A browser game about AI safety
Hi. I’m Nick, longtime LW reader and Berkeley meetup attendee. My interests include not dying in an AI apocalypse and making games, sometimes professionally. I’ve wanted to make an incremental game for a while but never got around to it. When I saw the Future of Life Institute’s Keep The Future Human contest, it seemed like a good time.
Play the game here: TheChoiceBeforeUs.com
My hope is that this is something you can share with people who aren’t already familiar with alignment concepts. Friends or family who’ve heard the term but don’t really get why it matters. Sometimes a game gets through where an essay wouldn’t.
[Minor spoilers ahead – nothing that will prevent enjoyment]
In the game, you run a small AI research lab. As you grow, a rival appears and the race kicks in. Build too slow and they steal your resources. Build too fast without alignment and everyone loses. By late game, a potential AI safety framework emerges. Your actions can choose to support or oppose it. If the framework passes, your rival is likely to “cheat” and get shut down, but the pressure is not off. By this point you are creating such amazing tech that the world has begun to depend on the wonders you’re creating (medicine, materials, climate, etc). The challenge is threading the needle. Serve humanity without destroying it.
My hope for these mechanics was to represent real challenges: race dynamics, coordination problems, and the tension of creating powerful but controlable tools. The original essay has some thoughtful solutions (go read it!!) The game focuses on winning by building “Tool AI” with limited in scope. Tool AI designed to aid humans rather than replace them.
Overall, I think the game is optimistic. You can create some amazing gifts for humanity.
The game is optimized to be a game. It’s not a simulation and I made tradeoffs for fun. But I tried to stay faithful to the spirit of reality. In fiction, the good guys win because they’re the good guys, or there needs to be a satisfying plot for the human readers. Reality is more unforgiving. I wanted the game to feel that way.
Feedback welcome!
One more thing: as your AI gets more powerful, the UI may start misbehaving. This is intentional. I wanted “misaligned AI” to feel real and a little unsettling, not abstract.
Play it here: TheChoiceBeforeUs.com
I loved the original universal paperclips. Your game is cool! I particularly liked the small realisation that I had been unlocking each of A, G, and I separately and boosting them as ways to “level up” (can’t figure out how to do spoilers, sorry)
Very Slightly More In-Depth feedback:
I have been thinking “someone(s) should make a ‘Universal Paperclips’, but a bit more realistic or framed around today’s situation.”
I think I would be initially somewhat confused about what’s going on. (Universal Paperclips was also kinda confusing about what’s going on at first, which was part of it’s charm, but, there’s at least one central thing which was “produce paperclips”)
I’m not sure what that translates into, I also haven’t gotten far enough to have an opinion about the overall arc of the game (literally just clicked a few buttons, sharing first impressions before I get back to work :P). But, that was my first ~minute worth of experience.
First nitpick but easily-fixed reaction: the text appears on the screen slowly (which is always a bit annoying for fast readers), but also reveals itself horizontally rather than vertically, so I can’t even finish reading a line as it appears.
Good feedback, thanks for checking it out!