Hacker News Reader

Comment by win311fwg | Hacker News Reader

> Can you describe the last time you asked your compiler how you should best approach an interpersonal problem

Sure. I wrote out specification that described the problem. It compiled that into a description of how to deal with it.

> People aren't just using AI for writing code

Compilers have never been about producing code in particular, though, even if that is where they are most commonly used. A LaTeX compiler, for example, outputs books. To think they are about code exclusively is a grave misunderstanding of what a compiler is.

View on HN

> Compilers have never been about producing code in particular, though, even if that is where they are most commonly used. A LaTeX compiler, for example, outputs books. To think they are about code exclusively is a grave misunderstanding of what a compiler is.

And what, pray tell, are those books output by TeX? That's right, PDF or PostScript code.

> Sure. I wrote out specification that described the problem.

The LLM usage I see from most people has the LLM write out the specification, asking clarifying questions as needed. If you're using LLMs as you describe, you're using them ineffectively.

> The LLM usage I see from most people has the LLM write out the specification

If the compiler is writing the specification, what purpose does the person serve? That is not what I'm seeing out in the wild, though. Perhaps what is happening in research labs aside, compilers, even those implemented as an LLM, don't do anything until the user specifies their intent.

In fact, you literally recognized that in your earlier comment, asking about my action to specify my problem. Not sure why the story is changing now? There is probably a day where the systems solve your problems before you even know you have them, but we're not there yet. The technology hasn't advanced that much in the span of a couple of hours.

> asking clarifying questions as needed.

Sure. Even the earliest, most primitive compilers would seek clarification if there was something that it felt required more information. That's expected.

> If you're using LLMs as you describe, you're using them ineffectively.

LLMs are just compilers so they can't do anything else. Maybe you're thinking of some technology that is layered on top of a compiler like an IDE, REPL, or something to that effect?

> Sure. Even the earliest, most primitive compilers would seek clarification if there was something that it felt required more information. That's expected.

Can you tell me what the last question GCC asked you was, and can you rate its coherence in comparison to your colleagues questions?

> Sure. I wrote out specification that described the problem. It compiled that into a description of how to deal with it.

I might just not know enough compilers, but didn't you write the specification of how to approach the problem, not just the problem itself?

I don't think I knew how to approach the problem. There would have been no benefit in using the compiler if I did. Kind of like, if we want to talk about code, many coders don't know how to approach making a hashmap. They don't need to. The compiler figures it out for them, as it did me for my problems. I'd call the input I gave a specification of the problem, but words are in the eye of the beholder, I suppose.

If we take "write the specification of how to approach the problem" at literal value, you might be narrowly thinking of a low-level compiler, or what sometimes gets called a transpiler? However, compilers encompass a much wider spectrum than that. Generally they enable users to operate at higher levels of abstraction.

I think you're talking about more than just a compiler, though standard libraries are often distributed with the compiler. Take Rust for example, it's not like the compiler has special integration with hash maps, it's just that the standard library has a standard implementation and the compiler lets you refer to that implementation without having to repeat it yourself. In a way you're just including parts of another spec in yours. Specifications do this all the time, for example by using English. You don't have to define all the levels of abstraction, only the ones that matter.

Even so, I'd argue that you still have to define how to use those abstractions rather than only the result you'd like to achieve with them. Sure, maybe the abstraction has a ".achieve_the_result()" method, and calling that method is how you achieve it, but often there's a bunch of stuff you fit together to achieve a solution to your problem and the compiler doesn't necessarily do that for you.

I think there actually may be some compilers that get closer to this, for example some problems can be expressed quite elegantly in Haskell or APL and the compiler will do a lot of black magic behind the scenes to get you a performant executable (somehow). But for most traditional compilers you are still going to have to define how to approach your problem, even if only at a higher level thanks to libraries.

Reply on news.ycombinator.com