Building hierarchical abstractions on top of code is IMO the only way to truly enable AI to write code beyond better autocomplete. Kevin's work with DreamCoder shows that hierarchical abstractions can be built automatically in the code domain.

Once these abstractions exist side-by-side with natural language (essentially a code-natural language world model), it'll enable arbitrarily complex code generation from descriptions of the outcomes/results.

Sadness, the "Supplementary materials" url <https://web.mit.edu/ellisk/www/dreamcodersupplement.pdf> is 404 as well as the 2 wayback captures of it are also 404

While trying to track it down, I did find a seemingly related set of GH repos

https://github.com/ellisk42/ec/blob/master/dreamcoder/dreamc...

https://github.com/ellisk42/ecPaper/blob/master/talk.pdf

(2020)

How's the wake-sleep pattern different from the exploration/exploitation in Deep Reinforcement Learning?