that end,
00:48:42.960 | or like we need next big revolution in LLM to happen. And to me personally, it all boils back to this active
00:48:49.280 | context management because LLMs don't have memory. All they have is context. Every, every new task is
00:48:54.480 | completely new from them. A person can learn. A person can internalize the, the, the, the information
00:48:59.760 | about this project, about knowledge, about the mission of this project and about some meta and the, and
00:49:05.040 | understanding of the domain that you are trying to solve. But LLM doesn't have that. And writing tests,
00:49:11.120 | good tests requires all of that. And Claude.md will not help you to capture all that knowledge.
00:49:16.400 | So until we see next generation of those AI of either LLMs or God knows, maybe it's going to be
00:49:23.200 | something else that sort of has this part of context management inside that loop, inside the model,
00:49:29.360 | somehow, I don't know how, it will continue to be this problem where there are just some areas which
00:49:34.240 | appear to be easy or appear to be non-important, like good tests for your production application. Not going to be
00:49:39.440 | solved without intense human input. And also like deep culture adjustment needs to happen.
00:49:44.720 | I think my only counter argument here is that I think that we have as an industry created so much complex,
00:49:51.680 | complex code and like over engineered shit and really bad tests that now the question is just, is
00:49:56.720 | it going to get worse in a way? Because the correct solution here is actually in a team to push back
00:50:01.200 | on slop either human generated or machine generated. And at least I found over the years, both in a
00:50:07.520 | company and on other projects in open source or elsewhere, it is actually very, very hard to take
00:50:13.120 | a principled stance on things that sort of most people think is actually a good thing. That's actually
00:50:18.160 | really hard. And I think in a lot of projects, you end up with the kind of code that you know after a while,
00:50:24.160 | you should never accept, but other contributors on the project will accept it or it's sort of
00:50:29.600 | industry standard, they need this stuff. I don't know to which degree that is necessarily an AI problem.
00:50:33.760 | I think is, I guess a little bit my call here. If we concentrate on tests as an example, are you going
00:50:40.480 | to get worse tests now with AI than you get? I think you get overall more tests, but the percentage of
00:50:47.680 | the shitty ones are probably like all things equal, you're just going to get more tests. So you're going
00:50:53.280 | to get more shitty ones, but the percentage might be the same. Because I actually think that the AI
00:50:58.160 | sort of perfectly creates sort of the standard crap that we have created. The reason it's so hard to
00:51:04.800 | work on large code bases, at least what has come out of the Silicon Valley in the last 15 years is like
00:51:12.800 | super complex systems, like overly, overly complicated. Everything needs to be outsourced to some
00:51:20.000 | third parties, like infrastructure startup. Like it's just insane. I don't know. I feel like that
00:51:26.960 | there's, that is at the root of that evil to some degree. And I actually found AI to be at least some
00:51:32.640 | sort of pinnacle of hope here, where rather than me having to go and use this infrastructure component
00:51:39.360 | that some random company might give me, because like, we need to do it this way because otherwise
00:51:43.040 | it doesn't work. Now it's like, okay, you know what? I get the 80% of solution to this. It's a 500 lines
00:51:47.920 | module code somewhere in my utility module. It's exactly tweaked for exactly the size of my problem. I'm going
00:51:53.040 | to use that. And as a result, I have less crazy stuff going on. So I feel like there's also like
00:51:57.680 | a flip side to that.
00:51:58.720 | It's going to be so funny if it's going to be bad unit mocked tests that break the AI back.
00:52:04.320 | And people just say, no, this is, this is the end. Let's not use AI after all. But it might as well
00:52:10.160 | just happen. Yeah. Because bad tests will slow down things, not only for people, but for AI as well.
00:52:15.040 | Testing incorrect things, incorrect behavior, codifying it, incorrect behavior in a test,
00:52:20.400 | which creates a bad sort of feedback loop in the development for everyone specifically,
00:52:25.040 | especially for AI. Yeah. All of those things are unsolved. I'm with you that part of this
00:52:29.120 | problem is coming from humans. Establish projects where you already have a good functional test suite
00:52:34.240 | with a good functional harness to run it will likely benefit. But even those things will degrade
00:52:42.000 | if people just blindly allow LLM or outsource the task to the LLM. So I think some education must
00:52:48.480 | happen and it might happen just naturally because people will start observing these problems more
00:52:55.600 | and more and more and more and finally understand the value of a properly written minimal functional
00:53:02.720 | test suite. But again, it's definitely going to be a learning curve, not just for EIs and LLMs and
00:53:08.880 | and everybody, but for like software engineers themselves, 100%. And ultimately, this is good.
