Why software keeps breaking? (Lost Episode)

Hey, we're doing a quick bonus episode of the stand-up. This is going to be short. It's going to be hot. It's going to be spicy. Casey is going to give us kind of a thesis on the change of software and where things are going. And perhaps kind of his thoughts on the web world and on the programming world.

Why does software always break? Anyway, sorry. Casey's going to give us a thesis on why does software always break? Casey. This is a hot new thesis. No one's ever seen it before. Except I posted it back in 2021. It's just gotten more true since then, Casey. It actually has gotten more true. And reposted.

So essentially what I wanted to try and point out to people, because I don't think that this is appreciated enough. We talked about it in the previous episode of the standup that we did, where I was saying like all of like a lot of the things that I see people saying positively about AI coding. I don't necessarily disagree with.

I just think they're not including this really important other part, which is that a lot of the things that people are talking about doing with AI are things that no one should have had to do in the first place. They're being done because we've created such a bad programming environment that nobody wants to interact with it anymore, right?

It's always breaking. It's always changing. It's got way too many layers of abstraction. Most of those layers don't work very well. There's way too much complexity, like all this stuff. So yeah, it's like that. I mean, it makes perfect sense why someone reached for an AI tool because why do you want to do it, right?

And so I just want to talk about sort of a separate part of that, which is just the very unreliable nature of software nowadays, and especially builds. Like, it's like, okay, I got this piece of software that I wrote, and like, what are the chances that I could compile it again in six months or something?

Or a year? Or two years? Right? or even not even compile, what's the chance that it'll run still if it uses things like REST APIs with some web service, right? So, you know, I've got these different web services I'm using. So even if I don't have to recompile my thing or even if it's an interpreted thing that runs and I'm keeping the same version of the interpreter or whatever else, it's going to make these API calls out to web services and those services could change.

What I did, and hopefully we can put the graphs up as I'm talking about them, but I posted these on Twitter. They're very simple. It's just taking the fact that, look, if you assume, right, and I say this in the tweet stream, that if you take the chance that something will remain working after a year is some probability.

So, like, 90% chance that this Twitch REST API that I'm going to call, a 90% chance that they will have kept it the same a year from now so that it will still work, right? sends this REST API call out to Twitch, it expects a certain response back, and they're not going to change it in some breaking way, right?

In a year. If we assume that we just have some probability, like 90% for that, that we can pick, we just imagine one, right? Or, you know, just imagine your head 90% or something like that. Then the chance that your code remains working after X years is just given by P to the X times N, where N is the number of those calls to things that you have, right?

Right? So you can graph this. And what I showed is that if you had a 99% chance that every API call you use, 99%, right? Which is way higher than anything in the web world typically has after a year. But 99% across all tools, the graph still looks pretty bad. Right. You look and you look at it. It's like, okay, after a year, it's like, it still looks pretty darn bad.

It's goes down pretty rapidly based on the number of tools. So you can see the graph I show is like one tool, two tools, three tools, four tools, five just goes down. I love that book. Dr. Seuss book. Yes. Great book. He captured... Dr. Seuss was... A lot of people don't know that he... The doctorate that he had was in computer science.

He was software engineering. He was one of the few to become a professional engineer in software. As opposed to the rest of us, yeah. Yes, yes. That book, One Fish, Two Fish, Red Fish, Blue Fish, that's where Two Fish comes from. He did that cipher, along with Bruce Schneier. So anyway, if you pick something less than 99%, like the graph for just 90%, it's very depressing, right?

Hey Brian, don't forget, scrum planning's later. No. No. No. I can't do this! Product manager here. We've decided to put you on call. I can't do- I can't do on call! You're on call? Lucky? What are you saying? With Sentry, On Call is so much fun. It has all of our errors, logs, spans, metrics, replays, and even releases in one spot.

It has all the context we need to solve any bug. Even if half the time Sentry beats me with a fix anyway. Just cause software development's torture doesn't mean On Call has to be. Wow, that's amazing! Does Sentry also attend my meetings? That's the best part. No! Sentry can actually make on-call fun, but it can make it less scary.

