Show HN: Detective Stories -Lateral thinking detective game with Deepseek player | Hacker News Reader

Show HN: Detective Stories -Lateral thinking detective game with Deepseek player

detective-stories.com

13 pointsemurph557 comments

I have had this lateral thinking game online for a while. The jist of the game is that one player asks a "host" a series of questions in order to reach a full understanding of what happened in a story, based on an initial premise.

I have tried in the past at adding AI players but could never get reliable answers. However, as AI has improved, I decided to give it another go. At the moment it is connected to deepseek v3 with some basic prompts, and it looks to handle the task pretty well. Not yet perfect but it definitely does form an impressive understanding of the stories, and the game itself.

It tends to reveal a lot of extra information, even when you're playing it straight rather than trying to trick the AI.

Like, when you ask the "does the reason the window is open matter?"-style questions that should be bread and butter for this kind of game it'll often go beyond a "YES" into "YES, the reason the window is open provides crucial context to what they were thinking". The worst case was me asking whether somebody died via foul play, and the AI answered "NO. Your question is whether he was intentionally killed rather than it being a <spoiler> accident".

that's pretty funny. Yeah, its far from perfect in terms of implementation at the moment. I was more curious to see if it was capable of understanding the game more than anything for the moment. Previous attempts with other models had proved to be disappointing. I will tweak the prompts to resolve that issue soon enough though. Like you say, it should be bread and butter for this.

The wording of the questions really impacts the AI answers. For the elevator one,

"Did Colin start on a lower floor than 6 to start?" "Yes"

"Did Colin Press the up button on the elevator?" "I don't know"

What’s Wrong? “I don’t know”

I tried several of these but never really got see the lateral thinking because the solutions were trivial on the first guess. It feels like no effort went into the contents of the mysteries themselves and they were sourced using AI.

Which is to say, that they're all extremely shop-worn cliches. I think this shows the false allure of the AI product. The creator's low standards for content take what could be something interesting and a perfectly nice looking site and turns it into something which only superficially appears to be interesting but has no value.

I've noticed this a lot with AI-based products: even their creators don't find the experience good, but they are imagining there exists an audience which doesn't value graphics and doesn't value the written word.

Any stories that have been written by AI are marked as such (they are categorized as user submission stories where the author is credited as the AI model). Interestingly the lowest ranked story on the site is one of those, although that was created with an older model. I had never thought of the site as being an "AI product" personally. I just wanted an excuse to play around with web sockets initially, but if I can improve the AI player maybe it will eventually go more in that direction.

Reply on news.ycombinator.com