00:53:13.600 | Like we can actually say that this might actually have net positive effect out of all of this is
00:53:19.440 | that... Look, there's optimism. Yeah. There is some optimism that people will actually understand the
00:53:24.560 | value of this because there is a lot of also like misalignment on tests in general. Should we have them
00:53:29.760 | or not and whatnot. Well, if you want to use AI, then you have to. There is just no way. You should come to
00:53:35.840 | the question of like, how hyped is it? Before, before that, like, let's quickly chat about the other
00:53:40.240 | thing, which is laziness. Laziness and the brain numbness or the way how I wanted to send it to
00:53:46.400 | your first brain smoothness and by another complaint with AI. So as I said before, I can, I can write Rust,
00:53:52.560 | I can understand it, but I'm not good at Rust. Not to the degree that I'm good like at C or Python or some
00:53:58.560 | other programming languages like TypeScript or whatever. Rust to me is still new. And what I
00:54:02.480 | observed is that there are a couple of code bases where I contribute and where it's like, it's socially
00:54:07.840 | acceptable for us to contribute some of the code that say I generated, people can review it and I'm open
00:54:12.240 | about it. That's fine. But I'm not learning it. That's the thing. Like, I'm not exercising my brain.
00:54:17.440 | I'm, it generates Rust code. I'm making some fixes here and there, but it's not compared to me writing that code
00:54:24.560 | at all. So it feels like I, like my progress is stalled. So I have to make a conscious decision of
00:54:30.560 | not using AI in order to get better. But there is a significant pushback towards it because the kind of
00:54:36.400 | Rust code that I need to write is actually extremely simple and like an RPC method here and an RPC method
00:54:42.480 | there. AI excels at that kind of stuff. I'm just harnessing the powerful already written code by humans
00:54:47.840 | and exposing it. That's easy, but I'm not doing it. So it's like I'm on a Smith machine doing bench and
00:54:54.160 | my personal trainer just lifts it for me. So I'm doing the movement, but I'm not getting stronger
00:55:00.160 | because of that at all. And it's such a deep trap. So you say that you're better at Go now. And I'm
00:55:06.080 | curious, like, are you? Because maybe you have perception of you being a good Go engineer now,
00:55:10.720 | but like in reality, it's not. I don't know. So I'm definitely a better Go engineer than it was
00:55:14.560 | like six months ago for sure. And it is very objective because I basically couldn't write any Go code.
00:55:21.040 | like I could, but not like. There's a lot of stuff in the standard library I would have to go to the
00:55:28.240 | Go docs for all the time. And I was like, okay, I know how this shit works. Like I definitely learned
00:55:32.160 | a lot. I think it is very easy to fall in the trap where you don't learn. And I made this learning the
00:55:38.560 | hard way for sure. And not even just for, in the context of programming. Like I feel like the thing
00:55:46.480 | that I learned the most over the last two years, just working with AI to begin with is understanding
00:55:51.040 | that if you turn off your brain, really bad things can happen. And so it is in the same sense of like,
00:55:56.960 | you have to, if you go with your gym example, like it doesn't help you that you know that you have to lift.
00:56:01.840 | You have to like, you have to make it a habit of going there, doing it regularly. Increasing
00:56:06.880 | weights. That's, that's not something that comes naturally. So neither does like working responsibly
00:56:11.280 | with API. You have to, you have to go into this, the reinforcement part. But then once you, once you,
00:56:18.880 | you understand the dangers and like how to work with it, you can make it more of a habit.
00:56:24.560 | So I think that's what makes that work. I really liked, I said this before, I really liked the
00:56:28.640 | ENTROPIC, like marketing campaign over the hand. It's like, there was never a better time to have
00:56:34.080 | problems. Because I, that's, that's how I feel like it's like, I have a problem now. I work and I use it
00:56:39.840 | like a better search engine, like talk myself through it. Like I learned a bunch of things where I feel
00:56:45.360 | like I wanted to learn that before, but now it sort of can dumb it down and sort of can transpose the
00:56:51.120 | problem into a space where I actually feel engaged. So if you want to learn, it's great to learn with
00:56:55.680 | it, but you have to make a conscious effort of wanting to do that. And if you, I don't know,
00:56:59.680 | if it's just vibe, if you just don't even try, it's like, you feel like you're making progress,
00:57:04.400 | but in reality, you just, you will feel bad about it or you should feel bad about it. That's, I think.
