Note that we don’t need to compete with any kind of “best possible” memory system. We to:
a. Compete with the benchmark AI.
b. Remain corrigible.
For [a] the plan is to co-opt whatever mechanism the benchmark AI uses to decide what information to store. This involves some subtleties, since we need to construct a benign analog, but it doesn’t require anything like knowing what computations we’d want to run in the future or which cognitive architecture we will adopt in the future.
For [b] the demands on memory seem quite weak. Probably we could get by with no memory at all.
Note that we don’t need to compete with any kind of “best possible” memory system. We to:
a. Compete with the benchmark AI. b. Remain corrigible.
For [a] the plan is to co-opt whatever mechanism the benchmark AI uses to decide what information to store. This involves some subtleties, since we need to construct a benign analog, but it doesn’t require anything like knowing what computations we’d want to run in the future or which cognitive architecture we will adopt in the future.
For [b] the demands on memory seem quite weak. Probably we could get by with no memory at all.