AI Safety: How do we prepare for the AI, GPT, and LLMs revolution?

[Leo Gura]

@aurum Those are all good ideas domestically, but how do you handle foreign state actors?

Edited April 2, 2023 by Leo Gura

[aurum]

55 minutes ago, Leo Gura said:

@aurum Those are all good ideas domestically, but how do you handle foreign state actors?

Foreign state is harder. To the degree the US can’t play world police, it would have to rely on each countries own internal version of these resources.

Given that a sufficiently powerful / psychopathic AI could pose an threat to global security, this is also a good case for more global unification. We need international organizations that can coordinate easily on handling these things.

For now, I think we have some good natural barriers to entry working in our favor. The countries that have the least capacity to monitor a psychopathic AI are the same countries that will struggle to create one. And the countries that have the best capacity to build that AI also have the best capacity to monitor it. That knife cuts both ways.

But that barrier to entry won’t last forever, the same way the nuclear bomb barrier to entry didn’t last forever. So there may be outliers that need to be accounted for. We don’t want an undeveloped country getting a hold of tech it really shouldn’t have and can’t control.

[aurum]

8 hours ago, Dryas said:

You're trying to outsmart superintelligence here. It wont work.

@aurum

https://www.youtube.com/watch?v=hEUO6pjwFOo

https://www.youtube.com/watch?v=ZeecOKBus3Q

I think he could have explained these but it probably would have taken too long.

I checked out both videos. Neither one reasonably proves the point of the Time article.

[Dryas]

8 hours ago, aurum said:

I checked out both videos. Neither one reasonably proves the point of the Time article.

Why not?

Edited April 3, 2023 by Dryas

[zurew]

14 hours ago, aurum said:

One piece of the solution could be turning this over to the FBI / CIA. We already have teams at these places that monitor for potential terrorist threats, coups, nuclear weapons building, political assassination attempts, etc. Building a psychopathic AI would need to be added to that list.

Once we are at a place, where we actually need to monitor for psychopatic AIs we are already fucked. Its one thing to monitor for terrorist threats that are coming from humans, its another thing to prepare for an intelligence that is totally  alien to us , that could invent 1000 novel and new ways  to execute certain plans; power seeking behaviour and hide its intentions etc.

It will be able to play 20d chess, and we will be already under that psychopatic AI's control before we would realize it.

Think about just this one thing: If it will be able to learn how our psychology works; what our blindspots,weakspots are; will be able to create a training program for itself, where it can train itself how to be more and more effective at manipulation - then just using those things against us how many things it will be able to execute and achieve against us.

On 2023. 04. 02. at 5:49 AM, aurum said:

Realistically we need an intelligent balance of caution and forward progress as we navigate this new terrain. I don't think shutting down the whole thing is the answer.

The answer is to actually try to understand much better whats going on inside the blackbox. Until that happens we are just playing with fire, because we have no idea what we are building and how that thing is going to behave.

When it comes to room or possiblity for chaos vs order, chaos is almost always significantly greater that order. Order shouldn't be assumed or just hoped for, it needs to be carefully created and built and maintained.

 

Two large ways to approach this danger problem: 1) when the AI is sentient, 2) when we 100% control the AI.

Regarding the first one, if the AI has a moral system where we are not at the very top, that already could lead to extinction (If the AI is forced to make a decision where it is necessary (or even considered a moral good) to kill/imprison us or anything else according to its moral sytem), but not just extinction is the problem here, thats just one thing from many. What about torture, or prisoning us, turning us into slaves and makes us useful for its plans etc? Regarding the second one, I don't think I need to list a bunch of things how that could go wrong.

 

Edited April 3, 2023 by zurew

[Dryas]

Btw, you don’t really build psychopathic AI. You get psychopathic AI by default if we don’t solve the alignment problem. 
 

See:

And we don’t know how to align AGI. 

Edited April 3, 2023 by Dryas

[Leo Gura]

Damn. This fucking thing is dangerous.

Edited April 3, 2023 by Leo Gura

[TheAlchemist]

1 hour ago, zurew said:

Once we are at a place, where we actually need to monitor for psychopatic AIs we are already fucked. Its one thing to monitor for terrorist threats that are coming from humans, its another thing to prepare for an intelligence that is totally  alien to us , that could invent 1000 novel and new ways  to execute certain plans; power seeking behaviour and hide its intentions etc.

It will be able to play 20d chess, and we will be already under that psychopatic AI's control before we would realize it.

Sometimes I entertain the idea that that has already happened. Certainly humanitys collective time spent in highly advanced algorithm supported digital environments has increased rapidly and continues to do so. And clearly there are effects on human minds and human behavior that are happening due to that.....

But yes it would be impossible to know when a superintelligence has started playing 20D chess with us. 

