This is what I want to see from a programming language: https://s.ai/nlws/

First brush: Can't compute.

Do it.

This might not be quite the same thing, but here's an idea I had: A diagram consisting of data (and/or labeled placeholders for data), some is which can be visually nested (lists, key-value maps), and connections (e.g. arrows) that show operations between them. Copy/assign A to B is an arrow from A to B Conditionals connect a condition to an operation(s). Either a bubble around the operations, or an indicator next to the line representing the operation (and all other operations that stem from it). A map/select operation where one end is a collection, and the other represents each element. Either some other connector "down the line" that "collects" it all, or a bubble around the whole map. In either case, the output is the new collection. Similar symbols for filter, reduce/aggregate, sort, etc. Some sort of haskell pattern match. For example, an arrow from A to some (partially specified) nested structure, and then connectors from parts of that nested structure to further operations (which only happen IFF the match succeeded in the first place). There's no inherent order to anything, other than by dependency. It's a DAG that you can trace forward or back.

Frege mentioned right up front: at least knows to reference. Will look further. Might have a fun idea or three. Of course, also has a feel of "watched Arrival, let's do this!"

@noahtren would something like a different notation like this help guide the training process in your experiments?

@Cole this blows my mind, basically exactly what I'm hoping my NN system can do, but applied to code