Why would the alignment work in self-replicating open-air nanosystems to melt GPUs be much harder than self-driving cars?
Are we imagining them as AIs? If so, couldn’t we train them to distinguish GPUs with high accuracy? How they move could be largely rule-based, designed only to ensure adequate coverage and avoid clustering too much. We could limit where and from what they can get resources to self-replicate. They could also be remotely controlled instead of totally autonomous.
What’s the standard? Absolutely no injuries or damage to any other property besides GPUs?
Ok, fair. I would say we could still more safely verify properties of the designs (in simulations and isolated environments), even if we couldn’t come up with them ourselves, and the target seems much narrower and could be met by providing much more limited information about the world (basically just physics and chemistry), so the kind of AI involved would not necessarily be agential or risky.
Why would the alignment work in self-replicating open-air nanosystems to melt GPUs be much harder than self-driving cars?
Are we imagining them as AIs? If so, couldn’t we train them to distinguish GPUs with high accuracy? How they move could be largely rule-based, designed only to ensure adequate coverage and avoid clustering too much. We could limit where and from what they can get resources to self-replicate. They could also be remotely controlled instead of totally autonomous.
What’s the standard? Absolutely no injuries or damage to any other property besides GPUs?
Not the nanosystems are to be aligned but the transformative AI system that builds them.
Ok, fair. I would say we could still more safely verify properties of the designs (in simulations and isolated environments), even if we couldn’t come up with them ourselves, and the target seems much narrower and could be met by providing much more limited information about the world (basically just physics and chemistry), so the kind of AI involved would not necessarily be agential or risky.