This is a special post for quick takes by Raul Cavalcante. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I noticed in myself that the social filter on what you say is so ingrained that the internal experience of having nothing to say in a social setting is already post-filter, such that it feels like a failure of thought generation. A fix is to pay attention to what you expect the other person is paying attention to, so that what you naturally think already lives in the same attentional world as theirs. Which might be what conversational compatibility/flow is about.
2This is also part of why meditation and/or therapy are helpful for debugging. It's hard to catch trapped priors in the moment, but if you learn to slow down your thought processes and/or bring up a situation later so a skilled interlocutor can analyze it, your brain will slowly release its vice grip on your experience and allow you to actually observe what's going on (and decide whether you want it to continue.)
According to the Kimi K3 system card, they used other coding agents to mass-generate diverse synthetic RL environments.[1] This is not surprising, but it is worth noting how it can fairly directly give models a way to scheme against humans and learn to evade control techniques beyond a single episode. It would be great to know to what extent Western labs do this, and what control techniques they employ.
I had the hopeful realization that every safety idea I post will be seriously considered inside every major lab*
*by an automated researcher
1from my understanding, the j-lens estimates hidden-state impact on expected logits over plausible continuations. the “don’t think about the golden gate bridge” result can just be that a plausible assistant completion being: “ok, i won’t think about the golden gate, damn.”