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.
[ Parent
| on HN ]
There are no comments on this comment yet.