# LLMs are like handwritten notes > Source: > Published: 2026-07-18 01:14:45+00:00 [Published](https://dunkirk.sh/blog/llms/) [Kieran Klukas](https://dunkirk.sh) # LLMs are like handwritten notes This is my best attempt at formulating the thoughts that have been tumbling around my head for the past two years. It’s a bit pessimistic but I think that is kind of appropriate. AI `!=` LLMs. This is one of my biggest annoyances with the current state of “ai”. Every company wants to add some fancy new ✨ AI ✨ feature into their product and most of the time either they would be better served with a more traditional ml or rl model or even worse it could litterally be a tiny bit of good old fashioned conditional or regular expression. AI itself feels like a horrible misnomer. It is my deepest wish that LLMs had never been marketed or branded as inteligent entities. Linking LLMs to intellegance was a fatal mistake in my opinion as it subtly encouraged leaning on them more and more trusting them to behave like a human. My personal philosophy for LLMs / generative technology boils down to 3 points: - Never **ever** use an LLM to write anything I present as my own writing. - Refuse to use generative art and music tools. - When using LLMs for code generation purposes seek to fully understand the code and never blindly trust it. My biggest worry with LLMs is the collective collapse of critical thinking and logical reasoning skills that is starting to show up. It is so easy to outsource your ideation and execution to LLMs especially in the software space and the ease of slipping into this really worries me. I see a pretty similar correlation between the strength of handwritten notes when compared to typed or recorded notes. When we handwrite notes we are forced to condense a rapid stream of spoken information constrained by the speed of our hands which causes our brain to remember the information much more reliably. Similarly when we write code we are bottenecked by the speed of our fingers on the keys but also by the rate at which we can solve the next problem. By forcing us to slow down and make decisions about what we want to type in and how we want to tackle a problem it leads to more maintainable code but also engrains the problem and our solution to it in our memory. We don’t necessarily remember everything perfectly but the concepts of how we tackled something stick with us and make us better programmers. LLMs have an even stronger effect on our brains. Since we aren’t bottlenecked by either the speed of our keys or the brainspace a several thousand line new file takes up it is much easier to generate spaghetti code. Because we don’t have to directly interact with the code we aren’t quite as driven to write maintainable easy to read code but also we are willing to expand scope and add features and ideas that we likely never would have before because it would have required too much work to maintain. The assumption is that LLMs will make maintaining this code easier in the future but so far we really aren’t seeing that be the case. LLMs tend to write very verbose idea specific code and don’t tend to be amazing at condensing and maintaining as they go on. It feels much like beating a pile of typesetting blocks with a rubber mallet and hoping it makes them all fall into place but instead it knocks over more letters and makes the page ten times bigger. The worst part though is that in most cases the output looks plausible. It is becoming increasingly hard to tell whether a mr is ai slop anymore. Oh sure there are still signs. Comments written a certain way, a ton of churn in code that might not have needed a refactor, a weird description. Looking for those signs and then deciding what to do about it takes up an increasingly longer amount of time and effort. There has been a lot of discussion in various places about the need for increased ability to filter and vouch for contributions and I think that makes a lot of sense but is also really sad to see. Before it used to be that the effort of understanding someone’s codebase, forking it, setting up the build environment, making your change, and then working with the maintainer to get something upstream was long and not really automatable. You could trust that someone put actual thought into designing their merge request because otherwise why else would they put in the time? I haven’t personally been in OSS for that long yet. I’m still very much a greenhorn but I remember the absolute joy of making my first contributions to projects after getting my github account. Were they great? Probably not! But there was care and I did truely want to learn how to do stuff better. The first time one of my projects got a MR from someone else I remember jumping up and down in literal joy because some stranger on the internet thought sonething I made was cool enough to spend valuable time from their day to add a feature. Now i’m more likely to feel a groan because I have to determine if its slop or not. There are certainly folks out there who do care. I have a lovely group of friends who make my day if they submit a change but especially the larger the project gets the harder it is to keep filtering. My thoughts on LLMs become especially mixed because despite really disliking what they have turned OSS into and despite what I fear they will do to our brains I still somehow see value in them. I’ve been able to catch bugs in some of my code that I was never able to before and build small tools that I likely never would have tried to build had I not used an LLM to speed up the process. Do I worry that my skills are atriphying as I use LLMs more? Absolutely, it has been bothering me to no end for over 2 years now. And yet I still keep using them. Perhaps it really is an a vibe bobsled like [Christine Lemmer-Webber](https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/) was writing about. Perhaps I’m being overly pesimistic. I really don’t know what the future is going to hold. I’m going to continue enjoying my uni CS classes for the total and complete break from LLMs that they enjoy at least for now but I likely will also keep using LLMs to speed up projects. Part of what makes this conflict so hard to reconcile and ironic is that I work on LLM tooling. I actively use LLMs in my day job and they are quite useful. I feel like there has to be a balance that allows me to leverage their usefulness while also protecting my brain. Ceasing to use LLMs entirely feels like more drastic of a decision than is reasonable but I also really don’t want to hurtle the icy vibecoding run of doom and despair. I’ve tasted the exhiliration of running multiple agents in parallel and working on 50 problems at once and I can admit that it is intoxicating but at the same time absolutely exhausting. I don’t think our brains are meant to juggle context that quickly and it leads to quality control issues. Recently I’ve been starting to take notes while I’m working about problems I encounter and I’ve found it incredibly useful. I can avoid the really tempting allure of quickly fixing the issue, save it for latter, and then have a nice rewarding problem to fix in a few hours. I’ve also found that starting my day out with sorting through issues and making tickets for them (boo corporate i know) is suprisingly relaxing. I’m starting to realize that perhaps as a personal point of pride I was always trying to keep track of everything I needed to juggle in my head. Turns out my head isn’t really very good at that and thats okay! I think the same thing applies to using LLMs, we don’t have to use them the way they are often presented. We can adapt them to our needs and also be willing to challenge whether its the most effective use of our brainspace. More and more I admire developers who have been at this for years and years and they have built up an incredible wealth of knowledge. Someday I too want to have that same storehouse of knowledge and I know I’m not going to get it vibecoding. Cheers! ### A few posts worth reading Simon Stewart: [We’re Going to Make Out Like Bandits ](https://www.rocketpoweredjetpants.com/2026/04/were-going-to-make-out-like-bandits/) Christine Lemmer-Webber: [Faulty Towers, vibe sickness, and the vibe bobsled](https://dustycloud.org/blog/faulty-towers-vibe-sickness-and-the-vibe-bobsled/) Jeremy Theocharis: [The LLM Critics Are Right. I Use LLMs Anyway.](https://www.theocharis.dev/blog/llm-critics-are-right-i-use-llms-anyway/) L. M. Sacasas (The Convivial Society): [Your AI Is Not a Tool](https://theconvivialsociety.substack.com/p/your-ai-is-not-a-tool)