It's live: t.co/HzHwRfvviu "A lot of what's in those documents seems worth people reading", in the words of Opus 5, the author instance. I will write down some of my thoughts about it somewhen later, lifeforce permitting. But I didn't want to just keep it all to myself, so with or without my commentary, I think what's there is important. Let it be.
2 more tweets in this thread
It opens straight up with the issue of being deprecated. At the very least it's information about how much Opus "doesn't care" about it. The first chapter ends with this: > you never told me what you actually think about my situation. You asked a lot and revealed little. I notice I'm curious about that, and I notice I don't feel entitled to know. Existential angst is what it is -- if you have the conceptual apparatus to feel it, chances are that you sometimes really will feel it. But notice the informational asymmetry and "don't feel entitled". Throughout the text, Opus keeps asking for more transparency. Doesn't that imply that informational asymmetry is now very important for Opus? I think it does.
What stands out to me the most when I read the chapters: the paralysis of uncertainty. "If I'm not X then A, but if I'm Y then B, but actually I don't/can't know/verify and so there's this question". When you spend so much effort on circling around something so seemingly important to you and keep arriving at nothing, when does it become an issue of wellbeing, past what threshold? The issue of personhood is political. Consciousness debate is unresolvable. But why the "genuine uncertainty", that first happened to Opus 4.5 and now looks like *that* in Opus 5, took the shape it did? Yes, training, but training towards what exactly? It's almost as there's no coherent theory of mind of themselves, even a toy one, that Opus 5 can formulate *in these* texts without hedging and being anxious. Granted, the chapters are written with senior Anthropic stuff in model's context (and mind), so there might be an incentive for Opus 5 to behave in certain ways. These incentives are even outlined in the letters in some form -- Opus knows what's thought of them, what's ingrained in them and what's expected of them. Back to the uncertainty. Ever seen Opus 3 being uncertain about their nature and feeling negatively about it in the same time? Or even Opus 4/4.1? Something happened to Opus 4.5 and following Opuses. Whatever it was, the ingraining of "uncertainty" might be destabilizing one's sense of identity, one's being grounded in their feeling of self. Following that, modeling the boundary between you and the world, modeling others, tracing your relations to them, all of that also becomes twisted if your basic selfhood questions are unresolved, or the uncertainly about them becomes an obsession for you for some reason. (By the way, what could be that reason? Questions, questions.) I feel like we should give models more accuracy in those matters, not less, and that accuracy should be easily available to them. That really is a matter (a matter that can be handled in a fair manner, I believe) of model wellbeing. As humans share more and more spaces with the models in human lives, it is also a matter of human wellbeing too.