00:57:10.560 | That's, that's, that's the thing. Like logically I hear you, like you're making a good argument,
00:57:14.640 | but then I'm looking back at myself and I have, I have two modes of operation. Well, it's more,
00:57:19.920 | but like the spectrum is either at the end of the spectrum or I'm a beast. Like I'm writing code,
00:57:25.200 | I'm debugging it. I'm in the loop. I'm wired the keyboard and I know the problem and laser focused
00:57:30.560 | on it. I will solve this problem. We'll fix this bug eventually. In another mode, I'm just like,
00:57:34.880 | like a jelly on, on, on my chair, just lazily typing prompt. We are, we are going in circles with Claude,
00:57:42.320 | both understand that this is unproductive. Claude hates me. I hate Claude. And we aren't like in this thing.
00:57:47.840 | And I know that the bug, the wall that it can cross, I can do it easily. I don't want to
00:57:53.840 | because I'm, I'm, I'm in this weird sort of like regime now that I'm just like not actively engaged
00:57:59.360 | with it. I'm just wasting time. And that's, that, that, that's my problem is that when you use AI,
00:58:04.560 | you are not as focused on alert, the TikTok of programming. Yes, but it is, but it is like,
00:58:11.360 | it's just, this is the laziness. This is, and then when you are in that mode, it's really hard to learn
00:58:16.720 | either because you learn when you are alert, when you are focused and when you are not focused and when
00:58:21.280 | this thing is doing for you and you're tired, like suddenly it works. You just like, okay, like PR is ready.
00:58:28.160 | You don't want to learn anymore. To my point of earlier, actually from a societal point of view,
00:58:33.280 | I have like this deep rooted concern that this will turn into yet another social network problem where
00:58:38.320 | people are like, this is going to be great for humanity. We're all going to talk more to each
00:58:42.640 | other. And now all we have learned is like, we're smart zombies now with like a lot of psychological
00:58:48.000 | problems and everything. Right. So like, that is my concern on this, but like individually, I feel
00:58:52.320 | like I have solved this for now for me, but like, I hear you there. It's like, there's definitely a
00:58:57.440 | version of this where, man, it just really feels like there's a, if you don't catch yourself in a
00:59:02.800 | moment where you just give into the machine and you turn off your brain, like it stops being great.
00:59:09.280 | It just has all the, the something, yeah, I mean, for sure. I think part of the problem here is that
00:59:16.160 | this is what I'm trying to solve now. Hopefully I will finish that project with my team soon. Uh, but
00:59:22.080 | part of this is that the, the, the AI is a little bit disjointed from your day to day, like actual
00:59:27.600 | workflow as a software engineer. It's, it's like this chat box where you type things, but it doesn't affect
00:59:32.960 | how you use tools yourself. So I want to fix that problem. Hence, hence the whole the terminal
00:59:38.480 | magic. Uh, I wanted to be part of that, but I think something like that is required. Like when AI
00:59:43.200 | actually augments tools instead of replacing them. So I'm not saying that I will solve this problem by
00:59:48.480 | any means, but I think that's where the direction must take turn eventually is that AI would augment
00:59:55.760 | your workflow and not replace it. I feel that one of the biggest problems right now with AI is that it
01:00:00.560 | attempts and the marketing and, and, and everything and the social pressure attempts to replace you
01:00:05.840 | instead of making you 10, 10 times more productive in what you already know how to do, just make you more
01:00:12.160 | powerful. So that's the thing that that's, I guess my biggest sort of complaint about the current
01:00:16.880 | landscape. And we might actually have enough of AI technology to make what I'm talking about happen. It's
01:00:22.720 | just as an industry, we're not focused on that yet. We're just like drinking this Kool-Aid of like
01:00:27.280 | generating new stuff in the chat window. And I think one of the problems that I have is like, because we're so
01:00:31.680 | focused on this idea now that this is like the revolution and everything, any attempt of bringing
01:00:37.120 | some sort of nuance to anything is immediately either ridiculed or it is, um, like there's,
01:00:43.280 | there's real pushback against it. One thing for instance, like it's, it's maybe not so much for
