Twitxy Saved

wassname (with fallback)

Verified account This account is verified. Learn more

@wassname

Michael J Clark / wassname. Let's align AI better than humans.

h+, curiosity, and the good ending.

anon feedback: admonymous.co/michael-c
Perth, Australia wassname.org Joined September 2009
1.9K Following 435 Followers 3.2K Posts

Latest posts 5

wassname (with fallback) Verified account @wassname
@Sauers_ so all your experiments are ability LLM introspection, what is your thinking so far?

It seems:
- They can change their own residual stream on an aggregate level
- They can read other models residual streams, even other families
(I guess this show activation oracle and NLA make sense)

and I guess
- They imagine a body during role playing
- single neuron steering is funny (but it's mostly cherry picked, or am I wrong)

Probably more?