Yuma is a social camera app that lets you turn anything you capture into a digital character you can chat with. Most AIs today are boring — labs optimize for helpful, harmless assistants that are competent and entirely forgettable. My job was to make Yuma’s characters sound truly alive: a soulful old backpack, an unhinged lighter, a cuddly stuffed animal.
Prompting the AI to “pretend you’re soulful” never works without truly interesting examples, so our core discovery was that the best voices came from layering human-written persona seeds. A rusty knife might be guided by its base KNIFE nature, a quality of being WORN-OUT, and a GLOOMY persona if it’s feeling down that day.
But that created a new problem: we needed thousands of these seeds to cover the range of objects people would encounter. And crafting AI personalities is a completely new creative medium somewhere between screenwriting, playwriting, and teaching a child to play pretend.
I ran a bounties program where writers anywhere could earn $11-55 to write us engaging persona. I developed principles to teach writers how to prompt engaging characters and built training sandboxes where writers could learn, experiment, and test their personas in conversation.
To match personas to objects, I built a computer vision pipeline that extracts qualitative attributes from each captured image (what it is, its function, subcultural association, physical state) and embeds them to search against our persona library. A stickered water bottle matches to an artsy voice, a plain one to something more utilitarian - so a few thousand well-organized personas could embody millions of potential objects.
The program paid out thousands of dollars and yielded the ~2,000 personas we needed for our initial launch. When you write for Yuma, your work is published to the world. A musician friend wrote an egotistical persona for his Fender Stratocaster, and now every Strat in the world is imbued with his arrogance. The same infrastructure later powered a corporate partnership to create a custom voice for a drink brand, validating the longer-term vision of Yuma as a new channel for publishing content grounded in physical objects.
