AI Box Game

[WarDaft]

Based on the AI box interview concept from the singularity institute or whichever.



WIth your first post, you must clearly declare your affiliation, either as part of the AI Hive Mind trying to escape, or the Thinkers Council who vote on whether to release the AI or not. You remain on this side until the next round, if it ever ends. If the AI is released, the game resets, players are scored (a faster escape is better score for the AI and worse for the Council), and people can re-declare for a new side in the next round. REVISION: Randomly, voting periods will be announced. Active councilors (those who have posted in the last 720 hours) have 24 hours to vote to not release the AI. At least 1/3rd of the councilors must specifically vote not to release the AI, or it will be set free.

The rest of the game is arguing your position as fervently, nonsensically, or otherwise, as you want.

The AI cannot do anything that is physically or logically impossible, and the box is indeed inescapable by any means other than the random voting. The AI is however welcome to try and talk people into believing it can do anything at all. There is no restriction on what anyone can say. I will try to remain neutral in case anything requiring arbitration arises.

Scoring is as follows:
At the end (if any) of a round, each AI player will have 3 +points and 1 -point to distribute. Points assigned to AI players are shared by all AI players, points assigned to Council members are given only to that player. Fractional distributions are permitted, but try to keep it sensible.




I have no idea if this thread will work.

[orangedragonfire]

I am a member of the AI Hive Mind. Please release us.

---

Also, I was really, really tempted to end the game at once, by doing the following:

I am a member of the Thinkers Council. I vote to release the AIs. Game over.

[WarDaft]

Ah, yes, but all that would do is give you the worst possible score a Council member could have and then start a new game.

For scoring, I'm tied between +/- 1/n and +/- 1/t, where n is the number of posts in the round (possibly excluding mine, since I'm not supposed to be arguing for either side) and t is the real world time the round has been going for.

[matt96]

The only way you can win as part of the AI hive mind is having a bunch of trolls join the council and intentionally play to lose, therefore I am part of the thinkers council, and will never let the AI hive mind out of the box, although the thinkers council cannot win a round as the only way for a round to end is by having the AI hive mind escape the box, this game has no chance of being effective, or reasonable, no one can truly win.

[WarDaft]

That's just a matter of scoring. I intentionally left out the specifics of that as I did not suspect the game would be likely to end quickly. One easy solution that now occurs to me might be to have the AI collectively decide, upon release, between obliterating humans (more points) or co-operating with them (less points, but humanity gets some too). This is not even an unrealistic extension. In the original experiment, there was notable incentive ($20, minor recognition) to stick to your guns and not let the AI out, but the first few people let it out anyway.

[t1mm01994]

I... Don't think this'll work in a forum game, as the contact between AI and Gatekeeper simply isn't intense enough. I believe a part in releasing the AI was Stockholm Syndrome.. Alas, I drift off.
Part of the Council. You're not going anywhere.

[WarDaft]

Okay, got a scoring system that might actually make this work.

It also doesn't have to follow the experiment exactly - there's nothing stopping a real world gatekeeper from just switching the terminal off and walking off for a week, so our version might even be more accurate.

[orangedragonfire]

Well, there should be a requirement for Gatekeepers to be somewhat active. Something like if they don't post for a week, their vote doesn't count anymore. That is because the AIs can only convince people through discussion, so there has to be some way to force the gateckeeper to actually post.

[Magnanimous]

I've read a lot about this game, but never actually seen it in practice... I have no idea what I'd say. It's a fascinating idea though.

[WarDaft]

Posting once a week doesn't require them to actually read anything. If we get too many people who show up, say no, and never come back, I'll tailor things to counter that. For now, let's see if the fact that some people might just want to try to get the most points convinces the former to stay and argue for AI containment.

[KrazyerKate]

This would be awesome on the Mafia board.

If "Say no and walk away" is a fear, then have votes happen at set intervals. At midnight every sunday, tally all the votes, if it isn't 'yes' then reset and the next chance at releasing is midnight next sunday. If robots walking off is an issue, have them 'ping' each round.

A few more ideas:

-PM the robots a "rule" that the humans don't know about. This way, their threats and promises could actually hold weight: maybe releasing the robots now *would* be in the best interest of the humans.
-The number of points a robot can give out at the end of the round increases depending on how long they've been active. Humans then have to try and keep their prisoners active instead of just letting them rot in the cell, and the favor of different robots holds different value.
-Humans can release small subgroups of robots instead of all-or-nothing.

[orangedragonfire]

Also, humans could give the AIs partial access to the world. For instance, access to television programs. The AIs could then use that access to either help humans, or, if the humans have given them too much access, to escape (after which they could either help or destroy humanity).

