What I learned building an AI coding agent for a year | Hacker News Reader

What I learned building an AI coding agent for a year

jamesgrugett.com

30 pointsvinhnx10 comments

why has nobody solved the code editing in a robust way yet. I think all the implementations so far have been hacky. I also had to write my own basic one a few times[1].

I guess diffusion-based models can prove good for this usecase?

1. https://github.com/asadm/vibemode/blob/main/source/editor.js

> why has nobody solved the code editing in a robust way yet.

Mostly because code editing is not the problem. When coding the solution exists out of the coding space. Code only remove ambiguity. It may conflict with earlier interpretations or the current interpretation is flaky, which leads to bugs, aka actual behavior differs from expected behavior (which also exists out of the coding space).

So trying to solve things within the coding space is an incorrect approach since the beginning of computation. And trying to merge natural languages (great for exploring problems) and formal languages (great for specifying instructions) was seen as foolishness by Dijkstra [0].

The reason natural languages are great for problem solving is that we can redefine what things means easily, changing the semantic of terms as our understanding evolves. And when we've settled on a set of semantics and a process, we translate that to formal notation so it stays fixed. An analogy is sketching (where we freely edit lines and just try stuff) and oil painting where every brush stroke is purposeful.

[0]: https://www.cs.utexas.edu/~EWD/transcriptions/EWD06xx/EWD667...

The most robust way is not to index. Then you can't go wrong, but it is slower. People seem to be ok with it since the ML part takes longer -- for typical codebases, at least.

https://news.ycombinator.com/item?id=44106944

I don't see how indexing is related to this? The question is about how to get the LLM to reliably apply the edit it wants to make. Even when the full current version of the file is in the context, this is one of the flakiest bits of the current LLM workflows.

In my mind the concern was about the LLMs mental model of what the files look like, which affects edits. I see where you're coming from too.

Looks really cool! But I have to able to bring my own LLM. In a space moving this fast by the time my enterprise has finished vetting a tool that wants to be it's own LLM/process my data itself, it's already out of date.

I'd love to be able to connect Azure AI and Vertex to this (for the full range of models it uses)

Why do blog post authors decide to use AI art with nonsense text in it? It's a dead giveaway and looks sloppy, IMO. Unless the new normal is that images with gibberish text are fine now.

I immediately stop reading whatever I'm reading when I see this. I'm left to assume that the post I'm reading was also heavily AI-assisted and isn't a true representation of the author's writing ability or their actual thoughts.

I'd much rather see a post full of grammatical errors and stock art from an author with a story to tell that actually put the work in than a grammatically/syntactically milquetoast piece in the style of a million other milquetoast pieces.

TL;DR: "If they can't be bothered to actually write this, then I can't be bothered to read it," is what I think when I see AI slop art.

I wrote every word! Though I did have some suggestions from Opus haha

Oh no achivement acquired!

jahooma failed the turing test as a human :'( The human text is seen as ai text.

I don't know what to think about this.

Current AI's could never have such deep thoughts haha

Reply on news.ycombinator.com