AI writes code extremely well - unless you do this (The "Paw Mirror" problem)

Added: 2026-06-02

[00:00:00]So, I came across this hilarious shy girl controversy recently and it struck me as a perfect example to explain to a non-technical person how AI-assisted coding can work great, but also how it can completely fail and what, you know, can trigger these kind of fails. And you know, to recap what happened here, so a woman wrote a book called Shy Girl and it got picked up immediately by publisher and I can understand why because it seems like from the plot the perfect intersection between like 50 Shades of Grey and the new kind of animal beast trend that seems to be going on and that I know about from a Meat Canyon video. Anyway, so it's about a woman who like becomes this living girlfriend of some guy who trains her to act completely submissive and then she slowly turns into a dog over the course of the book. So, this got picked up by a publisher and made headlines, but then more and more people found, "Hey, there's a lot of evidence here that this was written by AI." And yeah, there were some of the usual telltale signs like over obsession with a word like here the AI used a lot of sharp and quietly humming and stuff like that. But, the main problem was that the AI lost track of the slow gradual conversion of the woman into a dog. And I understand it's like a first-person narrative, so towards the end for example, there is apparently a scene where he says to the woman who's now supposedly completely in dog form, "Here, look at how disgusting you are."

[00:01:45]And the first-person narrator picks up a mirror to look at her face and of course, if she turned into a dog, how could she pick up a handheld mirror with her paw? And to me, that's such a perfect example that happens within the AI if you give it too much context. And any person who thinks about, "Okay, let me write about a woman turning into a dog, you know, using AI." You know, they might feed it, you know, the prompt, very detailed prompt about what should happen, and then, you know, AI writes the first chapter, then they feed the first chapter and says, "Here, write chapter two, and I want this and this and this to happen." And so on. And then, as the context grows and there are more and more chapters, AI doesn't know what's important, what's not. And that's like the one thing where with a code base or novel, humans are still um vastly superior over LLMs because we can pick out exactly what's important and then ensure that's the context. You know, you don't have to know every word of the first five chapters to write the sixth chapter, but there are certain points you have to know, like the current shape of that woman or the dog traits she already has or not has. That is absolutely important because as a human author, you know that the human reader will picture the scenes in their head, and then if there are inconsistencies, that will break the illusion. So, for the same reason, I feel that yes, AI can do coding for us, senior devs, but it still needs the senior dev to narrow down the context to what's important.

[00:03:25]And then you just give the prompt to it and let it run, and it does a fantastic job. But that prompt is super important.

[00:03:31]The narrowing down to the context info that's crucial, that's super important.

[00:03:37]And I suspect virtually all times when you hear that some programming project failed, it was because the human developers did what the author here apparently did, you know, just just feed the whole code base, you know, like feeding the whole chapters to the AI and say, here write me the next piece. You just can't do that. Those are LLM's.

[00:03:57]They have a certain context window and even when they have a context window of a million tokens, that doesn't mean you can put a code base that's a million tokens in the LLM and it knows what to do. You are much better off narrowing it down. And then especially if you're me and you know you have to save tokens and radically keep cost down for AI driven development. You have to be super selective in what you pick. And the good thing is you can actually use AI to help you keep track of it. But it starts with that first step, what's important. And the author here could have used a prompt such as uh I envision this novel to be about this woman who slowly turns into a dog. For each chapter, let's nail down exactly what her status is at that point in time and then write it knowing that context. But I guess that would have taken a lot of thinking, you know, a lot of creativity, a lot of okay, will will her hands morph first or will her voice morph first or her hair or you know what what will be the most appealing to the reader. And you know, a lot of people want to skip that important step. And a lot of senior devs or devs in general when they see how AI can write great code, they might be skipping that very important step. They give it the whole code base and say here, you know, just debug this problem, you know, throw it against the wall, you know, 10 times until it works. And that's just not a good idea. So AI coding or anything AI is about us humans picking the context, which can be time consuming but not nearly as time consuming as coding. And if you give the AI the right context, the right instructions, then it does actually just to me it does a bulletproof job. Especially if you have it write up a plan, which Claude now does routinely on its own and then you read through the plan and then when it's ready you have it go forward. So whenever you hear an example of a developer who says, "Oh, AI is so stupid. You know, I already had a library here that does certain things and then it just created a whole mass of code reinventing the wheel and recreating lot of very similar methods to that library instead of just using my other library." Whenever you hear complaints like that, you can't unleash AI on a big codebase, then that's some clueless developer who doesn't know that you have to pick the context and then if you want AI to use a certain library, then you have to say, "Okay, use this library here." You know, implement it that way or you you just simply let it do its thing because very often it picks it up on its own and you read through and if it says, "I'm going to write this library." You know, then as a senior dev you have to know, "Hey, wait a second.

