passing this on: asking AI to design a programming language for AI

GOALCODE, I am guessing it needs an LLM to execute it though

Forth? Assembler?

I don't think that AI has critically thought about it though. This is like the first idea of every programmer, I'll make code more human readable (e.g. BASIC). But it never works out like that. Even though its lines are goal directed, the algorithm itself is procedural, so its inconsistent, but I enjoy hearing what the AI thinks is a good idea even if it isn't.

One thing I like about Lisp syntax is that it has “scoping” built in. I’m using OhmJS a lot and finding that I’m using “+” and “*” less and recursion more. OTOH, I’ve been taking Forth Haiku apart (https://feelingofcomputing.slack.com/archives/CCL5VVBAN/p1772568421543419) and marvelling at how concatenative languages can make building compilers very simple.

One thing I have often wished for in Lisp: that links between a variable and the scope defining it were more apparent. Would make many tree walking macro use cases much simpler.

Have you seen the Racket IDE? I saw someone using it and it appeared to do exactly this.

yeah I was thinking that racket does that

I'm thinking more of the macro writing experience. I want to see that information in the output of

(read ...)

But yeah I won't be surprised if Racket has that too. The syntax/datum divide there is itself quite powerful.

Why should we trust what Copilot is saying? I mean as a guess, sure the shared link is a good place to start. The question is how to check. And I think there's a simple answer: (1) have an agent write programs in the language with desired output handy, (2) have another instance of the agent try to run them, (3) iterate until the two match. I use a similar technique to do context engineering. It works pretty well.

Interesting point. What jumped out at me is the issue of token efficiency - something I hadn’t considered before. As it happens another article was directed at me at about the same time. This one seems to argue that Clojure is better for LLMs (wrt token efficiency). I’m currently interested in Forth-like and concatenative things like Forth Haiku and advances shown in recent svfig youtubes like this one and wondering how that might relate to modern programming and LLM-ery.

Lisp is very good for token efficiency, but the winner is Forth. There were some papers six or seven years ago that showed it was the easiest to learn for a class of then state of the art models.

Not SQL? I find these things tend to use SQL like it was just looking around the room.