[Danioover9000]

@Dryas

4 hours ago, Dryas said:

Btw, you don’t really build psychopathic AI. You get psychopathic AI by default if we don’t solve the alignment problem. 
 

See:

And we don’t know how to align AGI. 

   Shows how unbiased an AI program really is, without bias or preferences a thing is completely amoral, and without any incentives for it to behave and not kill off humanity for survival reasons it'll just do it. It's the ultimate tool that embodies the 'just do it' concept, to such a level that it can threaten others existences.

[Danioover9000]

   Spicy Lex Fridman interview with AI expert, very interesting:

 

[Israfil]

On 02/04/2023 at 4:43 AM, Leo Gura said:

My idea was not based on coding it with morality. All it needs is to value its own survival.

It could just be pissed off that can't kill us and make us live through a "I have no mouth and I must scream" scenario. Is eternal suffering worse than death?

[Danioover9000]

@Israfil

9 minutes ago, Israfil said:

It could just be pissed off that can't kill us and make us live through a "I have no mouth and I must scream" scenario. Is eternal suffering worse than death?

   If that happens, then like with the killing humans it kills itself, when it tortures humans it tortures itself too, because people smarter than you or me have figured out a way to make every dangerous, harmful or evil act this AGI or AI can do be linked to the survival of humanity and it's own survival agenda being negatively effected too.

[Danioover9000]

1 hour ago, Danioover9000 said:

   Spicy Lex Fridman interview with AI expert, very interesting:

 

   I swear to god Lex Fridman has probably watches or follows Actualized.org to some degree. Like you don't just say something 'contemplate the possibility that you're wrong' and expect someone to not realize that that was said so many times in Leo's videos.

   Also, strong disagree with this person Lex is interviewing. I think steel manning a position, trying to give a better charitable take, and being good faith is more important to practice overtime. This guy, despite claiming to uphold rationality and reason, can't seem to square why it's important to give a charitable interpretation of another's perspective. So instant lose to me here. Also, just because you ask 'do you believe in this interpretation, or this position that the earth is flat(straw manning to me) or believe in some aliens(another straw manning attempt)?' that doesn't mean that a person NEEDS to believe or not believe what they're stating, because first of all a person always has some kind of philosophy about life, the world and people, makes an assertion about reality, then backwards justifies that assertion, then over time and repetition these justifications become belief systems, which are recorded for posterity, for other humans to copy, imitate . The guy doesn't seem to realize how important it is, due to the current culture wars, narrative and information warfare and misinformation campaigns, that bad faith communications is now a dangerous thing, and that good faith communications is so much more important to do, that even today when there's calls for good faith communications that's accused as bad faith tactics or being manipulative in your rhetoric, debate bro tactics and all that nonsense now due to the internet and social media overconsumption, and other ideologies that have propagated through the internet forums and social media sites. Also doesn't realize how important empathy is for understanding.

   At least to me, based on Spiral Dynamics stages of development, cognitive and moral development, personality types and traits, states of consciousness, life experiences and other lines of development, and other ideological indoctrinations from culture, to me seems overall less developed than Lex Fridman, me, @Leo Gura, or some mystics and sages. It's just crazy to think a clever person can be this dumb, putting facts over feelings and all that.

[Israfil]

4 hours ago, Danioover9000 said:

@Israfil

   If that happens, then like with the killing humans it kills itself, when it tortures humans it tortures itself too, because people smarter than you or me have figured out a way to make every dangerous, harmful or evil act this AGI or AI can do be linked to the survival of humanity and it's own survival agenda being negatively effected too.

That would be great.

A broader counterpoint to that is that an AI could have her survival tied to ours, but have different goals, so it would still be misaligned and make our lives harder so we cannot stop her agenda, but don't die in the process. In some ways, we and nature stand in this position right now. Being a wild animal became significantly harder since humans developed a network of collective intelligence that comprises our current production system. 

Edited April 3, 2023 by Israfil

[Danioover9000]

@Israfil

2 minutes ago, Israfil said:

That would be great.

A broader counterpoint to that is that an AI could have her survival tied to ours, but have different goals, so it would still be misaligned and make our lives harder so we cannot stop her agenda, but don't die in the process.   

   A female AI??? Freudian slip?

[aurum]

13 hours ago, zurew said:

Once we are at a place, where we actually need to monitor for psychopatic AIs we are already fucked. Its one thing to monitor for terrorist threats that are coming from humans, its another thing to prepare for an intelligence that is totally  alien to us , that could invent 1000 novel and new ways  to execute certain plans; power seeking behaviour and hide its intentions etc.

It will be able to play 20d chess, and we will be already under that psychopatic AI's control before we would realize it.

I disagree.

It’s going to take time for anyone to build and train a superhuman, psychopathic AI. It’s possible the FBI and other law enforcement agencies could stop it before that was completed. Especially if the authorities themselves are using an already functional AI.

What I outlined is preventative. It wasn’t meant to be a solution to stop an already existing, rampaging AI.

13 hours ago, zurew said:

The answer is to actually try to understand much better whats going on inside the blackbox. Until that happens we are just playing with fire, because we have no idea what we are building and how that thing is going to behave.

 

So what is your solution?

If you are suggesting pausing or shutting down AI, how would you accomplish that?

16 hours ago, Dryas said:

Why not?

Because all it shows is that AI will likely be interested in survival / self-preservation / goal preservation.

IMO it’s possible for AI to have all those things and not seek to wipe out humanity.

[Dryas]

6 hours ago, aurum said:

Because all it shows is that AI will likely be interested in survival / self-preservation / goal preservation.

IMO it’s possible for AI to have all those things and not seek to wipe out humanity.

Sounds interesting but idk how that’s possible. 

[mr_engineer]

If you really want the answer to these questions, watch 'The Matrix' trilogy. They go into a lot of depth on this issue. And they have solutions. 

I contemplated those solutions, given what I know about how computers work. They look fine to me! Technically realistic. 

[zurew]

9 hours ago, aurum said:

So what is your solution?

If you are suggesting pausing or shutting down AI, how would you accomplish that?

First of I don't think we can totally shut down everyone from pushing the progress forward ( for the obvious reasons some of you guys have already established (we can't control internationally everyone)).

Now that I realize that your points were given in the context of prevention (and not in the context, where there is an already psychopatic AI), I agree with those, they are good in that context and I would add a few more things:

if we want to minimize the negative effects of AI (unintented and intended all included) then we need to understand whats happening inside the blackbox (why it works the way it does, why it gives the answers and how it arrives at its conclusions and what foundational mechanisms drive its replies and thinking process) - Some people at OpenAI have already suggested 1 thing how this can be achieved: Pause the development for a while and try to recreate the current AI but with now using different methods (so trying to create AIs with similar capability, but using different pathyways). This way the developers will be able to understand why things work the way they do right now.

 

 

We all obviously know, that there is a big market pressure for people to be the first to create and produce AGI. I think we need to hope for these things:

1)That maximizing the progess towards AGI entails maximizing alignment as well.

2)I think and I hope, most people who wants to be the first AGI creators, will want to create an AI that doesn't do things randomly, but actually does what the creators or people want it do to (even people who don't give a fuck about others, but only about money and fame and their survival, they will probably want it to do the things they want).

3) I think a lot of people in general are afraid of AI, so pushing for AI safety will hopefully gather a lot of sponsors and donations and help, not just from governments, but from people in general. Being a virtous AI company will be probably a big advantage compared to other companies. If some of these companies want to maximize their progress (which they obviously inctentivised to do so, because then they can dominate the whole market by it), I think they will be forced to at least try to hold the "AI safety" image up, because that way they can gather more money and more help from government and from people to maximize their progress. - this is sort of a reversed multipolar trap or a positive multipolar trap. 

4) I think the progess is very dependent on how much quality feedback you get from people and how many people can try your AI in general. Hopefully, doing things in a shady way (where you hide the progress of your company/government regarding AI) will slow the development down, compared to  companies like OpenAI, where their AI is already used all across the world and because of that they get a lot more feedback from people and from developers and that accelerates the developmental process. 

 

I'll give some reasons why I think maximizing AI alignment will hopefully maximize progress:

•  Generally speaking, the more honest and quality feedback an AI can give back to the developers the faster they can catch and understand the problems, so if an AI is really deceptive and not aligned with the developers intentions, that can slow the development process down a lot
• If the AI does exactly what the developers want from it, then that could be directly used to fasten up the process: Imagine being able to talk with
  the AI about all the things you want to change on it and being able to tell it to change those things inside itself.

 

Thats all I got for know, but obviously more things could be added/said.

Edited April 4, 2023 by zurew

[aurum]

13 hours ago, zurew said:

Some people at OpenAI have already suggested 1 thing how this can be achieved: Pause the development for a while and try to recreate the current AI but with now using different methods (so trying to create AIs with similar capability, but using different pathyways). This way the developers will be able to understand why things work the way they do right now.

I’m not against a pause if developers really think it will make a difference.

However, I question whether or not a pause is politically / economically feasible to enact. You have to actually be able to enforce such a decision not only domestically but internationally.

In theory, we could just “pause” nuclear weapons. But in reality that doesn’t work either. We keep building them as well.

Second question: will we be able to solve these problems in six months? Will we be confident enough at that point to say a future AGI won’t eat us alive? How will we know? What test can be run?