Toy Model of Activation Obfuscation
A concrete example of why the most forbidden technique is forbidden
Toy demonstration of "the most forbidden technique" using linear probes
This work was done as part of my work for the BlueDot Impact Technical AI Safety Project, and is part 2 of a series here. Recall that we consider a residual stream MLP …
Activation Steering with Greedy Coordinate Gradients
Tricking a probe to get high scores with an adversarially controlled prompt prefix
Analytic nonlinear feature obfuscation with two MLP blocks
This work was done as part of my work for the BlueDot Impact Technical AI Safety Project, and is part 1 of a series here. Recall that we consider a residual stream MLP …