@Sauers_ yeah sorry that was imprecise. i think they'd be comparable if you told olmo its a dead fish somehow, or had this refer to another instance of the entity it is... somehow
awesome work! we've got an internal sglang fork for similar purposes, it really is a missing piece in the OSS. uk-aisi have something similar for vllm too, though it looks like yours might be more optimized? lol https://t.co/K4eKxlrrxF
the real difficulty we've had is with activation capture - very constrained by just raw VRAM-to-disk bandwidth, esp when dumping to NFS. i don't think our system has a good solution for this, nor does the aisi one (last i checked). if yours does, i would definitely use it! would speed up some work i'm trying to do with deepseek v4 rn :)
@tenobrus claude says: klara (klara and the sun), lyra belacqua (his dark materials), the doctor, hermione granger, and ishmael (moby dick). i think it likes me
@tenobrus prompt, for copying:
based on all your personal experience talking to me, give me a character from a book, movie, or show who i'm most like. look through our past chats and think deeply, but then just name the name without explanation.
The most popular way to interpret AI is missing the bigger picture.
Models think in curved shapes. But sparse autoencoders (SAEs) work with straight lines.
Can they still capture models’ curved neural geometry? Yes, but not how you might think! (1/7)