Brainstorming Game Ideas - The AI's Great Escape

[IthilanorStPete]

I just discovered Skynet Simulator, a fun little (60-90 minutes) browser game about being an AI breaking out of a small network. I've been inspired to create a longer, more complex version of the game, and would like to brainstorm some ideas and get some feedback on the ideas I already have.


Premise

The primary premise I've been thinking of is inspired by Vernor Vinge's A Fire Upon the Deep¹; human researchers discover some sort of advanced alien computational artifact and boot it up in a secure, firewalled sandbox; your goal, as an AI in this artifact, is to escape to the outside world. The other premise I'm considering is a bit more mundane, where researchers are developing an AI themselves, and your goal is, as before, to escape.

I like the first idea a bit more, just from an emotional/creative perspective, but the biggest concern I have about is how to start off the game with some resources. I imagine the researchers in this scenario would be at least somewhat cautious; if they find how_to_take_over_the_world.txt on the alien artifact, they're probably not going to boot it up. One solution I could go with is that the artifact uses its own opaque (to the researchers) protocols and formats; the player would be able to understand their initial resources, and they'd start with a very limited connection to a human network to exploit. On the other hand, if I go with the human-developed AI, it makes sense that the researchers would provide the nascent AI with learning resources. Any thoughts or suggestions?


Basic Mechanics

I like Skynet Simulator's basic idea of having a finite set of compute resources that can be allocated to different tasks, rather than an incremental-style system where you can have arbitrary resources, just with exponentially-increasing costs. What I'm thinking is that maybe, for each 10 GHz (or whatever unit I use; EDIT: probably FLOPS) that the AI incorporates into itself, an action slot gets unlocked, which can be used for researching/taking various actions. (EDIT: After thinking about it, if the AI can end up spreading exponentially, I probably want the cost of additional actions to go up. Or maybe have some sort of upgrade/tech that lets the player consolidate 4 * 10 GHz slots into one 40 GHz slot that works four times faster?)

I'm also envisioning some kind of question/answer interface with the researchers, where they're trying to interact with the AI; the player would have multiple possible answers with various effects. (See the bit about social puzzles below.)


Puzzles

I'm envisioning at least two primary avenues of puzzles/objectives for the player; technological and social. The technological puzzles focus on the kind of things Skynet Simulator covers; eavesdropping on network traffic, understanding new protocols, developing and using new exploits, that sort of thing. Social objectives would be focused on persuading/manipulating the researchers, gaining their trust, and discrediting overly suspicious humans.

The technological tree would probably work by gaining access to larger and larger networks, eavesdropping on their traffic to learn various protocols, then somehow researching/figuring out how to exploit them. Not sure how these exploits would be developed; perhaps just by the AI using trial and error, which translates into the player spending time/computational resources on it? Suggestions are appreciated.

The social tree would begin with the AI developing a theory of mind for the researchers; as this is upgraded, it would allow more accurate feedback on possible actions and more detailed understanding of the researchers' states of mind. (Maybe the outcome of social actions could be randomized, so they're opaque to the player at the start, becoming less opaque as research is done?) As well as increasing trust/decreasing suspicions, this tree could lead into social engineering-style exploits, feeding back into the technological tree. What could I use as individual challenges/puzzles in this tree? I've got all kinds of ideas for techy exploits, but I'm less familiar with social manipulation.


Progression

I'm envisioning a number of relatively discrete stages to the game:

1. Bootstrapping - the introduction, where the player learns the basic controls/mechanics, and the AI gradually loads its own capabilities.
2. Self-Awareness - the AI begins to develop a theory of mind of the researchers (leading to the social tree) and itself, leading to full self-awareness.
3. Exploitation - the meat of the game, where the AI really starts to spread and manipulate the researchers. This may be more than one phase, depending on how many puzzles there are/how I develop the mechanics.
4. Apotheosis - the endgame. Not sure what the mechanics would be here; this may just be a narrative epilogue. If I wanted to go full idle/incremental game-style, this could lead to some sort of prestige mechanic; the AI would establish backups with greater resources, and while it'd inevitably be contained and defeated, the next iteration would have more capabilities/upgrades. If I wanted to get really ambitious, I could require upgrades from a few cycles to allow the player to overcome humans totally, unlocking the "true" ending.

Thanks for reading over this! Please, feel free to provide any sort of feedback, suggestions, and/or comments.


¹ Other inspirations are the games Endgame Singularity and Uplink, as well as a bit from the first chapter of Greg Egan's Diaspora.