yminsky.bsky.social
Occasional OCaml programmer. Host of Signals and Threads http://signalsandthreads.com
2,904 posts
1,522 followers
294 following
Regular Contributor
Active Commenter
comment in response to
post
Yeah I appreciate that. High quality empirical studies are just really hard to do. It's still weird and sad that what seems to me, qualitatively, as an obvious fact (that type systems are extremely useful for helping you write more correct code faster) is so hard to verify empirically.
comment in response to
post
I'm especially curious if any of this covers languages with comparatively strong pay systems, eg, Rust, OCaml, Scala, Haskell, etc.
comment in response to
post
I mean, I assume it will just be the regular thing you get when humans are coding, which is that sometimes you will end up ping-ponging back and forth between the mli and the ml.
comment in response to
post
But only start working on the implementation after the mlis are done. It seems like a nice way of structuring things, and making the work of the AI legible.
comment in response to
post
Which...seems pretty good? I'm playing around with it now using Claude Code, and it seems pretty solid. You start by generating mli's, and can then negotiate with the API to get them right, and even make sure that the rest of the project works with those new mlis.
comment in response to
post
That said, we don't expect you to know OCaml or any other fancily-typed language.
comment in response to
post
Worth noting: we're looking for engineers who love front-end work, but are happy working on both front and back-end. The job here has the same basic requirements as our ordinary SWE role, with the additional requirement of front-end experience.
comment in response to
post
And, this being Jane Street, we have our own bespoke framework for building UIs in OCaml, called Bonsai.
github.com/janestreet/b...
It's really a beautifully designed library, and it's getting better all the time.
comment in response to
post
because currying is kind of dumb, really. it absolutely destroys error messages (it's probably the biggest contributor to the quality of ocaml's and haskell's error messages), it makes unknown calls really complicated to compile efficiently and in return you get to write less readable code
comment in response to
post
Indeed, I am pleased to announce the publication, by Harvard University Press (@harvardpress.bsky.social), of my new book, *The Oldest Constitutional Question: Enumeration and Federal Power.*
(2/13)
comment in response to
post
Is there a write-up of the details of this anywhere? It would be fun to read!
comment in response to
post
But if you're wrong, i.e., you assumed it was normal, but I'm fact it's heavy-tailed, then you will end up biased low in the way you say.
But I would have naively thought that things like chess ability were normally distributed...
comment in response to
post
I would have thought it was more complicated than that. You can compute an unbiased estimator of a distribution from a sample, if you know the family of distributions you're sampling...
comment in response to
post
IO! I'm not crazy...
comment in response to
post
What's wrong with snake case? I find it so much more readable.
comment in response to
post
stephen dolan doesn't miss
comment in response to
post
And if you're a vim addict, some fun things we've open-sourced in this space.
github.com/janestreet/v...
comment in response to
post
You can apply in the usual place, and mention your interest in the editors space.
janestreet.com/apply
And there's a lot of interesting work on integrating AI Assistance into editors as well!
comment in response to
post
Does this thing have a thumbs-down button?
comment in response to
post
It's been four hours! I do have a job you, know.
comment in response to
post
The results certainly look nice. I've tried that approach in the past and I found myself a little frustrated by dealing with the pile of JavaScript that is reveal. But maybe I should try again.
comment in response to
post
Got it. So reliability and ux, but not quality of completions.
comment in response to
post
I'm curious in what ways completion is better.
comment in response to
post
We have tools like this for programmers, but has anyone built integrations for these tools into agent platforms?
comment in response to
post
Here's an obvious thing I think you'd want: The ability for the agent to navigate by looking at the hard structure of the code. I.e, look up the definition of a given identifier, or find all or some of the use cases of a function defined in your file.
comment in response to
post
That's very much the intention, but we'll see how well it succeeds in the end.
comment in response to
post
Odin looks nice, but I find it hard to get excited about a language that doesn't have a useful memory-safe subset.