Ask HN: Are tab completion models dead? | Hacker News Reader

Ask HN: Are tab completion models dead?

22 pointssandruso17 comments

I think there is still place for this type of coding where you autocomplete predictable chunks and review it as you go.

But I haven't seen much development in this space. Are there any notable models optimized for this task or full/semi-autonomous coding has taken over?

View on HN

I'm craving something similar, but a smidge more than tab-completion. I'd like to treat the agent as an ephemeral pair programmer. Give the editor some keybind to mark and comment places in the code, and then I edit a prompt telling it what focused changes to make where I've marked. After a round of changes is good and complete, the next batch starts with a fresh context and no memory of the previous one.

Aider in watching mode is almost this, but I'd have to insert and save actual "AI:" comments for it to replace. It wouldn't surprise me if VS Code already has something like I'm thinking, but VS Code is painful to me in so many other ways I refuse to use it.

I should probably just vibe code a vim or emacs extension for exactly what I want, but I haven't yet.

Ctrl+I kind of does this: https://code.visualstudio.com/docs/chat/inline-chat

I disabled tab completion. When I open a file in an IDE, it is because I don’t want an AI edit; I’m fixing the confusing documentation, or changing a configuration, or making a small edit to the code. I do code changes through the agent, because it is more accurate and thorough.

I don’t think that I would use tab completions again. I bet that there will be some innovation in how code is visualized, though.

Much like other uses of LLM, it never worked right, and everyone went back to the original, working, way of doing it.

You do not, and never did, need AI for tab complete.

I don't know, I found AI autocomple pretty useful, e.g. if I'm coding manually, but still don't want to type tedious, repetitive log messages or other boilerplate myself.

The two Issues I saw were:

- If it says "tab completion", than that's what it should do, and not more. I eventually disabled the copilot model, not because it wasn't useful, but because VSCode's integration was overzealous and kept on suggesting all kinds of things even if I never asked for it, pulling me out of focus every time. Unlike any of their other completions, btw...

- It was not integrated with the existing completions at all, often creating chaos: The standard completions show you available classes, methods, parameters etc, in your current context, based on the information from the language server. The AI completion happily ignored the language server and showed you suggestions for nonexistent fantasy methods instead.

In general, I think acceptance of AI could be better if we offered more "lower-case AI" options, that let people offload the tedious parts of their workflow instead of taking away the entire process of creating.

As an example, there are lots of image generation models that take a short text prompt and will give you a complete finished work, making thousands of creative decisions in the process, according to their training data. Those models do not have artists as their target group - they are basically targeted at everyone but artists and replace their role. Hence the predictable backlash.

But where are the tiny, locally runnable models that, say, turn a sketch into a lineart, or apply coloring and shading according to given reference images?

I think the OG github copilot version felt magical in that time. But I've never tried working like that with current frontier models. Considering how they train it now it might never work without changes how model respond.

Personally I haven't opened an IDE in 6 months.

Do you check AI generated code directly in the PR page then?

Disclaimer: I wouldn't do this for a production application I charge for.

In about three days I can go from an idea to a fully polished internal tool, implemented in approximately 150,000 lines of code.

I've given up on line-by-line reviews in those projects.

Reviewing 150k loc line-by-line is not practical. I think we need a new set of tools to facilitate such tasks.

Otherwise, we will be spending most of our time monitoring AI's code writing process.

This worries me the most. Will all the tools we used before AI become obsolete and disappear. Is Ruby heading for the graveyard because LLM’s seem to prefer Python. If we aren’t careful we could end up with a lot less choice of tools.

Yeah, AI has really taken the fun out of niche languages for me. Not that Ruby is all that niche, but just learning different programming languages feels a bit pointless now since those are abstractions we built for people, not for bots. Sure, nothing is stopping me from writing code by hand for hobby project, but a big reason why I enjoyed learning new languages so much was precisely because of the productivity boost I might have gotten of that new language. It felt good to become an expert in domains few even knew about. Now, it doesn't matter if your app is written in Python, Ruby, Elixir, Rust, whatever, the AI can achieve it all the same.

The point is that writing code using tabs is only useful when the change is specific; otherwise, you may end up spending more time reading the suggestions than actually making the change. It was an early approach to automated programming that was phased out with the introduction of direct agent calling. It’s unlikely that this approach will be revived, as it almost always results in a negative balance between efficiency and effectiveness.

I still want it for small edits. When I'm renaming a field or filling in a match arm, opening an agent is slower than hitting tab. I'd use agents for any changes that require entire codebase exploration and changing multiple files, and perfer completion for the 2 line change to my portfolio website update for example.

I still want it for small edits. When I'm renaming a field or filling in a match arm, opening an agent is slower than hitting tab. I use agents for anything that required codebase exploration and changes across multiple files , and prefer completion for changing 2 lines in my portfolio website for example.

Zed still actively maintains their Zeta model

https://zed.dev/blog/zeta2

macOS ships an optional autocomplete model with Xcode, but I have no idea if it's actively worked on in any way

Reply on news.ycombinator.com