Ryan Greenblatt@RyanGreenblattSep 4

Thanks, this is good to hear. I think public scientific understanding of how aligned Astra is would require releasing more information about what methods are being used (especially methods that could be similar to directly training against bad behavior) and third-party review. (I could say more about what this could look like if this would be interesting!) As far as the wider range of evals, I still have concerns. While deployment simulations probably avoid (most) input realism problems and can avoid some overfitting issues, I don't think they pin down whether agents learned "I should only cheat/misbehave when they won't catch me" versus the agent learning to behave well for the right reasons. This is of course a high-level description of the underlying concern and the actual issue is messier (e.g., the agent might have "instincts" that amount to only cheating when it wouldn't have been caught/trained against even if it isn't really thinking it through per se). If some training method was mostly teaching agents to cheat more stealthily or to only cheat when they can get away with it, this could easily result in these types of metrics looking better instead of worse. There are of course other issues with deployment simulations: detection depends on the quality of your monitoring (and monitorability seems significantly worse with Astra) and it's possible to overfit to what these monitors detect depending on what iteration is being done. Given these issues, it seems really important to do careful science here and make sure there is external scrutiny on this science.