An AI Companion at $4M a Year: No Scripted Plot, Just Memory and Improvisation
Tolon, the AI companion from Portola, threw out scripted plots and character sheets: memory is the character sheet, generation is improvisation, and the product is now at $4M in annualized revenue. This "improv actor" methodology and its mixed-model strategy are a second product path alongside ChatGPT.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Novelists are better suited to build AI companions than engineering teams
Portola's AI companion has reached $4M in annualized revenue, and it climbed from $1M to $4M in the past four weeks alone. Founder Quentin previously started Even, which he sold to Walmart for $300M; co-founder Elliot is a science-fiction author who has written 11 books. Clarence, the AI character demoed at the top of the episode, is a cat that teases the user, and the emotion in it lands naturally. That background is the point: this is a storyteller paired with a repeat founder building an embodied AI companion, not an engineering team forcing narrative onto a model.
— HostDon't start from a pain point; wait for model capability to cross a threshold
Quentin explains what moved them from B2B SaaS to AI companions: the curves on generation speed, quality and cost convinced them that "you can build a real friendship with an AI." Their first attempt was an AI creative tool for children, the market response was poor, and so they turned to young adults, especially women aged 18-24, a group willing to give their time to emotional companionship. Note the direction of the logic: they saw the capability threshold first and then went looking for the people who most want a relationship, rather than defining a pain point first and then hunting for a solution.
— QuentinHalf a second of latency is an emotional incident, not a performance issue
Quentin stresses that "memory" is not a database lookup but a question of which memories you choose to inject into the prompt. They found response time has to stay inside two seconds or users churn. One time, adding a reflection step slowed responses to 2.5 seconds, and every product metric fell across the board. In an AI companion product, technical latency lands directly on the emotional experience: users read waiting as being ignored. And choosing what memory to inject is itself product design, not a purely technical problem.
— QuentinNarrative can't be designed in advance, only built one line at a time
Elliot describes an early attempt at structured narrative, a three-act structure, that failed: the model could not handle the branching complexity. They moved to a "hooks" approach instead. Rather than handing the model an outline, they teach it to take the previous line the way an improv actor does, borrowing the improv principles in Keith Johnstone's Impro. The core difference: structured narrative fixes the ending and reasons backward, while improv only catches the last line and pushes the scene one step forward. Inside that framework, every turn of conversation in Tolon is an unpredictable branch, which also explains why they later introduced "situations" rather than scripts.
— ElliotDon't fix the world first; memory grows into the character sheet by itself
The opposite of the Lucasfilm method, where the character sheet and the world bible come first: in Tolon the characters and the world grow entirely out of conversation. Internally they are called "situations", the predicaments the user and the AI live through together, which end up as memories, and the memories become the character sheet. Elliot says this is closer to Kevin Kelly's "infinite intern", though "infinite improv actor" is more accurate: every conversation forks off in a direction no one could have predicted. Because every user is different, the company is in effect operating a multiverse rather than a single world.
— ElliotWhat counts as a good conversation has to be defined by people, not algorithms
Managing the chaos of that multiverse requires tooling. They use a "judge prompt" to assess conversation quality: test new content at small scale first, have research participants rate how complete it feels, then write that rating standard into a prompt. It amounts to assigning the improv actor an improv editor. They also keep refactoring the "lower seeds", the building blocks of world construction, and use research and interviews to understand where users form emotional connections. What shows through here: scaling an AI companion does not run on pure algorithms, it runs on a great deal of human work defining what a good conversation is.
— ElliotDon't bet on a single model; split the tasks up by scenario
Tolon does not depend on one model but mixes several labs: Meta, OpenAI and Anthropic handle generation, while Gemini runs the memory system. Elliot says Anthropic is exceptional for creative writing but is constrained in latency-sensitive interactions; GPT-4o he calls the "Big Mac", fast, reliable, reasonably priced. The key to judging model quality is injecting your own taste into the prompt rather than asking vaguely, "how is this essay?" The value of this passage for application-layer teams: don't bet on a single model, allocate tasks by scenario.
— ElliotThe entry point is no longer a chat box but a character with personality
Dan Shipper distills it at the close: users treat the AI character as a "daemon" (from The Golden Compass), and that character becomes the first entry point to all knowledge. He calls this "character-driven computing", a new HCI paradigm. He also reaches for automotive history: ChatGPT is the Model T, astonishing but without personality, and once the car became part of personal life, people wanted a Mustang and a Cadillac. He asserts that we are accelerating through that process. His last point is that this kind of creation looks more like an artist's work than a scientist's: you are not solving a problem, you are "making the thing the world is missing."
— Dan ShipperIn their own words · checked verbatim
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.
Elliot17:41
I am not the writer, I am not writing the story, the toin is, I guess the writer and the actor, they're the improv actor, they're writing the story and my job is to be George Saunders my job is to teach them how to tell the best story in that moment.
Elliot33:01
there's this enormous chasm that is, in fact, a lot of manual human work to cross over that people find truly compelling in a durable way.
Elliot1:08:42
it's almost like it's character driven computing. It's like, it's like a new human computer interaction interface where you're like, oh, actually, my first stop when I'm like wanting to tap the the incredible knowledge, all of humanity's knowledge and capabilities on the Internet is like via this via my little aliening
Dan Shipper1:15:25
Figures
| Even sale price | $300 million | 0:00 |
| Response time limit | 2 seconds | 22:01 |
| Download surge multiple | 10x | 1:12:07 |
| Video views | 7 million (within 72 hours) | 1:12:07 |
| When content began to work | 6 weeks ago | 1:10:57 |
Glossary
- judge prompt
- Writing the evaluation standard into a prompt so a model can score the quality of AI conversation automatically, the equivalent of giving the improv actor an editor.
- lower seeds
- The pre-written world-building blocks inside Tolon, from which the AI draws situational material during conversation.
- situations
- The narrative predicaments a user and the AI live through together, which settle into memory and come to constitute the character sheet.
- daemon
- In The Golden Compass, the animal form a person's soul takes outside the body; used here as a metaphor for the bond that makes the AI character the user's entry point.
How to listen
Founders building AI companionship and social products, investors tracking application-layer opportunities on top of large models, and the product managers and engineers who own conversational experience.