Show HN: Uncracked City, detective game where you question LLM suspects that lie

I’m just curious, are you old enough to have been inspired by Where in the World is Carmen Sandiego? That was my first thought when reading this. I loved the computer game growing up and I wanted to be on the game show so bad.

Oh gosh, the memories.

Also, it seems that you and I are dinosaurs.

Ha, I'm old enough but grew up in another part of the world (Vietnam) where the game show never made it there. This is the first time I've heard of it and it's interesting that my games share some ideas with it :)

It started as a video game series all the way back in 1985

https://en.wikipedia.org/wiki/Where_in_the_World_Is_Carmen_S...

This looks very interesting, but the Italian copy on the website is poorly translated. For instance, nobody would write an explanation of what you need to do speaking to the player in the courtesy third person (Perquisisca i luoghi, Smascheri la bugia). Other sentences are written poorly, for instance "Nel gioco, «Segnalare» sotto una battuta ce la invia direttamente.". I haven't played the game, but if the translations are not up to par, it might be worth to publish it in English only for now

Thank you so much for your feedback. To be honest, I am a solo founder and dont speak Italian natively. So I have to base on the translation by AI completely. Once I have more players and more financial backup, I'll have those texts verified by native speaker.

I think that passing the copy from your site to whatever recent model and asking if it sounds native would already catch a lot of issues

Just imagine that all the margins on your site are misaligned and text overflows the elements. Buttons still work, but it all looks ass.

That's the feeling bad translation gives.

The French one is not bad, except for some elemnts (namely the last sentence on the page which I did not understand). Funny coming from a French company, 10 km away from my place :)

When I was growing up there was a wonderfully made interactive “novel” game on early Macintosh called Sherlock Holmes. I’ve always wondered when we’ll get to a point where this can be done ad-hoc with any potential crime thriller. LA Noir introduced nuanced facial movements and dialog chains but I keep going back to that 1990s era Sherlock Holmes Consulting Detective.

This one? https://www.macintoshrepository.org/3485-sherlock-holmes-con...

Thats the one but you'll need to print the newspapers

People are definitely doing "free form" interactive fiction now, though I think the one that really interests me is classic style but with an LLM interface that makes the interaction less rigid, eg

https://intfiction.org/t/glulxe-with-llm-based-user-input-pa...

Thanks for the feedback. The case is fixed and ensured that it is solvable. I have multiple models to play it before going live. Only the conversation part is free, but it's guarded against the timeline and facts that each suspect knows. In that way, they can lie but can not change the story. By the way, I'm going to take a look at Consulting Detective

this is perfect!

You should probably put some better examples of writing in the first trailer, especially if the game involves a lot of reading. Asides from just sounding horribly bland, there's only two sentences of dialogue shown and the character says "honestly" in BOTH of them.

Hi, fair point. Indeed, the trailer will get a more careful review from now on ;)

The speech in the first sentence is weird too. "Detective? Salem? The Ashcom house!"

Maybe the AI thinks you're saying "Salaam" to the detective?

I wonder if the cases in the game are created by AI as well. From my experiments LLMs outputs are rather mediocre, without "creativity", suspenses or plot twists

Thank you for your feedback. It's a fair point, LLM stories are often flat, I totally agree. And yes, the cases are written by AI, but through a pipeline rather than from scratch. I pick the idea and the city, then we develop the pitch together via Claude (I'm alone on this for now), then, only after my validation, the engine generates the case from it. After that, it makes sure the case is solvable by having multiple models play it as the detective, simulating different player profiles.

"It's a fair point"

At least don't use AI to reply. I get this whole game is entirely AI made and that's fine. But I just wonder how much effort did you put to make sure it's fun and human playable.

This didn’t strike me as an AI reply, just OP trying to be considerate when replying. I would have said the same thing. (Also kudos to OP for building a game with NPCs managed by an LLM, I’ve been toying with this idea for a while too, ever since I saw an amazing Skyrim mod where you could freely chat to each NPC and learn their back stories etc). Would love to know more about what it takes to put this type of approach into production and how to constrain the discussions.

Thank you. There's a harness layer that keeps each suspect from deviating from the current storyline. For the moment, I've deployed it to Lambda, where Haiku answers the questions in about 1–2 s.

No, it's me replying, for the record. I read and reply to all the comments myself :) Honestly, that's a month of work with intensive use of AI. I always keep in mind that the game is for humans, so every step, every case is validated and played by me.

Interesting idea. I was toying with a similar concept of a investigative journalist in a noirish town who's tasked with uncovering the secret identity of a vigilante.

It always struck me as absurd how obvious the fact that Clark Kent is Superman was to the whole audience. The characters in the story never seemed to understand it.

Ha, interesting, the Clark Kent problem. And your journalist idea sounds fun, I'd try it. For long term vision, I've already prepared an MCP so that people can build their own stories and start a private game, for example.

I also used to think it was silly, but then I read another take - in-universe, Superman is known to be an alien, with a crystal fortress in Antartica or whatever - there's not much reason to believe that he has a secret identity at all. If you aren't already used to the premise of "superheroes have secret identities," it probably wouldn't occur to you that he'd have any reason to want to work in the reporter's office 8 hours a day and commute in traffic.

