The LLMs write bad python as easily as any other language.

To make it good, you need to review and interate.

I think this means reviewing is the main thing with AI, and therefore the language to use should be one where reviewing is easy, for humans.

indeed, from my perspective doing heavy agentic dev for six months, the language all is implemented should be easy. its api - also. like Swift is super easy to read, but the APIs of underlying libs - not so much. Python is reasonably easy to read, but with GIL and everything is a very slow choice. Zig seems a nice apporach to readability, Rust is definitely not so readable.

we may want to invent something akin to concise math notion, but not so much, in between pseudo-code and math.

It is very hard to manually review AI generated code. May be we have to use another LLM to review and assume everything is good.

This hasn't been my experience. I find LLM code about as hard to review as human code, perhaps a little easier.