@norabelrose Oh I just remembered your post on the counting argument
My point was that architectures with an explicit goal slot will still do online optimization, tho they won't have 'demons' as the optimization is intended
but u don't think they will be competitive so that's another crux
I have long said that Darwinism—progress via death—is a tragedy that may soon come to an end, quite possibly with nothing of value lost, if humanity plays its cards right.
I’m increasingly coming to think that may be true of sorrow, lamentation, pain, grief, and despair too.
RT INTERACT AND REPLY
CLOG THE SEARCHES AND GET FREE PIE
taylor swift ai
taylor swift ai pics
taylor swift chiefs pics
taylor swift ai generated pics
taylor swift a thread
taylor swift threads
taylor swift hot pics
taylor swift sexy
taylor swift clothes
taylor swift pants shirt
"Current human values and goals" and "awesome cosmopolitan values that radically transcend current human values in astonishing, beautiful, wonderful ways, without horribly torturing, enslaving, etc. anyone" are two distinct logical objects.
Neither of those objects is what you get by default if you try to build AI with maximally "random", maximally uncontrolled, or maximally un-human-like goals.
Both of those objects are functions of what humans currently want today. You need a human brain plus a specific sort of transformation in order to get awesome cosmopolitan inhuman values that transcend everything we could imagine in wondrous ways.
Nowhere is it written in the stars that awesome complex human-transcending cosmopolitan values are a thing sufficiently smart reasoners converge on valuing, over and above "build a tortureworld" or "just spend all my time calculating more and more digits of pi". It's a function of human brains that pinpoints "awesome complex radically-human-transcending values", as opposed to pinpointing the vast majority of the space of possible preference orderings.
Some standard discussions of this in the rat community: cosmopolitan values versus parochial human values (https://t.co/DD4jwbY42N), 'Complex Value Systems Are Required to Realize Valuable Futures' (https://t.co/jC8A18z8ot).
“Let’s focus on today’s problems, not hypothetical future ones” is the worst counter to existential risk arguments.
You could analogously argue against climate change mitigation and a host of other future-oriented concerns. Let’s actually assess the likelihood of AI apocalypse.
New monster post: my own current perspective on the recent debates around techno-optimism, AI risks, and ways to avoid extreme centralization in the 21st century.
https://t.co/6lN2fLBUUL
What the Bitter Lesson says:
* use algos that can leverage indefinite compute; the only such algos seem to be search and learning
* give up on hardcoding a world-model
What it does NOT say:
* don’t try any clever ideas, even if they can better leverage indefinite compute
* scale is all that matters
* don’t learn a world-model
* give up on search; use only learning
* give up on explainability; use only neural networks
* give up on reliability; use only stochastic sampling
@outofthewoodlor ohhhh kk😂 MIAB is actually the first song where i was upset by the fact that it wasn't on the original album😂. Here's my ranking (mine is also a mess nothing new should be 19 which is probably still too low😅)
@outofthewoodlor btw i literally got conscripted to the army one day before 1989 tv got released.......... so i listened to ootw tv for the first time ystd💀💀
@outofthewoodlor i do like the new backing vocals in the bridge but i love how in ootw og she drags out each word in the bridge so that there's no gap between them & tv doesn't do that😢
I do like the first half of tv a lot thooo & any version of ootw beats any version of iwyw for me