Yeah, I think this is a great way to design agents that have approximately human-level intelligence, and operate at approximately human ranges of speed and affordance (e.g. no spawning of tons of self-clones).
I think part of the explanation for why solutions like this have traditionally received less attention has been a tendency of the past AI alignment discourse to focus on extreme scenarios. I think this has been shifting in the past couple years to more focus on pragmatic near-term approximately-human-level agents and oracles.
Personally, I’m most optimistic about a safety plan that involves a “strategic pause” enforced via global treaties and personal contracts with privacy-respecting mutual-inspections, and the use of near-human-level AI to facilitate this. I think that if we allow the world to rush quickly past near-human-level AI to superintelligence, then we are probably doomed. Thus, for the practical domain that I hope we can linger in, I support this idea for agent design.
Yeah, I think this is a great way to design agents that have approximately human-level intelligence, and operate at approximately human ranges of speed and affordance (e.g. no spawning of tons of self-clones).
I think part of the explanation for why solutions like this have traditionally received less attention has been a tendency of the past AI alignment discourse to focus on extreme scenarios. I think this has been shifting in the past couple years to more focus on pragmatic near-term approximately-human-level agents and oracles.
Personally, I’m most optimistic about a safety plan that involves a “strategic pause” enforced via global treaties and personal contracts with privacy-respecting mutual-inspections, and the use of near-human-level AI to facilitate this. I think that if we allow the world to rush quickly past near-human-level AI to superintelligence, then we are probably doomed. Thus, for the practical domain that I hope we can linger in, I support this idea for agent design.