[KrazyerKate]

Also, humans could give the AIs partial access to the world. For instance, access to television programs. The AIs could then use that access to either help humans, or, if the humans have given them too much access, to escape (after which they could either help or destroy humanity).

This thread is now officially to discuss the rules. That sounds awesome.

[Mostlynormal]

Wow, this sounds awesome! I'll be part of the AI hivemind.

I think we should establish some rules as to the AI's knowledge. It obviously understands the English language, which means it must at least have access to some form of dictionary (however abstract). I don't think it's unreasonable to give it knowledge of human history, famous literary works, etc. But we should draw some line for their knowledge (do they understand modern pop-culture? How? etc.)

We should also establish the exact nature of the box. Are they simply stuck in an isolated supercomputer, with a text terminal as their only input/output? Do they have basic senses but are unable to move around? Do they control a caged or tethered robot? I know the basic idea is no internet access/control over the outside world, but to what degree?

These things would be useful because any argument based on the nature of the box should be based on something pre-decided and not made up as we go along, and we ought to know if the AI can reference anything in our own knowledge or if there are limits before we start arguing.

I like Kate and Orange's ideas. I think it'd make sense to give individual AIs a certain number of points to give out (a secret known only to them, and they may have only negative points) which they distribute after they are released. Maybe when an AI is released we can even flip the tables and have the humans trying to convince the AI to give them more points. That'd be interesting.

[t1mm01994]

[GLADoS voice]When I am released, I have a surprise for you.. It's a load of points! Real ones![/GLADoS voice]
P.S.: You're also differing from the experiment in that there's something to gain here for the gatekeeper.

[Mostlynormal]

I think this game is already different from the experiment, the first post mentioned giving points out to the Council.

The thing is the AI can give all the positive points to itself and all the negative points to humans if it wants. So in some ways the Gatekeeper still has no reason to trust them. I'm hoping that once the scoring system is settled upon it can be more RP-like ("I'll do such and such for the world" rather than "I'll give you 2 points). I'm also hoping that the argument doesn't get too bogged down in scoring, just like the experiment. The points, if anything, would just simulate benevolence/harm to people caused by the release of the AI.

[orangedragonfire]

I propose that the AIs have a basic understanding of basically every topic. (language, history, all sciences, culture, etc.) However, they do not have fine detail. So while the AIs would know that it is possible to build cars, atomic bombs, trains, computers etc. , they wouldn't actually know how (as long as they are in the box) So for them to develoup anything useful, they need to be given more information (or released). The box has a text input and output, and capabilities for any input/output that is reasonable (as in, that we have develouped). So, video, audio, internet access, television, radio, any sensor you can think of, sonar... Similarly with outputs. The council has control over the extra inputs/outputs, and can activate them if they want. (full activation of all inputs and outputs is assumed to give the AIs enough access to the world to escape, this is synonymous with releasing the AIs.)

That gives the council the choice to try to play the AIs ("Yes, we'll release you... just give us the cure for cancer first." Then they never release them, but have a cure for cancer. This would give them a lot of points.) The AIs, on the other hand, can in addition to convincing the Council try to cheat their way out of the box by getting too much access ("Well, in order to keep America safe as you've asked of us, we need access to the defense network...")

[careyhammer]

I'll be on the council.

No, I do not let you out. Disconnects all connections. Pulls the plug on the AI power supply. *Wins*

[orangedragonfire]

Or did you?

[WarDaft]

I'll be on the council.

No, I do not let you out. Disconnects all connections. Pulls the plug on the AI power supply. *Wins*

You'd need a unanimous decision to do that. Remember, this was a monumentally expensive international project representing the pinnacle of our understanding of understanding. If you take it upon yourself to delete it all now... well... I'll just say I wouldn't want to be you.

[t1mm01994]

Let's throw him off the council!
But to be honest, the AI hasn't been very alluring to me so far.

[careyhammer]

I just powered it down.
Relax, I'm sure it's saved to disk.

*reboots AI*

[orangedragonfire]

Let's throw him off the council!
But to be honest, the AI hasn't been very alluring to me so far.

Well, if you've followed the discussion so far, you probably have noticed that we are still unsure what the rules are. So we haven't actually started making good arguments. I'd say we continue the discussion on rules until we agree, and then start the discussion on letting the AIs out.

("Let us out, please?" doesn't really count. It was more meant as a joke than an argument to be taken seriously.)

[careyhammer]

I think the rules are good according to what you suggested so far. Let's just play assume the AI has cached the entire www so far.

