(Late reply, sorry)

Well currently you can't get at the AST, only (with some hairy code) bytecode. (The use cases are post-compilation)

But suppose this work did enable that. What Paul is saying is that there are intermediate representations between bytecode and AST that are more helpful for these GPU / SQL / whatever runtime compilers. Representations that capture dataflows, for example. Chance are these compilers would transform an AST to something like this anyway.

See https://mlir.llvm.org/ which is referenced in that talk.