Skip to content

What Glamin taught me about AI-native language design

Context: So it's not that Glamin doesn't work as a programming language. It does. It's just going to be a super awkward deployment, even for training scenarios. This isn't failure so much as valuable information.

I think the solution is to do this a little differently. Instead of FAISS style embeddings, the better play is trained embeddings which effectively gives me a frozen model I can use to do the actual code execution.

So you gain all the benefits of having embeddings, but none of the flexibility of FAISS style embeddings, which you actually want. Originally, my feeling was that embeddings should be flexible, and code should be fluid.

But the lesson I learned, well, one of them, was that code shouldn't be fluid. It should be durable. Sort of. Maybe I've been putting fluidity on the wrong layer. That's interesting. We'll put a pin in that.

The syntax does not have to mirror the backend

Next thought: The other thing I learned, mainly from working on the Ruby project was that the syntax of the language itself doesn't necessarily have to reflect the back end that processes it in any direct way.

Compilers are flexible.
Also interesting.

That gives you some freedom to think about the things that are important. Ease of use, and good priors. Ideally, some fluidity.

There's that word again.

But when I think of fluidity, if that's the goal, which it should be, to put it on the right layer... my mind goes to dynamic languages. Things that already exist, or used to. Ruby's okay, but honestly, Ruby's boring. Perl is fluid, very fluid, but not your best candidate for meta programming that gets compiled into vector space.

And that leaves us with one viable candidate.

An older version of PHP.
Maybe v4, before the mailing list got all self conscious about wishing it was Java?

It makes sense, if fluid means meta-programmable.

It would need some cleanup.
But it could work.

What a strange animal that would be.
Like a little purple elephant with a flame thrower and agent context.

If I could get it to do what Glamin did, it would represent a really interesting way to write workflows, and use tools that nobody else is doing seriously yet.

Need to think this out.