Posted by nivter 9 hours ago
One can use broadcasting semantics similar to NumPy and PyTorch in a visual setting (imagine creating a list of circles where one dim corresponds to radius and another to the center). One can also use backpropagation, run gradient descent or visualize vector fields. Almost everything is reactive so changing a variable updates all of the downstream geometry. It also allows anyone to write and load their own visualization, which can be broadcasted and differentiated through.
- When I open an example, I expected to actually... see an example. I'm not gonna read the wall of text. I don't even understand what this is yet, that's why I tried to see an example.
The person it benefits the most is the author, when they are building it and the errors-per-use are as high as they’ll (hopefully) ever be.