The relation to AI is an assumption that AI and metaprogramming are connected. This is probably debatable, but it makes sense to me (and others) that AI development would benefit from being able to easily generate code and modify itself, much like you or I would learn a new skill.
If what you want is a system that rewrites itself, then you definitely want homoiconicity. It also helps if everything is one big expression.
Let's say you want to write an evolutionary algorithm that rearranges fragments of its code and runs it. More successful functions make it to the new generation. You want LISP for that.
Realistically, you could probably do this on any abstract syntax tree. But the problem is that unless your language is homoiconic, what you get back might not actually be valid in your language. So, homoiconicity gives you bidirectional support for rearranging the AST and still getting something you can read.
It's also just a lot easier to write something that looks at the code and does computation on it when that code is also data.
But, like I said, I'm not sure many people are actually doing this these days. It was popular back in the day. It's not clear that it's the right way to do things.
Scheme is a popular extension language. You could use it the same way Torch uses Lua.
Can you elaborate on this? Why are or what about homoiconic lnagunages make them more suitable for metaprogramming and AI in general?
Thanks