I haven't worked in OCaml but I have worked a bit in F# and found it to be a pleasant experience.

One thing I am wondering about in the age of LLMs is if we should all take a harder look at functional languages again. My thought is that if FP languages like OCaml / Haskell / etc. let us compress a lot of information into a small amount of text, then that's better for the context window.

Possibly we might be able to put much denser programs into the model and one-shot larger changes than is achievable in languages like Java / C# / Ruby / etc?

That was my optimistic take before I started working on a large Haskell code base.

Aside from the obvious problem that there's not enough FP in the training corpus, it seems like terser languages don't work all that well with LLMs.

My guess is that verbosity actually helps the generation self-correct... if it predicts some "bad" tokens it can pivot more easily and still produce working code.

My completely non-objective experiment of writing a simple CLI game in C++ and Haskell shows that the lines of code were indeed less in case of Haskell.. but the number of words were roughly the same, meaning the Haskell code just "wider" instead of "higher".

And then I didn't even make this "experiment" with Java or another managed, more imperative language which could have shed some weight due to not caring about manual memory management.

So not sure how much truth is in there - I think it differs based on the given program: some lend itself better for an imperative style, others prefer a more functional one.

If LLMs get a little better at writing code, we might want to use really powerful type systems and effect systems to limit what they can do and ensure it is correct.

For instance, dependent types allow us to say something like "this function will return a sorted list", or even "this function will return a valid Sudoku solution", and these things will be checked at compile time--again, at compile time.

Combine this with an effect system and we can suddenly say things like "this function will return a valid Sudoku solution, and it will not access the network or filesystem", and then you let the LLM run wild. You don't even have to review the LLM output, if it produces code that compiles, you know it works, and you know it doesn't access the network or filesystem.

Of course, if LLMs get a lot better, they can probably just do all this in Python just as well, but if they only get a little better, then we might want to build better deterministic systems around the unreliable LLMs to make them reliable.

I have found that Haskell has two good things going for it when it comes to LLM code generation. Both have to do with correctness.

The expressive type system catches a lot of mistakes, and the fact that they are compile errors which can be fed right into the LLM again means that incorrect code is caught early.

The second is property based testing. With it I have had the LLM generate amazingly efficient, correct code, by iteratively making it more and more efficient – running quickcheck on each pass. The LLM is not super good at writing the tests, but if you add some yourself, you quickly root out any mistakes in the generated code.

In Scala, I've had excellent luck using LLMs to speed up development when I'm using cats-effect, an effects library.

My experience in the past with something like cats-effect has been that there are straightforward things that aren't obvious, and if you haven't been using it recently, and maybe even if you've been using it but haven't solved a similar problem recently, you can get stuck trawling through the docs squinting at type signatures looking for what turns out to be, in hindsight, an elegant and simple solution. LLMs have vastly reduced this kind of friction. I just ask, "In cats-effect, how do I...?" and 80% of the time the answer gets me immediately unstuck. The other 20% of the time I provide clarifying context or ask a different LLM.

I haven't done enough maintenance coding yet to know if this will radically shift my view of the cost/benefit of functional programming with effects, but I'm very excited. Writing cats-effect code has always been satisfying and frustrating in equal measure, and so far, I'm getting the confidence and correctness with a fraction of the frustration.

I haven't unleashed Claude Code on any cats-effect code yet. I'm curious to see how well it will do.

I think that functional languages do actually have some advantages when it comes to LLM's, but not due to terseness.

Rather, immutability/purity is a huge advantage because it plays better with the small context window of LLM's. An LLM then doesn't have to worry about side effects or mutable references to data outside the scope currently being considered.

> My thought is that if FP languages like OCaml / Haskell / etc. let us compress a lot of information into a small amount of text, then that's better for the context window.

Claude Code’s Haskell style is very verbose; if-then-elsey, lots of nested case-ofs, do-blocks at multiple levels of intension, very little naming things at top-level.

Given a sample of a simple API client, and a request to do the same but for another API, it did very well.

I concluded that I just have more opinions about Haskell than Java or Rust. If it doesn’t look nice, why even bother with Haskell.

I reckon that you could seed it with style examples that take up very little context space. Also, remind it to not enable language pragmas per file when they’re already in .cabal, and similar.

I think LLMs benefit from training examples, static typing, and an LSP implementation more than terseness.

To be completely honest, I currently only use LLMs to assist me in writing documentation (and translating articles), but I know that other people are looking into it: https://anil.recoil.org/wiki?t=%23projects

I think this is putting the cart before the horse. Programs are generally harder to read than they are to write, so optimizing for concise output to benefit the tool at the potential expense of the human isn't a trade I'd personally make.

Granted, this may just be an argument for being more comfortable reading/writing code in a particular style, but even without the advantages of LLMs adoption of functional paradigms and tools has been a struggle.

Procedures can be much more concise in functional/ML syntax, but many things are not -- dependency injection in languages like C# for example are able to be much less verbose because of really excellent DI libraries and (arguably more sane) instance constructor syntax.