Khaled Messai

2026 · game engine

I want a world where people can speak a game into existence. You sit down with a model, you talk about the game you want, and the game comes into being. Crescent is the tooling to get there. Early on a lot of it was bandages for what models couldn't do yet: levels depicted as ASCII grids, because a model reads ASCII natively and can't read a scene editor. Later it became a collection of tools designed for a language model's cognition first. Trees, buildings and terrain can be described analytically, and tools like SpeedTree, Gaea and Houdini do exactly that, but they are native applications a model can't get into. A node graph is something a model pilots very well if you just make it JSON. So we made our own.

Oct 2025 – Aug 2026 · video understanding

For physical security, worksite safety, and autonomous management of physical platforms.

MetaNews

2026 · media analysis

I want to do something to help people fight back against the constant war that's being done across their minds, and mine. About 300 outlets in 22 languages, read side by side every day. Like Snopes, but trying to track them all at the same time. It traces back to my work in defense, and I genuinely believe it to be an issue.

March 2026 · music

People say a language model can't write music. It can, and it's beautiful. A series of experiments from March: we built a model, we tested the model, we built some canonical references, and we tried a whole bunch of stuff. It's all written down.

Motion studies

2026 · animation

I love working with a model on the things people say it's worst at. Claude can't animate, they say. So we measure it: every joint's path traced through the air, every part's speed charted frame by frame.