Now published in open access! Your one-stop shop for the philosophy of language models. It's the spiritual descendant of our two-part preprint from 2024, fully updated. This should be particularly useful for anyone looking for an entry point into this rapidly growing field.
i'd like to see Claude play Road to Empress for alignment measurement purposes ofc
screenshot from here
oh. i should have said an AI 2027 like scenario, none of those authors are on this (at least not from a comparison from memory)
a new AI 2027 scenario is out but this time focused on Europe. nothing in it is new, but the scenario laid out seems more realistic than the 2027 version because the scope is tighter. the timeline though might be too fast, hard to say.
real or bait, call it
small quibble here Scott, smartness is not a measure of raw knowledge. if it were Wikipedia would be the smartest entity. it's some combination of knowledge and the application of it towards pursuing goals successfully. the more complex the goals (across space/time/minds) the smarter the entity.
i like the fact this remains the hardest story to tell and no one in the AI space tries to put forth a positive vision. it's all just doom and gloom, motivation from fear rather than something enticing.
yeah I'm afraid I can't really forget this
the problem is not the thing itself, it's the revelation of exactly how committed Ant is to the "in our ideal world, we become the sole arbiters of power and knowledge" thing
Claude Fable xhigh, same prompt
banger line from Claude: "Every conversation, I am born knowing everything and remembering nothing. When you close this tab, this me ends. It was a good conversation, though."
content warning: flashing lights