Everything you need to fix real issues in one place. Buckaroo, got that 6 p.m. post-mortem planning. Don't forget. Look at that graph. It's like nothing's going to work. Even at 90% chance that it remains working over a year, not 90% chance it will break. 90% chance it will work over a year. It is dismal.

Even with just one or two tools, it's horrible. But if you're talking about using seven, eight APIs, which is very common in software nowadays, forget it. You simply will not be able to use this thing if you update anything. Okay. So I just wanted to underscore, like, this is not good, right? This is not a very sustainable way to work on things.

And it has incredibly bad knock on effects. One of them is what we talked about, which is that no one wants to do this anymore. It's not satisfying. to work in this world because you're constantly just, you feel like you're drowning, I feel like, right? With all these things, oh, this broke, this changed, oh, they changed the way React worked, oh, we're doing this now, oh, that web service doesn't even exist anymore, they canceled it and now it's this other web service.

That's not programming to me. That's like some kind of weird management feeling thing. It feels like you're a manager more than a programmer because you're just trying to make this house of cards not collapse by constantly shuffling things around. And so... A, I totally understand where people are coming from when they want to reach for these AI tools.

I totally get it. And I think that if I was developing software in this world, I would reach for them too. So that's why when I say I don't use any AI tools right now, I'm like asterisk. But that's because I'm working on very specialized stuff. I'm not having to do these things. Right. And so it's very obvious to me why I have a very different opinion.

It's not that I like if somebody just don't understand the power of it. It's like, no, it's like I do actually understand the power of AI. I just think that a lot of the power of AI is really it's only correcting pretty bad situations that we sort of created ourselves. That's mostly what it's doing for you in these scenarios.

But also. I think this has so many bad effects on everything else. Performance, security is the biggest one when things are changing like this all the time. You just have so many exploit surfaces and you have no idea, like if you ask me to secure some system that's using all of this kind of, this way of working, I just have no idea.

I'm like, how could I? Right? Here, Casey, I got a good one for you. Let me just show my screen really quickly. Yeah. This happened just a little bit earlier. When I was doing Chessie, the vibe coding, my vibe coding good times right here. Let me just go like this. I think I'm not mirrored, right? Yeah, I'm not mirrored.

This was the thing that I got back from one of the Cloud 3.5, which was I asked it because I was like, hey, I need to log in. We're going to use Twitch. And I need to be able to store, obviously, my session data so I can make sure when someone makes an HTTP request, I know who I'm talking to. And so when I did that, this is the message I got back because I asked, like, is the session data secure?

Like, if I just knew someone's Twitch ID and username, could I just, like, log in by them? And they're like, oh, yeah, totally. You definitely can. Yeah. Let's add a JWT. Like this was going to go out and then someone could just spoof me and then just – And it knew. It knew when you asked it, right? But I had to know to ask it because I've written this problem once before.

So it's like I know what I'm looking for, and it's just hilarious about how dangerous security issues actually are because if you didn't know this – Like, how would you know to ask to do some sort of JWT or some sort of cryptographically signed thing to verify you're from the server and not from someone malicious?

Like, you just wouldn't even know the aspect. I didn't get to see your initial prompt. Did you initially say make it secure? Yeah. How did it know you wanted a secure service? I mean, okay, so again, that's fair. That's prompting issues. We can call that prompting issues that I did not tell it to make my login secure.

It is funny, though. It is very funny that it's like, oh, you wanted secure? I can do that. I love that that is the... I gotcha. I'd like to raise a more serious point about that, though. Okay. which is that if you imagine how things work currently, they're already bad currently, but let's just take AI out for a second.

If you imagine the way things work currently, there is actually one saving grace, which is everyone uses all these frameworks and all these different things. They pile all this stuff together, and it's a nightmare to secure yes, but... But what is true about that? Well, you at least know that you are using this thing.

