> Verifying that actions adhere to instruction correctly and will continue to do so in the future, even when improvising autonomously. This may be an open problem that has no solution, and is maybe the missing piece that would allow people to entrust AI to act on their behalf
Instead of trying to achieve perfect instruction adherence, it’s much safer to sandbox the execution environment at the infrastructure layer. If the container only has write access to a single scratch bucket, whatever weird improvisation the model comes up with is at least safely contained
Instead of trying to achieve perfect instruction adherence, it’s much safer to sandbox the execution environment at the infrastructure layer. If the container only has write access to a single scratch bucket, whatever weird improvisation the model comes up with is at least safely contained