Evidence about risk should be transparent by Ajeya

Note: This post was cross­posted from Planned Ob­so­les­cence by the Fo­rum team, with the au­thor’s per­mis­sion. The au­thor may not see or re­spond to com­ments on this post.Subti­tle: We can’t de­velop safety stan­dards if we have to rely on opaque judgmentAll views are my own and do not rep­re­sent my em­ployer.In the wake of the re­cent wave of mis­al­ign­ment in­ci­dents, both OpenAI and An­thropic have re­ported slow­ing down RL train­ing to im­prove safety. Th­ese in­ci­dents, com­bined...