martinveron

Catching the monitor in a lie by comparing what it perceives against what it reports. TL;DR * I trained a linear probe to directly perceive code back-doors, which are represented linearly in activation space, using honest behaviour that can be elicited even from a scheming model. * I then trained...