Interview with game developer Mike Hein about Ridl-E

I recently had the opportunity to interview Mike Hein, the developer behind the unique game Ridl-E: The AI Riddle Bot. This innovative app uses artificial intelligence to transform the world around you into a series of cryptic puzzles. In our conversation, we discussed his creative process, the technical challenges of training AI, and his vision for the future of interactive scavenger hunts.

Mike, what was the primary inspiration behind developing an AI riddle bot?

Ridl-E had a number of inspirations. I wanted to create a game that captured the feeling of a riddle-solving contest, like Bilbo's "Riddles in the Dark" encounter with Gollum in The Hobbit. It was also inspired in part by Geocaching, where you explore real-world locations to find hidden objects. I felt that being able to easily create or participate in a scavenger hunt in a public space like a mall or Disneyland would be a blast. As a father of two boys, I’m always looking for fun, stimulating, kinetic games to play with them. I felt that being able to pull out my phone and create a riddle scavenger hunt in seconds during moments of downtime would be quite the superpower! I also enjoy playing with new technologies and was interested in exploring how new AI technologies could create a unique experience that would otherwise be impossible.

What was the greatest technical challenge in training the AI to create riddles that are both cryptic and solvable?

The quality and balance of the generated riddles resulted from extensive experimentation with different combinations of AI models and prompts. For example, a riddle may be impossible to solve if the image recognition fails to pick out the intended object or misidentifies it. It may be impossible to solve if the object is underspecified, causing the riddle to invent generic or incorrect details rather than describing the exact object captured. It may be impossible to solve if the riddle generation uses nonsensical logic that no human would follow. On the other hand, a riddle will be too easy if the object name is directly referenced in the riddle without obfuscation. A riddle could be too easy if there is too little misdirection or too few logical leaps required to figure out the answer. The current balance in the game required hundreds of iterations with different prompt and model combinations. I tried hard to find the best balance possible given current technology, but I’m sure it can be improved even more in the future as AI technologies evolve.

Which AI model powers the riddles and how do you ensure they remain unique?

I ultimately settled on a flow that takes advantage of multiple different models, each well-suited and cost-efficient for a specific subtask. For instance, image recognition tasks are generally handled by a cost-efficient Google Gemini model, which I found to be particularly good at accurately identifying objects and picking out details that could be used to flavor the riddles. Once an object is recognized, I use a different cost-efficient model from OpenAI that I found to be particularly good at creating fun and logical riddles. Of course, the models used differ if the objects are captured from a video, if you request a harder riddle, or if you request a hint, etc. In all these different cases, I’ve tried to select the best and most cost-efficient model paired with a prompt optimized through trial and error. Inputs and outputs from all these models are stitched together server-side with Google Firebase in a way that is seamless to the user and allows me to try out different models easily.

What new features or expansions can players expect in the near future?

As new AI frontier models come out and existing models become more cost-effective, I would love to experiment with them to create even higher-quality, more challenging riddles. On top of that, I’d like to flesh out World Mode, where players can see riddles generated by other players on a public map, with RPG elements and cooperative/competitive group scavenger hunts. Beyond this, player feedback will guide future updates.

How do you see the role of AI evolving within the puzzle game genre in the coming years?

The role of AI in puzzle games is real-time content generation. At the moment (and hopefully always), an expert human puzzle designer is better at hand-crafting puzzles than an AI. These puzzles need to be crafted in advance and placed within a context the designer has meticulously planned. However, what if we want to add a puzzle to a context that is emergent? Ridl-E is an example of this, where a puzzle is generated on the spot based on real-time context that a human designer cannot know. One could imagine a game so open-ended and so infinitely replayable that individually designing for all possible permutations would be impossible to anticipate upfront. This is where AIs can be used to create some exciting, never-before-possible experiences.

A special thanks goes out to Jesse Lennox of Inari PR & Consulting for facilitating this interview.