LLM Neuroanatomy II: Modern LLM Hacking and Hints of a Universal Language? (opens in new tab)

(dnhkng.github.io)

7 pointszdw2d ago1 comments

1 comments

Pretty sweet hack as it's orthogonal to quantisation. And while it uses more compute, it doesn't require more VRAM.

Maybe in the future circuits will become modular and composable like models are today?

j / k navigate · click thread line to collapse

(dnhkng.github.io)

7 pointszdw2d ago1 comments

Pretty sweet hack as it's orthogonal to quantisation. And while it uses more compute, it doesn't require more VRAM.

Maybe in the future circuits will become modular and composable like models are today?

j / k navigate · click thread line to collapse