.png)
4 hours ago
2
To preface this: if you are someone who writes much of the same code everyday, or someone who is incredibly experienced in the field (from the mutuals I have on twitter/x, I can think of maybe one or two), then this article probably isn't about you, but feel free to read regardless!
Let me set the stage: You are new to some concept, so you go to learn more about it by making a toy project, and to aide in this, you enlist the help of the mystical oracles we call llms.
You might be prompting this model in a chat window, maybe in your terminal, maybe some other third method I didn't think of while writing this. As you move towards building this toy project to learn, you constantly turn to the model for help, be it advice, or using it as a search engine for documentation. The output seems fine, and you move on all the less aware of the fact that you've slowly been fed the wrong information, or perhaps you have to deal with a dozen small syntax failures from the model, or just downright wrong code, but you just use the llm to fix it, that's what it's there for!
But you came into this project to learn, to actually get down in the mines and gain some knowledge. Yet you've slowly been offloading all of the process of learning to an llm, probably without being aware, or thinking you're not entirely doing it. The issue here is that for you to truly be able to learn these concepts and tools, you need to be able to internalize the flow of thought that goes into it. It's easy to read 20 lines of code and think "hmm yes I will remember this and apply it next time I reach a problem of this shape." But the issue is that you didn't actually think about what the shape of that problem is, you didn't even play with it in your head or on paper to actually understand it's contours, you might not even remember the code after 20 minutes.
The most important part of learning anything is the ability to internalize concepts and build a mental model, and I don't think llms are particularly good at helping you with this (at least without prompting them so much that you end up spending more time than if you just studied it yourself.) In the current software climate, the big names all preach to "ship! ship! ship!", and this leads young people to construct a delusional path to success (either rushing towards vc funded b2b SaaS or to get ready for a job), and through this they start using llms to their advantage (and why wouldn't they?) to catch up to these slop-peddlers, not that there's much catching up to do to most of these people xD. I dislike seeing people led into this hole, it ends up hurting them almost every time, and when it doesn't, it ends up wasting so much of their time.
I have fallen victim to this many times, it's an easy mistake to make, especially with new concepts, as you can probably extract some signal from the high level components of the concept, but as you dig deeper, you begin piecing together a puzzle with parts that aren't even from the same box. And while it's not the end of the world, it is really annoying, it might cost you a few hours now to go back and fix those holes in your mental model, but the real concern here is what if these holes weren't fixed now? What if you continued with these incorrect solutions and ended up using them in some critical code? While unlikely, it is something you should be on the lookout for.
I don't mean to scare you off of using llms, they are wonderful things and can be the most useful tools. The only real practical takeaway here is to stay aware of the balance of your llm usage and actual conscious studying, it really pays to become familiar with your tools, even if it means you have to spend more time gaining that familiarity.
For the time being, these machines cant entirely replace your job, even as they get more intelligent and capable, you will still need to provide some support at some level of complexity to ensure you're writing good code.
If you think I am entirely wrong here, or perhaps I'm missing something, please message me on X @ aryvyo :)