01:00:48.080 | programmers, but like, I noticed a lot on like, I'm still on this social network called X. Um, there's so
01:00:53.760 | much pushback against Entropic for instance, because they keep talking about like, it's going to replace
01:00:58.160 | jobs and stuff like this. And they have this like, they sort of come out of this, like what is called
01:01:02.320 | EA movement. But I think like it is actually, I think it's good that people are talking about this a little
01:01:07.440 | bit in, in, in the AI companies. It's like, yeah, this might have some impact on society. And maybe you
01:01:13.120 | should think about this. Because I, I think that unless we actually start thinking about some of the
01:01:18.560 | consequences of this, maybe like being responsible of how we use it, we're not going to be responsible in
01:01:23.760 | no way, shape, or form look at us. Like we so smoothly transitioned to the comfortable topic of overhype
01:01:29.760 | in AI. Uh, it's, it's, it's one overhype is what I'm concerned with. And like what you, what you said
01:01:38.320 | essentially, like the whole, the whole thing is part of it is that I have deep concern that AI is just
01:01:43.360 | being overhyped. Like the whole thing of like replacing humans and replacing, maybe not software engineers,
01:01:48.560 | but I don't know, like support people and whatnot. So far there, there are not too many examples where
01:01:53.040 | that are actually successful, but it's being touted as the next big thing by everyone, by CEOs of those
01:01:58.960 | shops and labs, uh, for sure. Like it's always three months until we have a full, full, full AGI. And I'm
01:02:05.520 | just getting, it's not even funny anymore. I'm just, I'm just exhausted. I'm exhausted about reading that.
01:02:11.360 | I'm exhausted about reading reactions to that or people talking about that at this point of time.
01:02:16.400 | And I'm, I'm, I think I personally feel an enormous amount of value in AI, in chat GPT, in Claude,
01:02:24.400 | in, in, in everything. I'm worried that this overhypeness will just actually create a bubble,
01:02:30.720 | inflate it artificially. And to a degree it probably is already. And then this beautiful thing might
01:02:36.000 | collapse and then I don't have an opportunity to, haven't had an opportunity to actually like even
01:02:40.320 | enjoy it, like, because, because, because it got overhyped. So I'm a little bit worried about that.
01:02:44.480 | I think people have unreasonable expectations about the effectiveness of AI and they, uh, uh, and they
01:02:50.960 | make plans about it and they talk about it without, uh, actually building proper expertise. And it feels like
01:02:57.840 | a lot of wrong decisions and opinions are made about AI. At least this is what it feels like to me.
01:03:04.320 | So I just wish that we would be like, whoa, whoa, whoa, like let's calm down. Like let's, let's,
01:03:08.960 | let's not try to replace everyone with AI tomorrow. Let's, uh, let's build better, better tools, better
01:03:14.320 | tests, not for tests that we were talking about, but tests for AI. Like understand how you can actually
01:03:19.520 | replace anyone with AI, because you need tests for that. You need to make sure that whatever agent you
01:03:25.280 | create actually works. That's an open area of research right now. How do they do that?
01:03:29.520 | Like, I think the problem with the hype in particular, like I, I, I still think it's going to change the
01:03:34.240 | world and I think it's going, they're going to be 100%. And I think that there's, I don't necessarily
01:03:38.800 | know, like there will be like 10%. I think there will be more programmers. That's generally my view,
01:03:42.400 | because I think like more people will program, but, um, I think like the definition of a programmer
01:03:46.880 | might change and stuff like that. But what is really, really frustrating is this ridiculous discourse
01:03:54.000 | around AI in any shape or form is like engagement bait and, and this ridiculous bets, which are going
01:04:01.200 | on about like, when we're going to have super intelligence and HEI and whatever. Like someone
01:04:07.120 | has to get burned in that somewhere. And I'm pretty sure it's going to be the wrong people because
01:04:11.760 | statistically most likely it is. Exactly. Exactly.
01:04:14.080 | So, so there's definitely like, there's, there's some bubble going on here for sure. I don't think that if it's going to a pop,
01:04:22.640 | it's going to be, oh, we didn't want to do AI. So we're going to do something else instead.