Yeah. I suppose so but with heroes who wear masks, there is always the question who they are (Batman, Flash, Spider man etc.). Isn't there?

I have LLMs as players in a hidden-information word game. Models would often leak hidden info in their reasoning, so you had to filter what each side sees. Sometimes you have to be aware that these LLMs can be sneaky at times.

The game is not bad. I tried it out and solved the Paris case. I liked the style and the locations.

I don't think I had to actually use the LLM chat at all though? It felt like suggested questions were all that was required.

There's an element of "just go everywhere and click on everything" too. For example having to return to a previous location to find someone now has a vital clue for you on how to solve a new puzzle. I'm not sure how you are supposed to know you should to back there. That kind of gameplay isn't that much fun.

Thank you for playing it through. The Paris case is the first of a five-level training pack, it's meant to guide new players with a tutorial. The free questions really matter in the real cases after that, where the suggestions don't cover what you'll need. And very good point about going back to a location, will definitely fix it in the next version.

Very cool - happy to feature this in my AI and gaming newsletter. Let me know how I can contact you!

MCP's in video games seems very interesting. I've played a lot of games with level builders and they can be insanely complicated. That could definitely cut down quite a bit and allow a lot more people to create cool stuff.

“Each suspect answers in their own voice. Some of them lie.”

EVERYBODY LIES.

Good catch, and that's by design. Someone lying doesn't necessarily mean they're the culprit. Some lies are not related to the case, some are.

haha very neat! Just curious how are you planning to make it free with the inference cost? Even with a luna or haiku its still a non zero cost to run per session.

Good point, I cover the inference cost myself for now, the running cost is a few thousand a year. There'll be premium features in the future, but that depends on reach and retention, and that'd be the second phase of the product.

I love this kind of game and I'm excited for AI to make them more individual and I really hope that sometime soon we'll get to a point where these AI-requests don't have to go to a datacenter but can be done locally on one's own machine.

Seems really interesting, i am actually most excited about this side of AI. Like making your own mini simulation and having ai live out their lives or something along the lines of it

Cool concept! It reminds me of the 1985 videogame Where in the World Is Carmen Sandiego. What a blast!

Thank you for your feedback. You're the third person to mention it today. And to be honest, I've never heard of it before. I'm gonna check it tonight ;)

I had too much fun with this acting like a schizophrenic batman asking everyone who Harvey dent was and to SWEAR TO ME

Thank you very much for taking the time to test the game. Which case did you play, and did the suspects keep a straight face?

Talking to lying LLMs is what I do all day for work. Can’t imagine I’d want to do the same in my free time.

Same. Also, staring at a screen.

I wonder if you could do it all over voice, say while alone in the car.

Hi, half of it already works. You can question the suspects by voice and they answer in their own voices. There are still interactions that need the screen though, for now. A voice-only mode is a really fun idea.

I think it's a bit unfortunate that you used AI for all the non-core elements of the game like the artwork. I do think that a game with judicious use of gen-AI allowing you to go "off the rails" has potential, but it's going to take a lot of hard work to make it OK and not chock full of all the AI clichés. It can be a good game, but you need to avoid not only to make it not look like slop, but avoid it BEING slop, and it's like you don't quite appreciate the problem when using AI for things you don't even need to use AI for.

I totally understand your point. But as I mentioned, I work alone on the project and that's the only way for now to have playable content. The very next thing, once I can hire someone, would be the artwork and the storytelling.

Nice idea! i'll give this a go tonight :)

I am curious - are you using on-device LLM for this?

No browser version? :(

Why does the Android app need to collect device ID?

Thanks for the feedback. It doesn't collect the hardware ID or the advertising ID. It collects the random player id the app is given on first launch, which lets you keep your progress without an account, and the push token if you allow notifications.

Sounds like work to me.

I went to the website to play the game, and was disappointed that it was only a marketing web site for a phone app. Why not let me play the game?

Thank you. You've saved me a click.

Hi, that's exactly what I'm working on right now. There was a technical barrier to porting the game to the web, and given the slowness of the App Store review process, I'll definitely release a web version very soon. Please stay tuned!

Any reason this requires an app? A quick look suggests nothing at all that cannot be done on a simple website.

add me as a beta tester

testflight me

I think you should disclose on the website that the game is using LLMs

I don’t know how much more explicit it could possibly be honestly. The graphics and voiceover and writing are all painfully stuffed with tells.

By making clear in the copy on the website that it is using LLMs

Thank you. Fair enough, I'll add a line saying that the suspects are played by a language model.

Thanks!

I enjoy asking ai to play with me Where in the World Is Carmen Sandiego. You can ask for different variations, difficult levels and ask it to add images.

My first thought as well! In less then a second, when I saw the Map with the airplane line I remembered Carmen Sandiego! What a good memory!

no thank you

Show HN: Uncracked City, detective game where you question LLM suspects that lie

No, no, no. The LLM suspects can't lie. That requires intent. They're "hallucinating." Yeah, that's the ticket.