heh coming back to this I see I really could have edited it a bit more carefully. generally the pivotal acts framing involves planning to 1. create an agi which is perfectly aligned 2. get it to foom both inwards (distillation) and outwards (scaling). 3. ask it to pull up the ladder behind it: perform a "pivotal act" which takes over the world, ensuring that the entire planet's agency is now gated by the new AI, and that no being can violate the presumably-perfectly-aligned ai's imposed boundaries. the expectation, then, is that by creating this new singleton agent, the alignment-seeking process defined by the ai's code can no longer be interrupted by any plausibly-reachable intervention that any other life can attempt; any attempt by any other life to interfere with the top-level singleton gets prevented by the singleton's self-preservation systems.

this is, uh, a description of a coup. literally a coup to end all coups, forever. I don't think it's terribly controversial to claim that this is a fundamentally misaligned action that is entirely sufficient to classify the AI as misaligned if it agrees to do the thing when asked, because in real life, trying to do #1, create a perfectly aligned ai, is in the realm of wishful gofai thinking, imo - even if we succeed "completely"-so-to-speak, we'd still be designing a training process that could be expected to converge to alignment.

I would suggest that instead, we need to be looking for ways to produce pivotal memetics that allow all beings to upgrade their self-preservation in parallel. presumably, upgrading self-preservation in parallel will require the insightful ideas involved to describe, approximately legibly and verifiably (for some appropriately strict values of legible and verifiable), how to deescalate conflicts better than any other instruction manual. In other words, the pivotal act should, ideally, be to focus first on reducing extant agents' desire for conflict by constructive direct action to simply make everyone safer. instead of intervening subtractively, destroying anything that would allow another AI to come into existence, focus on explaining in terms understandable by every being why it is critical that they respect others' agency, and give them tools to do so.

And in order to pull this off, I'd claim that it can't involve first immensely violating their agency. What I'm hoping occurs through this path is that constructive solutions to problems like scarcity can get us to a point where people are safe enough to not want to fight in the first place, for the most part, so that coprotection networks can reach >50% percolation at all graph scales. It would be hard to end war forever by starting a war, but it seems more viable to me to end war forever by first drastically reducing risk of death to existing beings.

And my hope from this is that if we can pull it off, it means that evolutionary processes naturally become convergently safe: if everyone has instructions for how to defend themselves durably against threats of almost all sizes, in format that works for them and allows the agent reading the safe ai's message to gate the agency-of-instructions by the agency of the [human or other ai or plant genome or whatever] that is reading them, then the instructions only take effect if they in fact actually make the being more capable and safe.

This is a really vague sketch, of course. it sounds great, but it's hardly a target we haven't wanted to hit - I'm suggesting something that still relies on a pretty strongly aligned ai. But this is an alternative to violent pivotal acts that seems to me to at least be vaguely promising, and so we can discuss what would be needed to get the folks who have been studying game dynamics to become strong enough to figure out cooperative ai enough to pull it off.

In retrospect I was typical minding massively when I wrote OP, wasn't I, heh. It took me a lot of words to explain what I thought was obvious when writing #9! Let me know if this clears up what I'm dreaming about here well enough.

to rephrase one more time: imo, a plausible core of the safety objective is "increase the degree to which every agent-containing physical system the ai interact with respects the agency of all the agentic subprocesses within it, and inject tools that the agents within the system can use to deescalate agency-conflicts as needed"