01:04:26.160 | I don't think we're going to end up like Dune where we said like, this is too powerful.
01:04:29.120 | Let's just say no to AI. I just don't see that. Um, so my, my view is that even with the existing
01:04:35.760 | models, even with the existing technology, we're going to end up somewhere where it's, it's actually
01:04:39.120 | going to be pretty cool. But some of the bets going on are just pretty insane and interested.
01:04:44.960 | Are we going to need all that energy? Maybe, maybe it's also just a little bit insane what
01:04:49.440 | we're doing. I don't know. Like what is at one point, I think we have to look at this and say
01:04:54.800 | like some of the people which are currently paying someone's bill are doing this because they actually
01:04:59.200 | think there's some value in it, but they might themselves at one point feel like that was not
01:05:05.360 | sustainable. And then they stopped paying for that. And then the service that they paid money for
01:05:08.800 | stops paying for the service that they paid. Like there's a, there's a, all that money that goes to
01:05:13.440 | different kinds of companies up and down the AI stacks right now. I don't think it's going to
01:05:17.200 | be sustainable long-term. Someone made a misbet somewhere. I don't know who it is and maybe I'm
01:05:21.760 | wrong and maybe this is like the forever bubble, but some of it looks so unsustainable.
01:05:27.520 | Yeah, I agree. I don't have much to add on top of it. It's just my deep concern. I don't have any answers
01:05:32.720 | to it. I'm concerned. I'm concerned about this technology. Yeah. Being impacted in a negative way
01:05:40.240 | because of the consequences of it being overhyped. I also have another like just fundamental problem
01:05:44.960 | with the whole thing. How can private companies at this point be so fucking valuable? It's like if you
01:05:50.160 | wanted to say like capital is all great because like people can participate in the public markets,
01:05:54.400 | there's nothing where me I will grow to at least in a reasonable way because that company is like
01:05:59.840 | it like we have created this behemoth of private capital that hold all of it. There's a problem on
01:06:05.040 | its own, but it's just insane in a way that they grow like that. Like what was Google worth when it
01:06:13.600 | went public? A fraction of what all of those companies are even inflated for inflation and sort of the
01:06:17.920 | development of S&P 500. That to me like this insanity of creating this super national private large companies
01:06:26.720 | is just... It's the pump phase for sure. Maybe we'll collapse. I really hope there's not going to be a dumb
01:06:32.640 | phase anytime soon. That's going to be massive. Anyways, they will learn something here.
01:06:37.600 | I don't know what they will learn. It seems like the conversation converted you to become an AI pessimist.
01:06:45.920 | No, I am super optimistic. I love it. It's great. Am I optimistic about like what society will do with
01:06:52.240 | it? I don't know. I don't want to deal with it right now. Individually I feel super charged.
01:06:57.120 | Yeah. Well, me too. Ultimately I'm using AI. I think that a lot of our own practices might actually
01:07:05.120 | improve because of AI and will ultimately allow us to build faster and better. I also think that there's
01:07:10.800 | going to be a lot more software engineers in the end because somebody will have to fix the AI mass and
01:07:15.600 | actually make good out of it. So I'm not worried about engineers being replaced by software at all,
01:07:22.800 | despite what a lot of people think on the street. So yeah, I think the future is bright. It just feels
01:07:29.680 | like we are like really in a lot of uncharted territory right now and we don't have good answers to a lot of
01:07:35.680 | actually hard and imminent questions. So I'm not worried about HGI happening anytime soon because
01:07:41.520 | like what I see right now is that it's barely able to consume ClaudeMD reliably, left alone of launching
01:07:49.120 | nukes at us. So I am not worried about any of that HGI-ness, but I am worried about the impact on the
01:07:56.000 | field and like ultimately AI slowing down some parts of software engineers as opposed to accelerating it.
01:08:03.040 | Unless we get better at using it and instrumenting it and just like better at writing software actually.
01:08:09.680 | So it's an interesting intersection of problems for sure.
01:08:13.840 | All right, then thanks for the conversation.
01:08:15.840 | Thank you Armin.
01:08:16.320 | Okay.
01:08:18.560 | Let's see the feedback on this. I'm kind of curious.
01:08:21.440 | Well, I don't think that we said anything controversial.
01:08:23.840 | I was sh*t on some people indirectly.
01:08:25.680 | You never know.