General Intuition

Most AI company sites open with a promise and a waitlist. This one opens with an argument: models trained on the archive of human language inherit its limits, which the page calls planar abstraction without embodiment, prediction without participation. You have to read three paragraphs before anything is offered to you. For a frontier lab that is the correct order, because the product being sold is judgement, and judgement can only be demonstrated by argument.

The essay does something else that is easy to miss. By citing Johan Huizinga's Homo Ludens from 1938 and building the case that play is how learning gets cheap, it filters its readers hard. A recruiter skims past it. A researcher who has thought about exploration and exploitation stops. That is a more honest audience filter than any "who is this for" section, and it costs the page nothing because the people it loses were never going to apply.

Then the claim gets made checkable. MIRA is a playable multiplayer world model, and the page offers the live demo, the technical report, the code on GitHub and the blog post next to each other. Ten thousand hours of data, four-player Rocket League dynamics, twenty frames per second: every one of those is a number someone can dispute, which is exactly why stating them beats any adjective. Naming Kyutai and Epic Games does the same job from the other direction.

The quietest strong move is explaining where the data comes from. Building on Medal, a platform for gamer moments, gives the lab action-labelled video across countless environments, which is the one thing a competitor cannot simply buy compute for. Saying it out loud turns a corporate detail into the argument for why this lab and not another. All of it on Webflow, white on black in a single typeface, with the restraint you find across the sites in minimal design: when the writing is the product, the design's job is to get out of the way.