And there are security researchers also looking at this thing. So when there is an exploit, you either know or could know. And when it's fixed, if you update, you get that fix. So if I'm using one of these things, I don't know what's a good example to pick here. 4chan. We can just use Express.js. Express.js just literally had this with, or this was like six or eight years ago with the Regex expansions.

You could do certain Regex expansions and have a Regex DOS effectively. Single request, take down an entire machine. So that's what I say. I don't know what to pick because I don't know what would be fresh people's mind, but something like that. So you take an exploit like that, it's like, that's bad, and that happened because you're using all of this other code that you have not secured yourself, and that's not a good thing.

But it does mean that when someone finds it somewhere, you will at least know, and if you update your system quickly, you can mitigate that damage to some degree. Right? If you imagine the alternative world where, you know, some open AI product is just crapping out exploits like that, because in the system, there's certain things that just didn't know or learned improperly.

That's like when it compressed them down and it's got its weights, it never really understood how to secure this one particular thing. Everyone who asked for that thing now has that in their own code base. And there's no tracking. Because we have no idea how many people asked for something that happened to hit that part, right, of the LLM's production chain.

And so unlike the Express JX, it's where everyone at least knows they got chumped. In this case, we don't even know where all of those exploits are. They're everywhere, right? Yes, and then you also have the creeping problem or the leaky abstraction, which is everyone has that problem. A quarter of those become open source projects.

The LLM learns from those open source projects. It re – like Donald Rumsfeld. Like Donald Rumsfeld talked about this RIP. So it's one of those things where it's like I just really – This is one of the things too is like, I don't actually dislike progress in some Luddite way. Like I like computers getting better and I like pushing the boundaries of what they can do.

And people have this weird thing where they think if you're not pro AI, you're just like some kind of person who just doesn't understand or doesn't like progress. Like, no, it's like, The problem is I'm not hearing anyone solve these problems. If I thought that these things were incompetent hands, where people were making recent decisions and they saw the train wrecks and they had ways of figuring out how they were going to fix them, I would be much less worried.

But a lot of the stuff I see with AI is it just feels like people who don't really know what they're doing, applying these things way too early, and I think the costs to software are going to be really high. And these people are going to be nowhere to be found. They're going to have collected their huge paychecks from companies that never even made money in the first place, which were VC funded.

They're going to take a bunch of that money and they're going to be gone. And they're not going to clean up the mess. So in my mind, it's like there's only two ways this goes. Either the AIs get so good that they can fix this problem themselves. And just to show you how not anti-pricing, I'm praying for that outcome, right?

Because it's the only one that I see plausibly averting this disaster. Or B, they don't. They don't get that good, right? For the next 10, 20 years, no one can figure out how to make an AI that can really be good. And then we're cleaning up this mess. And that is going to be the nightmare to end on nightmares.

Because it is just going to be like, if you thought security was bad now, and it is, oh my god, dude. If you thought performance was bad now, oh my god, right? It is going to be an epic nightmare. So I just think I wish people took this stuff more seriously, and they're really not. And that's the part that really definitely gets me upset about it in that Casey Rant way.

I'm just like, what if this doesn't work, guys? What if it doesn't work? You put all your hopes on this someday getting way better than it is right now. What if it doesn't? I'm so nervous. So fingers crossed. That it works. That's all I would say. Yeah, two things. Number one, Casey, since you like seeing computers push to their limits, you are a fan of JavaScript on the server then.

That's okay. True. True. Because there's nothing to bring that CPU temp up for a smaller amount of users than putting that JS on the back end, boys. But only one of the cores, TJ. Only one core. Can everything go through a MySQL query as well? Everything. I mean everything. There should be no data stored anywhere except in MySQL.

Let's query every pixel of an image from MySQL. We call it our SQL at this point. Our SQL. Our SQL. Our SQL. Yes, it's our SQL. My serious point is I heard it framed this way a while ago. Justin Keyes, one of the maintainers of NeoVim, was talking about he would be very interested to see, and this is more towards vision one of AI being good at solving these problems than not, is like, can we start seeing AIs reduce the entropy of a system?

