3. As a corollary, explain why I mostly reject non-frontier-pushing principles (and explain what would cause Anthropic to stop scaling, besides RSP stuff), and clarify that I do not plan to abide by past specifically-non-frontier-pushing commitments/vibes (but continue following and updating the RSP of course);
4. Encourage Anthropic staff members who were around in 2022 to talk about the commitments/vibes from that time.
I’m kinda confused about what you mean by both of these. For #3, can you say it again in different words? For #4, which particular thing are you getting at?
3. (a) explain why I think it’s fine to release frontier models, and explain what this belief depends on, and (b) note that maybe Anthropic made commitments about this in the past but clarify that they now have no force.
4. Currently my impression is that Anthropic folks are discouraged from publicly talking about Anthropic policies. This is maybe reasonable to avoid the situation an Anthropic staff member says something incorrect/nonpublic and this causes confusion and makes Anthropic look bad. But if Anthropic clarified that it renounces possible past non-frontier-pushing commitments, then it could let staff members publicly talk about stuff with the goal of figuring out who told whom what around 2022 without risking mistakes about policies.
I’m kinda confused about what you mean by both of these. For #3, can you say it again in different words? For #4, which particular thing are you getting at?
3. (a) explain why I think it’s fine to release frontier models, and explain what this belief depends on, and (b) note that maybe Anthropic made commitments about this in the past but clarify that they now have no force.
4. Currently my impression is that Anthropic folks are discouraged from publicly talking about Anthropic policies. This is maybe reasonable to avoid the situation an Anthropic staff member says something incorrect/nonpublic and this causes confusion and makes Anthropic look bad. But if Anthropic clarified that it renounces possible past non-frontier-pushing commitments, then it could let staff members publicly talk about stuff with the goal of figuring out who told whom what around 2022 without risking mistakes about policies.
Ah makes sense. Point #4 is interesting. Probably not really scalable/repeatable without things getting weird but might work as a one-of.