On July 30th, we announced Anacreon: our new human simulation model which achieves state-of-the-art accuracy in predicting individuals' answers to new questions. Read the announcement to learn more.
Today, we announce Semilattice Custom, the product offering for Anacreon. Custom gives you your own Anacreon model of your users, your customers, or any segment you can define for $1,799/mo. This post explains what you can do with a custom model, how custom models are validated, how they're built, and what else a subscription includes.
Fix problems during development, not after launch
As a product or marketing person, your job is to know your users. You constantly have to make judgements about what people will do and what they'll think—will they use this feature, will this UX solve their problem, will this message resonate. Sometimes there's good data, but usually there isn't.
At Semilattice, we're trying to make it easier to do great product and marketing work by applying simulation to this problem. If you can simulate users accurately, you can start to predict how they'll react to a new feature, how they'll interact with a new UI, or how they'll interpret a new campaign. But the reason this makes a difference is that these simulations take minutes and cost almost nothing to run, which means the question of whether a new feature, new UX, or new campaign will succeed can be answered while it's still a work in progress.
This changes everything. If you can find out whether something will succeed while you're still working on it, you can change it and make it better when it's still easy and cheap to do so. You can fix a UI label which is confusing, or reprioritise a feature people care about more, or change an image to one which is more understandable. You can go into a creative, iterative cycle with the future: testing new ideas and new decisions until you find one which is predicted to be a success.
We call this "user in the loop" development, and a custom model of your users is all you need to get started.
Semilattice custom models let you do two kinds of user-in-the-loop simulations. User research simulations are like surveys: you ask your users a set of questions. User journey simulations are like user tests: you put your users through an interface and observe their behaviour. (Note: the accuracy validation below applies only to user research simulations at this time. We are working to validate the accuracy of user journey simulations, but our customers still love these simulations because they often surface self-evidently true findings.)
To make running simulations easy, our product is built around what we call the planner. The planner is an agent which knows how to apply user simulation to your product and marketing work. You can simply describe a feature, an update, a campaign, or a tactic you're working on and the planner will construct a set of simulations designed to predict the consequences of shipping that work. You can also just upload a PRD, prototype, messaging doc, or any other piece of work and the planner will use that to run simulations.
Once simulations are run, you get actionable recommendations backed up by data from your simulated users. The recommendations come as a report which can be easily exported for AI assistants to action immediately or for collaborators to review.
State-of-the-art accuracy
We test the accuracy of custom Anacreon models by getting them to predict the answers to test questions. Before we deliver your model, we hold back a set of real answers from the data it's built on. We then ask the model to predict them and compare its predictions to the answers real individuals gave. Every custom model ships with its own accuracy report, so you know how much to trust your model before you run your first simulation.
We measure test question accuracy using ordinal alignment: a score of how closely predicted answers line up with real ones. Anacreon has achieved an ordinal alignment of 0.775, ahead of every published academic baseline. Your model's score may vary slightly depending on your audience and the data it's built from.
What makes this score especially robust is that it's measured at the level of the individual. Every simulated individual answers many held-out test questions, so the score is built from thousands of individual comparisons between predicted and real answers, not from averages across the whole audience.
A mixture of minds
Most simulation methods build a single model of your whole audience. That can work at the aggregate level, but it breaks down at the level of the individual: one model biases towards the average, so results skew towards the preferences of majority groups and minority opinions get averaged away.
Anacreon takes the opposite approach. A custom model is a population of individual simulacra, each one a model of a person in your audience—a mixture of minds. Because every simulated person is distinct, diverse and minority opinions survive, and results can be segmented and re-cut any way you need.
We build your model from real, unstructured human data: qualitative interviews, reviews, forum and community threads, open-ended survey responses, and any proprietary data you can share. And because your users are a moving target, we can retrain your model on new data to keep it current.
Semilattice Custom subscription features
For $1,799/mo, you get:
- a custom Anacreon model for your target audience
- unlimited simulations
- unlimited seats
- organisation features
- API and MCP access
Semilattice Custom is a partnership, not a licence. Alongside the product, we meet with you regularly to help you point your model at the decisions you're actually making. If you'd like a custom model of your users, book a call.