Hacker News Reader

Comment by locusofself | Hacker News Reader

Every day for the last few months I've been wishing I had a cabin in the woods with zero modern technology. A big shelf of books, my acoustic guitars and some comfy places to sleep.

View on HN

My internet died earlier this week. I put on a vinyl record (Paul Desmond) and sat on the couch enjoying the brief break from my digital life. Hoping to enjoy more simple times soon.

I’m literally planning to do this in a few days to get some solid focus time in. The thought of the disadvantage of not having an AI model to support already has crossed my mind (no internet access in the woods)

It’s noted that this editor supports offline - does this mean that the AI features also run offline? Or a limited version?

Unfortunately, everything but the AI works offline. Though, maybe that's a feature if you're planning a more mellow retreat :)

Have you considered a limited LLM that could run locally?

> planning a more mellow retreat

The objective here is to forcefully going to where internet is impossible (no phone reception, I don’t have starlink) with the objective of focused productive output with limited distractions.

The idea came to mind after reading about John Carmack doing this for a week, diving into AI using nothing but classic text books and papers as reference material to work off.

EDIT: here is the HN thread on Carmack’s week long retreat:

https://news.ycombinator.com/item?id=16518726

> Have you considered a limited LLM that could run locally?

I think there are two main issues here. LLM are large (the name even hints at it ;) ) and the smaller ones (still, multiple GB) are really, really bad.

Edit: and uses a ton of memory, either RAM if CPU or VRAM if GPU.

Are they all that bad? [1]. I’d be ok using a few 100gb on my laptop, given that storage is so cheap these days.

[1] https://github.com/nat/openplayground

Compared to GPT-4, most of them are not super great, yeah. I've tested out most of the ones released for the last weeks and nothing have been getting the same quality of results, even the medium sized (30GB and up) models that require >24GB of VRAM to run GPU. I have yet to acquire hardware to run the absolute biggest of models, but I haven't seen any reports that they are much better either for general workloads.

gotta love the fact that this is the top comment here

Winter is coming...

On a serious note, I hope AI will help me write documentation. But at that point, do we still need documentation?

If you use AI to write the documentation, then it's unlikely to be better than the documentation people could generate themselves with GPT-4. But it will be a while before it's better than documentation written by professional writers that have all the context, target audience in mind and know what they are documentation inside and out.

The big advantage of doing it ahead of time is that you can proofread and edit it. LLMs are good, but it'll be a while before they get to a point where a bit of human editing doesn't help.

Reply on news.ycombinator.com