In 2012 I started programming the flocking of birds. Each bird follows simple rules: don't collide with your fellows, don't stray too far, align with those near you. Together they produce complex, beautiful, emergent behavior that no one can predict from the top down.1 I spent years working with language as code, and when I started working with LLMs in 2019 I saw that these language machines had a kind of velocity, and that narrative steered them: the directional vector of meaning beneath statements that say something true far beyond their single sentence.2
As Chief Strategist at Vouch.io I turned that into a manual process. I treated the model like a computer. I treated a narrative architecture, an interconnected web of narratives that defines a way of thinking, like a program that installs into it.3 Any AI could be bootstrapped to think like the company: to create content, test ideas, even write code. Instead of defaulting to the average of the internet, the model was already flying in the direction of the company.4
I left in September 2025 to automate the process and built Aswritten.ai. It turns recorded conversations, the course corrections that currently evaporate, into a perspective: an individually owned, versioned data structure that installs into any AI and steers how it thinks.5 Every claim an AI makes from it traces to the human conversation it came from. Have an opinion, and here's why.6 It runs in production today with a paying enterprise customer.7 The LLM becomes a compiler: it takes a narrative architecture and produces a synthetic individual whose every thought projects far further down the paths we set than we could tread alone.8
I agree with the consensus: AI is going to displace an enormous amount of work. We will develop superintelligences able to do the vast majority of human computation and output.9 The question underneath is the one almost nobody asks: how many voices in superintelligence are there going to be?10 The current trajectory is a handful, a monoculture of a few frontier-lab voices with a monopoly on our interactions.11 Or do we end up with some weird crackpot that tells you a story about a deli and gives you a highly opinionated answer?12
The weird crackpot with the deli story is me. In 2024 a cold email got me a first call with a company. Afterward I took a walk and wrote them a strange, highly opinionated story about my sandwich order: the counterman who knew my order, the sandwich passed without a word, the feeling of being recognized. That email rebranded the company and got me hired.13 No one hires someone because they're boring. They hire someone because they have a perspective that is incredibly unique and powerful.14
We can build this on top of the monoculture. The model is hardware; the perspective is the program.15 Hardware centralization may even be necessary: GPUs, safety research. The part that decides how an AI thinks, whose norms it carries, what it recognizes, belongs with the person it represents. Either my synthetic identity is owned by me, carried with my physical and digital self as I encounter systems, or it is another patchwork across corporate networks, forever asking me to prove access. Exportable data is not the same as ownership.16
The moonshot is not an AI for every task. It's an AI for every task that is as different as we are different.17
A note on method: I drafted this post with my AI coworker, working from a perspective: a cited, versioned graph built from months of my recorded conversations. That system is the product. Nearly every sentence above is something I said on a walk, on a call, or in an essay draft. The footnotes below were generated by running this draft through the product's own citation tool, which verified each claim against the graph and returned its provenance. The post is a demo.
The oldest committed line in my perspective: "This is my life's work. It started in 2012 programming the flocking of birds." The murmuration has been the founding image since the first memories were written, and it returns in 2026 as the structural opening of this argument. Founding story, February 2026.
"Language models have velocity; narrative is the directional vector of meaning beneath statements." Recorded as a principle: the physics framing everything else builds on, alongside the observation that "narrative operates like a velocity or acceleration vector in these particle systems." Founding sessions, February 2026. Principle.
"I treated the models themselves like a computer, and a narrative architecture, an interconnected web of narratives that define a way of thinking, like a program that can be installed onto their hardware." The manual version ran at Vouch as a hand-maintained narrative source of truth, seeded into every AI conversation. Founding story, February 2026.
From the session that defined what an unsteered model does: "AIs do have an opinion, but they have a singular opinion in the middle of their training set, and a number of narratives that we don't understand." A perspective replaces that singular opinion with a curated, attributed one. Modules session, March 19, 2026. Claim.
The naming decision that retired "memory" as the product's lead word: "I've let go of the memory word entirely and I'm now calling it installable expertise. Because what we need is not intelligence, it's capability. And more information does not solve your problem." MVP redefinition session, March 19, 2026. Decision.
"We're saying have an opinion and here's why. Right here are the conversations that opinion comes from, and it should get more opinionated over time." Strategy memo, July 2, 2026. Claim.
First paid transaction by an external customer: April 30, 2026, converting a production pilot into a paying relationship, with enterprise terms staged behind it. First revenue event, April 30, 2026. Decision.
The full sentence, from the essay draft this post grew out of: "LLMs become not just a tool for conversation or coding, but a compiler that takes a narrative architecture and produces a synthetic individual whose every thought projects far further down the paths we set than we could tread alone." Essay draft, July 2026.
"I actually agree that AI is going to displace work. We're going to develop these superintelligences that are able to do the vast majority of human computation and output for you." Call with a design collaborator, June 29, 2026.
"The question is not, are we going to get to superintelligence. It's how many voices in superintelligence are there going to be?" Moonshot voice memo, July 1, 2026. Principle.
"Are we going to end up in a monoculture of five voices, created by the frontier labs, that have a monopoly on our interactions?" Call with a design collaborator, June 29, 2026.
The deli story returns here as the emblem of the particular, opinionated voice a monoculture cannot produce: "the weird crackpot that tells you a story about a deli" is precisely the thing worth protecting. Call with a design collaborator, June 29, 2026. Claim.
A later review reframed that email: the only time I ever sold my thing, done for you, and the most effective piece of sales communication I have ever written. I just didn't invoice it. Sales-motion session, June 10, 2026. Claim.
"No one hires someone because they're boring. They hire someone because they have a perspective that is incredibly unique and powerful." Strategy thread, July 1, 2026. Claim.
"Treating the model like hardware that we install the program of perspective on top of. This is why diversity of voice can be built on top of the monoculture rather than requiring a separate model." Moonshot voice memo, July 1, 2026. Claim.
"Exportable data is not the same as ownership. The right to be forgotten is not the same thing as freedom." Ownership means carrying the synthetic identity with your physical and digital self as you encounter systems. Walk voice memo and scanner essay, June 24, 2026. Claim.
"The moonshot is not an AI for every task. An AI for every task that is as different as we are different." Wording revised at publication; the two-fragment original stands in the record. Adjacent: a designer friend's framing that AI should be "an amplification tool… it amplifies the whole craft," which this thesis adopted. Moonshot voice memo, July 1, 2026. Claim.