zargon

Sarcasm?

What is the word for this type of troll that feigns lack of understanding?

"Both" requires allocating limited die capacity to products that have over an order of magnitude less margin.

An LLM can be useful to help you figure out stuff like that.

These days, most things that I run that aren't from my distro's repos get their own bubblewrap. On top of this, I use opensnitch. Even if I trust the application (uncommon), I never trust npm, pypi, etc. any longer. It's tedious to set this up and OSes should be helping make this easy.

Never make the mistake of anthropomorphizing David Ellison.

I promise I searched on my phone before posting that, but it looks like I misinterpreted an hf search result that was actually Qwen-Image-2512 but the url was truncated to Qwen-Image-2 due to my screen width (ugh). For some reason I was thinking Qwen Image 2.1 was the one that had been available for a while but not released, but you're right that was 2.0.

It would make more sense to compare with Qwen Image 2, since that was the last open weights Qwen model.

Edit: This is wrong.

Memory allocated for HBM is memory taken away from DDR production.

"Your Brain Is Not an Onion With a Tiny Reptile Inside" is a really stupid and pretentious article arguing against a straw man.

I have a tendency to use documentation as a quality bar. Byre might be great but I can't continue reading this slop.

Fun for about 5 minutes but screeches to a halt when ships start to teleport. Only takes a few minutes to be able to get to wave 12 and then there's no progress after that.

I didn't think to check for mods. Thanks.

This perspective that consciousness cannot be programmed can only make sense if you're a dualist. We don't know how consciousness arises. If you're a naturalist it can't be ruled out based on the simplicity of the algorithm.

Unpopular opinion, but for me, no amount of puzzles/story can offset the utter obnoxious frustration of the time loop.

> I've been using AI increasingly more and more

> genuinely

We can tell.

The Glove80 looks close to what I want, but that thumb cluster looks basically as bad as the Ergodox one. Too many keys overall, especially pinky ones.

AGPLv3 is awesome, great choice!

Aris Mpitziopoulos at cybenetics.com

I'm not sure whether this comment is serious or not. Poe's law strikes again.

> Do not give yourself diabetes by eating vast amounts of sugar. If you are doing this, stop. Not having diabetes decreases CVD risk.

Not eating vast quantities of sugar is good advice, but it's not the direct cause of T2 diabetes. The best evidence suggests that T2 diabetes is caused by the accumulation of fat in the liver and pancreas. See the twin cycle hypothesis. To prevent diabetes, one needs to maintain a weight low enough where the body isn't storing fat in the liver and pancreas (everyone has their own individual threshold for this). If you're pre-diabetic, lose enough weight and most people will regain insulin sensitivity.

Yes, llama.cpp does that.

KL divergence is nothing close to a replacement for benchmarks. As flawed as benchmarks are, KL divergence is a barely useful signal. The fact that Unsloth only just started publishing KL divergences shows how unserious the quantization space is.

As a replacement for ollama, LM Studio is probably the most appropriate. Though I haven't tried Unsloth Studio yet. You get more control using llama.cpp directly, but there's more of a learning curve.

vLLM has the best performance if you can fit your entire model into VRAM. llama.cpp is the usual go-to if you're partially loading into RAM. LM Studio is a sensible front-end to llama.cpp. ik_llama.cpp has more advanced CPU quantization strategies than llama.cpp. If you're running super large models mostly from RAM, ktransformers can sometimes be the highest performer, if it works for your model.

> ollama's model index typically only distinguishes variants by their size

Don't use ollama. The entire project is just a series of stupid decisions like this.

I have hearing loss and movie theaters are deafening.

V4 Flash fits entirely in two RTX Pro 6000s without any quantization at all.

> The point of local models is privacy, offline use, and maybe no guard rails.

But also, control and consistency. A local model cannot be changed out under your feet like an API model can be.

> emerging practice of using Q4 quants and Q8 KV cache for local inference

That's not an emerging practice, it's a tested strategy that is these days only used as a last resort by those desperate to fit a model in memory. Some models do better than others, but generally the model quality suffers greatly under those conditions.