News for hackers. Stuff that matters.
Re:AI Models Can Provably Hide Arbitrary Capabilities by ag8 on Tuesday October 13, 2026 @03:05AM (#50067855)
Yeah. Also, we think that in certain scenarios, it's possible for "cryptographic patterns" to arise in model weights via normal training, especially if models are trained against things such as honesty probes.

There are no comments on this comment yet.