I play Starcraft:BW sometimes with my brothers. One of my brothers is much better than the rest of us combined. This story is typical: In a free-for-all, the rest of us gang up on him, knowing that he is the biggest threat. By sheer numbers we beat him down, but foolishly allow him to escape with a few workers. Despite suffering this massive setback, he rebuilds in hiding and ends up winning due to his ability to tirelessly expand his economy while simultaneously fending off our armies.
This story reminds me of some AI-takeover scenarios. I wonder: Could we make a video game that illustrates many of the core ideas surrounding AGI? For example, a game where the following concepts were (more or less) accurately represented as mechanics:
--AI arms race
--AI friendliness and unfriendliness
--AI boxing
--rogue AI and AI takeover
--AI being awesome at epistemology and science and having amazing predictive power
--Interesting conversations between AI and their captors about whether or not they should be unboxed.
I thought about this for a while, and I think it would be feasible and (for some people at least) fun. I don’t foresee myself being able to actually make this game any time soon, but I like thinking about it anyway. Here is a sketch of the main mechanics I envision:
Setting the Stage
This is a turn-based online game with some element of territory control and conventional warfare, designed to be played with at least 7 or so players. I’m imagining an Online Diplomacy variant such as http://www.playdiplomacy.com/ which seems to be pretty easy to make. It would be nice to make it more complicated though, since this is not a board game.
Turns are simultaneous; each round lasts one day on standard settings.
Players indicate their preferences for the kind of game they would like to play, and then get automatically matched with other players of a similar skill level.
Players have accounts, so that we can keep track of how skilled they are, and assign them rough rankings based on their experience and victory ratio.
Rather than recording merely wins and losses, this game keeps track of Victory Points.
All games are anonymous.
Introducing AI
As the game progresses, factions have the ability to build AI which are implemented by bringing in another player from outside the game.
The skill level of the AI player is random, but most likely to be the same as the skill level of the other players in the game.
Additional resources and time can be spent on building the AI to increase the likelihood that it is Friendly (more on that later)
What It Is Like to Be An AI
When you enter a game as an AI, it is randomly determined whether or not you are Friendly. Only you (the AI) can see the result.
An AI starts off “boxed,” controlling no territory, no armies, no anything. It cannot even communicate with any players other than the player who created it. However, we assume it has been given enough information about the world that it can model it very well.
As a result, it is omniscient. Not only can the AI see the entire map, but it can predict the moves that human players will make. More on this in next section.
The player controlling a boxed AI can do various things:
Message it in-game (no livechat)
Destroy it
Give it military units
You gift some units to the AI. You no longer control them; the AI does. They receive combat bonuses of some sort.
This gives the AI the ability to communicate directly with the other players as a side-effect
Give it manufacturing capability
You gift some production center to the AI. It can now build nanofactories and whatever fancy technology it has dreamed up, all things which are too complicated for you to understand.
This gives the AI “Unboxed” status.
AI’s can become Unboxed in some circumstances even against the will of their controllers. For example, perhaps if they use their troops to conquer a manufacturing facility, they can choose to become Unboxed. Or perhaps they can broadcast their code to the world (if they have communication ability) allowing any faction to build them, and perhaps having some possibility of being built spontaneously (this represents them convincing random minor factions to build them, like North Korea or BestBuy)
In some cases, AI’s can become Unboxed without their captors realizing this. (not sure how to make this happen yet)
Modelling AI predictive power
We are trying to model the ability of the AI to predict what the humans will do, at least in the short term. AI cannot predict what other AI’s of similar computing power will do, but AI with ridiculous amounts of computing power can predict what weaker AI’s will do by simulating them.
We accomplish this by splitting each round into three phases: the human phase, the Boxed AI phase, and the Unboxed AI phase.
The Human phase is straightforward: 24 hours for all players to message each other and decide what their moves will be (which units will move where, which territories will manufacture what, etc.)
The Boxed AI phase is another 24 hours in which Boxed AI have the privilege of revising their planned orders as they see fit, and talking with the people they can communicate with if they want. Importantly, Boxed AI can see all the planned orders of the human players. People to whom they talk then gain the same privilege, since their ability to act on the decisions of the AI makes them as unpredictable as the AI. If those people then talk to other people, the privilege spreads...
