I agree that synbio has some very nasty and rapidly emerging capabilities. However, with respect to your last paragraph are you also assuming that defenses don’t improve? Fancy biotech enables better detectors and rapid creation of tailored countermeasures (including counter-organisms). Surveillance tech restricts what students can get away with, sterilization and isolation of environments becomes easier, etc.
However, with respect to your last paragraph are you also assuming that defenses don’t improve?
The statements I made were agnostic as to the likelihood of a given event, as opposed to the raw capability of the devices—that is, beyond saying that it would become a non-zero percent chance. Furthermore; it is generally true that defense is “harder” than offense when it comes to weapons-tech.
Even if better technology means defenses can improve, does that mean they will improve at a fast enough pace? I don’t understand why your same logic wouldn’t also imply the belief that it will be easier to make AI friendly when we understand more about AGI.
I don’t understand why your same logic wouldn’t also imply the belief that it will be easier to make AI friendly when we understand more about AGI.
Ceteris paribus, that argument does go through: for any given project, success is easier with more AGI understanding. That doesn’t mean that we should expect AI to be safe, or that interventions to shift the curves don’t matter. Likewise, the considerations I mentioned with respect to synbio make us safer to some extent, and I was curious as to Logos’ evaluation of their magnitudes.
Okay, thanks for the clarification. If we would expect the magnitudes for synbio to be significantly higher (or lower) than for AGI, then I would be curious as to what differentiates the two situations (I could easily imagine that there is a difference, I just think it would be a good exercise to characterize it as precisely as possible).
ETA: Actually, I think there are some plausible arguments as to why AGI progress would be less relevant to AGI safety than one would expect naievely (due to the decoupling of beliefs and utility functions in Bayesian decision theory—being an AGI hinges mostly on the belief part, whereas being an FAI hinges mostly on the utility function part). But I currently have a non-trivial degree of uncertainty over how correct these arguments are.
I agree that synbio has some very nasty and rapidly emerging capabilities. However, with respect to your last paragraph are you also assuming that defenses don’t improve? Fancy biotech enables better detectors and rapid creation of tailored countermeasures (including counter-organisms). Surveillance tech restricts what students can get away with, sterilization and isolation of environments becomes easier, etc.
The statements I made were agnostic as to the likelihood of a given event, as opposed to the raw capability of the devices—that is, beyond saying that it would become a non-zero percent chance. Furthermore; it is generally true that defense is “harder” than offense when it comes to weapons-tech.
Even if better technology means defenses can improve, does that mean they will improve at a fast enough pace? I don’t understand why your same logic wouldn’t also imply the belief that it will be easier to make AI friendly when we understand more about AGI.
Ceteris paribus, that argument does go through: for any given project, success is easier with more AGI understanding. That doesn’t mean that we should expect AI to be safe, or that interventions to shift the curves don’t matter. Likewise, the considerations I mentioned with respect to synbio make us safer to some extent, and I was curious as to Logos’ evaluation of their magnitudes.
Okay, thanks for the clarification. If we would expect the magnitudes for synbio to be significantly higher (or lower) than for AGI, then I would be curious as to what differentiates the two situations (I could easily imagine that there is a difference, I just think it would be a good exercise to characterize it as precisely as possible).
ETA: Actually, I think there are some plausible arguments as to why AGI progress would be less relevant to AGI safety than one would expect naievely (due to the decoupling of beliefs and utility functions in Bayesian decision theory—being an AGI hinges mostly on the belief part, whereas being an FAI hinges mostly on the utility function part). But I currently have a non-trivial degree of uncertainty over how correct these arguments are.