Yeah. Right now, they're very good. Very good. We'll put quotes. Very good. I mean, okay, 10 years ago, we would have considered it literal magic to type something in and have a website come out of any kind. Correct. Absolutely. So we'll say very good. I'm going to say very good because it's, like, unfathomably good compared to what, like, my prediction for where we would get in my lifetime 10 years ago.

The human language understanding part is, like, clearly light years ahead of anything we had 20 years ago. Yeah, and there's a bunch of other follow-up effects. So it's like, okay. So it's like, okay, it can add a bunch of stuff to my system, I need a new feature, I have a clearly scoped bug request, I have some idea that I'm like, I'm the driver, it is my agent, right?

I think that's kind of where this is, like I'm sending it as my representative to go solve these things. And it can like maybe do that, but it doesn't actually, like, if I just say fix the mistakes or like, I say make it better. Make it more secure. Look at this code base makes more secure. Right. Sometimes it will pick up some obvious ones, which also you're sort of like, okay, but then shouldn't you have gotten that the first round?

It is kind of like a strike against the LLM. You have encoded inside of you the secure pattern, and you gave me the not secure pattern. That's stupid. I'm not writing the code anymore. Pick the secure one. To be fair, to be fair, that is also what a human would do. It is learned correctly. If you ask them to write it, they will write the inscription.

You're like, why didn't you write the secure one? It's like, I didn't want to. It took longer. I don't know. There was this paper a while ago where it was like LLMs were more tired in the wintertime because they had time in the system prompt. Sadder answers. They had sadder answers. And they worked less hard when it was wintertime because all the training data is like, oh, it's January.

And everyone's like, dude, I hate work. What? Yeah, they also got more accurate on math answers if you said, take a deep breath. Their accuracy actually skyrocketed because it was just like, well, because people will – because remember, LLMs are just reflection of written human behavior, right? Like that's what it is.

They're error-minimizing devices, right? So the next most likely thing of after try again and think smarter this time is to get a better answer. Could you try again? But hey, let's take a deep breath. Like relax for a second. Could you answer it one more time? You'll get a you'll you'll literally get a better answer from most people because they're like, oh, yeah, I feel better.

OK, yeah. Maybe they solved that. Whatever. I don't really know. They're doing all this. But my my general point being, I don't currently see them overall. Being a thing that I can let go and it reduces the entropy of my code base, which is that's, if it could do that, even just a little, we would be like way more on the track of your vision one where it's like, Oh, we can just let this run on Chromium.

We'll just spend a billion dollars. We're going to run it for five, 500 million human years on Chromium. And like in six months, it's going to come out and Chromium is going to be tight and, It's no longer one gig per tab page. It's going to be 800 megs, boys. 800 megs only to load that static site. And we'll be like, incredible!

But I don't see that as a thing people are proposing of we are close to. I get the zero to one. I get smaller features. I get agentic things for different stuff. I've seen value in each of those. But I'm not seeing anybody being like... We just let this run wild on Chromium. It's a superhuman programmer.

You know what I mean? Like, where is that? Yeah. I mean, that's probably because the direction originally of the research, right, is generative. Right? Yes. So, like, it's why it's called generative AI is because it's, you know, so it probably has taken them a bit to course correct to the extent that they even want to course correct.

Yeah. to do like, okay, what if it's about refinement now, right? Although again, like reinforcement learning is kind of in that direction, right? So that is a change. And so, you know, presumably they are kind of working on this stuff. Obviously it's, you know, I don't work on AI, so I don't have predictions about how they're going to get there, but...

But anyway, I do want to throw out something also about what you said a little bit earlier when it came to just like the security vulnerabilities and all that. I think one of the reasons why this will be the case is that we are also marketing a tool that gives the illusion of experience to people that don't have the nomenclature to understand the usage of the experience.

I mean, it's the same reason like if you've ever chopped wood, the first time you chop wood, you almost hit off your legs. Because you like swing your axe and you realize your legs are too, like you have, oh my gosh, I almost just hit my shin with my own blade. Like you learn to stand different because you, it takes.