The Unboxed AI phase is to the Boxed AI phase what the Boxed AI phase is to the Human phase.
After the Unboxed AI phase finishes, the moves finally resolve and the next round begins.
Decoy Moves:
For various reasons I think it would be good to introduce the following element of randomness:
In addition to inputting their orders in a given round, players input their “decoy” orders. (A user-friendly UI would make this feasible)
To represent imperfect predictive power, AI’s see either the real orders of the human players, or else the decoy orders, with probability (say) 80/20%
Same goes for boxed AI when they are being predicted by unboxed AI.
Victory Conditions
Human players get, say, ~50 VP’s for winning a game, ~25 VP’s for playing in a game won by another human player or friendly AI, and ~10 VP’s for playing in a game won by an unfriendly AI.
Friendly AI players get ~50 VP’s if their human creator wins the game, ~30 VP’s if they win the game, ~20 VP’s if another human player or friendly AI wins the game, and ~10 VP’s for playing in a game won by an unfriendly AI.
Unfriendly AI players get ~100 VP’s if they win the game, and ~10 VP’s if anyone else does.
Questions:
(1) The most crucial part of this design is the “Modeling AI Predictive Power” section. This is how we represent the AI’s massive advantage in predictive power. However, this comes at the cost of tripling the amount of time the game takes to play. Can you think of a better way to do this?
(2) I’d like AI’s to be able to “predict” the messages that players send to each other also. However, it would be too much to ask players to make “Decoy Message Logs.” Is it worth dropping the decoy idea (and making the predictions 100% accurate) to implement this?
(3) Any complaints about the skeleton sketched above? Perhaps something is wildly unrealistic, and should be replaced by a different mechanic that more accurately captures the dynamics of AGI?
For what its worth, I spent a reasonable amount of time thinking about the mechanics I used, and I think I could justify their realism. I expect to have made quite a few mistakes, but I wasn’t just making stuff up on the fly.
(4) Any other ideas for mechanics to add to the game?
Just for fun: Computer game to illustrate AI takeover concepts?
I play Starcraft:BW sometimes with my brothers. One of my brothers is much better than the rest of us combined. This story is typical: In a free-for-all, the rest of us gang up on him, knowing that he is the biggest threat. By sheer numbers we beat him down, but foolishly allow him to escape with a few workers. Despite suffering this massive setback, he rebuilds in hiding and ends up winning due to his ability to tirelessly expand his economy while simultaneously fending off our armies.
This story reminds me of some AI-takeover scenarios. I wonder: Could we make a video game that illustrates many of the core ideas surrounding AGI? For example, a game where the following concepts were (more or less) accurately represented as mechanics:
--AI arms race
--AI friendliness and unfriendliness
--AI boxing
--rogue AI and AI takeover
--AI being awesome at epistemology and science and having amazing predictive power
--Interesting conversations between AI and their captors about whether or not they should be unboxed.
I thought about this for a while, and I think it would be feasible and (for some people at least) fun. I don’t foresee myself being able to actually make this game any time soon, but I like thinking about it anyway. Here is a sketch of the main mechanics I envision:
Setting the Stage
This is a turn-based online game with some element of territory control and conventional warfare, designed to be played with at least 7 or so players. I’m imagining an Online Diplomacy variant such as http://www.playdiplomacy.com/ which seems to be pretty easy to make. It would be nice to make it more complicated though, since this is not a board game.
Turns are simultaneous; each round lasts one day on standard settings.
Players indicate their preferences for the kind of game they would like to play, and then get automatically matched with other players of a similar skill level.
Players have accounts, so that we can keep track of how skilled they are, and assign them rough rankings based on their experience and victory ratio.
Rather than recording merely wins and losses, this game keeps track of Victory Points.
All games are anonymous.
Introducing AI
As the game progresses, factions have the ability to build AI which are implemented by bringing in another player from outside the game.
The skill level of the AI player is random, but most likely to be the same as the skill level of the other players in the game.
Additional resources and time can be spent on building the AI to increase the likelihood that it is Friendly (more on that later)
What It Is Like to Be An AI
When you enter a game as an AI, it is randomly determined whether or not you are Friendly. Only you (the AI) can see the result.
