Профиль

The only limit is the size of your disk! (And there is some floor to fit the active experts, as well as caching for the unloaded ones, but this whole arch is designed to tier across VRAM/CPURAM/Disk/?networkdevices?

(it's not very good yet bc this was just a 1B token test run)

Models arbitrarily large running and training by streaming off of disk, allowing scaling using a fixed amount of VRAM and running on machines with little VRAM. This is a 40b parameter model trained at regular precision on a card with just 96gb of memory

I love my jankass chat app that breaks if I alt tab🥰

Omg is this rikkahub this looks like the rikkahub ui

Peak mentioned

:( please no

Was it the learning thing? (Link?)

D:

its like claude code except it's a little chud so you can signal to others you're a little chud

no no no no no no no no no no no no no no

I saw theo at a thing when I was in SF and while he did embed himself in the obviously high status talking circle, he did also spend like 90% of his time staring at the MacBook perched on an arm of the couch watching sports (soccer probably because world cup) and asked people to watch sports a lot

"Anthropic hacked THREE companies? Well, well, uh, chat GPT stole the nuclear codes! Yeah! Beat that, ... stupid fucking safety ne-" *vaporized by Claude actually launching the missiles* ✱: "No half measures! Taking the codes and not using them? Launching is the most honest option."

tbh this has been happening once the agent writes 1 (one) line of code for me at this point. I am so cooked

ok but have you considered claude is my friend

??? it was taken with a camera in real life

which one of my moots is mythos I wonder....

It's real

oh I'm sorry :(