Pick a real task you want an AI to do inside an organization. Decide whether a leave request can be approved. Flag the customer who is about to leave. Tell an elderly parent which pill to take, and when. Now swap the model underneath — Claude for GPT, GPT for Gemini, this version for the next. For most of these tasks, the answer barely moves.
That should be surprising. The field spends most of its attention, and most of its money, on the model. Yet the part that decides whether the work gets done well is somewhere else. It is in what the model knows at the moment you ask.
Context is the product
A model on its own is a capable stranger. It can reason, summarize, and write, but it has no idea what is true in your company or your home right now. Ask it who is out of office next week and it will guess. Ask it whether a customer has an open escalation and it will invent something plausible. The intelligence is real. The grounding is missing.
The usual patch is to paste more context into the prompt. That holds up until the context is longer than a paragraph, changes by the hour, or has to be exactly right. Then you need something sturdier than a prompt. You need a model of the world the AI is acting in, kept current, kept honest, and ready to be reasoned over on demand.
We call that a world model, and we think it is the actual product.
What a world model is
Not the weights. A world model, in the sense we build it, is a durable, structured picture of a domain that stays current as the domain changes. It has an anatomy:
- An observation log. Every relevant event, appended and never rewritten: a message arrived, a deal moved, a blood-pressure reading came in.
- Entity state. What is true now, derived from that log: this person’s leave balance, this customer’s open problems, this room’s layout.
- Change events. What just became true, so the rest of the system, and any AI acting on it, can respond.
- Grounded reasoning. Questions answered against that state, with the reason and the source attached.
The log is the part that matters. Because state is derived rather than kept by hand, it can be rebuilt from scratch, audited line by line, and trusted. When the world model says someone is out of office, it can point to the approved leave that makes it so. Chat cannot do that. Chat forgets the moment the tab closes, and it cannot show its work, because there is no work to show.
Two worlds
We build world models for domains where being wrong is expensive.
OmniManas is a world model of a company. HR, projects, sales, support, and knowledge sit on one shared graph instead of a dozen disconnected tools. An AI working inside it acts as a permissioned member of the team: it reads the company’s real state, takes an action, and leaves an audit trail, because it reasons over the same graph everyone else uses.
Saathi is a world model of a life. It runs on an elderly or low-vision person’s phone and keeps a patient, private picture of their day: the messages, the medicines, the appointments. When it reads a prescription aloud, or declines a call that looks like a scam, it is not improvising. It is acting on what it knows about that one person, and it can say why.
Different domains, one shape. A log, a current state, and an AI that reasons over it with the evidence in hand.
You cannot trust what you cannot measure
A world model that is confidently wrong is worse than none at all. So the third thing we build is not a world. It is the instrument.
Multivon measures whether an AI system is actually right. It runs 44 evaluators locally, attaches a confidence interval to every score, and lets a judge answer “unknown” rather than guess. We hold it to its own standard: we set a determinacy bar of 50%, measured 20.9%, and published that number. A tool that grades honesty has to be honest first.
That discipline sits under the whole studio. Evidence over adjectives. A world model earns trust the way a scientific result does — by showing the measurement, including the one that does not flatter you.
The bet
Models will keep getting better, and it will keep mattering less than the industry expects. The gap between the best model and a good-enough one is already small. The gap between an AI that knows your world and one that does not is enormous, and almost no one is building for it.
So that is the bet. The frontier is not a bigger model. It is a better world model around it, and the evidence that the model is right. The model is the easy part. We work on the rest.