An AI starts off “boxed,” controlling no territory, no armies, no anything. It cannot even communicate with any players other than the player who created it. However, we assume it has been given enough information about the world that it can model it very well.
As a result, it is omniscient. Not only can the AI see the entire map, but it can predict the moves that human players will make. More on this in next section.
The player controlling a boxed AI can do various things:
Message it in-game (no livechat)
Destroy it
Give it military units
You gift some units to the AI. You no longer control them; the AI does. They receive combat bonuses of some sort.
This gives the AI the ability to communicate directly with the other players as a side-effect
Give it manufacturing capability
You gift some production center to the AI. It can now build nanofactories and whatever fancy technology it has dreamed up, all things which are too complicated for you to understand.
This gives the AI “Unboxed” status.
AI’s can become Unboxed in some circumstances even against the will of their controllers. For example, perhaps if they use their troops to conquer a manufacturing facility, they can choose to become Unboxed. Or perhaps they can broadcast their code to the world (if they have communication ability) allowing any faction to build them, and perhaps having some possibility of being built spontaneously (this represents them convincing random minor factions to build them, like North Korea or BestBuy)
In some cases, AI’s can become Unboxed without their captors realizing this. (not sure how to make this happen yet)
Modelling AI predictive power
We are trying to model the ability of the AI to predict what the humans will do, at least in the short term. AI cannot predict what other AI’s of similar computing power will do, but AI with ridiculous amounts of computing power can predict what weaker AI’s will do by simulating them.
We accomplish this by splitting each round into three phases: the human phase, the Boxed AI phase, and the Unboxed AI phase.
The Human phase is straightforward: 24 hours for all players to message each other and decide what their moves will be (which units will move where, which territories will manufacture what, etc.)
The Boxed AI phase is another 24 hours in which Boxed AI have the privilege of revising their planned orders as they see fit, and talking with the people they can communicate with if they want. Importantly, Boxed AI can see all the planned orders of the human players. People to whom they talk then gain the same privilege, since their ability to act on the decisions of the AI makes them as unpredictable as the AI. If those people then talk to other people, the privilege spreads...
The Unboxed AI phase is to the Boxed AI phase what the Boxed AI phase is to the Human phase.
After the Unboxed AI phase finishes, the moves finally resolve and the next round begins.
Decoy Moves:
For various reasons I think it would be good to introduce the following element of randomness:
In addition to inputting their orders in a given round, players input their “decoy” orders. (A user-friendly UI would make this feasible)
To represent imperfect predictive power, AI’s see either the real orders of the human players, or else the decoy orders, with probability (say) 80/20%
Same goes for boxed AI when they are being predicted by unboxed AI.
Victory Conditions
Human players get, say, ~50 VP’s for winning a game, ~25 VP’s for playing in a game won by another human player or friendly AI, and ~10 VP’s for playing in a game won by an unfriendly AI.
Friendly AI players get ~50 VP’s if their human creator wins the game, ~30 VP’s if they win the game, ~20 VP’s if another human player or friendly AI wins the game, and ~10 VP’s for playing in a game won by an unfriendly AI.
Unfriendly AI players get ~100 VP’s if they win the game, and ~10 VP’s if anyone else does.
Questions:
(1) The most crucial part of this design is the “Modeling AI Predictive Power” section. This is how we represent the AI’s massive advantage in predictive power. However, this comes at the cost of tripling the amount of time the game takes to play. Can you think of a better way to do this?
(2) I’d like AI’s to be able to “predict” the messages that players send to each other also. However, it would be too much to ask players to make “Decoy Message Logs.” Is it worth dropping the decoy idea (and making the predictions 100% accurate) to implement this?
(3) Any complaints about the skeleton sketched above? Perhaps something is wildly unrealistic, and should be replaced by a different mechanic that more accurately captures the dynamics of AGI?
For what its worth, I spent a reasonable amount of time thinking about the mechanics I used, and I think I could justify their realism. I expect to have made quite a few mistakes, but I wasn’t just making stuff up on the fly.
(4) Any other ideas for mechanics to add to the game?