AI compression is the future
>>110393872why does it take a week, do they read 100mb a trillion times
>>110393872>a weeklol
>>110393882It likely just trains a tiny LLM to the point where it's so fried it can only reproduce the training data 1:1 and nothing else.
Wow, since the AI stocks are crashing, the catalogue have been spammed nonstop with even more shills.Noooooticing.
>>110393882>>110393911GPUs are built to crunch shaders really fast that's it. They suck at general purpose stuff suited to a CPU. That's why AV1 on GPU looks like ass compared to DAV1D.
>"A dozen hybrid compressors can beat its score, but none relies on a neural network alone"So this is just pure AI wank with little practical purpose
>>110393931Your 401k depends on that AI wank
>>110393872Can I extract an arbitrary page and read it though?
>>110393872Reminds me of the guy that "compressed images" reducing them to a prompt.
>>110393872i cant wait for the 5MB 1080p rip of The Odyssey where i wait 3 weeks to decompress before playback
>>110393872futazip?
I have a better algorithm>compression process: delete the data>decompression process: wait for the universe to give rise to the circumstances that will produce an exact copy of the input data
>>110393939yes it will be ready in about a week
>>110393882>do they read 100mb a trillion timesYou jest but that is literally how transformers work. All weights are read for each token generated (mixture of experts changes this but a 100mb model is probably dense).
>>110393872rar stocks are crashingthanks a lot nvidia
>week per direction with 2x 1000W TDP gpustruly magical and efficient
yeah this was proposed in like 2023, before people realized how fucking expensive and slow itd beunfortunately the AI scene is now full of dumb cunts so theyre trying it again
Why don't they just make it into a text file
>>110393922
>>110393872>TuFa zip>FuTa zip
>>110393872wikipedia is mostly text, zip can already do that in a few seconds
>>110393872Compression & communication are outdated technologies. Instead train your computer to already know everything without needing input
>>110393872seems kinda excessive. can't they make a model that's 20x faster with 2-3x worse compression ratio that actually has a practical usecase?
>>110393926b200 is not for shaders but for matmul it doesn't even have display out ports
>>110393872>compress>wikipediahow much shit can you fit in a 5lb bag
>>110394514this is not meant as a viable product for people to use.>"A dozen hybrid compressors can beat its score, but none relies on a neural network alone"It's like trying to get the highest score while restricting yourself to some specific tool in a vidya. Just to see how it goes.
>>110393882the 100mb is the llm, it takes a week bc its bruteforcing the wiki based on its sha
>hey guys check out my new compression NisPe.zip!
>>110393872You laugh, but when shit hits the fan and all you have left is on your person a 128MB flash drive and two Nvidia B200 GPUs and 400kWh of power, you'll be wishing you had just compressed it. At least I'll be prepared. Maybe if you agree to be my concubine I'll lend you a little knowledge.
>>110393872> a week
>>110393986Mambo futa...
>>110393926What? GPU encoders are fixed function units, all the shaders do is scaling and maybe noise removal. There are commercial encoders that do use the GPU cores but they are way slower.