Hacker News Reader

Comment by jeffhuys | Hacker News Reader

I'm not really sure actually. But to be honest, I'd rather see these tools be public instead of private; you can't really block this kind of thing anyway. Better have it out in the open where others can benefit/learn...

This whole thing is a Pandora's box. We can regulate, forbid, anything, but we all already have models downloaded locally (you did too, right..?). So unless there's some client-side "computer says no" we will never be able to block this anymore.

View on HN

No, I don't have any models downloaded locally. I've tried a couple of models in the past and found they aren't that useful for me

How long ago?

The local models were mostly unusably weak until about six months ago when they suddenly got useful: Qwen Coder 2.5, Llama 3.3 70B, Mistral Small 3 and Gemma 3 have all really impressed me on a 64GB Mac and I expect Mistral Small 3 would work in 32GB.

Meanwhile this years's Gemini Pro 2.5, Claude 3.7 Sonnet and the most recent GPT-4o API models (or o3-mini high for coding) are significantly better than what we were using last year.

I don't know, 5-6 months ago? I used them, found them acceptable at making human sounding text, and haven't really used them since. I've not found a use-case for them that is useful to me. If I don't know how to program something, I'll google it, or just work it out myself

I wrote about my own processes using LLMs for code here (mainly aimed at people who aren't finding LLMs useful for coding yet): https://simonwillison.net/2025/Mar/11/using-llms-for-code/

It feels like a lot of the advice I'm seeing is aimed at people who just want to have something that works. I program because I like programming, not because I want a finished thing now. The stuff I program at work is well within my abilities, and uses libraries I'm pretty familiar with, so I don't need a program to write it for me, and when I'm doing stuff in my free time, its because I want to program, so I don't want it to be done for me. The best use I've found for LLMs is generating place names in DnD

One of the things I've been appreciating most about LLMs is how they accelerate my exploration of other languages.

I'm fluent in Python and JavaScript, but these days I'm using LLMs to help me get started writing code in AppleScript, Bash, Go, jq, ffmpeg (that command-line interface is practically a programming language just on its own) and more. I'm learning a ton along the way - previously I wouldn't have been able to get up the energy (or time) to climb the initial learning curve for all of those.

You have to take Simon with a grain of salt he confuses demonstrations with solutions. I was just looking at some other code he generated using vibe techniques and it's generally not suitable for anything remotely robust and for some reason the models still think markup languages are regular and can be handled with regular expressions, among many other foot guns present in just about everything. But don't worry! It is us sane people who don't get the joke they'll say and that it is "good enough" lol. I have a really hard time with LLM people who still consider telling people who dissent to go fuck themselves when their work is so insulting and ignorant.

"... for some reason the models still think markup languages are regular and can be handled with regular expressions"

Presumably you mean this code here? https://github.com/simonw/llm-hacker-news/blob/e945c84e825f4...

I reviewed it. For this particular application (turning JSON from https://hn.algolia.com/api/v1/items/43615912 into a more concise format suitable for feeding into an LLM like https://github.com/simonw/llm-hacker-news/blob/e945c84e825f4...) it's perfectly adequate (or "good enough" to quote your comment). You're welcome to convince me otherwise!

Reply on news.ycombinator.com