Ostension

Helsinki · FR8 · 2026

Assistants model the task. We model the person asking.

Ostension trains models that infer what a user believes, wants and is trying to do — so a system can act on what was meant, not only on what was said.

Definition

ostension n.
The act of conveying meaning by pointing at a thing. Every request is an ostensive act: the words are the gesture, the intent is what they point at.

The gap

Language underdetermines intent.

“Don't worry about the deck” can mean I have it handled, stop interfering, or tell me it's fine. Nothing in the sentence settles which. People resolve this constantly by running a model of the other party — what they know, what they want, what they are in the middle of — and reading the utterance against it.

Assistants trained to follow instructions skip that step. They map an utterance to a task and then optimise the task, which is why they are competent and still miss: the failure is rarely execution, it is having inferred the wrong goal and pursued it well.

Reading

What a user model reads.

The same utterance, resolved against a model of the speaker rather than against the sentence alone.

Inferred state

Don't worry about the deck, I'll handle it.

Belief
Thinks you are already overloaded with the launch checklist.
Desire
Wants the deck done well, and wants not to be a burden about it.
Intention
Taking the task off you — not asking for permission or for help.
Common ground
You both know the deck is due Thursday and is currently three slides.

Task-only readingBegins drafting deck slides.

Intent-aware readingStands down on the deck; surfaces the Thursday deadline once, then leaves it alone.

What we build

A theory of mind, learned rather than prompted.

We train models whose objective is the user's mental state — belief, desire, intention, and the common ground between user and system — as a first-class representation rather than a by-product of instruction following. That state persists across a relationship, updates on evidence, and is what downstream behaviour conditions on.

This is a training problem, not a prompting one. The representation has to be supervised, evaluated against held-out behaviour, and made robust to the fact that people are inconsistent, change their minds, and say the opposite of what they mean.

Commitments

Three things we hold to.

  • Learned, not elicited

    A user model written into a system prompt is a description of a user. We are interested in the thing that is trained, measured against what the user actually does next, and wrong in ways we can count.

  • Legible by construction

    The inferred state is inspectable and correctable by the person it describes. A system holding a wrong model of you should be something you can see and edit, not something you discover through its behaviour.

  • Calibrated over confident

    Knowing when the model of the user is uncertain matters more than being sure. Asking and abstaining are valid actions, and a system that infers intent is one that can be confidently wrong about a person — which is the failure mode we are most careful about.

Contact

sam@ostension.ai

Ostension is early. If you work on pragmatics, user modelling, personalisation or alignment — or you want to try what we are building — write to us.