ctmbottomtext retweeted
There's so many "insignificant changes" that it gets to the point where it feels less like "nitpicking" and more like this:
ctmbottomtext retweeted
en el uber empezó a sonar LA VECINITA TIENE ANTOJO y me tuve que aguantar la risa pq me acordé de este video
ctmbottomtext retweeted
RAM STICKS SHOULD BE USED EXCLUSIVELY FOR PLAYING VIDEOGAMES AND RENDERING BLENDER OVERWATCH PORN
I Agree with Sam Altman.
First of all, local models are useless, they're dumb and slow
You'll spend ~$1000 on 128 GB RAM to run a local model
And you won't get anything done with it, dumb models at 8tps are useless
A $20 subscription will get more things done
And for 95% of the time, RAM is not being used so it's doing less work there compared to in a data centre where utility is more
Readers added context they thought people might want to know
Local models on suitable 128GB hardware often run at 30+ tokens/second and offer practical uses; 128GB DDR5 RAM kits currently start above $2200, not ~$1000.
datahardware.ai/tools/benchmar…
theramageddon.com/128gb-ddr5-ram…
ctmbottomtext retweeted
THIS THING CAME FROM THE CINEMA SON. YOU EVER BEEN TO THE CINEMAS? YOU EVER BEEN IN THIS SHIT? WELL MY DAD WAS
AND HE LEFT ME THIS DUNE POPCORN BUCKET
