@Buck was so kind as to discuss with me my confusions about Redwood Research (RR) and their level of transparency. You can read my original comment here. I am relaying Buck's response to the best of my ability, but probably am getting some things wrong.

TLDR of what I said: It seems very easy for people new to AIS to get confused, and do useless things. Maybe RR should be more transparent, and highlight the nonobvious strategic considerations which inform their research agenda. Also, transparency might allow for better external feedback to RR, e.g. about control preventing warning shots.
TLDR of Buck's reply: It's not that bad for people who don't think about strategy to do useless stuff, and RR's overall strategic considerations over time have been somewhat consistent. Those considerations are shared a few paragraphs into blogposts that people who want to understand RR can read past the first paragraph. Also, HF was confusing but not a huge update for RR, which had already considered the warning-shot-reduction externalities of control.

I took notes during our discussion, and below is Buck's reply in more detail, to the best of my understanding. That may not be a very strong understanding and I expect I unintentionally misrepresent Buck at least once:

• Some people probably do unstrategic control research because they are inspired by RR's control research but don't understand what makes it useful. This isn't important, and probably doesn't have a significant effect on the labs' control mechanisms, or anything else for that matter.
• RR's views on control are pretty consistent with Buck and Ryan's original thoughts on control.
• RR always emphasized that internal deployments, rather than models within evals or RL runs, are of central concern. This is because of the great affordances allowed to them and thus the great risk they pose of takeover.
• RR has long felt that compared to misalignment, misuse is not a very big deal.
• RR also hasn't been working on control mechanisms since March. This is because they've noticed the labs start to implement somewhat sophisticated control mechanisms. So, they work on the red teaming and eval methodology of control instead, as well as trying to surface evidence of misalignment risk.
• RR views on "conceptual uplift stuff" aren't yet public, but they are currently working on writing up their thoughts about this. It is unclear to me whether this will then be made public, but it seems it will at least be available to in group people who ask to see it.
• Buck is confused about the lessons from the Hugging Face incident, and confused about the fact that the diffusion of RR's earlier control work could have prevented it, but isn't updating that much, or isn't sure how much to update (“I haven’t even changed my mind very much, its more of a vibes thing”).
• Even in light of his confusion around the HF incident, Buck doesn't think that there was really a crucial consideration or argument about close calls that RR hadn't considered beforehand.
• In an ideal world, RR would have infinite time to write up all of their strategy considerations and confusions and share these and get feedback (and I assume include lots of info about the associated strategy considerations within all of their technical work, although we didn't discuss this). Unfortunately, we don't live in a perfect world and they're busy guys.
• Specifically, Buck feels "moderately confused about a lot of things related to all of this" and wishes he had time to write a "glorious post" about whether all of the prosaic/scalable oversight/RSI stuff is good (I don't know exactly what he meant by prosaic/scalable oversight/RSI stuff, but I assume he is pointing at the tradeoffs of research which focuses on aligning transformative AGIs in the hopes that they will help with the harder ASI alignment problem).

I'm pretty satisfied by this. I previously overrated RR's strategic uncertainty in the wake of the HF incident, and probably overrated how helpful more external feedback on their strategic thinking would be. Still, I would prefer if RR shared their strategic thinking more broadly than they do now, and think this could be pretty cheap. Specifically, when publishing research it would be pretty easy to highlight the relevant strategy writing which informs it (if it exists, which Buck seemed to say it mostly did). For example, in their 2024 paper AI CONTROL: IMPROVING SAFETY DESPITE INTENTIONAL SUBVERSION—and I don't mean to pick on a paper from two years ago, but I think it's an important example—the authors only mention in section 5.1.2 that they expect control to be useful despite the fact that they do NOT expect control to scale to ASIs:
"Therefore, as more powerful untrusted models are developed, and the gap between these powerful untrusted models and trusted labor grows, AI control will become a decreasingly effective approach to safely using powerful AI. We think control is a valuable research direction anyway, because it can be applied to reduce the risk from using AIs at capabilities levels which we think would otherwise pose a lot of the total risk. Applying control to these powerful AIs might allow for productively and somewhat safely using these AIs for key AI safety applications like further AI control research or alignment research..."
They do not mention these considerations in the abstract, introduction, or conclusion of the paper, nor do they denote where people can go to read about the relevant strategy considerations in further detail.

I would also still like if RR were spicier, and posted their beliefs more often. I see how this could have some important costs in terms of status outside of the AI safety bubble and RR's time. I also think it would have a very nice effect on the Constellation-epistemic-bubble (and relatedly, I broadly agree with the ~40% that I understand of Alex's thoughts here). I'm not sure if this is worthwhile at the margin, because I don't have a good sense of the costs, but Gosh it would be nice.