I launched this at midnight prior to logging off and going to bed, funnily enough a massive storm hit us around 3am and I woke up and noticed my OpenAI account failed to fund lol. So we ran out of funds and it caused things to break for some of you.
Just waking up here on the East coast, once my daughter is occupied I will try to answer more of your comments!
Very neat idea, and overall it is more well-done than I assumed a vibe-coded game would be. Unfortunately I ran out of time and now the game locks me out of doing anything and I don't know who the murderer is :( I wish there was just a way to have it revealed.
I've always thought that a game like Civilization should allow the user to interact with different opposing world leaders, but in a non-deterministic way as opposed to a subset of fixed interactions. Given what you know about this leader's temperament, goals, and so on, you can conduct diplomacy in a way that can be way more immersive and crafty than what the game currently gives you, which hasn't seem to evolve much in decades.
I played various murder mystery games this summer and one of them had this feature: you could call a suspect and ask arbitrary questions for a limited amount of time (three minutes or something). The problem was of course that the model hallucinated an answer to one of the questions, suggesting that other suspects were related by blood even though they clearly weren't.
Congrats, I have not had the chance to try it yet, but what I saw looked awesome and I'll give it a go, for sure. I'll update this comment when I do.
I've been working on a similar concept for 6 months now (https://game.selna.ai, portuguese only for now, sorry!), and it's been a journey. I've been working with VR professionally for some 8 years now and creating what I call a "social virtual reality" seemed simple enough, but has turned out to be much more complex than I initially imagined.
In the same way that in normal VR a single badly scaled object can break immersion, in "social VR" a single contradiction kills the experience. But when stars align and eveything works, it's magical.
I started with whodunnits but have expanded it into linear and non-linear storytelling. At this point the generation pipeline has some 40+ separate steps, the output of ones feeding others, until it all comes down to single prompt the NPC uses to talk using the realtime API.
The platform can create a complete narrative (including cast personalities and images) in some 20 minutes, and in the process generates the equivalent to a big book of text.
I've been working on this just because I can and I am a bit obsessed with getting it working perfectly, and I don't really know what I'll do with it when it does.
Sorry for the self-reply, I couldn't figure out how to edit the comment. I tried it now, it's very good. I love the fact that you can just move from one suspect to the other quickly.
You'll need to generate and give the characters a canonical current date and time, that will fix a lot of different subtle issues.
You have a huge break of character or the fourth wall whenever the player makes an accusation, the character goes like "you're making a serious claim, let me formally validate it before I respond" and then they refuse to reply to anything else while "the response is being validated".
Neat! I've been wanting to make something like this for a long time, well before LLMs came around. It was one of the first things I tried out with an LLM in the early days of ChatGPT. My son and I made little mysteries for each other and just pasted them in a chatgpt conversation with instructions for it to pretend to be the other characters and whatnot, then handed the keyboard to the player. Mine didn't work that well, I don't think he ever solved it; partially because chatgpt kinda led him astray. When I started his, I just point blank asked the mystery and it told me the solution. I felt bad about that, he had put a good amount of effort into it, but I had reminded him to tell it not to reveal the answer and he didn't. Anyway, I figured someone would make this eventually! Excited to give it a try!
Even the tag-line, "one poisoned patriarch, one locked-in truth" sounds very AI, and makes me think that I'd struggle to enjoy this. But I'll give it a go, because the idea is excellent.
I think this would work very very well in combination with human writers.
Not really. It is true in a way, but the question is if this game is made to collect data, or just to present a game, where the recordings are then deleted, in good faith.
The developer's homepage gives the impression that this is more for the tech and fun aspect, rather than collecting speech samples.
this uses a proprietary gpt model and so the data is sent to closedai. even if the developer is in good faith you can expect your voice to be part of training data.
sounds good, but it won't get past Dispatch confirmation. You should specify what networking permissions or whatever is needed. My It goes red and the text to read disappears.
This is a really fun take on a detective game. One small thing I noticed: during the initial detective briefing, the time always seems to show 0:00. Not sure if that’s intentional or a bug.
Very understandable… I originally didn’t have accounts when sharing it to friends. The account is really just there as a barrier to try and circumvent bots and bad actors as much as possible, even though there are still flaws. The goal was simply to have a way to force a max timer on someone so that more people can try this out
even without trying it, its so cool this exists. I think this kind of voice interaction has been a dream of adventure game creators for so long, and now its possible. there was a game many years ago (in the days of messengers like trillium being popular), I forget the name, where one of the characters would make a landline phone call to you. The voice acting was great and it felt novel and exciting to see it work.