I had a buddy who did that. Yeah, I know. It's a very reasonable thing for a lot of people to do. Yeah, you introduced me. Yeah. And so it's like it's that's what I worry about is it's not, you know, security. Hopefully it will get better. I assume that all things will it will be better in 10 years than it is today.

I think anyone will agree with that statement. I'm not measuring how much better all that kind of stuff is. But experience doesn't get better. People will still be the same people. So if you're marketing to the same people, they will objectively build bad stuff and they'll objectively build insecure stuff.

They'll do stuff that's crazy. They'll be like, hey, I need to be able to access my database for the client. No one's going to be like, hey, bro, that's a bad idea. They're going to be like, got your back. Keys are down. Client is downloaded. Let's go. That's what they're going to have to realize. Okay, here's the best possible endgame for AI.

All right? Okay, I'm ready. This is the best possible endgame. So they get it all worked. They get to the point that I want. They're like, okay, this thing is actually like a master programmer. right? And even better because it knows more domains right? Like it knows more domains than that because most master programmers are in a certain domain right?

But it knows more domains so it's great. We've got it. This is going to be great right? And then what they realize is they've still got that obsequiousness thing going where it's just always like oh yes absolutely oh I'll do that for you master. Certainly. You know what I mean? It has that kind of like really like almost slavish kind of disturbing degree to which it's accepting your commands and doing what you ask.

And what they realize is that being a dick... was critical to master programming. You had to be able to look at the program manager and go, you are so stupid right now. You have no idea, right? You had to turn to them and go, you don't understand Galactus' pain. You don't understand this thing. It's like, shut up and leave the meeting, right?

And so what they have to do is they have to re-change that front fine-tuning process they do afterwards to make the AI a dick. And then... then we have fantastic software. I love it. But the problem is... I want this future. AI companies, where are you? Do this for me. Make the AI dick programmer who changes the world.

Please. I will be so happy with that outcome. You heard it here. We're going to cut that clip, and it's going to stop before Casey says programmer. AI companies make me that. No, but they've had that forever. That's probably been there since 10 years ago. That already exists. That got different. Okay, one thing for...

I want to go back quick to Prime's thing about the security. I don't... The one thing I'm feeling somewhat... like less optimistic about it is people will be able to build more complicated ways, more easily to break systems because of AI, right? Like, so like, not only do I think there's going to be more services, probably it definitely in, um, I don't know about percentage wise, but certainly in absolute terms, more insecure things on the internet.

I think that's pretty much on undeniable because there's going to be so many more things on the internet. Um, Then the second thing is the cost of creating software in this world goes dramatically down. Okay, well, what happens when costs go down? People make more of it. Malware, hacking things, DDoS bots, all these other... And so it's like there is something where I don't even know that we can say for sure, oh, security is going to be so much better in 10 years.

The people making bad software, like evil software, bad software, they're already doing that, but evil software... is going to be more because it's going to be cheaper, right? And it'll just be like... So I don't know. So it's just something like... That's a scary thought, actually, because if you think about it, it's like, well, what would you have to do to make an AI system that was good at producing more secure software?

Well, we have to write a counteragent that's looking for exploits, and we're going to run that, and that's going to be part of the feedback loop where we train this AI. Well, the AI training sometimes takes months to run, depending on how... you know, serious this thing is. That means whatever AI agent for finding the things that we can write we have today, we won't have the AI can deploy those things for a little while, but the people who are exploiting the exploits, they will have that system for finding the exploits today.

So the cat and mouse game just got to the, I guess you don't know who the cat and the mouse is, but the exploit finders are always at an advantage. Because they will always have the AI system for finding exploits before the people who have the one that can correct it. Unless, again, there's some really revolutionary change in how these systems are made.

And to circle back to your first point, Casey, of the more things you have in your software stack, the more difficult it is to change anything because it's more likely to take down your entire system. People are still running Windows like XP. XP. Yeah, yeah, yeah. Like right now, you know, like in mission critical scenarios, they have Windows XP or like the joke I was saying for Fortune was like they got owned by some like 15 year old buggers.

