But the real problem with satisficers isn’t that they “want” to become maximisers; the real problem is that their behaviour is undefined.
Unless they are maximizers with complicated utility functions. There are NP-hard problems where we can get within 5% of the optimum in polynomial time. You can be a satisficer happy with 95% of the solution. On the other hand, you can incorporate computer time into your utility function and maximize accordingly.
For example, let A be a satisficing agent. It has a utility u that is quadratic in the number of paperclips it builds, except that after building 10100, it gets a special extra exponential reward, until 101000, where the extra reward becomes logarithmic, and after 1010000, it also gets utility in the number of human frowns divided by 3↑↑↑3 (unless someone gets tortured by dust specks for 50 years).
Using units of utility here is going to throw you off. By definition, a unit of utility already takes into account all of A’s preferences. For example, if A got paid in dollars per paperclip, then we could weigh the benefit of making extra paperclips against A’s utility in taking the night off and heading off to the pub. However, this would make less sense if we assumed quadratic returns on utility.
All these actions are possible for a satisficer, and are completely compatible with its motivations. This is why satisficers are un(der)defined, and why any behaviour we want from it—such as “minimum required” impact—has to be put in deliberately.
Why would A do any of those other things unless it gained some utility thereby? If A had a utility function that valued writing Harry Potter fanfic but didn’t want to lose out on the paycheck from the paperclip job, standard utility theory would predict that A expend the minimum effort on paperclips necessary to keep its job. A’s marginal utility decreases sharply after passing this minimum and would explain the satisficing.
Unless they are maximizers with complicated utility functions. There are NP-hard problems where we can get within 5% of the optimum in polynomial time. You can be a satisficer happy with 95% of the solution. On the other hand, you can incorporate computer time into your utility function and maximize accordingly.
Using units of utility here is going to throw you off. By definition, a unit of utility already takes into account all of A’s preferences. For example, if A got paid in dollars per paperclip, then we could weigh the benefit of making extra paperclips against A’s utility in taking the night off and heading off to the pub. However, this would make less sense if we assumed quadratic returns on utility.
Why would A do any of those other things unless it gained some utility thereby? If A had a utility function that valued writing Harry Potter fanfic but didn’t want to lose out on the paycheck from the paperclip job, standard utility theory would predict that A expend the minimum effort on paperclips necessary to keep its job. A’s marginal utility decreases sharply after passing this minimum and would explain the satisficing.