[ivan.moony]

Sentient beings seem all about input/output. A being perceives some input and responds to that by some output. Then it is all together fed back to input, which generates some other output, and so on, while the being lives.

The current situation is that we have a way to train artificial neural networks on supercomputers that can be interpreted on average computers, providing an output for a certain input. But certain outputs are unacceptable, or ethically questionable, so it wouldn't be wise to always pick the first output from the stack. The solution might be in programming a counterpart to ANNs, which we could call "God". The God's purpose would be to judge if a certain output will be performed, or another output would be requested to judge about.

Now we have two entities to worry about: a sentient being and a God. The sentient being seems more-or-less solved, it could be simulated by ANN, but the God... That should be a real complication, shouldn't it? The question I would like to rise is: "How to program an universal God that handles any ANN trained instance?"

[frankinstien]

You mean a politically correct entity. 

Actually here's an example of how emotions work to solve that problem. Where the AI imprints on humanity, it forms a sense of devotion which then always looks out for the best interests of humanity, something like a sibling and/or mother and/or father notion of allegiance and concern.

[MagnusWootton]

Azimov's 3 laws of robotics would work, but are impossible to code,  maybe not impossible but its a very hard mission full of exceptions to rules.   The easy way to motivate things is to monitor the robots eye,   but anything that isnt just ordinary geometry,  some invisible concept, is impossible to detect.

[ivan.moony]

You mean a politically correct entity. 

Actually here's an example of how emotions work to solve that problem. Where the AI imprints on humanity, it forms a sense of devotion which then always looks out for the best interests of humanity, something like a sibling and/or mother and/or father notion of allegiance and concern.

Oh no, then it would be what we, humans want. But do we really want the right things for everyone? And more important, do we want to use our creation and throw it away when it is not useful to us anymore? Do we really want an artificial slave?

Azimov's 3 laws of robotics would work, but are impossible to code,  maybe not impossible but its a very hard mission full of exceptions to rules.   The easy way to motivate things is to monitor the robots eye,   but anything that isnt just ordinary geometry,  some invisible concept, is impossible to detect.

You know how ANNs could be trained separately on tone friendliness, and be filtered out to pass only the ones above a certain level? I believe the same could be done with Asimov's laws, or anything else. Then we would have one general ANN to shoot out action ideas as possible outputs, and the other one, the God, to filter and control those ideas. But as ANN is error prone, the filter may fail. And the God shouldn't ever fail.


More I think about it, if God should never fail, it should be aware of each entity it deals with, and the only way to do it is to have a strictly closed system, no interfacing to the outer world. I'm talking about a closed Little Universe, guesting AIs, and knowing all about them to make the right judgement about their actions. And if we don't really have a material use of that kind of Universe (because it is a closed system), if we really have no motives to do anything wrong to its inhabitants... what is left to us? I'm telling you, if such thing would be ever made, it would be a result of a pure love. But does merely a dream deserve such a love?


I better reduce these coffees, this is not going to end up well...

[MagnusWootton]

It could all disappear in an instant, and we'll never know why.

How is God even supposed to even know if hes invincibly immortal or not?   (more than immortal,  invincible...)   The whole thing could disappear in a puff of smoke,  or we all get banished to never ending torture.  hell on earth.   what do we do?  panic to hell about the possibilities we can never know?

About Asimov's laws,  you could get close to getting them to happen,  but there would be cracks in the system because your not going about it the correct way if you're just noticing things happening on its sensor,  its something else....

[ivan.moony]

ANNs are great at estimating algorithms, even our own behavior. When we get example outputs, if we answer enough times with yes/no, ANN will find out regularities and continue our agenda. But it will never be perfect, it will always be an estimation. So, if you are willing to risk that much, there you go, spend a couple of months/years filtering the output with red/green button yourself, and you can be considered a seed for the God for that ANN instance. Then you can stitch it to a web interface and see what it does across the web. It should be a kind of a copy of you, meaning, it will inherit all your virtues and flaws (beware of flaws).

[MagnusWootton]

Yes,  but we need an unflawed method,  its kinda important to do things properly, without a doubt.

[MikeB]

The training data is the Bible, the person who wrote it is the God of the AI...

[MagnusWootton]

The training data is the Bible, the person who wrote it is the God of the AI...

What if you just wrote the a.i. to put money on the table, continuously,  why super ai is possibly stronger than a stock market predictor, if you want to abuse the situation.

[frankinstien]

Oh no, then it would be what we, humans want. But do we really want the right things for everyone? And more important, do we want to use our creation and throw it away when it is not useful to us anymore? Do we really want an artificial slave?

Who is going to decide what morality to implement? I mean there are those that believe that if you think a bad thing then there's the possibility to act on that idea, so, to prevent rape, murder, physical abuses, we should as a society screen for thoughts and punish those that think such thoughts!

All machines are slaves, they are created to serve humanity.

[yotamarker]

I've tried something like that :
https://www.yotamarker.com/t46-yotamarker-artificial-intelligence-walkthrough-and-source-code?highlight=bees+knees#65

when the program feeds on its own output you simply jam it up
it cant learn, and it eventually loses "steam" and becomes silent

(.net makes me wanna puke)

[MagnusWootton]

computers require feedback from the output back to input to be a finished machine, otherwise your stuck as a feedforward perceptron.   (even tho that works anyway.)

when your preventing rape and abuse with abuse, it isnt the solution...

[ivan.moony]

Abuse is a common answer to abuse. I wish there is some another way.

[HS]

I think a sentient/conscious ANN would be better off learning from the feedback of the environment itself. Having an unconscious program limiting the actions of a conscious being seems like it would be dystopian. But if we're talking about an ANN that's just an idea generator, then a "God" could be a good solution. Maybe the "God" could apply a version of Maslow's Hierarchy to each sentient being within its sphere of influence. Then it could let through outputs from the ANN with a high probability of optimizing each estimated hierarchy.

[LOCKSUIT]

Sentient beings seem all about input/output. A being perceives some input and responds to that by some output. Then it is all together fed back to input, which generates some other output, and so on, while the being lives.

The current situation is that we have a way to train artificial neural networks on supercomputers that can be interpreted on average computers, providing an output for a certain input. But certain outputs are unacceptable, or ethically questionable, so it wouldn't be wise to always pick the first output from the stack. The solution might be in programming a counterpart to ANNs, which we could call "God". The God's purpose would be to judge if a certain output will be performed, or another output would be requested to judge about.

Now we have two entities to worry about: a sentient being and a God. The sentient being seems more-or-less solved, it could be simulated by ANN, but the God... That should be a real complication, shouldn't it? The question I would like to rise is: "How to program an universal God that handles any ANN trained instance?"

You may have seen GPT-3, DALL-E, JUKEBOX, CODEX on openAI.com, and now NUWA made by Microsoft, and you probably have not had the time to fully understand how they work and how legit they are, but if you also see Facebook's Blender, you would see it is GPT-X but has the ability to guide what to say and avoid what to say. One similar project called it taming, or forcing the model. It's not god, it's just another part to the ANN. Just like embeds, recency priming, etc.