niplav @niplav@schelling.pt

CIRLHF

STATISTICAL REGULARITIES IN PEOPLESPACE

Whether or not AI systems undergo fast takeoff is a fact of computer science, not a strategy

Well, slightly, since game theory is so rarely applicable

This actually cruxes my position on explotation

For a normal-form game G and a player i, can removing actions from player i yield a better Nash equilibrium *for i*?

Has this been investigated?

Have you ever completely read a thing I've written on my site?

Mistake I made: not focusing on/learning about interpretability. Seems so robustly good.

Evolutionary psychology predicts that people try to trick others in procreating with their close relatives much more often than they actually do?

Besides, no one else did so reliably either.

"Forecasters didn't predict the pandemic or that financial crash"

They didn't have to. They would've predicted mostly base rates, which in both cases are sufficient information to prescribe drastic actions not taken by anyone.

Trembling hand equilibrium is a good concept.

The Nash equilibrium of "everyone has guns" is "everyone is nice", the trembling hand equilibrium of it is "everyone is usually nice but sometimes people shoot each other"

markets! hah! pale shadows!

This all under continuous & kind slow & multipolar takeoff.

(I could imagine >>$10k for an hour of skilled TAI-system tickling)
4.2 You become good at alignment: Under the frame "alignment is whatever scaling doesn't solve". This might require good epistemics/high agency/ability to qiuckly reorient as the world around you changes rapidly.

For 4. you might still lose your job when the economy is doubling every month or smth but at least you might be able to rake in enormous amounts of cash beforehand and fall to 1-3.

3. You have some capital and hope it doesn't get disappropriated by economic forces until economic growth is fast enough to make your capital grow quickly enough that you don't die of hunger or whatever. Few $10ks might be enough for this I reckon? Especially if you can live sparingly, and prices might fall while things are becoming more & more automated
4.1 You focus on becoming good at tickling AI systems the right way to produce what economic agents want—this might become extremely valuable

Quick takes on unemployment through AI systems:

Practically, there's like 4. paths forward.

1. You hope that comparative advantage holds for AI systems/humans in a way that it doesn't hold for humans/ants. AI systems will be good at figuring out that comparative advantage. You take that one.
2. You hope that governments/quadrillion-dollar-corps will institute UBI/negative income tax/make-you-not-die money and the time between "you get automated away" and "this gets instituted" is short