While we are at it:
caffeine
meditation
music
mood
social interaction
Also, which hacks are available to better interface our mental hardware with the real world:
information presentation
automated information filtering
While we are at it:
caffeine
meditation
music
mood
social interaction
Also, which hacks are available to better interface our mental hardware with the real world:
information presentation
automated information filtering
I have this idea in my mind that my value function differs significantly from that of Elizier. In particular I cannot agree to blowing up Huygens in that Baby-Eater Scenario presented.
To summarize shortly: He gives a scenario which includes the following problem:
Some species A in the universe has as a core value the creation of unspeakable pain in their newborn. Some species B has as core value removal of all pain from the universe. And there is humanity.
In particular there are (besides others) two possible actions: (1): Enable B to kill off all of A, without touching humanity, but kill off a some humans in the process. (2): Block all access between all three species, leading to a continuation of the acts of A, kill significantly less humans in the process.
Elizier claims action 1 is superior over 2, and I cannot agree.
First, some reason why my intuition tells me that Elizier got it wrong: Consider the situation with wild animals, say in Africa. Lions killing gazelles in the thousands. And we are not talking about clean, nice killing, we are talking about taking bites out of living animals. We are talking about slow, agonizing death. And we can be pretty certain about the qualia of that experience, by just extrapolating from brain similarity and our own small painful experiences. Yet, I don’t see anybody trying to stop the lions, and I think that is right.
For me the only argument for killing off special A goes like: “I do not like Pain” → “Pain has negative utility” → “Incredible pain got incredible negative utility” → “Incredible pain needs to be removed from the Universe” That sounds wrong to me at the last step. Namely, I feel that our value function ought to (actually does) include a term which discounts things happening far away from us. In particular I think that the value of things happening somewhere in the universe which are (by the scenario) guaranteed not to have any effects on me, are exactly zero.
But more importantly it sounds wrong at the second to last step, claiming that incredible pain has incredible negative utility. Why do we dislike our own pain? Because it is the hardware response closing the feedback loop for our brain in the case of stupidity. It’s evolution’s way of telling us “don’t do that”. Why do we dislike pain in other people. Due to sympathy, i.e. due to reduces efficiency of said people in our world.
Do I feel more sympathy towards mammals than towards insects, yes. Do I feel more sympathy towards apes than towards other mammals, again yes. So the trend seems to indicate that I feel sympathy towards complex thinking things.
Maybe that’s only because I am a complex thinking thing, but then again, maybe I just value possible computation. Computation generally leads to knowledge, and knowledge leads to more action possibilities. And more diversity in the things carrying out computation will probably lead to more diversity in knowledge, which I consider A Good Thing. Hence, I opt for saving species A, thus creating a lot more of pain, but also some more computation.
As you can probably tell, my line of reasoning is not quite clear yet, but I feel that I got a term in my value function here, that some other people seem to lack, and I wonder whether that’s because of misunderstanding or because of different value functions.
I disagree strongly with you about who is better, if there are artificial pyramids on Mars.
If you have received sufficient evidence to be pretty certain, you are acting rational on rejecting the notion of pyramids on Mars up to the point (and only up to there) you receive more convincing evidence to the fact that there are pyramids on Mars. In that case you should a) gather more evidence and decide which point is correct b) switch pretty immediately in case there really are pyramids.
In particular I claim you have basically zero evidence against martian pyramids except the general heuristic of occam’s razor.
Also, abstaining from making public your (uninformed) opinion on martian pyramids would reduce you credibility loss in case there are any.
Finally, science will not just turn to “wrong” just because there are martian pyramids, most of it still stands as it is.
The reason you can’t profit on your possibly newly aquired knowledge of the local lake is that you don’t know whether the other traders did or did not already include said lake in their calculations.
However, long term traders know which information they did factor in and would probably be interested in buying the information from you, iff it is new information to them. So, they should in theory be willing to pay you a price for your piece of information, even if you don’t have the global view of the market.
