Ask a person how they learned to ride a bike, play chess or read a room, and the answer is never a list of facts. They were somewhere, wanting something, and the world pushed back.
Most of what we call intelligence is that: being somewhere, seeing what happens, and changing what you do. Language and pictures are how we describe it afterwards. The doing happens in a world.
We think the next stretch of progress in AI depends on the same thing. Models read an enormous amount. They've been given far fewer places to act, to be wrong in a way that gets measured, and to try again. So we're building those places.
A world is more than a scene
A picture of a kitchen is not a kitchen. A world has things in it that can change, rules for how they change, and a way to tell whether what you did worked. Take any one away and there's nothing to learn from.
That's the test we hold ourselves to. A world is only a world if someone can act in it, the rules are clear, and you can check the result without taking anyone's word for it.
- StateWhat exists right now.
- ObservationWhat the player or agent is allowed to see.
- ActionWhat it can do next.
- RulesWhat the world allows, and what follows.
- OutcomeWhat changed, scored against an objective.
- VerificationWhether anyone can reproduce and check it.
Why bricks first
We started with something small enough to get exactly right. A real brick has a shape, a way of clicking to its neighbours, a weight and a place it can sit. You can't hang a brick in mid-air. A wall with no overlap topples. A design can either be built by hand, or it can't.
That makes bricks an unusually honest world. The goal is easy to state, the rules are physical, and a finished model can be checked against them. People love it, too. A photo becomes a set you can hold, and the people who design them get better as they go.
Who decides what gets learned
Whoever writes the goal decides what is learned. Reward a model for finishing fast and it will learn to hurry. Reward it for a design that stands, can be built and still looks like the picture, and it learns something closer to craft.
That's why we don't want to be a black box that hands out scores. The rules of a world should be readable. A person should be able to open it, see what counts, change it, and play it too. Worlds that people can make and remix teach us far more than worlds only a lab can build.
Two halves of one company
The first half is for people. Studio is where people and agents edit the same construction. Play is where those worlds become games. Runtime gives developers the same worlds through an API, Marketplace lets people publish, sell and remix, and Worlds covers structured places, including real ones drawn from map data.
The second half is research and infrastructure: spatial intelligence, game intelligence, learning environments, benchmarks and evaluation. It exists because the first half exists. Real people making and playing real things is where good environments come from.
What is real today, and what is not
Create is live: you can turn a picture into a brick kit you can build. The workbench is live and under heavy development. The scoring environment, the agent comparisons and the wider platform are research and plans, not finished products, and we will say so wherever they appear.
When data from people's play or design is used for research, it is with their consent and under a licence that covers it. That is not negotiable, and it is not something we will trade for speed.
Why this matters to us
We think the most useful machines will be ones that learned by acting in places that are well made, fair and legible. We'd like people to be able to build those places, and to play in them first.