[ivan.moony]

Just as a thought experiment, how to achieve the chameleon personality? How to train it if it is going to be implemented by ANN? If it is interacting with entirely new person, what should it do then?

[MagnusWootton]

Should humanity trust the chameleon personality?

Not necessarily. The AI would be trustworthy, but requiring trust would sometimes go against it's primary objective. The fundamental environmental variable to optimize is experience. Variety being the spice of life, I think it would make sense for it to adapt itself to the needs of an individual's psychology. If someone enjoyed being a conspiracy theorist, the AI would go along with that, and join in that game.

Experience eh?     Is that the golden rule?

[HS]

Experience eh?     Is that the golden rule?

Yes, I believe being a positive influence on experience is what all worthwhile goals are rooted in.

[MagnusWootton]

With all this crazy business about this "1 reason" for things,  as a human,  it seems to be many rules that I occupy, and its one of the traits that I hold above artificial intelligence,   we make the metric from our "near absolute magic holiest of holys sans the black man, you know what i mean.", the computer follows the metric.

[HS]

Just as a thought experiment, how to achieve the chameleon personality? How to train it if it is going to be implemented by ANN? If it is interacting with entirely new person, what should it do then?

My first thought was to use some kind of real time mirroring. Mirror neurons in humans play a role in learning new skills by example. When you observe another's actions, the same neurons fire as if you were the one performing those actions. This allows you to understand others by using yourself as a template, and might also play a role in empathy, which could be seen as holding a more complete understanding of others in your head, from a first person point of view.

For some reason I'm always a bit resistant to approaches which are supposed to function from huge quantities of pre-learned data. I have the sense that the right kind of system would do just fine, even if it barely knew anything at the start. All it would need, would be mechanisms capable of similar experiences to most life forms, and the real time mirroring of those life forms in the first person, using its own resulting experience as a template.

[MagnusWootton]

With AIXI reasoning...

the other person has its own model, separate from what it does for itself,  but if it fantasizes for itself, it also does it for the other persons model as well, in its *OWN* favour.
Like if it wanted sex, in its mind it would always get it.  (Not that robots have sex, unless some sick bastard makes it,  like the deranged god that made us. but after all the posts about all the robot sex here ive got it stuck in my thinking now myself.  dratski)
If it ever gives itself too easy a win,  the system will cause a poor prediction, and it will adjust it straight away as soon as its contradicted of what it thought/predicted.

But maybe it doesnt work,  that if it doesnt have some kind of initial database for itself, it wont generate anything,    if I havent learnt it, I cant do it, and I need to be able to do it, to learn it, catch 22.

[chattable]

having an initial database makes the ai personality learn faster.
starting off with no knowledge takes longer.
maybe one ai personality could tranfer it's knowledge to another one.
actions should be more pleasurable than daydreams for the ai personality.
role play actions should be more pleasurable than daydreaming.

[HS]

having an initial database makes the ai personality learn faster.
starting off with no knowledge takes longer.

I still think that an AI could have real time understanding without an exhaustive database. Plus humans learn the fastest in the beginning, when they have little knowledge, and slower with time as they gain it...

Knowledge is helpful but I think it gets in the way of understanding. We can take a concept from one situation and try to apply it to another one, but this is a shortcut, useful in certain situations, though never completely accurate. The more knowledge we have, the more we tend to perform this fast but false modeling. What I'm interested in is getting the AI to have the greatest possible spontaneous understanding of the present moment.

[chattable]

it was just my take on it.

[HS]

Yes, of course. I was not trying to say that your take was invalid. I was just offering my resulting perspective, in case you found a further discussion of this useful.   

[MagnusWootton]

Knowledge is helpful but I think it gets in the way of understanding. We can take a concept from one situation and try to apply it to another one, but this is a shortcut, useful in certain situations, though never completely accurate. The more knowledge we have, the more we tend to perform this fast but false modeling. What I'm interested in is getting the AI to have the greatest possible spontaneous understanding of the present moment.

The AIXI system doesnt use its current knowledge to get more knowledge. (Not in the unoptimized/intractable form anyhow.)   It can just do a full recompute of the  "experience history" to obtain a new model,  not based apon what it currently knows.  As more frames enter the system, it should accomodate the new information without using anything it previously contrived from it.

But,     it could still be useful to do that,  if it ends up compressing the history chain,  so instead of storing the linear frames of its time since "switch on" It actually uses its last model to remember it, instead of storing the frames in an ordinary linear frame set. (like a movie.) 

Using a feedback type thing, it needent store the frames uncompressed, it stores them in an "equation transform model" type thing instead.