And I think there is a way for you to profit, given that the transactions on the market are essentially cost-less. Namely, bet on whatever your information predicts, totally irrespective of what the market looks like, then publish your information to all traders. Wait for the market to settle at the new value (now guaranteed to include you info), zero out your positions.
If your information was already known to all traders, well, you have to zero your positions without any gain, and if transaction costs are zero, also without loss. (Random market fluctuations should even out).
Fun fact—better strategy for memory: You play memory 1 vs. 1, and it’s your move.
If you know a pair, take it (this much is obvious).
Chose an unknown tile at random.
If you know the match take it (also clear).
Take a tile known to both of you (if there is doubt, take one your oponent knows).
People don’t do step 4 right usually.
Apparently, the term you are searching for is “Second Order Probability”.
See here for a paper: www.dodccrp.org/events/2000_CCRTS/html/pdf_papers/Track_4/124.pdf
Why do you count (ROC, Data) Pairs?
Clearly, if all people rely on the same data, but all use different (but quite sound) cognitive rituals, there is still only one data set.
I’d think that you should first compute the meaning of the data by averaging over all (apparently sound) ROC results, and then update based on that outcome. I.e. if only lunatics saw something, and they all say it means A, then that counts for nothing. If a bunch of trustworthy bayesians see something and they all conclude B, then that counts like one trustworthy bayesian for B. If some trustworthy bayesians and a similar number of (apparently “winning” ⇒ implying sound ROC) aliens, who deny to update a la Bayes, say it’s B there is still one vote only.
If the aliens and the bayesians saw different data though, that’ll make two votes.
Actually “If A then B. And B. Hence A.” is probably a bias in your very sense, namely heavily selected for due to heuristic value.
Example: You know: “Approaching Danger ⇒ Friendly People Crying” You perceive: “Friendly People Crying”. Do you think there is danger?
Well, the friendly people might be practical jokers, on the other hand, people who take the “might be practical jokers” train of thought have a tendency to get their genes removed from the pool.
Anybody got a good reason for adopting a certain utility function versus some other one?
Because I can’t find one, and now I feel weird, cause without a decent utility function, rationalism gives you knowledge, but no wisdom.
“Buffalo buffalo buffalo buffalo buffalo buffalo buffalo buffalo.” Although the ambiguity is all in the syntax.
Actually, I disagree with some possible connotations of “better” here. It might be that many thoughts are most accurately expressed using ambiguous sentences, but that is probably because the thoughts are ambiguous as well (and I see that as pretty likely).
Example: “I’d like some icecream.” seems quite like a typical thought. I have thought of that form as qualia, and they seem kind of useful. However, they are ambiguous all over the place.
They omit a lot of (my) default assumptions: I’d like it now. I’d like it for eating. I’d like to consume it in company of some peers. I’d best like citron.
They imply (but not state) some course of action.
They imply (but not state) a request for other people around me to join in my pursuit for icecream.
They signal (if uttered) a lot of stuff to the people around, e.g. relaxedness
But “better” also has the connotation of “optimal” in a utilitarian sense. And it might be that thoughts should best not be expressed in ambiguous terms, because that creates weird priming-effects which depend on your culture. If you think “I’d like some icecream” and your language lacks a proper distinction between icecream-type-A and icecream-type-B you are priming also for the wrong icecream type everytime you process that thought.
Now, one might come to the conclusion that you should just avoid verbalizing thoughts, but that will make communication difficult to say the least, as communication will in general involve passing categories of things. But maybe verbalizing a bit less internally might remove some of the cultural effects our respective languages bring.
Slide 24: “your the map” is probably a typo.
In a perfectly rational agent, no n-th order wants should exist.
Your problems with mountain dew might account for −1 util, you being awake for 2 utils, then you “want” to drink that stuff. Shut up and add.
The only source of multi-level desires I can see is an imperfect caching algorithm, which spews forth “Do not drink mountain dew” although the overall utility would be positive.