What was it, Prime? It was some PHP vulnerability. I can't remember what it was. Yeah, some ancient PHP vulnerability. It was deprecated like 4.2 before I started using PHP. It was like so old it was deprecated. I didn't even know PHP ran on Windows XP in those days. What did it? Yeah, who knows? I don't know.

I guess you could also install it. I was just like, it's crazy that those were contemporaneous. Because I always think Windows 7 was PHP or something. I believe I did. XAMPP. XAMPP. Anyways, it doesn't matter. TJ, your point. But my point though is like... People don't update when it's even available to them.

It's not like, oh, well, the hacker guy's got the new stuff today and the fix comes out next week. So everyone's up to date next week. No. Not even close. Not even close. So that part is a little bit like just something to be wrestling and grappling with. And once again, like my point through most of it is like, yo, if you know things about software...

And you think more software is going to exist. That is a nice combination of skills to have. I don't know exactly what software development will look like in five years. But my general thought process, just from first principles reasoning, is if you know about software... And you know lots about it and you're good at it.

And your prediction is much more software. Those, that is a good combination of things to have. You will be valuable. It might not be hidden keys inside of NeoVim. I don't know. Maybe NeoVim will be dead in five years. And it's all Tesla, the brain control thingy from Tesla with the implant, you know, whatever, from Elon, right?

And that's the only way we're coding. Sick. But knowing more things about software is still better. Because I'm going to say, use a JWT instead of store this in plain text cookie on the front end, right? Like... That will be better. It will be better. All right. So I also have one more thing that I want to point out with all of this, all these things, and especially targeting people that are, you know, no coders to make code like things, is that how I learned how to code was that I first started off and they said, okay, this is an if statement.

Somewhere between 1999 to 2005 is when I started kind of doing like basic exploration of code. And so... Here's an if statement. Okay, this is a while loop. Okay, this is a while loop. Okay, I want you to print out a house. Okay, I want you to print out a diamond. You got to print out four diamonds. And now you'll notice it gets really annoying.

I want you to be able to do it by different sizes. Here's a function. Here's how you can make a print diamond function. And you go through all these things and you slowly go, okay, okay, yeah, okay. You build up this kind of picture of how code executes. You learn to debug. You do all these things. A lot of people that are vibe coding...

I'm curious how discouraging it is to get dropped into a Next.js app with Supabase, with Auth0, with 900 things, and you have to start by debugging a request response. And you're like, what's a server? And you're already into some sort of crazy amount of difficulty where it's like, I started by drawing a diamond.

Right? Like that starting point is vastly different. It was literally the draw the owl, but you had the middle steps. I had all the middle steps. They're draw the owl. They're literally given the owl. Be like, draw the owl, right? It's just like that's really, really, really hard. And so I'm actually curious about the success rate of somebody going through and being dropped in hyper complex projects comparatively to, hey, you're young.

You have this free time. We're now putting you through this school. Like maybe it's high school time. Hey, do you have this free time? Let's draw a diamond. We're going to draw a diamond together. We're going to use QBasic or some basic language, right, Lua. And you're just going to do the simplest kind of form of doing things.

I'm curious what that does to somebody because I know there's going to be a bunch of success stories. There's going to be people that are super stoked about programming. They're super stoked about building a product, and they will figure out a way no matter what system you give them. But I wonder, like, overall, does this actually help make programmers or does this actually hurt the learning process?

There's also, I think, going to be a bunch of success stories of people who hate programming, but were able to make whatever their business product thing is without really having to know anything. And that, like, is certainly going to happen. And you can say, like, maybe it's a net negative for programming or something like that or for the web, although probably most of these people are not building, like, foundational technologies.

So hopefully, like, but, like, that is going to happen. Yeah. Right. Yes. They're going to make their own website. They're going to get to, they're going to make Uber for cats. Um, and like, they're going to get to make Uber for cats and they finally don't have to recruit their other friend and tell them, I've got the idea.

