Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

kurt comments on Are wireheads happy? - Less Wrong

109 Post author: Yvain 01 January 2010 04:41PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (95)

You are viewing a single comment's thread. Show more comments above.

Comment author: kurt 30 December 2016 04:41:17PM 1 point [-]

One could also note that 'liking' is something the mammalian and reptile brains are good at, and 'wanting' is often related to deliberation and executive system motivation. Though there are probably different wanting (drive) systems in lower brain layers and in the frontal lobe.

Also, some things we want because they produce pleasure, and some are just interim steps that we carry out 'because we decided'. Evolutionary history determines that we get rewarded when we get actual things far more than when we make one small step. We can sometimes utilize will power, which could be seen as an evolutionarily novel and therefore weak mechanism to provide short term reward for executing steps towards a longer term benefit. Steps that are not rewarded by the opioid system.

I think that the more we hack into the brain and the more we will discover that 'wanting' and 'liking' are umbrella terms completely unsuited to capture the subtlety and the sheer messiness of the spaghetti code of human motivation.

I am also going to comment on the idea that intelligent agents could have a 'crystalline' (ordered, deterministic) utility evaluation system. We went down that road trying to make AI work, i.e. making brittle systems based on IF/THEN - that stuff doesn't work.

So what makes us think that using the same type of approach will work for utility evaluation (which is a hard problem requiring a lot of intelligence)?

Humans are adaptable because they get bored, and try new things, and their utility function can change, and different drives can interact in novel ways as the person matures and grows wiser. That can be very dangerous in an AI.

But can we really avoid that danger? I am skeptical that we will be able to have a completely bayesian, deterministic utility function. Perhaps we are underestimating how big a chunk of intelligence the evaluation of payoffs really is; and thinking that it won't require the same kind of fine-grained, big-data type of messy, uncertain pattern smashing that we now know is necessary to do anything, like distinguish cars from trees.

We have insufficient information about the universe to judge the hedonic value of all actions in an accurate way, that's another reason to want the utility evaluation to be as plastic as possible. Chaos must be introduced into the system to avoid getting caught in locally optimal spaces. Dangerous yes, but possibly this necessity is what will allow the AI to eventually bypass human stupidity in constructing it.