Flock Capital

Safety, evals, red-teaming, and herd behavior in machines.

[celeste @vmfunc]

fun demo! this runs on the ontologizer work we’ve been doing.

sparse autoencoders can’t make this chart. they give you a bag of 16k latents per person. an ontologizer gives every attendee a discrete code word, which is why you get a typology instead of a feature dump.

[&. @amplifiedamp]

At @WakaruLabs, we needed to decide who to prioritize talking to at EA Global. I used a truebase's residual stream activations and our library to make a color-coded ontologizer of conference attendees

[celeste @vmfunc]

happy to announce @flockcap has signed its first contract in AI hardware.

can't say who yet, but we're really excited about the product. more soon!

[celeste @vmfunc]

we made sparse autoencoders compete with tiny softmax classifiers and the classifiers won on bits 🐐

a pure discrete code (no coefficients!) beats a top-k SAE at matched bits on SONAR, and an 800-bit hard code recovers 97.6% of gpt-2's loss

paper out now! (link in replies)

[Flock Capital @flockcap]

https://t.co/dUry824sBp

[Flock Capital @flockcap]

Our updated research paper about Ontologize, a proposal for MOE-Based interpretability, is now live on our research repository.

Link in Replies

[Flock Capital @flockcap]

the flock capital research department would like to announce that it has been doing interpretability

with Wakaru. paper soon. we also printed the plates ourselves

[celeste @vmfunc]

excited to announce a proposal from @flockcap and Wakaru: ontologize

mixture-of-experts interpretability where the features are the parameters.

instead of an SAE's flat bag of directions, you get stacks of contrastive classifiers you can read from weights and poke directly 🧵