Urk. That was supposed to be @clattner_llvm. Apologies!.

Actually, maybe it's the only game in town if you want any two of those three properties. .

I'm pretty excited about the possibilities with Mojo, and I think more people should check it out. If you want a language that is pleasant to write in, that lets you write speed-of-light kernels, and lets you write kernels portably, it's kind of the only game in town.

I think it's a great approach, and it turns out that it's something we're doing in OxCaml as well (though it hasn't quite hit production yet.) And you can see some early work in this direction with dialects like MetaOCaml:

Template error messages are pretty tough, and the compile times are rough too. Mojo provides typed metaprogramming as a first-class language feature, similar in some ways to the approach Zig takes.

In CUDA, libraries like Cutlass use template metaprogramming to do this adaptation. But that's a problematic approach!.

But the constraints when targetting GPUs are really different! A critical thing when targetting GPUs is being able to adapt your code not just to the shape of the problem, but also to the weird and idiosyncratic shape of the hardware!.

Mojo is very much a language designed for performance engineering, which is an idea that's close to my heart. Indeed, a lot of the work we've been doing on OCaml is about providing the kind of control you need for performance work.

A new episode of Signals and Threads just dropped! This one is an interview with @ChrisLattner, talking about Mojo, a new-ish language for GPU programming that's aiming to be an alternative to the CUDA stack.

Will's research doesn't attack these kinds of questions, and instead focuses on smaller scale questions that are easier to study in a rigorous way. Which is great! But I nevertheless pine for rigorous answers to the bigger questions.

I'm somewhat dismayed by the methodological difficulties here. E.g., how do the major improvements to PLs over the last 50 years, like memory safety, GC, rich type systems, type inference, etc, influence productivity? We don't know!.

This was a great talk from Will Crichton. I think Will's approach to approaching questions around language tooling and teaching is compelling, though I wonder how far the approach can scale!.

Here's an example of a talk I did that was mostly presenterm:. I also did some vibe-coded ocaml/html/javascript visualizations. I actually quite like having those be outside of the presentation proper, a bit like you would with a demo.

My two favorite ways of doing presentations are pretty clearly org-tree-slide in emacs, and presenterm in Kitty. Having a system that gives me just enough flexibility, but doesn't require (or allow!) much visual futzing is ideal for me.

If you're looking for a fun and challenging way of spending the summer, consider applying! .

The post itself is here: .

I just finished our yearly roundup of this year's intern projects, and it was a real bumper crop! Lots of fun projects, and some cameos from OxCaml.

RT @__protected: On my way to the Effect Handler Oriented Programming workshop. I spotted this commercial at Waverley, Edinburgh. Didn't….

RT @satnam6502: @yminsky Another consequence of the “If it compiles, it works” property mentioned in the article is the good match very str….