Regarding most of the lengthy examples of “philosophy” given by Stove:
Reading a text takes time, time can be spent acquiring utilions. Hence reading a text is only worth if the expected utilion win due to additional knowledge is grater than the expected utilions when using the time differently. This approach kills most of his examples dead in their tracks for me. This also implies positivism, if a text does not either generate utilions directly, i.e. fun reading fiction, then it needs to provide knowledge (in form of testable statements about the world), otherwise, how would I generate utilions from the “knowledge”?
Possibly, some thoughts are only valuable when more efficient methods of communication become available.
Solitary walks (if truly solitary) should not be usable for signaling, at least not directly.
What you might get from it with respect to signaling, is authentic experience you can later use while talking about solitary walks and thereby signaling.
But I feel something else might be at work as well. I claim (although I have only anecdotical evidence) that one’s brain also responds to signals from oneself, i.e. by observing that you do solitary walks, you start to believe that you are the kind of person doing solitary walks, thereby increasing your chances of successfully convincing others that you are the kind of person doing solitary walks.
Martial Arts training might be a candidate, not the choice of doing it, but where and when to punch / kick exactly. You signal via success of your strategy, but I think you do not signal about which way you achieved your success. This might obviously be different for different styles of MA, I only have experience with Ju-Jutzu. Also, I think this example might be interesting, because the mental effort is high compared to some of the other examples.
Regarding Nose-picking, I think it is a quite signal-high activities if done in public. I’m not sure what Nose-picking signals though, but since nearly all parents constantly tell their children not to do it in public, the probability of it having signaling effects is extremely high. So at the very least, not doing it signals “I have a good upbringing”.
I think transfer of mental models is often crucial for efficiently learning new procedural knowledge. Your skin oil problem is only a small example. Consider for example the task of adding up to numbers, the task of computer programming, or the task of designing a mechanical structure. Once you have been told about decimal places and carry, about memory and variables, about stress and strain and whatnot, reasoning about your problems suddenly becomes a whole lot easier and more successful.
Depending on the problem domain, supervised learning from an expert or the procedure in the post might be more appropriate.
Rice’s theorem says nothing about the impossibility to show properties of programs in special cases. It only shows that there are always some programs which will thwart your decision procedure. But your possible AI does not need to be one of them.
So where do we go? We go kicking Mu out of the picture:
For the most interesting aspect of an intelligent system is whether it is efficient in the very world it will occupy. So a reasonable test of it’s intelligence is to run it in a detailed simulation of said world. Of course, if your simulated world and reality differ (for example in the existence of Mu), the measure is pretty much worthless.
However, decent simulations of our world can be built.
Also, the solution to your puzzle is to copy the world (including Mu) into a new Matrix, testing whether the program reaches it’s goal, and using that as your efficiency measure. That way Mu cannot test the program’s efficiency either, and you get a fair chance.
How to design utility functions for safe AIs?
Make a utility function which will only emit positive values if the AI is disabled at the moment the solution to your precise problem is found. Ensure that the utility function will emit smaller values for solutions which took longer. Ensure the function will emit higher values for world which are more similar to the world as it would have been without the AI interfering.
This will not create friendly AI, but an AI which tries to minimize its interference with the world. Depending on the weights applied to the three parts, it might spontaneously deactivate though.
Minimize is not “reduce to zero”. If the weighting is correct, the optimal outcome might very well be just the solution to your problem and nothing else. Also, this gives you some room for experiments. Start with a function which only values non-interference, and then gradually restart the AI with functions which include ever larger weights for solution finding, until you arrive at the solution.
I’d be similarly interested in covering philosophical Daoism, the path to wisdom I follow, and believe to be mostly correct.
Things they get wrong: Some of them believe in rebirth, too much reverence for “ancient masters” without good reevaluation, some believe in weird miracles.
Things they get right: Meditation, purely causal view of the world, free will as local illusion, relaxed attitude to pretty much everything (-> less bias from social influence and fear of humiliation), the insight that akrasia is overcome best not by willpower but by adjusting yourself to feel that what you need to do is right, apparently ways to actually help you (at least me) with that, a decent way accept death as something natural.