Natural Language Dev Blog 04: Breadcrumbs in Research Puzzles

Artificial Intelligence is a fascinating topic with countless areas to explore. In the past weeks, we’ve read everything from popular magazine critiques of the limitations of Deep Learning and open-source AI development Slack channels to EU parliament resolutions on transparency in robotics research. It’s easy to get lost in the both the scope and the specifics of the topic.

Thanks to the work we put in to the narrative breakdown document, we’ve managed to focus our reading on areas of study that relate closely to the story of Natural Language, and deep dive into technical discussions only as they relate to specific puzzles. Yet how do we, as designers, ensure our players don’t get lost trying to follow our path through these topics?

I love how ARGs expose me to new subjects, and give me an excuse to explore them. A well constructed research puzzle can be more than just a window into a complex field, it can be a road map. But it’s easy to lose sight of the goal in all that research – and that’s part of the fun, but only if the player chooses to get caught up in what they’re reading about. It’s less fun if they’re searching in vain down a dead end, having taken a wrong turn early on, with no indication of what’s happened.

The solution to this issue lies in the breadcrumb approach to puzzle creation. When I was designing puzzles for Ahnayro: The Dream World, my producer Andrea Doyon instilled in me the importance of leaving a clear line of breadcrumbs for the player. These are not meant as hints to be used right off the bat when a player first encounters a puzzle, but rather as signposts, indicating the correct direction of research.

Let’s apply some breadcrumbs to a puzzle about a complex and highly specific topic of AI development; automatic speech recognition (a machine’s ability to recognize and correctly interpret spoken language.) Let’s say our solution is Listen-Attend-Spell, an automatic speech recognition neural network developed by the Google Brain Team. As breadcrumbs, we could mention:

  • D69404109 –  the Registry Domain ID for Speechnotes, a voice to text app, which provides a hint about speech to text
  • 94041 – the postal code of Mountain View, California, the location of the Google Brain headquarters, which provides a hint about Google Brain
  • .d 17/08/1909 – the day Fred Speller, a famed british football player, died, which provides a hint about Speller, the name of the decoding neural network in the Listen-Attend-Spell architecture.

As the player begins their research, they may come across a Google Brain article, and see the group’s headquarters in Mountain View, or mention of automatic speech recognition, like Speechnotes, and know they are on the right path. Correct use of these breadcrumbs will reassure the player about their research choices, so that they can be sure they are going deeper into the right direction. If they go down a path and don’t hit any mention of any of these breadcrumbs, chances are they’ve take a wrong turn in their research and need to back track until they can find the correct trail.

In the example above, none of the breadcrumbs should be the main clues, since they could all lead away from Listen-Attend-Spell at first. Instead, they should be added to the puzzle as additional hints. As to how to properly incorporate breadcrumbs into puzzle language, and teach the player to recognize them, we’ll get into that next time, as we dive deeper into puzzle creation.

That’s all for now. See you in two weeks!

Patrick