[sterlind]

the 1% is extremely important though, and the tools for it look very different.

I do algorithm design, a lot of which is seeing what's possible/feasible before committing. I prototyped a classical AI planning system for orchestrating maintenance. there's no battle-tested standard library to reach for here. I needed prolog, I chose picat because it had a good planning engine.

before this, I worked on an NP-hard graph problem that involved tons of topological transformations for heuristics. I wrote the damn thing in C#. it was a nightmare. it's so hard to iterate and experiment in a language like that. even using something like Python/Jupyter, nested loops and horrible messes of bidirectional maps and indices bog you down. I should have done functional. nobody touched my code for two years because it was pure math. I think using C# made it much harder to understand - too much scaffolding obscured the semantics.

now I'm doing another NP-hard graph problem. this time I'm using MiniZinc. because it's concise, because it's powerful enough that I can at least figure out how to specify my problem and see where scale falls down.

so I'm in this position of using a menagerie of bizarre languages not because I'm a hipster obsessed with pretty code, but because my problem domains are really weird and maintainability doesn't matter when you're not sure if what your goal is even possible. something like Lisp - a lingua franca for weird DSLs - would be great, for my very niche use case.

[taneq]

Absolutely true, and it sounds like you're doing really interesting stuff where the language and the way it maps to the problem are more important than the factors I listed above.

Do you stick with the same (maybe-)esoteric language for the production implementation? And if so, does this then drag the rest of the production implementation into that language as well, or do you do the hard bit in one language and the routine stuff in another more common language?

[sterlind]

I do the hard bit in the esoteric language, then the glue in a traditional language. for instance, I made a tool that exhaustively steps an implementation through every possible state of a TLA+ model. I wrote "serializers" to parse TLA+ expressions into C# classes and vice-versa, and used this to build a bridge. Parser combinators work great for this stuff, though that's another reason I'd be happier with a Lisp representation for everything.