I've been playing with some of the new micro/nano LLMs, and it has made me wonder whether Carl's old objection to large programming languages has acquired a new relevance.
We keep making LLMs larger partly because the software environments they have to understand are enormous.
But LLMs don't normally program in x86 assembler. Higher-level languages reduce the amount of machinery they have to reason about.
So how far does that go?
Rebol could compress a lot of implementation detail into a small problem-specific dialect. Glon is taking a similar idea to an extreme: Forth-like smallness, but with Rebol-style blocks and dialecting.
For a human, the attraction is familiar: reduce the program until the important part fits inside a human cognitive window.
But perhaps the same applies to small LLMs and their context/reasoning window.
Instead of spending more tokens and using a larger model, could we make the programming language more semantically efficient?
Don't spend more tokens. Make the language more efficient.
There ought to be a limit. Natural languages seem to trade information density against transmission rate: some communicate more information per syllable but are spoken more slowly, and vice versa. Perhaps programming languages have a similar trade-off between token count, semantic density and the reasoning required to unpack each abstraction.
It suggests an experiment: give the same nano LLM the same application task expressed first against a conventional API and then through a small Rebol/Glon dialect, and compare reliability and token/reasoning cost.
Maybe one use for the old Rebol idea of dialecting is to make useful programming possible for much smaller AI models.
Or perhaps we've just reinvented abstraction again.
I have often thought that Ren-C or another Redbol language might be good for ontology modeling. It’s one of the reasons I wanted to develop a flexible query dsl.
Ontologies matter because it’s usually better to feed the AI reasoning model / LLMs a more narrow context of the search domain instead of having them review the entire corpus with each query. Ontologies map meaning to categories, graphs then provide facts/evidence, allowing a tool like Jev to decide what is most likely to produce an accurate answer from the LLM.
There is actually a technique people use which is to speak "caveman" (as per @BlackATTR's avatar...) to the LLMs. This uses less tokens.
You're also supposed to not bother with "politeness". Say 'that worked' not "Thanks! I'm glad that finally works, I was worried it wouldn't...".
(Not sure I agree with this. But then again I'm one of the people who would edit out signatures and superfluous things from StackOverflow posts to get to the point faster.)
As I am making my OSM language, it sometimes surprises me how well Gemini and ChatGPT pick up on it.
At least to the point where they think it is some colorful language. I know exactly where they mess up. Easy fixes.
If the language is small and regular enough, the AIs stop needing to be spoon-fed every rule and start inferring the language from examples.
What has surprised me is how quickly ChatGPT, DeepSeek and Claude now pick up Glon code patterns once they have seen a little of it.
And I think Rebolek is starting to see the same thing from the other direction: he is reimplementing the seven-instruction VM in Recoil rather than just treating Glon as some opaque WASM blob.
That is the interesting bit for me — Glon is becoming a semantic target, not just one implementation.
