Love it! I’ve been thinking a lot recently about the role of hedonics in generally intelligent systems. afaik we don’t currently try to induce reward or punishment in any artifically intelligent system we try to build, we simply re-jig it until it produces the output we want. It might be that “re-jigging” does induce a hedonic state, but I see no reason assume it.
I can’t imagine how a meta optimiser might “create from scratch” a state which is intrinsicallyrewarding or adversive. In our own case I feel evolution must have recruited some property of the universe that was already lying around, something that is just axiomatically motivating to anything concious in the same way that the speed of light is just what is it.
“Pleasure” seems to be any kind of signal that causes currently engaged behaviors both to continue and to be replicated in similar situations in the future. Think of the “GO” pathway in the basal ganglia. (By “situation”, I mean some abstract pattern in the agent’s world model, corresponding to features extracted from its input stream, that allows it to reliably predict further inputs. Similar situations have similar statistical regularities in these patterns.)
Conversely, “pain” would be any signal that causes both the cessation of currently engaged behaviors and a decrease in probability of replicating those behaviors in similar situations in the future. Think of the “NOGO” pathway in the basal ganglia.
There is no fundamental law of physics underlying pain and pleasure for evolution to recruit. Nothing in the chemistry of dopamine makes it intrinsically rewarding; nothing in the chemistry of substance P makes it intrinsically painful. These molecules simply happen to mediate special meta-computations in the brain that happen to induce certain adjustments in behavioral policies that we happen to have labeled “pleasure” and “pain”. They’re just mechanisms that evolution stumbled upon that happened to steer organisms away from harm and toward resources, discovered by trial and error and refined through adversarial optimization over eons. If there’s some fundamental feature of reality being stumbled upon here, it probably has something to do with the statistics of free energy minimization, not any kind of “axiomatically motivating” “stuff”. There are No Universally Compelling Arguments, and reward and punishment signals only work for systems Created Already In Motion.
Love it! I’ve been thinking a lot recently about the role of hedonics in generally intelligent systems. afaik we don’t currently try to induce reward or punishment in any artifically intelligent system we try to build, we simply re-jig it until it produces the output we want. It might be that “re-jigging” does induce a hedonic state, but I see no reason assume it.
I can’t imagine how a meta optimiser might “create from scratch” a state which is intrinsically rewarding or adversive. In our own case I feel evolution must have recruited some property of the universe that was already lying around, something that is just axiomatically motivating to anything concious in the same way that the speed of light is just what is it.
“Pleasure” seems to be any kind of signal that causes currently engaged behaviors both to continue and to be replicated in similar situations in the future. Think of the “GO” pathway in the basal ganglia. (By “situation”, I mean some abstract pattern in the agent’s world model, corresponding to features extracted from its input stream, that allows it to reliably predict further inputs. Similar situations have similar statistical regularities in these patterns.)
Conversely, “pain” would be any signal that causes both the cessation of currently engaged behaviors and a decrease in probability of replicating those behaviors in similar situations in the future. Think of the “NOGO” pathway in the basal ganglia.
There is no fundamental law of physics underlying pain and pleasure for evolution to recruit. Nothing in the chemistry of dopamine makes it intrinsically rewarding; nothing in the chemistry of substance P makes it intrinsically painful. These molecules simply happen to mediate special meta-computations in the brain that happen to induce certain adjustments in behavioral policies that we happen to have labeled “pleasure” and “pain”. They’re just mechanisms that evolution stumbled upon that happened to steer organisms away from harm and toward resources, discovered by trial and error and refined through adversarial optimization over eons. If there’s some fundamental feature of reality being stumbled upon here, it probably has something to do with the statistics of free energy minimization, not any kind of “axiomatically motivating” “stuff”. There are No Universally Compelling Arguments, and reward and punishment signals only work for systems Created Already In Motion.
Thanks for your reply :) as in many things, QRI lays out my position on this better than I’m able to 😅
https://www.qualiaresearchinstitute.org/blog/a-primer-on-the-symmetry-theory-of-valence