You must log in or register to comment.
The article concludes with this:
We would be doing a disservice to our users and our downstream fellow developers in the same way that a visual artist would be doing a disservice to their viewers or a musician would be doing a disservice to their listeners if they served them auto-generated filler instead of their own creative output.
You can personally select and fell all the trees used to build your house. You can cut and sculpt each 2x4 with chisels and a hand plane like a true artisan.
Or you could just buy lumber in bulk, toss the boards that have flaws, and build your house that way. What’s great about that is that as long as you follow building codes your home won’t just collapse at the first storm. It won’t just burn down with the drop of a match. It won’t leak and ruin your stuff.
Using LLMs to code is like buying lumber in bulk. Some of the boards that come off the truck will have flaws. Sometimes they’re so bad you can break the 2x4 over your knee! But the job of the builder/framer is to make sure those bad boards don’t get used. They’re spotted, marked, and replaced.
Code can be art! Absolutely! But not all code needs to be art. Not only that, but your employer doesn’t give a shit about how well-written any given code is so why bother? Just give them the AI output that works/passes tests and collect that paycheck.
I have the feeling you only read the first and last few sentences of the article, or maybe used an LLM to give you a false summary?
Yes and no.
Differentiate. There are engineering and architectural tasks with a certain degree of artistic freedom. And there are many tasks which should be done by the book.
Nobody wants plumbing and electric wiring done by an artist. Painting work? Maybe.
LLMs are a tool, if a tool is useful people will use it. Programmers already were fine with using each others code when it fits the job, why would using a machine’s code be any different?
The amount of time an LLM saves is negated by fussing over which service provider and model to use, what to put in your agents.md file, setting it all up, maintaining it, and then arguing with its output since after all that trouble, the damn thing is still wrong 10% of the time. And then you’ve made yourself dependent on a service that is going to explode in price when the AI bubble pops.
It’s a tool whose pros are more flashy and prominent than its cons.
For the most part I’m just focusing on learning and making myself a better programmer. I’d rather do that by going to the source (forum and blog posts, tutorials, Stack Overflow, and documentation) rather than have that information mixed in a blender and removed of its context and nuance by a bot that always gives an answer even when it can’t find anything useful.
looking back at that reverse engineering attempt, I think the reason I turned to trying to use AI was largely because there was next to no information about reverse engineering windows drivers (specifically pcie device drivers) that was remotely easy to follow (the closea I remember finding involved using qemu with an option that isn’t built into it by default to I think buffer all dma accesses through a file).
they seem fairly useful for highly copyright ignoring reverse engineering, mostly because they have some understanding of most of the windows functions the driver I’m trying to dissect is using.
