Sharing my method to compress 16-bit video in real-time using almost no CPU or GPU time:
Follow
catid (parody)
@MrCatid
Engineer at Anduril Industries. Prior: Oculus VR, Game Closure, MSEE@GATech
catid (parody)’s Tweets
K-Planes: Explicit Radiance Fields in Space, Time, and Appearance
abs: arxiv.org/abs/2301.10241
project page: sarafridov.github.io/K-Planes/
32
174
If you didn’t buy two vibrating toothbrushes to clean your teeth twice as fast, what’s wrong? Do you not value your time?
2
1
Idea for next rev of Voyager Probe: Instead of a golden disc, attach a GPU cluster that can hold most human knowledge and answer questions using a LLM
1
2
If you could live on forever as a brain in a jar with only Neuralink input and output would you do it?
13
1
12
It's mindboggling that many years of research, engineering, community engagement, and meticulous benchmarking all distill into a single, elegant, plug-and-play function:
torch.compile()
You rarely get to see so much behind the scene in a single blog.
4
170
1,235
Show this thread
Hol up you mean his name wasn’t spelled “Tolkien”?
Oh my god all that was under the couch
Show this thread
Moving day!
2
1
Show this thread
Would love to see this in a VR headset to improve silent cooling froresystems.com
1
2
Well that GPGPU trade blockade lasted all of one month… hpcwire.com/2022/08/22/chi by the time Nvidia stops delivering current orders this startup will be shipping. Progress trumps politics
1
We're releasing our code for "driving any robot", so you can also try driving your robot using the general navigation model (GNM): github.com/PrieureDeSion/
Code goes with the GNM paper: sites.google.com/view/drive-any
Should work for locobot, hopefully convenient to hook up to any robot
GIF
14
186
879
Thanks to VC firms like Lux Capital, defense tech is at the leading edge of American Dynamism.
Made this Anduril montage, using MidJourney AI Outputs. One of the most exciting American companies
4
17
Petals model is giving me some sass:
1
1
My petals instances show up in the new health dashboard:
1
Allow users to upload their own scans and get a good default translation from the model in any language
Show this thread
sizing, etc. Upvote/downvote translation patches using github as a backend. All browser rendered so there's just one base graphic that is modified based on the patches and you can toggle off things you don't like or look at the original page.
1
Show this thread
Idea for a side project: BLOOM-176B powered manga translation website with collaborative feedback and correction. Original scans, Japanese text extracted automatically with human in the loop. Model translates to all other languages with proposed text replacement locations
1
Show this thread
I'm not <procrastinating>. I'm <waiting for GPT-4 to launch and write this for me>.
2
1
Used NeRF to make a "Bullet Time" effect for a friend's wedding.
We set up 15 iPhones to capture slow-motion video, then used Instant-NGP to train a bunch of NeRFs on the frames.
..need to work on improving the quality / resolution a bit more.
43
472
2,794
Seems like you could run BLOOM-176B at about 1 token per second using 5 CPU servers with 64 GB RAM each. Kind of interested in trying it on digitalocean or something
My advanced Prompt Engineering techniques have yielded a stunning result: Actual code!
1
1
Show this thread
BTW I am running 6 GPU instances on the Petals swarm. It's a little janky but seems to not crash most of the time
1
Show this thread
Petals provides the laziest AI chat responses:
1
3
Show this thread
"Song of Saya" VN reminds me a lot of "A Terrible Beauty" by Nancy Baker, rare type of story
Finally beat BROTATO as Lucky with the upgrade that grabs greens from everywhere on the map. It's nuts because every time you grab a green it damages another enemy - usually killing them, releasing more greens. So you get this chain reaction of map-wide death and 800+ credits
Red Letter Media has a cameo performance in High on Life, if you purchase the theater disc and warp it in, their commentary track plays over the whole movie
Someone please sell these as wearable displays: zdnet.com/article/this-n There is so much value being wasted on using the best wearable display technology for AR when it could be huge improvement for computing on laptops/cellphones.
3
Wasm2c seems good for reverse engineering but LLVM already provides an excellent IL for compilation? People must be joking using it for anything else
1
1
Nice Meta agreed that TensoRF is the best result to build on: hyperreel.github.io
