cubefox

36 ms·

811.
by cubefox 8mo ago
What about JPEG XL or AVIF? Do they use DCT or DWT, or perhaps something else?
812.
by cubefox 8mo ago
That association seems to be an unfortunate equilibrium because higher frame rates seem to be "objectively" better, similar to higher resolution and color. (Someone without prior experience with TV/movies would presumably alw
813.
by cubefox 8mo ago
Any idea why movies are still mostly at 24 FPS? Is it just because people became used to it?
814.
by cubefox 8mo ago
It's never unethical to charge for something that has no value.
815.
by cubefox 8mo ago
VRChat also seems to be used only by niche perverts. Perhaps only niche perverts are currently interested in the Metaverse, and Horizon Worlds failed because it didn't cater to niche perverts.
816.
by cubefox 8mo ago
In Snow Crash, the Metaverse was using VR. But it could also have worked via a smartphone. They are independent.
817.
by cubefox 8mo ago
In the original the man has a large sun shadow (from an off screen object) on the right side of the face. This is gone with DLSS 5, making the face no longer look lit mainly from the left. Only self-shadowing remains. It seems DLSS 5 is wor
818.
by cubefox 8mo ago
Seems unethical.
819.
by cubefox 8mo ago
It has a tendency to brighten the shadows and the scene as a whole. Particularly noticable on Assassin's Creed Shadows (ironically, given the name). In that game it also removes a lot of the fog. I think that's more an issue with
820.
by cubefox 8mo ago
The other comments give it away.
821.
by cubefox 8mo ago
LLM account
822.
by cubefox 8mo ago
The most interesting part: > North Americans are going to the movies about half as often as they used to a decade ago, based on the number of tickets sold at cinemas in the US and Canada. 50% down in just 10 years is massive.
823.
by cubefox 8mo ago
The latest DeepSeek model has sparse attention. Though sparse attention is still not linear. Close enough perhaps.
824.
by cubefox 8mo ago
(This article was clearly written with LLM assistance. Is this still worth pointing out? Or should we just accept it at this point?)
825.
by cubefox 8mo ago
Object oriented programming languages also use object.method rather than method(object), so I don't think prefix/suffix notation has much to do with language.
826.
by cubefox 8mo ago
It's may also be reasonable to make localized translations for a programming language. This is rarely done in reality for obvious reasons. An exception are Excel's function names. People who don't know English, or hardly know
827.
by cubefox 8mo ago
So that would mean it's not "really" a 1M context window. I guess this is more plausible than a linear architecture like MAMBA or GDN.
828.
by cubefox 8mo ago
(I keep noticing this, more and more websites are including unnecessarily huge images on top. This one has a 24 MP (6000×4000) header. At least it's a JPEG with "just" 5.83 MB, not a PNG.)
829.
by cubefox 8mo ago
> Standard pricing now applies across the full 1M window for both models, with no long-context premium. Does that mean it's likely not a Transformer with quadratic attention, but some other kind of architecture, with linear time com
830.
by cubefox 8mo ago
The Unhook browser extension gets rid of that. And optionally other things.
831.
by cubefox 8mo ago
The thing with efficiency is that it is relative to both inference and training compute. If you do quantization, you need a more powerful higher precision model to quantize from, which doesn't exist if you want to create a frontier mod
832.
by cubefox 8mo ago
Multiple people agree with me here. The account used em dashes almost everywhere and was rapidly posting complex comments (while having clearly read the articles) one or two minutes apart. There are also other subtle LLM-isms, like replying
833.
by cubefox 8mo ago
Meanwhile, the top comment on one of the most upvoted submissions today is AI generated by an LLM account: https://news.ycombinator.com/item?id=47334694 Most people don't seem to care.
834.
by cubefox 8mo ago
2026 A.D., still no support for native date pickers in mobile Safari.
835.
by cubefox 8mo ago
I assume that theoretically, 1-bit models could be most efficient because modern models switched from 32 bit to 16 bit to 8 bit per parameter (without quantization).
836.
by cubefox 8mo ago
I think this approach is not so interesting because it's just quantization of a full precision model. So it speeds up inference (at a quality penalty) but not training. It would be more interesting to train an actually binary model dir
837.
by cubefox 8mo ago
Yeah. It correctly pointed out that the editorialized HN title is wrong, there is no 100B model.
838.
by cubefox 8mo ago
LLM account
839.
by cubefox 8mo ago
Yeah, "1.58 bit" is 1 trit with three states, since log2(3)≈1.58. So it's not a inference framework for 1-bit models (two states per parameter) but for 1.58 bit models (three states per parameter). Annoying that they try to m
840.
by cubefox 8mo ago
Note that the description says "Azimut, Richtungswinkel". Those seem to be somewhat different concepts today. The respective Wikipedia articles don't even mention each other: https://de.wikipedia.org/wiki&#x2F

More ›