[00:06:45]It's planning to write this library in the plan and I already have a library like that in some other place and uh so then you just tell it to reuse. But yeah, nothing aggravates me more than people doing these, you know, superficial tests where they go like this author say, "Hey, I mean, write me a novel about a you know, woman who turns into a dog. You know, write me this big codebase and then let's throw on and add on 20 more features and then AI started screwing up half of the way so AI is useless and you just know that their process completely sucked. It was their process. It wasn't AI at fault. And there was one more aspect where this apparently mostly AI generated novel rubbed people the wrong way, which was that you didn't really get as a reader a good feeling of the inner world of the narrator. And again, you know, how can AI write a novel where the first-person narrator has a rich internal emotional world if that didn't get prompted. That doesn't mean AI can't do it. It just meant that the author, if she you know, insists on using AI, you know, at least sit down and really think of what should be in your novel and if it's a first person narrative, then hash out with the AI, okay, what is her emotional state at any point in time and you can't generate a chapter at a time.

[00:08:11]You're more like a few pages maybe at most. And there were some other, you know, hilarious examples, too, like she gets like a eviction notice on her door like handwritten from her landlord. So, more of a casual note, but it said something like a a quit notice. I don't know if that's a British thing or or something. So, apparently, uh move out or quit is like a legal phrase, but you know, to a reader reading the book, you know, having that as the note on your door didn't make sense, but AI was trained on, you know, legal documents of eviction notices. So, it used that for that handwritten note. So, so yes, you have to check the code and I would actually argue if you carefully review the plan, you don't even have to examine every line of of written code because the plan will tell you what you need to know. You don't need to inspect every single loop or wherever it created the method or whatever. AI is super great at structuring code, naming variables, naming methods, writing things in a reusable way, well engineered, but not over engineered. And even just the naming conventions thing or the naming of things. Have you ever seen like a human code base written before, you know, 3 years ago and how many humans like make stupid typos like in database fields or methods and it always drove me crazy because those will be then more or less what you are doing maybe for eternity. And then you have to know that when you search the code base for reinsurance, you have to search for resin insurance or whatever.

[00:09:48]And of course a lot of coders were not, you know, first language English English speakers, you know, just like I am not.

[00:09:56]But yeah, and I'm sounding like the alarm here in my videos because theoretically AI could already write 100% near 100% of the code in nearly all corporate projects which for the most part and throughout throughout my career I'm mostly like in the widest sense boilerplate. You know, maybe AI can't write something like truly truly innovative very low-level like C++ code or something where there are not many examples. But for virtually all SaaS apps, mobile apps, websites, APIs, cloud architecture, there are already so many examples out there. AI has been trained on it. It can do all the coding. And at least there personally I believe that tomorrow all corporations could get rid of anybody who is anti-AI or who can't and who can't prompt, anyone who can't understand the bigger picture of a code base which sadly there are a lot of developers and replace them with a few selected senior engineers who understand the large picture of an app and yes that takes like full stack knowledge and not necessarily very deep knowledge but broad knowledge and some experience of what's going on in the code base. And then very little to no manual coding would be happening on those corporate projects. And a lot of things like offshore teams would be greatly reduced.

[00:11:20]And yeah, right now the collective cognitive dissonance is that we have this AI hype bubble where the valuations already kind of baked all that in, that already happened. Because if the economic value of offshore teams manually [clears throat] writing code is, let's say, you know, going on a trillion, and then, you know, a coding model or a company providing AI coding services, if that is valued at a trillion, you know, that's already that's jumping way ahead. You know, we have to go through the phase in between where companies have to completely restructure their IT department. It's kind of interesting to me that, you know, even like a year ago, nobody was talking about developers being replaced by AI. And now suddenly you have like, okay, this tech company is laying off thousands of workers. Tomorrow it's that company, and then there's another company. And yeah, a lot of that was through over-hiring or for other reasons, but I feel it's an indication of there will be an immense change in how software is being written going forward. And now this turned into a long rant, which I didn't mean to, but, yeah, if you're a non-technical person overseeing a software development team, and they said we tried AI, we gave it our codebase, and what it wrote was absolute nonsense, so let's put AI aside, and it also was costing too much, then think of did they do what the author did, which was, you know, AI write me the next chapter, you know, after, you know, having 10 chapters, and then being surprised if if kind of garbage comes out that does not withstand scrutiny. And it costs a ton of tokens, so, yeah, all right, rant over.