[seanmcdirmid]

I don’t think we’ve really figured out how to use AI in coding yet, vibe coding doesn’t really feel like it’s it. Vibe coding and just generating code like how some people claim intellisense is just to save on typing, when it’s actually a great in-situ browse what members can be selected on a value of a certain type.

There is definitely a way to abuse AI in programming, but it doesn’t seem to be very compelling and I don’t think it will get people who do that very far (eg relying on intel sense to save on typing rather than just learning how to type).

ChatGPT is a great writing tool if you already know how to write. You can curate and modify on top of it, allowing you to write your paper faster with the same amount of quality. But again people just using it to write essays or paper without knowing how to write themselves aren’t going to get good results.

[pona-a]

I understand what you mean, but let's be honest: this is a rare kind of tool that's more useful to feign competence, deceive yourself and others, and produce industrial volumes of slop than it is to do better work. IntelliSense is just dynamic documentation, which has existed since at least Emacs — it doesn't do the thinking for you.

Professional tools, from music notation and art to typesetting and programming, are about translating an image inside your mind into something physical. When you know what you're doing, the lack of an interpretable mapping between prompt and generation means you spend more time trying to describe what you want to write instead of just writing it. I'd be much happier with code generation if it could take a formal specification and either return an error or something that provably implements it. Maybe interpretability research will one day change that, but as they are now, they're simply not tunable or reliable enough to be used as tools. And yes, prompting doesn't count when they increasingly disregard your instructions.

There are many valid uses: I have a tiny WolframAlpha-like script that lets me type some basic computations and the LLM translates that to Python. I sometimes use LLM completions to get some inspiration when writing prose — while I usually discard them, they still help me think. They can often act as better grammar checkers than LanguageTool, and they make a nice companion to smaller translation models, both having their own quirks.

But most of this doesn't need these larger and larger models; I haven't yet tried, but I think fine-tuning some mid-size open-weights LLMs will yield similar or better results. The industry sold the public AGI, not better auto-complete, a fuzzy parser, or a smarter translator, and now they're burning growing piles of money on a saturated research direction to maintain the delusion singularity is 5 months away.