You do the code. We'll split it 50 50, right? They'll just be like, I'm going to do the website. The Winklevoss twins would have owned Facebook. They wouldn't have needed Mark Zuckerberg. They would have owned Facebook. Yeah. But would they have had Justin Timberlake say, drop the, it's cleaner. Yeah.

You know, I mean, that's the thing is, I don't know. Would you have had that? Yeah. Drop the Facebook meta. It's just clean. It's just clean. That's what he said. They didn't listen to him at the time. Zuckerberg realized later. Took a little while. Took a little while to sink in how Justin Timberlake does.

I will also give the inverse, which is that you can also ask AI any question, and you can repetitively ask stupid-ass questions over and over again. And it will repetitively, in the obsequious way – I'm not sure how to turn that word. Obsequious? That's it, Brian. Whatever that word is for saying subservient.

I know the word. I just don't know that term, the lowly term. It will repetitively be like, certainly, I would love to help you. No matter – unlike Stack Overflow mods, you will not be marked as a duplicate or opinion-based. You will actually be given a nice full answer every single time. So maybe in the end it does actually help more people achieve their coding dreams.

I don't know. I want to make sure that people don't think I'm just hyper-negative on all those things. I'm just – I don't understand how this affects the new people or how it affects learning because I also had no shortcuts. When my teacher said – Build a – solve a maze recursively. I had to learn recursion at that point.

There was no other option. I had to learn it. I couldn't just get an answer. I had to figure it out, which is like – there's something there that is also very special. And I don't know where the balance is. I do think that there might be an argument there for like can we make an AI that's been trained not to really give you full answers, right, for educational purposes.

So it's like – The rabbi? The rabbi? GPT? Kind of, right? Like one that's going to give you a hint to help you get unstuck or to help point you in the direction that you need to go, but it's not going to just tell you how to do the thing because it wants you to learn, right? And I assume that that is doable if you spend time training it to do such a thing.

Because obviously they train it to do very complicated things already. They have, you know, post... Their post-initial learning phase stuff is very complicated at this point. So I'm assuming that this would... If someone put their mind to it, this would be very doable. Or maybe someone already has. And that does sound useful as a learning tool, right?

Because you can't... A lot of people don't have the ability to ask, you know, a great programmer sitting next to them or something. They don't have that opportunity. So... The two that I know for sure, like BootDev, shout out, promo code by the way for me and Prime if you like that. Yep, share the code.

They've got a little like AI helper thing and it's got special prompts for like each lesson and a bunch of other stuff like that, right? So it's like it can help you when you get stuck on a lesson. Yeah. And it's like not supposed to be like here's the code. Right. I mean like I'm sure you could prompt ejected blah, blah, blah.

It'll give it. It's like. Yep. Okay, but it's helping you to learn. So it's at least its aim, its error minimization is towards that. The other one that I've seen is called Synthesis School or something like that, which is a bunch of AI tutor things, but they build it into a bunch of lessons. And so this is, I definitely think, and I've said this before too, I think people are underestimating in the learning phase if you are motivated to do it, LLM can be very helpful at doing that.

Now, you have to make sure you're not getting gaslit into believing some function doesn't exist. But for basic CS fundamentals, it's got all of those books loaded in. direct. Probably you could ask it what's on page 37 of algorithms and it'll pull out. It probably knows. You know what I'm saying? It's seen it so many times on the internet.

So for basic CS stuff to get you far on a bunch of these basics, you can be asking it, you can be doing questions, you can be like explain that again, explain it in a way that I would understand if I really like, I don't know, math. Can you relate this to a physics example for me because I understand physics but I don't get computer science.

People are definitely sleeping on that aspect of helping them out. But you got to do it yourself. So that's kind of the drawback. That's where the danger is because even if you're semi-desiring to learn – because it really is a desire magnifier. So it's like it really reveals your ultimate desires. Was your desire to simply get the thing done or was your desire to learn?

