Comment by fsloth
5 days ago
I totally agree. The ”hard to control mech suit” is an excellent analogy.
When it works it’s brilliant.
There is a threshold point as part of the learning curve where you realize you are in a pile of spaghetti code and think it actually saves no time to use LLM assistant.
But then you learn to avoid the bad parts - thus they don’t take your time anymore - and the good parts start paying back in heaps of the time spent learning.
They are not zero effort tools.
There is a non-trivial learning cost involved.
The issue is we’re too early in the process to even have a solid education program for using LLMs. I use them all the time and continue to struggle finding an approach that works well. It’s easy to use them for documentation look up. Or filling in boilerplate. Sometimes they nail a transformation/translation task, other times they’re more trouble than they’re worth.
We need to understand what kind of guard rails to put these models on for optimal results.
” we’re too early in the process to even have a solid education program for using LLMs”
We don’t even have a solid education program for software engineering - possibly for the same reason.
The industry loves to run on the bleeding edge, rather than just think for a minute :)
when you stop to think, your fifteen (...thousand) competitors will all attempt a different version of the thing you're thinking about and one of them will be the about the thing you'll come up with, except it'll be built.
it might be ok since what you were thinking about is probably not a good idea in the first place for various reasons, but once in a while stars align to produce the unicorn, which you want to be if you're thinking about building something.
caveat: maybe you just want to build in a niche, it's fine to think hard in such places. usually.
1 reply →
> We don’t even have a solid education program for software engineering - possibly for the same reason.
There's an entire field called computer science. ACM provides curricular recommendations that it updates every few years. People spend years learning it. The same can't be said about the field of, prompting.
2 replies →
also, the agents are actually pretty good at cleaning up spaghetti if you do it one module at a time, use unit tests. And some of the models are smart enough to suggest good organization schemes!