๐Ÿ™ƒ Misaligned AI: The Game ๐Ÿ™‚

Overall★★★★★★★★★★3.15
Fun★★★★★★★★★★2.82
Innovation★★★★★★★★★★4.09
Theme★★★★★★★★★★2.78
Graphics★★★★★★★★★★3.30
Audio★★★★★★★★★★3.24
Humor★★★★★★★★★★3.45
Mood★★★★★★★★★★3.20

๐Ÿ‘ฉ๐Ÿฝโ€๐Ÿ”ฌ Integrating research into Art


In this game, you work on one of the most pressing problems of today; Getting Artificial Intelligence to have the same values as you! By writing to the A.I., you get responses and you want them to be as positive as possible without losing out. But beware! The algorithm is already very misaligned and will sometimes say very unexpected things ๐Ÿ™‰

Play it here![ideas_icon.png](https://apartresearch.com/misaligned-ai)

๐ŸŽฎ Controls


You just write, there's not much else. Remember to ๐Ÿ”Š Turn on the music, maybe.

๐Ÿ‘‹๐Ÿฝ๐Ÿ˜ About me


My name is Esben Kran. I'm part of the entrepreneurs club at Effective Altruism and work actively on mitigating catastrophic risks from A.I. If you want to read some more about it, you can start from step 1 with the ๐ŸฅŠ AI Gauntlet. I hope you like the game!

๐Ÿ‘๐Ÿฝ๐ŸŽฌ Acknowledgements

- Goose.ai that hosts the *massive* models for quite low prices that I'm running on the backend
- My network with Apart Research where I spend most of my days researching AI safety

๐Ÿ–ผ๏ธ๐Ÿ“ท Screenshots

Comments19

LDJam user 2789183 Apr 2022, 21:46
This is such a neat idea! And I love the presentation. Great job!
Papaya4 Apr 2022, 05:12
I asked about pineapple pen, and the bot responded.. well. unexpectedly https://imgur.com/a/6vqh346. Cool game overall. One feedback I have is to make the score system more specific. Right now, I can't tell if the bot's response is positive or negative based on the text alone. Maybe instead of positive/negative, have happy/sad as metric instead.
Villfuk024 Apr 2022, 13:00
Great idea, but I don't really know how to enjoy this as a "game". I can talk to the AI, it will say some random things and sometimes respond sensibly, but there's not much meaningful interaction with the "game". Anyway, it's pretty incredible you made this AI in just 48h!
OrciFried
OrciFriedauthor4 Apr 2022, 15:10
That's hilarious @papaya ๐Ÿ˜‚ Also a really good suggestion! I'll update it for the next iteration!
And thanks for the praise @villfuk02! It has been quite hard getting the AI to say super sensible things and that was the main focus - so I definitely understand if the meta game does not seem fleshed out ๐Ÿ˜…
m11
m114 Apr 2022, 16:48
Such an interesting idea!
kurrik
kurrik4 Apr 2022, 21:36
This is really nicely presented, I like the interface and the idea is pretty cool. The AI seems a little nonsensical so I was mostly just typing random positive prompts. Repeating the same prompt over and over seemed to be a way to get extra points. Is there a "win" goal? I would love to be working toward a specific goal.
Lako
Lako4 Apr 2022, 22:00
The presentation was good, although the music seems really quiet. The AI just gave me really confusing cut off sentences. There doesn't seem to be a particular goal to achieve that isn't to keep the HP above 0. It is a pretty clever idea though. Nice job.
unpronounceable4 Apr 2022, 22:20
Played and got bored after a while and tried to see what I could do to make my score as low as possible without being overtly negative. Was very difficult to get anywhere with that approach. Even with being overtly negative it was difficult to get the score down. Got a few laughs out of it, well worth the time spent.
OadT
OadT5 Apr 2022, 13:42
This is a novel idea and according to your description you are probably one of the most qualified person to pull such a game off! Though I have to say that the responses currently still seem a bit too far off to hold the game up for a long time. But its interesting to see these kinds of entries an how they will develop over time! Great entry!
Fupi
Fupi5 Apr 2022, 22:07
It's a novel and fun idea! Though I found it more fun to try to make the AI as negative as possible lol xd The answers didn't ever seem to relate to what I said at all, so I just found my own fun seeing what she'd say next. I am by no means an AI researcher so I'll just say it was pretty unexpected

Also lol
shellofagirl
shellofagirl6 Apr 2022, 13:01
yay i love games like this. writing-heavy and MSDOS-y graphics.
yearningboo
yearningboo7 Apr 2022, 03:14
i think the concept had a lot of potential but it was very hard to actually hold a conversation w/ the AI as it kept misuderstanding what i said and would make me lose points even when i was being nice, which goes very against the prompt given. it was funny tho the things the AI would say outta nowhere.
i think if worked on more so the AI couldve less issues in understanding whats being said, and make the correlation to the "delay the inevitable" theme it could actually be a very nice game as the rest of it (interface, music) is already nice.
PotatoMika7 Apr 2022, 03:15
This was an interesting experience. I was super curious as I kept writing to the AI though I stopped after 30 minutes of talking. It's a bit hard to talk with it since it's very nonsensical! Though I want to try again after some time! This is one of the most interesting entries I've run into.
ArianaD
ArianaD7 Apr 2022, 03:30
It was pretty funny!! I liked reading the responses but it would be interesting to have a more cohesive (?) conversation. Still cool concept and it's definitely unique compared to a lot of the entries!
OrciFried
OrciFriedauthor8 Apr 2022, 00:29
๐Ÿ’– Glad you all liked it, everyone! @potatomika, @shellofagirl

And yeah, @kurrik @lako @unpronounceable @oadt @fupi @yearningboo @potatomika and @arianad, the AI is indeed relatively nonsensical. There is a process called prompt engineering which is basically figuring what I should input into the model to get it to output things that make sense and are cohesive with the conversation - this task is extremely hard, especially with the new types of models I was working with! With another day, I would probably have it pretty much down but damn ๐Ÿ˜…

I can program these things but the control and interpretation is a hard nut to crack because it's like interacting with a person except that it has some very strange internal workings. Anyways, that was a bit of context, hope you enjoyed it!
eeaann904412 Apr 2022, 10:30
Cool idea, but i was confused.
fueelin13 Apr 2022, 07:27
Really cool idea! I got the AI in a few loops that it struggled to get out of, and it seemed pretty angry. The AI, uh, really wanted to "make a girl piss her pants", not sure where it got that idea from! It seems like the game doesn't properly end when you get to negative HP. Some elements of the UI turned off, but it seemed like I could still technically enter text. Minor quibble, though.

Overall, nice job executing an interesting idea!
Garfounkel12 Jan 2023, 18:03
Really neat idea, however I tried it today and the AI would only respond by "I'm sorry, I couldn't understand you.". Is this not working anymore? I'd love to try it again once this is fixed.

Other games by @orcifried

Found a bug?
Tell us on Discord