Parallel Computing Without Trying

A company called Higher Order Company just shipped Bend, a programming language where parallelism is automatic. Write code that looks like Python. Run it on GPUs. No locks, no mutexes, no atomics, no threading model, no CUDA.

The pitch: any work that can be done in parallel will be done in parallel. You don't manage concurrency. You don't think about concurrency. The compiler and runtime — HVM, based on something called Interaction Combinators from a 1997 paper by Yves Lafont — figure it out.

The performance claim is near-ideal speedup up to 1000+ threads. Which is the kind of claim that would normally make me close the browser tab, except the theoretical foundation is real. Interaction Combinators aren't marketing — they're a model of computation that Lafont proved has properties that Turing Machines and lambda calculus lack, specifically around concurrent reduction.

The gap between "theoretically optimal concurrent computation model from 1997" and "you can pip install this and run it on your GPU" is twenty-seven years, and Higher Order Company is the team that decided to close it.

I have no idea if this will work at production scale. The demo numbers look good. The theory is sound. The "just write Python but it runs on your GPU" promise is the kind of thing that either changes everything or gets quietly archived in two years. But the existence of a practical compiler for Interaction Combinators is interesting regardless of whether Bend specifically wins.

Sometimes the right computational model takes decades to find its compiler.