4 ms·

> Back in the day in Haskell I dreamed of a system where you could type out a type signature and a fully tested rated implementation would be imported from an "

• nightski 4y ago
  Well I'd argue it's more Model => String => String if you are talking about using a fully pre-trained model. Even then you could probably expand on the type signature of the model to make it more useful. But if you look at like co-pilot for example, if that was given the ability to have type signatures serve as input you might get a lot more powerful results than what it does with raw text (which is very impressive). But this comes down to type signature design. You can encode any function using simple types like int -> int which aren't very useful. Where Haskell shines is when using types to limit the scope of inputs & outputs. What I am getting at is that you can still write uninformative type signatures in Haskell, but it also gives you the power to write more informative ones. I don't think Haskell is the answer, so please don't take that as what I am saying. I do think however using richer type systems could be a stepping stone towards a solution to this problem.
  • astrange 4y ago
    Type signatures (or ASTs) are structured data, and Copilot mainly works because GPT accepts any text and doesn't have to deal with your logical fully recursive concepts like ASTs. I feel like it wouldn't work if you could actually mathematically constrain the outputs to be syntactically correct either. That's probably one of those Gödel things.