Why not have the LLM go straight to LLVM IR? What would a program look like when you remove all (or most) of the layers of abstraction needed by humans? Or are LLMs too contaminated by the training data to do this? I almost wish I could try this.
You probably could but the thing is that low level languages do not encode intent as well as high level constructs. You can trivially make an iterator in assembly but in a high level language a `for`, `map`, and `reduce` have specific meaning that helps catch bugs.
As I commented in the other post, it killed mine at work, because my boss is pushing "AI" really hard on the devs. Fortunately, he's now seeing enough evidence to counteract the hype, but it's still going to be present and dragging down my work. But it my off time, I only experiment with LLMs to see if they're getting better. Spoiler alert: they aren't, at least not for the kind of things I want to do.
I have an experimental project where I was asking various LLMs/tools (ChatGPT, Cursor, Google, Lovable) to implement an old game for me. They all failed spectacularly in various ways. For example, when trying to debug an issue, got into a loop making the same sets of mistakes over and over again. Or "solving" a problem by removing an implementation, or claiming something was fixed but all it did was stop checking the error. It's been disastrous.
I've had better success with LLMs as just a supercharged search engine, but only after I went through several rounds of adding instructions to prevent hallucinations and lies.
I also asked one to create a tutorial for me to follow in regards to a complicated game I'm trying to understand. It lied repeatedly, making up features and telling me to set options that just didn't exist.
My boss loves LLMs and claims it really improved his productivity, but the stuff he's talking about is JS stuff. When he (and I as well) try to use it with Java the viability of the results drops off dramatically.
I'm 62, and it's had the opposite effect on me. I've never stopped loving writing code, learning new things, trying random stuff, etc. I code all day, and spend more time playing with stuff in the evenings (the main difference is I'm sipping some scotch while I do it). Having to use LLM's at work has sucked most of the joy out of my work. Fighting with them, keeping them on track, catching hallucinations before they go too far, wasted effort...it's exhausting me like nothing else in my 40+ year career.
The worst excesses of "modern" web presentation, coupled with a complete lack of actual gardening info...I'm completely baffled. 1% "here's your zone", and 99% "your zone is almost no use for gardening"