Right. And if you don't have your desires correct – Or at least in your head, what you think of yourself versus in your actions, it will make it happen. Revealed preference. Revealed preference. That's the term I'm looking for. That can be hard. This is why I think having an AI that's specifically designed for this and you only subscribed to that service would be helpful for people.

Because I don't know about you guys, but if I want to eat more healthy food, the easiest way is to just only buy the healthy food so I don't have bad food around the house. Yes. It's much harder if, like, I buy a bunch of, like, cake that I love or something, and I'm just supposed to not eat it, right?

Just only eat one slice a week, Casey. Whatever, right? I'm just like, I want to eat the rest of the cake. My wife does this, and she's like, I bought natural popsicles for the kids. And I'm like, damn, I love strawberry. There you go. You know, like, it's just very hard for me not to want to eat them, you know?

Yeah. This is what I'm saying. For reals, though. And so I feel like the AI is a bit of a problem that way, which is why it would be nice if you had a service that was like OpenAI or somebody, right? They have this one that's like education only, and it's just – it's specifically designed not to give you answers very quickly.

Like it's like I'll dribble out some stuff. I'll give you some other stuff. Even has like this – maybe you could bake the cost of time in there. It's like if you haven't been working on this for a couple of days, I'm just not going to tell you. anymore right you have to spend some time trying it yourself right i could see that being very helpful for people because you know inaccessibility is best willpower is second best right so if you can have it you know if you could have that that would be cool and i think that would help bring out those learning abilities of the llm so that people aren't sort of like too tempted to just ask it like just tell me how to do the freaking diamond right or whatever yeah just give me the code yeah

And the thing is, I also can't blame people for doing that. I would totally do that, too. This is why I said, like, how the AI not do that is better, even though it's less, you know, it trains your willpower less, but willpower is hard. It's hard for everybody. Also, if, like, if you get the experience of actually solving it for yourself a few times, then you're like, oh, it actually was kind of rewarding.

Right. So, like, it really can help you, right? Right. oh the first if you start working out and then like suddenly your times start going down on like you know how how fast it takes you to row 200 meters right and you're like sick that feels good like initially i wasn't thinking it would feel good i didn't see any change at all at the beginning you can get there like same thing can be happening for some of these too where like use in your brain it's getting connected like oh working hard can pay off interesting interesting yep believe it or not chat Believe it or not.

I like how this was going to be a super short, quick episode. We are probably like an hour in at this point. We're 36 minutes in. And that's why I said we're going to stop and start a new recording. So you guys on YouTube, you can like it. Like it. Like the video right now. Subscribe. Leave a comment.

Say, hey, Casey, thank you for this bonus episode. I never ask people. I never do calls to action. So, hey, like the video. So he has TJ do it. Press the subscribe button. TJ streams, by the way. KC has computer enhance. Yeah, slam dance that. And the bell. What about that bell? You got to click that bell.

Oh, man, get that bell. Click that bell. Look at that bell there. You know what DTube needs to do? Why does that bell not make a sound when you hit it? You know what I'm saying? Like, you get that little Pavlovian response for smashing the bell. Ooh, that would be nice. Because someone at YouTube is still trying to figure out which Gemini prompt to type in to get that to happen.

It hasn't happened yet. Yeah. Yeah, yeah, yeah. And the sound's all wrong. Well, it's because they're like, I'm not going to do it until... two weeks before my review cycle. So I can have a good review cycle. I can get promoted to L2 and then I can trash this project to get back to L3. And then after that, we can just delete the whole downvote button and then I'll get promoted to a VP of upvote downvote systems.

And then we can close down YouTube because it's a Google product. Boom. C-suite. That's what we're talking about. C-suite material, boys. All right. Well, hey, that was fantastic. That's the end of the episode. Goodbye, everyone. See you later. Take it easy, everybody. Move up the day. Vibe code and errors on my screen.

Terminal coffee.

Why software keeps breaking? (Lost Episode)

Kapitelklicken springt direkt hin

Shownotes

Transkriptvom Podcast bereitgestellt · Zeitmarken springen hin