@CJHandmer Thank you, super cool! I have nature on my mind atm, and wonder--is there a nature angle here too? Eg could this work in a scheme to compete with (marginally) saving nature in the Amazon (or other lush+vanishing places), for instance, by making new wild forest in Nevada?
I get excited seeing two sides of a disagreement get together and come up with a new proposal that works for both! Thank you @deanwball and @DKokotajlo67142.
The proposal: much more transparency in advanced AI development.
In @TIME this morning, @DKokotajlo67142 and I make the case for frontier AI lab transparency—either as a voluntary commitment or as a law.
While we disagree quite a bit about the trajectory of AI, we concur here. Link to the piece, with much more detail, in the reply.
@weidai11 So in some sense, I disagree: we did the best with what we had, as far as we knew at the time.
In another sense, I agree: I think we have more ability to affect global trajectories than we realized, but doing so means giving up a sense of control, maybe losing a "hope of glory"
@weidai11 3) we don't yet know if we can find technical breakthroughs that make alignment and maybe even coordination dramatically easier.
We were great at technical thinking, inexperienced at political thinking, so we went with our strengths and explored technical research.
@wolflovesmelon @thegartsy Ah yeah, I think I get you that, given the system how it is, you don't trust developers, and don't want the public to trust them. I'm saying I want the system to change (eg. via this bill) so that the public, accurately not mistakenly, can trust developers more.
California's big AI bill SB1047 is on Gov. Newsom's desk to veto or sign. The compromises made between Anthropic and authors are the best we'll get for now. If it will bring more mutual trust between AI creators and the public, sign it, that's what we need. I think it does.
AI will advance massively again in the coming 5 years. We need more trust between AI creators and the public to keep the creation flowing. SB1047 looks like a step in that direction.
3 trusts we must increase for AI to advance well:
1. Public trust, that AI creators will protect the good we have.
2. Creator trust, that the public won't kill creation.
3. Creator trust, that their work will be used wisely.
This bill will affect all 3 of these, over years.
@jalammar in https://t.co/Cac9BbcniZ there's a typo (homonym) that kinda makes sense on its own but makes the sentence have a different meaning from what you intended: "peak" should be "peek" in "A normal self-attention block allows a position to peak at tokens to its right."