This is a special post for quick takes by Thomas Kwa. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
If we see misalignment incident announcements/leaks, we should pay attention to the actual date of the incident. The later it was, the more concerning about the adequacy of current safeguards. For OpenAI, incidents well after the July Hugging Face incident should be more worrying because they would have happened despite mitigations OpenAI announced in response to HF.
In general, if a company has a steady stream of severe incidents, say, a month apart, one can infer that the later incidents happened despite mitigations the company was willing/able to implement in a month after the earlier ones.
Conversely hearing about an incident that happened prior to previously known ones should drive an update that all the later ones happened despite any mitigations from the earlier one. That's also pretty bad.
That's only true if the subsequent incidents happened after detection of the previous incident. My understanding is that detection for many incidents has been disturbingly slow, so adding a few months delta is neccessary.