The world-model moment.

On 29 June, the Beijing Academy of Artificial Intelligence released Orca, a research system with an unusually philosophical title: The World is in Your Mind. Behind the title sits a serious engineering claim. Orca is a world foundation model: instead of learning to predict the next word in a sentence or the next frame in a video, it learns a single latent representation of a world’s state, and learns how that state changes. The authors are direct about what they think this means. Intelligence, they argue, is not next-token, next-frame or next-action prediction alone; it is the ability to model world states — to understand, predict, and act upon the world.

The idea has a lineage. Sutton’s Dyna sketched it in 1990; Ha and Schmidhuber’s World Models revived it in 2018; MuZero showed in 2020 that a system could learn to plan by modelling only what mattered for the decision, without ever being given the rules. A world model, in the technical sense, is a learned model of an environment’s dynamics: given a state and a candidate action, it predicts the next state and the outcome. And once a system has one, something genuinely new becomes possible. It can evaluate an action in imagination, before taking it in reality.

How a world gets learned

Orca learns its world in two ways, and the paper’s vocabulary for them is striking. Unconscious learning absorbs dense, natural state transitions from 125,000 hours of video: the world simply unfolding, frame after frame. Conscious learning absorbs sparse, meaningful transitions described in language: 160 million annotated events, each one a moment somebody thought worth naming. Watching everything, and being told what mattered.

Notice what made this possible. The physical world is, in a sense, lucky: it has footage. Billions of hours of it, recorded as a by-product of human life, dense enough for a model to learn how the world behaves just by watching. Whatever you think of any single paper, the direction is unmistakable, and Orca is one of several signs that the field’s centre of gravity is moving toward it: modelling the state of the world beneath its surfaces.

The worlds without footage

Here is the question that matters for banking. Some of the most consequential worlds AI will ever act in are not physical at all. A lending operation is a world in exactly the technical sense: it has states (a borrower, a portfolio, a macro backdrop), actions (approve at this price, restructure, schedule this contact within permitted hours), and consequences (cure, default, prepayment, complaint, recovery). Everything a world model needs — except the footage.

A lending operation’s footage would be the record of its behaviour: every action taken, the alternatives that were considered and rejected, the reasoning, and what happened next. That record does not exist. Core systems record the transaction and discard the reasoning behind it. The decisioning stack is a chain of single-purpose models, each seeing one slice of the lifecycle, stitched together by workflow software. And nothing in that stack logs the rejected choice set: the prices, limits and contact strategies that were evaluated and not chosen. The raw material a world model would learn from is discarded at the moment of decision, every day, at every lender.

The record comes first

This ordering is the heart of the matter, and it is why you cannot get to an institutional world model by pointing a bigger network at a bank’s data warehouse. The footage has to be produced, and producing it takes one runtime running the whole lifecycle — origination through servicing, collections, disputes — writing every action to one ledger with the choice set it was selected from, the reasoning, and the outcome. And because the world in question is a regulated lender, that record has to be produced under discipline: every action validated before it executes, so the runtime that generates the footage is also the gate that holds each action to the operation’s obligations before it runs.

Kovida: a world model with a jurisdiction

This is the thinking behind Kovida — the world lending model: the learned, action-conditioned model of a lending operation that becomes possible once, and only once, every action, its reasoning and its alternatives live on one validated ledger. Where Orca’s world is physical and its supervision is video, Kovida’s world is an operation and its supervision is the ledger KrimOS writes as it works. It is built going forward, and it is an active area of Krim research.

The near-term shape is deliberately disciplined: decision support grounded in the operation’s own recorded behaviour; short-horizon sequence decisions in servicing and collections, where outcomes arrive in weeks and a model’s claims can be checked against reality quickly; scenario exploration whose uncertainty widens honestly with the horizon. The discipline is the point. A world model of a regulated operation earns trust the way the operation itself does: one validated action, one reconciled outcome at a time.

Orca’s title is better than a slogan; it is a definition. A model’s grasp of a world lives in what it has absorbed of that world. For the physical world, that is a hundred and twenty-five thousand hours of video. For a lending operation, it is a ledger that remembers everything — including the roads not taken. The world-model moment has arrived in AI research. In lending, it begins with the record.