Hacker News Reader

Comment by yNeolh | Hacker News Reader

My take on LLMs is that it won't scale much of what we already see because this is "just" text prediction on steroids. I'm not an expert by any stretch of the imagination, but that's my opinion, and going through that path, no, I don't see this path as the best path for "autonomous development machines", only powerful autocompletion like we already see today.

View on HN

I agree. I've been using Copilot for several months now, and the only thing it (almost) consistently helps me with is predicting relatively trivial snippets.

Anecdotally, I've had it mispredict from very simple contexts, such as skipping numbers in series' where the pattern should've been extremely obvious.

I've had it sneak in sublte and obvious bugs on a regular basis, to an extent where I don't have much confidence beyond any code I can grasp in a single look and be confident it's correct. Sorry bros, I'm not on the hype train this time. Feels like crypto all over again.

I switched to Supermaven and in my experience it’s at least 2x Copilot. Might be worth a try.

There are many way how current LLM can be scaled in different dimensions and there are research around it e.g.:

1) Many different AI with different role: business analysis, tester, developer. You as developer are treated as customer and write simple prompt but business analysis AI will make a proper step by step prompt to Developer AI - so that you don't have to very good with prompt engineering

2) bigger context for LLM so you can feed up to date documentation and full repo

3) LLM having access to do RAG on web search to get up to date information

4) LLM having access to terminal and debugger so Tester/Developer AI can automatically see the flow how code is executed and variables states during execution

5) faster and cheaper LLM so that you give a task before you go to sleep and all those AI in a loop try to solve this task trying many different options until passed all tests.

Yes, but that doesn't change how they fundamentally work. You can replicate things fast, with many variations, almost like brute force but more "thoughtful." Don't get me wrong, this opens many possibilities, and I am even thinking about having a local AI machine for my stuff. However, working with multiple layers of knowledge connections is where I don't see LLMs arriving. Maybe some other technology is based on this, but following LLM evolution will be better data and "patches".

Reply on news.ycombinator.com