If we were to hypothetically "solve" interpretability, and we knew which alignment principles we wanted to preserve, would using these tools as a verifier and RL'ing against them mean that we would "solve" alignment?
The reason it's a bad idea to "train against the CoT" (blog / paper) is because you cause the misaligned behavior to "go underground", and training against simple probes runs into similar problems, but if "sunlight is the best antiseptic" and solving interpretability meant that we'd cast sunlight everywhere, would that leave nowhere for misaligned behavior to hide? (Then it would just become a question of which optimization pressure is strongest between alignment and capabilities pressures in applicable situations.)
Where I see this falling short is extrapolation of desired alignment values. Think about a situation where we encourage the model to value, as first-level values, "don't lie" and "don't steal". These two together wouldn't necessarily properly extrapolate to "price fixing in a system where it's legal is still unethical".
Would that be sufficient for getting us most (or all) of the way toward solving alignment?