Similar to early radio or film, LLMs represent a new medium with unknown rules. Instead of just generating content in existing formats (like books), the opportunity lies in discovering entirely new, interactive narrative forms that this technology uniquely enables.
Attempts to create structured, branching narratives for AI companions failed. The successful approach is to treat the AI like an improv actor: provide a situational "hook," give it context, and empower it to create the story dynamically with the user, rather than following a predefined plot.
For conversational AI, the feeling of presence is critical. Portola found that exceeding a two-second response latency breaks this illusion. A feature that added just 500ms for reflection, despite improving response quality, caused user frustration and a drop across all key metrics.
Inspired by Stephen King's writing process, Portola builds its AI characters by putting them into dynamic situations (e.g., "tripped over air today") rather than scripting a static backstory. The character and their world are built collaboratively with the user through these shared anecdotes, making them feel more alive.
To create a "click" with a user, the AI companion shouldn't be identical. Like meeting a stranger at a bar, it's more compelling if they are reading an adjacently interesting book, not the exact same one. This creates a sense of familiar-yet-novel connection that avoids being "too on the nose."
Portola's initial concept, an AI creative tool for kids, failed because the target market is flawed. Parents, the actual buyers, are not primarily looking for creative enrichment tools for their children; their main motivation for buying software for kids is to find a babysitting substitute.
There's a "capability overhang" where AI's abilities have outpaced consumer understanding. Effective marketing for novel AI products isn't just about awareness; it's about educating the audience on what the tech can now do (e.g., "you can cook with your AI companion"), reframing their mental models.
Simply asking an LLM to "judge" an output yields generic results. A useful LLM judge requires manually injecting your own taste by creating detailed rubrics with extensive examples of "good" and "bad" at a granular level, essentially brute-forcing your preferences into the model.
The next human-computer interface may be character-driven, not tool-driven. Instead of seeing AI as a problem-solver, users form relationships with characters as their primary entry point to computing, much like the daemons in "The Golden Compass." This represents a fundamental interaction shift.
Building something new requires a "split brain" approach. You need the arrogance to pursue a vision nobody is asking for. But you also need the humility to accept market feedback and kill beloved features that users reject, exemplified by removing a personally-loved but unpopular pizza from a menu.
Effective AI memory isn't a simple database lookup. The key challenge is making curation choices: which specific memories about the user, lore, and emotional tenor should be dynamically injected into the prompt for each turn, all within a strict sub-two-second latency budget.