Edit brushes and planet procedural (shader) brushes were the same thing and could blend. Here I took a hunk out of the ground and built a (Coder Art!) Hobbit-hole; modelled/cloned each brick, coded a door, made a table, made some apples, added collision. (some debug UI in here)

How scalable are these methods you used in your engine?

The brush representation is key. You get trivial streaming, orders-of-magnitude reduced dataset sizes, automatic LOD, manageable conflict resolution; content-addressed-hashing just "falls out" of this... this thing *scales* :)

And why did you abandon this? 🤔

Ran out of runway to pursue funding. We did the circuit for about 8 months, which is nothing compared to others. But seeing absolute travesties of products that you knew would never go anywhere, get $m's was utterly soul destroying. Didn't give me confidence in investors.

Does the collision detection use the SDFs directly, or do you generate regular collision geometry for each shape? (It kinda looked like there was some convex decomposition going on with the apples).

Both! Broadphase and narrowphase used convex decomposition from a generated voxel volume. SDFs were used to ensure volumes never interpenetrated too deeply. So failing through the world was a thing of the past, even with UGC.