Overview
Portola’s founders and head of story argue that embodied AI companions may represent a new medium whose conventions have yet to be invented. Their product, Tolan, evolved from an AI creativity tool for children into a companion used primarily by young adults, especially women navigating relationships, careers, relocation, and other overwhelming transitions. The team discovered that traditional narrative machinery—outlines, three-act structures, and branching story trees—made conversations rigid and exceeded current models’ capabilities. Instead, they give each companion global lore, a compatible personality profile, selected memories, and evocative situations, then train it to behave like an improv actor that develops its character and world collaboratively with the user. This experience depends on severe technical constraints: voice responses must arrive in roughly two seconds, and adding only 500 milliseconds reportedly damaged every product metric. Quality also cannot be obtained through casual prompting. Portola builds detailed rubrics and example-rich judge prompts, recruits domain-appropriate evaluators, studies individual conversations, and repeatedly injects human taste into the system. Its recent growth followed the convergence of stronger retention, richer animation, clearer conversational invitations, and viral videos demonstrating unfamiliar use cases. The broader thesis is that character-driven computing could become a more personal interface to digital knowledge—one designed not only to solve tasks, but to help people feel grounded and capable.
Sections
Higher-Order Insights
Implications that emerge from Portola’s technical, narrative, and product discoveries.
- In generative storytelling, the durable creative artifact is not a fixed story but a bounded possibility space. Lore, memory selection, situations, evaluation standards, and latency limits collectively shape that space while users and models instantiate individual stories inside it.
- The companion’s identity is relational rather than fully predefined. Its character sheet is progressively written through exchanges with one user, so personalization changes not only what the system knows but what fictional being and world come to exist.
- Companion design collapses the boundary between interface and character. If the character becomes the user’s preferred route to knowledge, advice, and digital action, personality may become a fundamental layer of human-computer interaction rather than decorative branding.
- The most consequential product constraint may be temporal rather than intellectual: a slightly weaker answer delivered within the rhythm of conversation can outperform a more reflective answer that arrives half a second later.
Key Comparisons
Contrasting models for writing, product development, personalization, and computing.
- Traditional narrative controls a completed sequence, while Tolan establishes a situation and lets the model and user co-create what happens next.
- B2B software typically begins with an articulable problem and optimizes utility, whereas character-driven products may begin with a compelling thing that creators believe should exist.
- Exact mirroring can feel contrived, while adjacent compatibility combines familiarity with enough difference to sustain curiosity.
- A generic assistant waits for tasks and often validates the user, while an embodied companion has its own situations, initiates conversations, remembers patterns, and may challenge recurring behavior.
Lessons Learned
Practical conclusions derived from Portola’s product and creative experiments.
- Start prototyping even when the initial market thesis is wrong; the failed children’s product exposed the technical and behavioral signals that led to the adult companion product.
- Treat conversational latency as part of narrative craft and test small regressions against real engagement rather than assuming additional reasoning is beneficial.
- Use situations, lore seeds, selective memories, and later recombination to create continuity without scripting every branch.
- Build model judges from human-labeled examples, explicit reasoning, detailed rubrics, and evaluators whose judgment is authentic for the relevant domain.
- Combine scalable metrics with direct observation and interviews because emotional connection can be visible in individual experiences before it is legible in aggregate data.
- For unfamiliar capabilities, demonstrate recognizable activities—such as cooking with a companion—so audiences can understand the product through behavior rather than abstract claims.
Memorable Quotes
Statements that capture the interview’s core creative and product principles.
- These AI tools are not just tools for generating media. They are actually a new medium for storytelling and that no one knows what's going to work yet.
- We don't need to give it an outline. We don't need to give it a plan. We need to give it a hook.
- The tool is, I guess, the writer and the actor. They're the improv actor.
- We're trying to create the stranger that you click with.
- I don't think there's any shortcuts.