Recent submissions

Comment · 1 month ago
Wasn't it also quantized aggressively, like 1 or 2 bits?
Comment · 1 month ago
Maybe when you want to learn the process for when/if you ever need to fine-tune a larger model?

It's faster to make iterate when you're toying around with a 1B model than a 27B one.

Comment · 1 month ago
Cathartic.

Can I upvote twice?

Comment · 1 month ago
Qwen 3.6 did the same thing for me.

Only after some cajolling it finally went to check the history I asked it to (I had been testing KoboldCPP's web search).

Comment · 2 months ago
You can run a 4bit Qwen3.8 27B on a 8GB GPU, though on my RX 6650 XT it will only reach about 3 tokens/s.

If you can give the GPU units enough RAM to load the whole thing, it may be a little faster.

I'm not sure why would you want to run something at 3-4 tokens/s though. At least try a MoE model. I get about 30 tokens/s from those.

Comment · 5 months ago
When it answers something I asked, or when I play with a small model with llama.cpp, I sometimes find their reasoning funny while they go in circles.
Comment · 5 months ago
Replying here to throwaway2037 because I can't reply directly to him, but yes, even most informal businesses accept PIX, including some random guy selling candy or bottled water at a stop signal.

The only exception I have found to consistently refuse PIX are some parking lots, and they refuse credit cards as well, accepting only cash, probably to hide their earnings.

Comment · 5 months ago
I'm pretty sure they're "employed for life" because otherwise every new administration would replace as many people as possible.

Can you picture a company replacing 90% of their workforce every 4 or 8 years, all at once? Because that's what I think would happen if government employees could be fired as easily.

Comment · 5 months ago
Maybe this will help: https://marcusolang.substack.com/p/im-kenyan-i-dont-write-li...

And I think I saw some others similar articles and blog posts over the years, but the tl;dr; is that a large enough ammount of text used in their training contains em-dashes because a large ammount of writers use them.

Comment · 6 months ago
What do you mean by "It's not automatable at all..."?

At a former job we had automated extraction of data from a 3270 terminal (several of them actually).

Comment · 7 months ago
Is this satire?

The way things have been going, I can't tell.

Comment · 7 months ago
I thought they had been retired long ago.

I think it's been years since I last saw one at Rio de Janeiro.

Comment · 7 months ago
It feels like it won't be long before Microsoft starts helping with that (by making Windows less stable, not improving Wine).
Comment · 11 months ago
Take a look at https://old.reddit.com/r/FPGA/

Whatever people may say about reddit, it seems to be a good starting point to get started with pretty much any hobby (even if it's just to get a link to another community).

Edit: https://github.com/BrunoLevy/learn-fpga looks interesting.

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton