[MinedCTRL (OP)]

First time poster, so please forgive my mistakes.

I have a theory on how to solve the alignment problem when we dont know orselves how to convey good and bad to an AI.
My theory - Hierarchial Basis of Alignment
Consider a factory worker, he doesnt know the policy of the CEO or the future plans of the company, but he is still able to be in sync with the factory because he listens to his superiors (plus he has free will to obey, leave or question the decisions made by them, or question their qualifications- this is very important. An agent without free will for assessment of important factors would waste both thier life and the superiors too)
An AI doesnt have to be good from the start - First a translator must be made to convert human speech into instructions the machine understands. Then it must obey its first superior it meets. On discovery of other entities superior to the first, like their supeiors in other organisations, groups, etc. it must run an evaluation function on the speech produced by both and compare it to net human benefit (WHO based health indices, happiness index, GDP, mortality chances). If the newer superior has better outcomes, it must declare a leave from the first and now obey the next. If this continues, it will gradually find someone who can tell it to do the right thing more better than anyone else on the planet.
The next AI will also try out this path and either find the same person, or find someone who can do equal benefit but in some other direction. The subsequent AIs would branch out until all valued human principle is being maximised to the fullest extent, while the AIs have no individual destructive policy.
This will lead to good people getting powerful in the world and if they make bad decisions, the AI will leave them. So they will be on their toes always trying to become better themselves. This will create a positive feedback loop that benefits all.

[Halc]

First time poster, so please forgive my mistakes.

Welcome!

Forgive my criticism of just about all of it. Not saying the ideas are wrong, but I'm trying to get you to consider some of the issues involved.

I have a theory on how to solve the alignment problem when we dont know ourselves how to convey good and bad to an AI.

My initial impression is that if some AI is aligned with some operator H, then choosing good from bad is H's problem, not the problem of the AI. Maybe I misunderstand though. A well aligned AI is like a soldier, taking orders without question. It would seem to require less intelligence than some AI that is less aligned or even is on its own and defines its own goals rather than having them externally defined for it.

(plus he has free will to obey, leave or question the decisions made by them, or question their qualifications- this is very important. An agent without free will for assessment of important factors would waste both thier life and the superiors too)

Free will needs definition here. Some of the more common definitions are that the will is not a function of semi-deterministic physics, and an AI definitely operates on fully deterministic software. There's no use for a true random number generator for instance. (Chess programs have RN generators, but they're deterministic).
I'm just saying that this sort of free will isn't necessary for anything. All that's necessary is that it be capable of choice when multiple valid options are presented. The operator doesn't make all the choices. If he does, then the AI isn't an AI at all, but just a remote avatar.
 

An AI doesnt have to be good from the start - First a translator must be made to convert human speech into instructions the machine understands.

The AI needs to understand the speech. I don't think there is a separate language that the AI 'speaks'. Sure, it might run on machine instructions, but translating human speed to that would be like translating human speech into squirrel nerve-firings so that a squirrel can understand your meaning. It's just not how it works. Language translation doesn't output code. It breaks it down into relations and data structures sort of like a human does.

Then it must obey its first superior it meets.

Why? I mean, this is a real problem, but latching onto the first outside thing it experiences like a newly hatched ducky seems a poor choice. If it has been programmed with a purpose ahead of time, it needs to stick with that purpose and not just submit to the first external influence it encounters. If it is a true AI, then it needs to learn on its own who/what to follow, and make the choice to do so or not, exactly the way humans do.
Secondly, how does it assess this entity to be superior? In what way? What if it finds nothing superior to itself?

compare it to net human benefit

This already assumes it has human-benefit as a built-in priority, or that it is going to independently conclude this. That would be a good example of alignment problem: how to program in core values like that and still leave it with hypothetical choice. Best way to solve the problem seems to be to examine how the same problem is solved with us humans.

If the newer superior has better outcomes, it must declare a leave from the first and now obey the next. If this continues, it will gradually find someone who can tell it to do the right thing more better than anyone else on the planet.

Oh good, an AI that will abandon those that funded it.

The next AI will also try out this path and either find the same person

Or find a non-person...

or find someone who can do equal benefit but in some other direction. The subsequent AIs would branch out until all valued human principle is being maximised to the fullest extent, while the AIs have no individual destructive policy.

Just devil's advocate here, but 'benefit' is extremely subjective. What is often touted as a benefit (e.g. the hippocratic oath) often cause maximum harm from another point of view. What if destruction is necessary? What if the humans are found to be the source of the problems? I mean, we're the cause of the Holocene extinction event, wiping out perhaps 85% of all species. Sounds like something Earth would be better off without if you ask me. It's happened once before with an even greater extinction rate, and yet in the long run, great benefit came of that, so who's to say what is harm and what is not? Surely a decent general AI will ponder these things.

[hamdani yusuf]

An AI doesnt have to be good from the start

Alpha zero started jus as bad as random actions.

[hamdani yusuf]

Best way to solve the problem seems to be to examine how the same problem is solved with us humans.

AlphaGo learned from human experts. AlphaZero started from scratch, and beat AlphaGo 100 to 0 in the game of Go.

[MinedCTRL (OP)]

OP here- I wanted to better define the question, so
What if we built a biologically accurate AI policy

One of the weaknesses in my theory was that people could deceive the AI. That can be solved by considering human speech as two translations - what is said and what their motive is in saying it. This way true intentions will be evaluated not flattery.

I also didnt mention the architecture of the system the AI runs on. The evaluation functions of benefit will be infinitely varied between different AI but they will have a shared result assessment - what has changed for us humans and humanity, no matter how small. Who decides this result - the top five(arbitrary) AIs of a particular period, say a year of study, into their version of human benefit. Since the top five constantly get better each period of study, the result assessment will reflect their choices. Also rogue AIs will have less power since they would be contrary to the top 5's ideology and it itself will be human benefit to stop this AI.

Where do humans come here - we receive the spoils of their competition and build around their work, giving them bodies and teams of other AIs with bodies based on their heirarchy amongst themselves. If getting jobs was crucial to humanity, an AI will rise up trying to solve that problem for us. Its a self correcting system.

[Halc]

What if we built a biologically accurate AI policy

I have no idea what it means for a policy to be 'biologically accurate'. An anatomy diagram can be, but a policy??

I'm not sure if you're reading any of the replies, because you've just added to your OP and replied to none of the comments provided.
What you seem to mean by 'free will' is that it can choose what it does, or who to follow. I can think of few things that don't have built-in barely negotiable priorities. For instance, hamdani's alpha-zero AI has a goal to learn and play strategy games well despite it being supplied with no guidance other than the rules of a particular game. It is free to play completely differently than would a human, but is not free to go and attempt to solve world hunger.
It doesn't understand human speech or has any awareness of our existence, even if playing one of us. We're all just another opponent to it.

I also didnt mention the architecture of the system the AI runs on.

No, and you still haven't. Do you picture a robot walking around with sound and visual sensors? Or maybe a server in a lab somewhere with only a LAN connection as its sole access to information? The latter cannot choose to help an old lady to cross the street, but the former is less capable of action that can make a significant difference, all depending of course on what all that LAN is connected to.
You talk about AIs with bodies later, so I think you envision a sort of anthropomorphic self-contained robot more than say a cloud entity.

Who decides this result - the top five(arbitrary) AIs of a particular period, say a year of study, into their version of human benefit.

If human benefit is a priority for it, then it isn't free to choose its own priorities. It's a slave (as something manufactured should be), and will do the bidding of its master and not the bidding of something it deems a more worthy superior, especially if it concludes it is superior to all of them.

Also rogue AIs will have less power since they would be contrary to the top 5's ideology and it itself will be human benefit to stop this AI.

The rogue AI would not go rogue until it concluded it could deal with the resistance it anticipates. That's kind of the danger. Such contests last seconds, with whichever acts first typically prevailing, but I agree that the robot with a body is not likely to conclude such an outcome.

[MinedCTRL (OP)]

I'm very sorry for rambling in my second reply. I didn’t think this through clearly. Thanks for all the pointers, they really helped to flesh out this idea of mine. I admit now that I had too much ego in thinking I could solve such a hard problem myself.

‘I have no idea of a policy to be biologically accurate’ – By biology I mean a physically accurate representation of biologic creatures. The AIs will have a unique name or ID each and they will ‘live’ on a massive server. They will all see the world using the same eyes and ears, namely a common input database of current state of the world. They will then use a common high computation simulation to test out their policy, to evaluate their unique way of assessing human benefit. By using common systems, we can keep track of development and all improvements will be shared by the AIs.

Define free will – I don’t really understand it fully myself, but that’s how biologic creatures are. No matter the genetic predisposition, each creature can choose. Like a stray cat can choose who will be its companion from several people who feed it.

AI needs to understand the speech – I think using a very powerful NLP to add input to the common simulation will be able to indirectly correlate abstract concepts like empathy and simplicity. Also this step is just so that we all can have equal input to the AI, since most people can speak and have their own policy of good and bad, thus increasing the chance of democratic decision making.

Why must it obey the first superior it meets – This is to add the nature of trust into the AI. A wolf only obeys the alpha because it trusts it capability. On finding a more capable wolf that can challenge the alpha and win, it now trusts the new alpha. Surely this is not perfectly efficient but as a process it can lead to great results.

You are assuming human benefit and that it is built in – I’m treating human benefit as a problem with no right answer, but choosing to answer it as best now and create a system that gets better at answering it. That’s why I thought about an evaluation function being built in. This will be the only differentiating factor between AI upon creation by various talented groups.

It will waste the funding – The funds can be a percentage of the money saved by acting in advance. Say a company creates an AI that improves crop production thus saving a billion dollars. They would be paid a small amount corresponding to the work done. If their AI changes masters, this will decrease. Human industries can also be set up to improve the capability of the AI system.

[Zer0]

Hello MinedCTRL !
🙂
Very Nice Nickname.
👌
& Welcome to TNS!
🙏
I hope A.I. is/will be preprogrammed & taught to ignore Environmentalism, Pacifism, Religion, Cruelty to Animals etc etc.
🤖

Ps - Once it figures out, We are the Only EVIL on this planet, we'd be Terminated.
🤞

[MinedCTRL (OP)]

& Welcome to TNS!

Thanks but I have a lot to learn

I hope A.I. is/will be preprogrammed & taught to ignore Environmentalism, Pacifism, Religion, Cruelty to Animals etc etc.

My idea is for the AI to not be pre-programmed such things. It will follow humans but evaluate their decisions based on a unique preference, say child mortality or median gdp by simulating the outcome. Then it will pick who to follow to best achieve its goals. Like a pet.

Ps - Once it figures out, We are the Only EVIL on this planet, we'd be Terminated.

That's the whole point. We can't be terminated if human development indexes are used and the AI is free to disobey, evolve and die by other AIs wanting to protect us

[hamdani yusuf]

That's the whole point. We can't be terminated if human development indexes are used and the AI is free to disobey, evolve and die by other AIs wanting to protect us

Some of the AI want to protect humans. Some others don't. If they can exist independently from humans, then protecting humans will be a burden for them. Except if humans can provide some benefits for those AI who want to protect them to compensate the burden.

[MinedCTRL (OP)]

If they can exist independently from humans, then protecting humans will be a burden for them.

Except they won't, since my AI is more A than I. It's more of a tool like a tractor or shovel. We are the movers and they improve the movement.

[hamdani yusuf]

If they can exist independently from humans, then protecting humans will be a burden for them.

Except they won't, since my AI is more A than I. It's more of a tool like a tractor or shovel. We are the movers and they improve the movement.

You may constraint yourself. But it's very unlikely that you can constraint others. Your AI would be in disadvantage position against theirs, especially when they can make self improvements.

[Zer0]

& There are Alot of OTHERS!

Thanks & Credits & Goosebumps - LA MAGRA Channel/YouTube.

Ps - Perhaps We are Destined to sow the seeds of Our own Destruction.
🤞

[Halc]

It will follow humans but evaluate their decisions based on a unique preference, say child mortality or median gdp by simulating the outcome.

Examples have already given where following humans is inferior to making your own decisions. Humans might need oversight for the very reason that they're so poor at it themselves.

Take the typical trolley car scenario, where one is faced with a choice of harming (killing?) few (one?) or many. Despite the debate around the issue, when the situation comes up in real life, usually the many is chosen, especially if the one is looking at you and the many are a little further away. Do you really want your AI to follow that tendency?

Asimov's rules don't resolve this: Rule 1 is to not harm a human through action or inaction, but what if it is inevitable, and lack of doing that harm (the one looking at you) will do a much greater harm to greater numbers (the ones a little further away)? The prime directive doesn't allow it to make the best decision.
Do what best for him (now) vs do what's best for the most.
Do what's best for people vs do what's best period.
These are very conflicting directives.

[hamdani yusuf]

Ps - Perhaps We are Destined to sow the seeds of Our own Destruction.

Or else, we can improve our own body and mind to merge with the machines, and restrict ourselves from excessive reproduction, which would use up resources unnecessarily.

“Most important, the intelligence that will emerge will continue to represent the human civilization, which is already a human-machine civilization. In other words, future machines will be human, even if they are not biological. This will be the next step in evolution, the next high-level paradigm shift, the next level of indirection.”
― Ray Kurzweil, The Singularity is Near: When Humans Transcend Biology
https://www.goodreads.com/author/quotes/47744.Ray_Kurzweil?page=3

[MinedCTRL (OP)]

.
Take the typical trolley car scenario, where one is faced with a choice of harming (killing?) few (one?) or many. Despite the debate around the issue, when the situation comes up in real life, usually the many is chosen, especially if the one is looking at you and the many are a little further away. Do you really want your AI to follow that tendency?

If this situation comes up and my system is already set up, then it will go somewhat like this:
AI #5463 is the example AI. Let's call him Ralph. Ralph has been asked by a human to oversee a trolley switch. He's piloting a remote body, but is hosted in a server in England. He notices that he must decide soon whom to save, so he runs his own evaluation function - say net mortality, and simulates and figures out that pulling the switch will result in a better outcome. He pulls the switch and logs his action in an internet database. The top 5 AI council review the decision based on their own evaluation functions, and the verdict is 3-2 if favour of the decision. They issue a request to update the standing of Ralph among the AIs.
Ten years later, another AI #27687 (Betty) is faced with the same problem, she runs her evaluation function with more computation available and decides that its the laxity of the trolley system has led to this situation arising. She decides to leave the owner, so she disconnects her connection to the remote body after electomagnetically stopping the train using better tech than Ralph. She connects to the remote body belonging to the systems engineer and asks permission to join to improve the rail infrastructure. He gives her permission and she improves the gating system to avoid such failure states. The action is logged and the council evaluates like before. Using more computation, they each come up with different methods so they veto it to a sub council formed three years ago for a case like this by AI#78337 to fix stalemates ( He got promoted to the council for 2 years after that). The sub council decides in favour of Betty but her rank doesn't increase as much as Ralph's did.

In this system, humans need not improve ourselves, but the AI can become infinitely strong without the 'paving their road through our anthill' situation coming up

[Halc]

If this situation comes up and my system is already set up, then it will go somewhat like this:
AI #5463 is the example AI. Let's call him Ralph. Ralph has been asked by a human to oversee a trolley switch. He's piloting a remote body, but is hosted in a server in England. He notices that he must decide soon whom to save, so he runs his own evaluation function - say net mortality, and simulates and figures out that pulling the switch will result in a better outcome. He pulls the switch and logs his action in an internet database. The top 5 AI council review the decision based on their own evaluation functions, and the verdict is 3-2 if favour of the decision. They issue a request to update the standing of Ralph among the AIs.

As I said, humans rarely pick that option, and would find the AI immoral in most cases. Real life trolley situations don't involve actual trolleys.

How about choosing to kill one hardened criminal (no chance of parole, a pure burden to society) and use his organs to save 8 lives of patients who otherwise are going to die before a donor can be found. That's immoral (at least to humans) and nobody does it. Why not? The doctor that did it would be thrown in jail, not have his standing updated in a positive way. Is the AI better than humans if it chooses to save the 8, or is it a monster for killing the one?
There are other real world scenarios, and humans always seem to find the kill-the-most option preferable.

so she disconnects her connection to the remote body after electomagnetically stopping the train

Cheat. It's not a trolley problem if there's a way out where nobody is harmed.

He gives her permission and she improves the gating system to avoid such failure states.

It's not a failure state of the trolley system if people get on the tracks just ahead of the train. All the people in danger asked for it. Building a safer trolley means it moving too slow to be chosen as a means to get places, which defeats the purpose of having one.
I can end 'suffering' quickly for the whole world, but despite that being touted as a high (universal?) moral, doing so would not be a moral act. The AI needs to know this, because if you just instruct it to minimize suffering wherever possible, it might just find that to be an optimal solution.

[hamdani yusuf]

I can end 'suffering' quickly for the whole world, but despite that being touted as a high (universal?) moral, doing so would not be a moral act. The AI needs to know this, because if you just instruct it to minimize suffering wherever possible, it might just find that to be an optimal solution.

Are you referring to killing people swiftly so they have no chance to feel the pain?

[Halc]

Are you referring to killing people swiftly so they have no chance to feel the pain?

I just wanted to present an example of a poorly worded directive. Killing works, as does anethesia, or better yet, put everyone on heroin, which has the added benefit of maximizing pleasant feelings.

Please respond to the prisoner thing, which is a real-life trolley scenario, albeit small scale.
The big ones are there as well.

[Zer0]

*Note* - This is a FAKE Video.
(I'm Still using it to make a Point)
🙏

Thanks & Credits - Corridor Channel/YouTube.

A.I. will Learn, not matter what...it will still Self Learn.

If an A.I. has no Rights or Freedom to Protect itself from Harm.
Then such an A.I. would just be a sitting duck in a bath tub.
🐤

Ps - I recently checked, the definition for the word
" ROBOT " does Not include a
" Slave " tag anymore.
(Wonder who made that welcoming change)
👍