J Bostock comments on How will we do SFT on models with opaque reasoning?

[J Bostock]

The technique you describe here seems like it’s very vulnerable to the decoder model colluding with the policy

Yes, I was claiming that this was likely, not that it was desirable.