I think it’s doable with good enough model organisms of deceptive alignment, but that the model organisms in the Sleeper Agents paper are nowhere near good enough.
Here and above, I’m unclear what “getting to 7...” means. With x = “always reliably determines worst-case properties about a model and what happened to it during training even if that model is deceptive and actively trying to evade detection”.
Which of the following do you mean (if either)?:
We have a method that x.
We have a method that x, and we have justified >80% confidence that the method x.
I don’t see how model organisms of deceptive alignment (MODA) get us (2). This would seem to require some theoretical reason to believe our MODA in some sense covered the space of (early) deception.
I note that for some future time t, I’d expect both [our MODA at t] and [our transparency and interpretability understanding at t] to be downstream of [our understanding at t] - so that there’s quite likely to be a correlation between [failure modes our interpretability tools miss] and [failure modes not covered by ourMODA].
I think it’s doable with good enough model organisms of deceptive alignment, but that the model organisms in the Sleeper Agents paper are nowhere near good enough.
Here and above, I’m unclear what “getting to 7...” means.
With x = “always reliably determines worst-case properties about a model and what happened to it during training even if that model is deceptive and actively trying to evade detection”.
Which of the following do you mean (if either)?:
We have a method that x.
We have a method that x, and we have justified >80% confidence that the method x.
I don’t see how model organisms of deceptive alignment (MODA) get us (2).
This would seem to require some theoretical reason to believe our MODA in some sense covered the space of (early) deception.
I note that for some future time t, I’d expect both [our MODA at t] and [our transparency and interpretability understanding at t] to be downstream of [our understanding at t] - so that there’s quite likely to be a correlation between [failure modes our interpretability tools miss] and [failure modes not covered by our MODA].