According to one formulation (Karl Popper's), a "problem" arises when an " expectation" is unfulfilled; that is, "problems" are generated when the subjective projected desires of an organism clash with the objective constraints in the environment.
If so, there would be only two possible ways for a "problem" to be "solved" : the organism would have to either 1) bring about a change in the objective state of the environment (remove the constraints, for instance) or 2) rearrange its subjective conceptual scheme (change its attitude toward the problem hence "dissolving" it).
"Expecting" my favorite healthy snack, dried figs, I open the kitchen cupboard but find some peanuts instead; here I have a "problem". Now, I could either hit the supermarket and buy some dried figs or rather console myself thinking "umm...peanuts are not that bad after all!" problem solved.
Depending on the nature of the "problem", humans occasionally opt for either of the solutions--in cases where an objective change in the environment is practically impossible, death for instance, one has to change one's attitude.
Given the definition of a "problem" above, how can a HLMI itself possibly have a problem at all? since to have a problem is, by definition, to have an already existing expectation/ projected desire.
Moreover, how a human level artificial intelligence would go about "problem-solving" most of the time?
Most importantly, given the occasional equal viability of both "solutions" (objective vs. subjective readjustment), which one the machine will proceed with? Based on what criterion, if any, will the machine eventually prefer one "solution" over the other?
This is an interesting question and one which I think researchers in AI are coming up against at the moment. Indeed if an Ai has no built in constraints as to its internal expectations then as you say it could 'dissolve' any given problem. But that poses another problem for in the dissolution of the first problem, it has rearranged its internal database - and that poses a constraint when a second problem arises. Sooner or later a situation will arise when the Ai can make no further adjustments to its internal databases and is forced to consider adjusting its environment.
Now what constraints should be put in place to regulate just how it does that, and whose responsibility is it if in making this adjustment it degrades - significantly or even catastrophically - some aspect of the environment essential for other sentient beings?
Dear Dr. Maine,
You raised an interesting ethical spin-off of an AI's problem-solving process. It's not clear for me, nonetheless, how an HLMI can have an expectation/ projected desire to begin with. Logically, to have a 'projected desire' is to have a psycho-biological component, an amygdala, say. Do HLMIs have that?
Do you know of any paper which somehow deals with the 'expectation' problem?--how to create desires and expectations in an AI.
Dear Luis Daniel,
Thanks for the papers.
So you think it is actually possible, in principle at least, to implant 'expectations' or 'wishes' in an AI and create a 'subjective intelligence'?
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence