lairv

21 ms·

121.
by lairv 3y ago
This is a new project from Jean-Baptiste Kempf, main contributor of VLC It sounds pretty insane, basically real time latency streaming for audio, video and inputs, where both the client and the server can run anywhere, on desktop, on mobile
122.
Kyber: A new approach for real-time video and controls streaming based on QUIC(youtube.com)
2 points by lairv 3y ago|2 comments
123.
by lairv 3y ago
Has he done any foundational science either ?
124.
by lairv 3y ago
If you ever feel like improving your matmul kernel you can check this blog https://siboehm.com/articles/22/CUDA-MMM
125.
Proof You Can Train an AI Model Without Slurping Copyrighted Content(wired.com)
4 points by lairv 3y ago|0 comments
126.
by lairv 3y ago
I would say it emphasises that training a good model is more than throwing random data and compute
127.
by lairv 3y ago
So how is it that tree-search algorithms like AlphaZero or MuZero don't drift ? I understand that the searching part allows to reduce the epsilon errors, but with tree search of bounded depth, drifting should still occur right ?
128.
by lairv 3y ago
Humans can very quickly pick up a new game and have strong priors on how to move, how to interact with the world etc. so its not unreasonable to expect an AI to be able to do the same
129.
by lairv 3y ago
I agree but a major selling point of React is its ecosystem, and libraries are not compatible with Preact by default
130.
by lairv 3y ago
Pretty exciting indeed to see they used flow matching, which have been unpopular for the last few years
131.
by lairv 3y ago
You can literally run 3D algorithms like NeRF or COLMAP on those videos (check the tweet I sent), it's not my opinion, those videos are sufficiently 3D consistent that you can extract 3D geometry from them Surely it's not perfect,
132.
by lairv 3y ago
I find it wild that this model does not have explicit 3D prior, yet learns to generate videos with such 3D consistency, you can directly train a 3D representation (NeRF-like) from those videos: https://twitter.com/BenMildenh
133.
by lairv 3y ago
The 3D consistency of those videos is insane compared to what has previously been done, they must have used some form of 3D regularization with depth or flow I think
134.
by lairv 3y ago
The name of the blog post is "Our next-generation model: Gemini 1.5", how am I supposed to infer from this that it is only the 1.5 pro and not ultra ?
135.
by lairv 3y ago
This naming is terrible, if I understand correctly this is the release of Gemini 1.5 Pro, but not Gemini 1.5 Ultra right ?
136.
by lairv 3y ago
You still need to get an invite / a waitlist to register or am I missing something ? I don't know how you can expect people to make the move when they're not even able to sign in
137.
by lairv 3y ago
Had to shorten the title, the original statement is: "#AlphaCode 2 is powered by Gemini and performs better than 85% of competition participants in 12 contests on Codeforces!" Technical report: https://storage.googleapi
138.
AlphaCode2 powered by Gemini performs better than 85% of programmers(twitter.com)
3 points by lairv 3y ago|3 comments
139.
by lairv 3y ago
Why would they do that, have you seen the numbers their tweets made ? Everyone was already looking for the trailer tomorrow I don't think it's beneficial
140.
by lairv 3y ago
They acted really quickly after the leak, I guess these studios already have plans in case of leaks
141.
A Guide to Cold Emailing (In Research)(eugenevinitsky.com)
1 points by lairv 3y ago|0 comments
142.
by lairv 3y ago
I think the bottleneck is data For single 3D object the biggest dataset is ObjaverseXL with 10M samples For full 3D scenes you could at best get ~1000 scenes with datasets like ScanNet I guess Text2Image models are trained on datasets with
143.
by lairv 3y ago
There was a really good video from a french youtuber about this 9 months ago: https://youtu.be/JcFRbecX6bk?feature=shared In French, we have a word that I find more accurate, which is "baratineur", which means &qu
144.
by lairv 3y ago
I think it's fair to want to keep an evaluation private so that it doesn't become part of a train set, but you should know that OpenAI uses users chat data to improve their models (not for entreprise)
145.
Advice for Prospective PhD Students(gonzales.science)
58 points by lairv 3y ago|92 comments
146.
by lairv 3y ago
How do you chose between ClickHouse and DuckDB ? It feels like they solve the same problem
147.
by lairv 3y ago
The "insult to life itself" indeed refers to the particular demo that they showed this day, but they also mention multiple times they want to make "computers that paint like humans" and HM doesn't show much more ent
148.
by lairv 3y ago
https://youtu.be/7EvnKYOuvWo?si=KjIPP5LtZvab_Ls5 Well he didn't seem very fond of it even before it was created
149.
by lairv 3y ago
The main pain point I still have with GPT4 is that it's hard to prevent it from being a yes-man "- I have problem A, what do you think of X ? - Yes X is great ... - But with X I'd encounter problem B, maybe I should go with Y
150.
Remix v2(remix.run)
9 points by lairv 3y ago|2 comments

More ›