>>109417576
Google has been trying Diffusion
https://deepmind.google/models/gemma/diffusiongemma/
Kind of dumb but fast

File: apu-82ghg3hgj3.gif (51 KB, 462x306)

51 KB GIF

Swap the router for a magic 8-ball.

I would replace all the text with emojis

convert the image to PNG and have anons use imagination to run the animation
follow me anonymous for more tips

File: zero.jpg (36 KB, 619x495)

36 KB JPG

>>109417576
Remove the entire process.
The process becomes divisible by zero when timed for efficiency.
Infinitely faster. Also infinitely slower.
And the toast never hits the floor.

>>109417795
>just pretend it's faster

This is correct. It's exactly what cutting edge tech networks do today to improve efficiency.

So mixture of experts is basically just a bundle of feed forwards instead of one, typical DL solution, just throw more layers onto it

File: moe-transformers-fast.gif (696 KB, 1116x1126)

696 KB GIF

>>109417576
>Can /g/ find a way to speed these up?
Yes

magick moe-transformers.gif -coalesce -delete 1,3,5,7,9,11 -layers optimize moe-transformers-fast.gif

You're welcome

>>109417834
The fact that they're already glossing shit with vague metaphors that mean nothing on a programming level is a sign huh?
>it's not a subprogram, a module, a library, a driver or a controller, it's an EXPERT

This is applications instead of portable compiled executables all over again.
If you told a current AI programmer:
>portable
>compiled
>executable
you would see steam coming out of their head.

And rightly too, because they're just as vague.

>>109417874
FASTER!

>jannu_flag.gif
There's a limit with gifs actually.

>>109417576
E=MC2+AI