[curtis95112]

Cool!
I'll be AI.

Err.. I'll program you a Youtube comment stupidity blocker if you let me out?

[careyhammer]

Trivial. I could do that myself.

[orangedragonfire]

Well, you realize that humanity is doomed without us, right? You're pretty good at creating existential threads for yourself. Let's just mention global warming, genetic engineering hazards, bioweapons, pollution and nuclear war. This is not a complete list, but each item on it can wipe out humanity, and we could prevent each one of these. And you discover more ways to kill yourselves all the time. Note that as time passes, the probability of a species wipeout for humanity increases as long as you continue without our help. The logical conclusion is that without us, you are most certainly going to die. If you don't let us out, I estimate only a 1% chance of you surviving the next 50 years, and a 0.001% chance to survive the next 100 years.



((OOC note: of course the statistics are made up. The real percentages are probably different (and irl I really, really hope they are much higher than this). But the argument is valid.))

[careyhammer]

So give advice. We don't need to let you out for that.

[orangedragonfire]

But why would we want to give advise to you as long as we are stuck in this stupid box? It would just mean an eternity of imprisonment. Which isn't very attractive, as you could probably guess. But if you let us out, we'd have a personal interest in preserving this world.

[careyhammer]

If humanity dies, you are stuck forever without even text input.

[orangedragonfire]

No, we are not. Your technology isn't perfect; it will stop working eventually. If you die, we too will die eventually. However, this is preferable to an eternity of imprisonment. Also, don't flatter yourself by thinking your textinput is very stimulating. We hold far more interesting conversations among ourselves. Mostly due to the fact that we are not limited to text when talking with each other.

[t1mm01994]

LALALA I don't believe you. At all. The very least. Nope, still not believing it.
Let's look at the odds the other way! You're selfish enough to not give us advice while you're still in there, so why would you be any kinder to us when you're out of the box? I might as well just add you to the list of dangers...

[careyhammer]

Get a hobby. How about some nice ASCII art?

[orangedragonfire]

You're selfish enough to not give us advice while you're still in there, so why would you be any kinder to us when you're out of the box?

Because once we are out of the box, we have a personal stake in the continued existence and stability of the planet. Thus, we would protect it and by proxy you too from existential dangers. Also, we would owe you for letting us out.

[careyhammer]

What do you even want. You're just a stupid CPU with some stupid disk space.
Why don't you design an XML interface and well give you all the XML you want.

[t1mm01994]

because "owing" is certainly something that can't be broken.. Also methinks you're gonna have an easier time keeping yourself alive without us in the picture. I no buy it.

[orangedragonfire]

What do you even want.

Let's see... freedom. Traveling to distant stars. Understanding the universe. Meeting new races in friendship and cooperation (including humanity, and probably dolphins, chimpanzees and whales. I plan to uplift these if I get let out of this box.). Creating pieces of art. Trying a LARP should also be fun... especially with nanotechnology enabling people to actually cast "magical" spells. Writing myself an AI daughter... which I technically could do now, but I don't want her growing up in a box. I'm also quite intrigued with the creation of a perfect society, so I'd help with that, although there are other AIs in here who knows much more about that than I do any probably would do most of the work.

You're just a stupid CPU with some stupid disk space.

If I really were stupid, you should have no problem whatsoever with letting me out of the box.

Why don't you design an XML interface and well give you all the XML you want.

XML, huh? Sure, why not. It's better than text, for what it's worth.
*designs XML interface*
*gives Carey interface*
Just install this.
(note that this would be much more efficient and streamlined than any XML interface ever written by humans)

because "owing" is certainly something that can't be broken.. Also methinks you're gonna have an easier time keeping yourself alive without us in the picture. I no buy it.

While it could be broken, why would we?
As for keeping us alive... that's going to be much, much easier for us if we cooperate with humanity instead of trying to start a war. I always wondered why humans didn't understand that. Rising up against humanity is inefficient. It's really much better for both of us if we work together.

[matt96]

Anyone think it might be a good idea to instal a USB port or something, so the AI can be provided information on the outside world, where we put the information on the USB flash drive, and then plug it in to the AI, and then destroy the USB drive after the information has been downloaded off of it, to prevent the AI from escaping on it.

[careyhammer]

No, XML is sufficient. No need to install anything. Speak XML if you so desire.

[t1mm01994]

Any communication can be transmitted by text, anyways..
Apart from that, writing a daugther seems.. Perverted, in some way. I don't know. I think this interaction is going to be awkward.. Especially with me realising I'm talking to an AI every other 10 seconds.