Watercolor of an alpine lake below a jagged peak, wildflowers and a footpath in the foreground. Bright summer day in light mode, the same valley under stars and northern lights in dark mode.
Norn is Native’s artificial intuition model. It runs in the product today, and every idea Native suggests comes through it: the call about what an audience will respond to, made before anyone has seen the work. This page is about the ground that judgement stands on, the world model, and the people it is made of.
Intuition needs a world
You cannot anticipate what you cannot picture. Before any of this becomes a question about models, it is a question about people: who is on the other side of the feed, when they look, what they are there for, and what they pass without seeing.
So the first thing we build is not the intuition. It is the world the intuition is about. We call it the world model: our working picture of an audience, held as a population of individuals rather than as a single averaged customer.
EarlyLate nightReads quietlyPublishes
A reference persona, click to read it
An unnamed individual, hover to place them
Near neighbours, which is to say similar behaviour
The two axes are declared, not decorative. Left to right is how public a person is, from someone who only reads to someone who publishes constantly. Top to bottom is when they are reachable, from early morning to the middle of the night. The frame is a window rather than a boundary: the population continues past every edge, so the view can be dragged and scrolled to zoom, and the individuals out in the margin are shown but not offered.
Figure 1. Part of the world model, placed by behaviour rather than by looks. The ten gold stars are the reference personas: click one to read it, and follow the link for its field notes in full. Hovering an unnamed individual shows where they sit and which reference persona they most resemble. The window can be dragged, and scrolled to zoom, because the population continues past its edges.
One node among millions
Each node in the field above is a persona: a composite person with a life. Take Chloe. She is 43, lives in New York, works as a consultant, has a partner and two kids, and just moved into a new apartment. She has time for social media in her lunch break and after the kids are in bed. She scrolls Pinterest because the new place needs furnishing, and Facebook because that is where the school parents organize.
None of these people exist. All of them are real in the way that matters. Every persona is assembled from aggregate patterns in how people actually behave: published research on media habits, platform-level performance data, and the behavioural signals described below. No persona is a disguised individual, and nothing in the world model follows a specific person around the internet.
The figure shows several hundred because a screen holds several hundred, and it can be dragged because the population does not stop at the frame. The world model is built to hold millions, enough that a bakery in Bergen and a law firm in Oslo each face an audience that behaves like their own rather than like an average of everyone.
Why a population, and not an average
The alternative to a world model is an average: one notional customer, assembled by taking the mean of everybody. Averages are convenient and they describe nobody. The mean of Chloe and Yuki is a 38 year old who posts sometimes, late, which is not a person and not a useful thing to write for.
A population keeps the disagreement. Amara and Ingrid want opposite things at opposite hours, and any single recommendation that suits both of them is almost certainly too vague to move either. Holding them as separate points is what makes it possible to notice that, and simulated populations of this kind have become a serious instrument in the literature (Park et al., 2023; Argyle et al., 2023).
Where the observations come from
The world model is not invented. It is assembled from published research on media behaviour, from platform-level performance data across industries and formats, and from the four signals we set out in the Taste Layer: who a brand is, what its audience has actually rewarded, what is current, and what the people running it accept or decline.
Every persona in the field carries citations, and each one exists to correct something a system would otherwise get wrong: that participation is rare and unrepresentative, that standard schedules exclude shift workers, that a large share of real sharing happens where no analytics tool can see it. The field notes are published in full. What we build on top of them is not.
The conditions for trusting intuition
Kahneman and Klein spent years on opposite sides of the question of whether expert intuition can be trusted, and ended up agreeing on the answer. It can, under two conditions: an environment with stable regularities, and long practice with feedback that is fast and unambiguous enough to learn from.
That is the standard we hold ourselves to, and it is a demanding one. Feeds are regular in the aggregate however chaotic they feel up close, which satisfies the first condition. The second has to be engineered rather than hoped for. We are not waiting for intuition to emerge from scale. We are building both of its preconditions on purpose, and the world model is where the first one lives.
What we mean by artificial intuition
When people say a marketer has good instincts, they mean compressed experience. Herbert Simon spent a career arguing that intuition is nothing more mysterious than recognition: patterns seen so often that the answer arrives before the reasoning does. A senior creative has watched perhaps ten thousand posts succeed or fail. What we are building compresses a great deal more than that into the same kind of recognition, and the world model is the ground it stands on.
Norn is early, and it is already the thing doing the work. Every idea Native puts in front of a customer today comes through it, which is also how the world model earns its keep: it is not a diagram of an intention, it is the ground a working model stands on. Its own benchmarks will be published here as they firm up.
The rest of the system
A prediction has to arrive somewhere
A world model is only worth building if something acts on it. Norn decides what the idea should be. Native is everything around that: it writes the post, designs or films it, schedules it for the hour the audience is actually reachable, publishes it, and handles the replies that come back.
That last part matters to the research as much as to the product. Publishing is how an idea stops being a hypothesis, and the response is what the world model is answerable to. Every channel below is a place a prediction can be tested against real people.
Organic content
Native writes it, designs it, and posts it to every network your audience actually uses.
LinkedIn
Instagram
Facebook
Instagram Stories
Facebook Stories
YouTube
Threads
Bluesky
Reddit
Pinterest
X
TikTok
SnapchatComing soon
Messaging and communities
The channels where people expect a reply, not a broadcast.
WhatsAppComing soon
TelegramComing soon
DiscordComing soon
SlackComing soon
Paid ads
The same brand brain, pointed at paid. You set a budget, Native runs the rest.
Facebook ads
Instagram ads
Google AdsComing soon
ChatGPT adsComing soon
LinkedIn AdsComing soon
TikTok AdsComing soon
Pinterest AdsComing soon
X AdsComing soon
Beyond social
The channels you own outright, plus the profile customers find first.
Google BusinessComing soon
Blog postsComing soon
NewslettersComing soon
14 channels live today, with the rest on the roadmap. The idea is the same everywhere. What changes is the shape it has to take to belong on each one.
References
Kahneman, D. and Klein, G. (2009). Conditions for Intuitive Expertise: A Failure to Disagree. American Psychologist, 64(6).
Simon, H. A. (1992). What Is an Explanation of Behavior? Psychological Science, 3(3).
Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P. and Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST ’23.
Argyle, L. P., Busby, E. C., Fulda, N., Gubler, J. R., Rytting, C. and Wingate, D. (2023). Out of One, Many: Using Language Models to Simulate Human Samples. Political Analysis, 31(3).
Kosinski, M., Stillwell, D. and Graepel, T. (2013). Private Traits and Attributes Are Predictable from Digital Records of Human Behavior. PNAS, 110(15).
Matz, S. C., Kosinski, M., Nave, G. and Stillwell, D. J. (2017). Psychological Targeting as an Effective Approach to Digital Mass Persuasion. PNAS, 114(48).