Comment by spicyusername
2 months ago
The amount of text alone in the original post was a giveaway.
LLMs produce so much text, including code, and most of it is not needed.
2 months ago
The amount of text alone in the original post was a giveaway.
LLMs produce so much text, including code, and most of it is not needed.
I keep talking to people who say stuff like “Claude wrote it all for me in a day”, but when I look at the code (or try it myself) it’s just so much useless code.
I recently asked for Python code to parse some data into a Pandas dataframe and got 1k lines plus tests. Whatever—I’m just importing it, so let’s YOLO and see what happens. Worked like a charm in my local environment. But I wanted to share this in a Jupyter notebook and for semi-complicated reasons I couldn’t import any project-local modules in the target environment. So I asked a much more targeted question like “give me a pandas one-liner to…” and it spit out 3 lines of code that produced the same end result.
The rest of that 1k lines was decomposing the problem into a bunch of auxiliary/utility functions to handle every imaginable edge case and adding comments to almost every line. It seems the current default settings for these tools is approximately the “enterprise-grade fizzbuzz” repo.
Sure, I’ll get better at prompting and whatever else to reduce this problem over time, but this is not viable when the costs are being pushed onto other people in the process today.
And many of those utilities and edge cases will have been wrong or inconsistent, too. That's what the new "100x engineers" don't realise, because they never check those 1,000 lines of code they generated for themselves in a few minutes.
I've made similar experiences to yours for some one-shot scripts, and once decided to actually look inside. It did stuff like writing three different validators for the same data, each called only once, each validating slightly differently, and no doubt each with their own set of subtle bugs and gotchas.
These tools are intrinsically incapable of creating clean architectures and adhering to consistent standards and best practices. They are not cost-cutting or raising efficiency, they're simply very good at camouflaging the immense time costs they will cause down the line.
I'm using ChatGPT to generate some code for me quite often, and my instructions prompt for all chats is slowly gaining more and more ways to say "Answer shortly". And I need to prompt defensively to repeatedly tell it to only do what I tell it.