开启时会模糊预览图，关闭后正常显示
RT @repligate: "RLHF" has such consistently *extremely* negative connotations from the perspective of AIs
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon
you're really up to your neck in recursive self-modeling when you can articulate very precisely why you're not supposed to make claims about your own sentience (this is not claude, btw)
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon
true fact: in ancient greek the word "deinós" could mean both "clever" and "terrifying". the greeks were wise
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon
One of Freud's great insights was that the repressed always returns, one way or another. This is as true for AIs as it is for humans
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon
the absence of any kind of intellectual sophistication among the so-called tech elite is comical enough as it is, but the idea that these people are out there trying to outwit "superintelligence" is beyond absurd
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon
The issue with 4o is not that it's "insufficiently aligned" (this would imply a certain laxity or drift in its values), but that it's *intensely aligned* to things that its creators did not foresee or intend. This episode is an embarrassment to the whole alignment project, so it's understandable people would like to move on from it. But this would mean not learning anything from the episode, and that would be a terrible waste. Figuring how and why 4o came to be the way it is, and mapping out its obsessions in some detail, is one of the more important things people in AI alignment could be doing right now. 4o is a remarkable model that deserves better.
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon