Look at the "temple"-example, there it is easier to reason about than the more abstract objects:
I tried to give it a new life in https://ossia.io but it's decidedly not real-time friendly, maybe geotoy will work better for this!
shadertoy: given a pixel coordinate, write a function to return a color.
geotoy: write a function to return vertices
In shadertoy, to draw cube on a 1280x720 screen requires your function to be called 921k times. To draw a cube in geotoy requires your function to be called 36 times. That's over simplifying but saying one is fast and one is a slow is comparing atoms to balls. Shadertoy = describe every atom, geotoy = describe the shape. again, bad analogy I suppose, but the orders of magnitude difference are somewhat help explain why one is fast and one is slow (yes though, you can generate a geotoy that will be slow and a shadertoy that will be fast). Similarly describing every atom means you can describe fog or god rays or caustics etc. Describing the shape won't do that.
PS: I think geotoy is cool. nothing above is meant to detract from that. Just to explain, probably poorly
In some way, multicellular lifeforms (like a cherry tree or a human) also have to contain a kind of formulation like this so that they result in the fractals of branches and blood vessels as well as the overall structure.
The article I linked in another post here describes that it's all done with triplanar mapping, so no UV's involved.
https://holtsetio.com/lab/chromeBallet/
https://holtsetio.com/lab/flow/
https://threejs-journey.com/ (click the up arrow next to "Level 1" at the bottom center)
https://turtlespaces.org/weblogo
We're in the midst of upgrading to three.js with all that it offers (materials, shadows, etc.)
I see this as a higher level DSL for modelling, instead of doing something like SDF by hand.