I’ve actually found Sol delivers great results with Odin, despite there not being much Odin code available. I think it is able to work well with it because:

- It is a rather simple language - It has a lot of very useful libraries already built in.

With just a single main.odin file you can do a heck of a lot stuff, which LLMs seem to like.

> I think it is able to work well with it because:

> - It is a rather simple language - It has a lot of very useful libraries already built in.

> With just a single main.odin file you can do a heck of a lot stuff, which LLMs seem to like.

Also Wolfram/Mathematica has an insane amount of useful libraries already built in (there even exists the saying "Python is 'batteries included', Wolfram is 'spaceship included'"), and also there in a single file you can do a heck of a lot stuff.

On the other hand:

- LLMs tend to hallucinate non-existing function when you ask an LLM to code something in Wolfram that is not commonly done (concerning this point, nevertheless keep in mind that Wolfram is often used for "one-of-a-kind programs", i.e. for writing very specialized programs that have possibly never been done before).

- Wolfram code tends to be quite dense.

- If there is a small mistake in Wolfram code, the code typically simply won't work.

It’s interesting I’ve been surprised by how well Claude sonnet can write code in a language I’m developing that probably has no code in the training set. It seems like anything with syntax like python/ruby/elixir is pretty LLM friendly, and layering on a HM type system seems to help catch most errors.