It is a nice idea ; I will try it later. Everyone can poop on AI, that’s fine, but this specifically states it is AI and how it is used, so just click on another HN link to cry about AI, not the ones that actually specifically mention it.
same here. js error though is 0oxvw2ytjke1_.js:1 POST https://api.openai.com/v1/realtime/calls 429 (Too Many Requests). and auth bearer token in clear js request here towards openai. someone could have just used the token for their own good. cant go directly to openai's apis from the website. huge security risk.
True, I have a dungeons and dragons combat tracker (https://topoftheround.com) where I let the user generate loot items with different settings for them like rarity or magical/non-magical etc. and I am using a firebase cloud function that talks to the actual AI and returns the result. I also have set a limit to how much that token is allowed to spent per month, hopefully OP has thought of that as well. Interpreting and generating voice probably uses a lot of tokens.
Meh, it's like the fourth such "interview AI to solve the mystery!" implementation I've seen. It's always so bland, uninspired and boring, and drags on for waaay too long.
I have no problem with AI generated text and story but I just can't stand AI voices. Even a hint it is AI narrated and I check out. I would encourage you to consider getting it voice narrated by real people. I'd rather listen to bad/amateur human narration than AI voices.
As a second point - this is very interesting using voice control as game input are there viable alternatives for realtime voice uasing your own GPU?
Please add a demo video/voice where you show someone asking questions and AI answering it. I wanted to see what it is like but dont want to signup.
I will do this, I was not sure how to do this cleanly, my recording I did was a little awkward haha
Hi everyone!!! Thanks so much for the support.
I launched this at midnight prior to logging off and going to bed, funnily enough a massive storm hit us around 3am and I woke up and noticed my OpenAI account failed to fund lol. So we ran out of funds and it caused things to break for some of you.
Just waking up here on the East coast, once my daughter is occupied I will try to answer more of your comments!
We are at $100 usage since posting this haha. I will probably cap it at $500 for you guys
Very neat idea, and overall it is more well-done than I assumed a vibe-coded game would be. Unfortunately I ran out of time and now the game locks me out of doing anything and I don't know who the murderer is :( I wish there was just a way to have it revealed.
I've always thought that a game like Civilization should allow the user to interact with different opposing world leaders, but in a non-deterministic way as opposed to a subset of fixed interactions. Given what you know about this leader's temperament, goals, and so on, you can conduct diplomacy in a way that can be way more immersive and crafty than what the game currently gives you, which hasn't seem to evolve much in decades.
I played various murder mystery games this summer and one of them had this feature: you could call a suspect and ask arbitrary questions for a limited amount of time (three minutes or something). The problem was of course that the model hallucinated an answer to one of the questions, suggesting that other suspects were related by blood even though they clearly weren't.
This is solved by having a shared knowledge base between the agents!
So it was very life like and realistic.
Not really, since it was something that particular character would clearly have known about and wouldn't have gained anything from lying about.
Again, sounds very realistic.
Congrats, I have not had the chance to try it yet, but what I saw looked awesome and I'll give it a go, for sure. I'll update this comment when I do.
I've been working on a similar concept for 6 months now (https://game.selna.ai, portuguese only for now, sorry!), and it's been a journey. I've been working with VR professionally for some 8 years now and creating what I call a "social virtual reality" seemed simple enough, but has turned out to be much more complex than I initially imagined.
In the same way that in normal VR a single badly scaled object can break immersion, in "social VR" a single contradiction kills the experience. But when stars align and eveything works, it's magical.
I started with whodunnits but have expanded it into linear and non-linear storytelling. At this point the generation pipeline has some 40+ separate steps, the output of ones feeding others, until it all comes down to single prompt the NPC uses to talk using the realtime API. The platform can create a complete narrative (including cast personalities and images) in some 20 minutes, and in the process generates the equivalent to a big book of text.
I've been working on this just because I can and I am a bit obsessed with getting it working perfectly, and I don't really know what I'll do with it when it does.
Sorry for the self-reply, I couldn't figure out how to edit the comment. I tried it now, it's very good. I love the fact that you can just move from one suspect to the other quickly.
You'll need to generate and give the characters a canonical current date and time, that will fix a lot of different subtle issues.
You have a huge break of character or the fourth wall whenever the player makes an accusation, the character goes like "you're making a serious claim, let me formally validate it before I respond" and then they refuse to reply to anything else while "the response is being validated".
