atorodius

9 ms·

61.
by atorodius 3y ago
Yes, next token prediction is exactly what you need for compression. Many papers on this, recent one https://arxiv.org/pdf/2309.10668.pdf
62.
by atorodius 3y ago
I clicked on the YouTube wikipedia article. TIL that that is the 2nd most viewed website in the world! Crazy
63.
by atorodius 3y ago
Same on iOS
64.
by atorodius 3y ago
Do you have example queries where Kagi shines? Just curious
65.
by atorodius 3y ago
Seen this on Twitter, very cool! > We present Marigold, a diffusion model and associated fine-tuning protocol for monocular depth estimation. Its core principle is to *leverage the rich visual knowledge stored in modern generative image
66.
by atorodius 3y ago
Can someone familir with the scene comment on how this compares eith Crossover?
67.
by atorodius 3y ago
Chips take a lot of time from design to roll out. Giant LLMs being something you want to serve is a very recent development and it will take time for the deployed hardware to catch up, no surprises here.
68.
by atorodius 3y ago
A true, I did the same and cancelled my spotify
69.
by atorodius 3y ago
YouTube Premium everyone… If you get the family version, 6ppl can watch YT without Ads from the main apps, imho a great deal (I don‘t have any other video subscriptions)
70.
by atorodius 3y ago
FWIW, my human brain did not understand your first query, in the second you conflate voice search with Google, and re the last, if I try it, it shows me a dropdown to select what I meant… (Generally just confused by how ppl always claim it’
71.
by atorodius 3y ago
182k employees, still! Impressive.
72.
by atorodius 3y ago
That's normal for ML
73.
by atorodius 3y ago
This is always my biggest gripe with these comparisons. But 14 is not enough, please show the 11 and 12.
74.
by atorodius 3y ago
Agreed with the overall point (small models will prove powerful) but I disagree with: > But having more fine-tuned control over model parameters is a pattern that will emerge This assumes parameters do something interpretable. IMO, it’s
75.
by atorodius 3y ago
> Question: why isn’t Anthropic using Google Cloud, given who their past investors include? They claimed to use it back in February: https://www.anthropic.com/index/anthropic-partners-with-goog... „Anthropic, an AI
76.
Self Organizing Textures (2021)(distill.pub)
3 points by atorodius 3y ago|0 comments
77.
iOS Human Interface Guidelines (2011) [pdf](tableless.github.io)
2 points by atorodius 3y ago|0 comments
78.
by atorodius 3y ago
Yeah I was thinking the same. Jax does this, but I did not see a comparison.
79.
by atorodius 3y ago
Being able to condition on a video vs. just text massively simplifies the task. Mainly because you get consistent camera motion and movement in the scene for free!
80.
by atorodius 3y ago
I learned to program in Objective C, it was fun. Spent many hours digging behind the scenes, but the only fun fact I remember is that it’s a strict superset of C, so any C program is a valid Objective C program. Always found this fascinatin
81.
by atorodius 3y ago
While I would question this statement in general (people definitely can still write formal and poetical content without mocking, eg recently eulogy here on HN even), but on the other hand it also seems this piece itself is mocking to some e
82.
by atorodius 3y ago
Hot take: we also just memorize icons, they just happen to me made out of 26 symbols. But to read fast, you absolutely do use some high level icon based parsing, IMO
83.
by atorodius 3y ago
Not sure if it counts but my Eurorack modular synth is definitly my most loved phyiscal thing and it’s kind of built (from modules). Still habe to get into soldering my own modules..
84.
by atorodius 3y ago
You don’t need these. I‘ve been viming for 10+ y and just using the normal shortcuts and ways of manipulating text is a major productivity gain.
85.
by atorodius 3y ago
fair, but can’t you just randomize the grayscale generation for training?
86.
by atorodius 3y ago
good point, I hadn’t realized that you only need to predict chroma! That actully greatly simplifies things re. chroma subsampling in training data: this is actually a big problem and a good generative model will absolutely learn to predict
87.
by atorodius 3y ago
This is true, but if you have some reference images, you can probably adapt some of the recent diffusion adaptation work such as DreamBooth, to tell the model „hey this period looked like this“, and finetune it. https://dreamboot
88.
by atorodius 3y ago
You can take arbitrary images and convert them to grayscale for training, and do conditional diffusion
89.
by atorodius 3y ago
yeah, you can use SOTA super res, but that tends to be generative too (even diffusion based on its own, or more commonly based on GANs). it can be a challenge to synthesize the right high res details. but that’s basically the stable diffusi
90.
by atorodius 3y ago
I meant more from a comoute standpoint, the models are expensive to run full res

More ›