Neat! I've been wanting to make something like this for a long time, well before LLMs came around. It was one of the first things I tried out with an LLM in the early days of ChatGPT. My son and I made little mysteries for each other and just pasted them in a chatgpt conversation with instructions for it to pretend to be the other characters and whatnot, then handed the keyboard to the player. Mine didn't work that well, I don't think he ever solved it; partially because chatgpt kinda led him astray. When I started his, I just point blank asked the mystery and it told me the solution. I felt bad about that, he had put a good amount of effort into it, but I had reminded him to tell it not to reveal the answer and he didn't. Anyway, I figured someone would make this eventually! Excited to give it a try!
AI: Yea I did it.
Even the tag-line, "one poisoned patriarch, one locked-in truth" sounds very AI, and makes me think that I'd struggle to enjoy this. But I'll give it a go, because the idea is excellent.
I think this would work very very well in combination with human writers.
Hiiiii, i would appreciate your feedback now that I resolved the openai issue. I think you will be surprised. I was!
Good way to get more people's voices into training dataset
I have no idea how to do that, but yes, as someone else noted, OpenAI might!
Is getting people’s voices difficult? Doesn’t YT have more than enough data already?
Maybe run a STT -> TTS filter in front of your mic?
I guess could work, real mic -> stt -> tts -> emit from a virtual audio device configured as the mic...
I only hope latency is not a problem.
The only relevant comment!
Not really. It is true in a way, but the question is if this game is made to collect data, or just to present a game, where the recordings are then deleted, in good faith.
The developer's homepage gives the impression that this is more for the tech and fun aspect, rather than collecting speech samples.
this uses a proprietary gpt model and so the data is sent to closedai. even if the developer is in good faith you can expect your voice to be part of training data.
sounds good, but it won't get past Dispatch confirmation. You should specify what networking permissions or whatever is needed. My It goes red and the text to read disappears.
OpenAI credits ran out! Resolved!
Same.
Vibecoding at its finest
This is a really fun take on a detective game. One small thing I noticed: during the initial detective briefing, the time always seems to show 0:00. Not sure if that’s intentional or a bug.
Oops, ran out of OpenAI credits
Maybe it only works for hn accounts created yesterday
sounds neat, would love to try it but I don't make accounts for websites anymore.
Very understandable… I originally didn’t have accounts when sharing it to friends. The account is really just there as a barrier to try and circumvent bots and bad actors as much as possible, even though there are still flaws. The goal was simply to have a way to force a max timer on someone so that more people can try this out
so if you got locked out of hn, you would just never comment again?
what a strange comment
I would assume it implies “websites that have no need to actually have an account”.
Although most time I cave in, I really understand the sentiment of the parent.
It makes perfect sense. You are being intentionally obtuse.
setting an arbitrary point in time where you no longer will sign up for accounts does not, in fact, make perfect sense to me.
even without trying it, its so cool this exists. I think this kind of voice interaction has been a dream of adventure game creators for so long, and now its possible. there was a game many years ago (in the days of messengers like trillium being popular), I forget the name, where one of the characters would make a landline phone call to you. The voice acting was great and it felt novel and exciting to see it work.
It was working great... But now it says it can't connect.
Did the server go down or did I hit one of your limits?
Ran out of OpenAI credits, resolved now
It is a nice idea ; I will try it later. Everyone can poop on AI, that’s fine, but this specifically states it is AI and how it is used, so just click on another HN link to cry about AI, not the ones that actually specifically mention it.
"voice connection failed"
Sorry! Ran out of OpenAI credits, resolved now
same here. js error though is 0oxvw2ytjke1_.js:1 POST https://api.openai.com/v1/realtime/calls 429 (Too Many Requests). and auth bearer token in clear js request here towards openai. someone could have just used the token for their own good. cant go directly to openai's apis from the website. huge security risk.
True, I have a dungeons and dragons combat tracker (https://topoftheround.com) where I let the user generate loot items with different settings for them like rarity or magical/non-magical etc. and I am using a firebase cloud function that talks to the actual AI and returns the result. I also have set a limit to how much that token is allowed to spent per month, hopefully OP has thought of that as well. Interpreting and generating voice probably uses a lot of tokens.
Meh, it's like the fourth such "interview AI to solve the mystery!" implementation I've seen. It's always so bland, uninspired and boring, and drags on for waaay too long.
I have no problem with AI generated text and story but I just can't stand AI voices. Even a hint it is AI narrated and I check out. I would encourage you to consider getting it voice narrated by real people. I'd rather listen to bad/amateur human narration than AI voices.
As a second point - this is very interesting using voice control as game input are there viable alternatives for realtime voice uasing your own GPU?
The entire game hinges on the premise of being able to narrate and understand dynamic text on demand. Pre-recorded messages don’t work for this.