I’ve been experimenting with Flet (a Python framework for Flutter) to quickly prototype cross-OS apps, mostly to easily get to iOS. It turns out Flet/Flutter is not great on iOS, and a number of packages don't work on iOS (PyTorch, Sentence-Transformer along with more usual ones).

I got frusterated and had Aider+Claude just rebuild my app in Swift (+ some server calls for sentence-transformer) and it worked surprisingly quickly (and almost well)!

This got me thinking ... maybe the future is building in janky expressive Python (or high level language of your choice) and using LLM enabled transpilers + profilers + unit tests to convert the bits that need performance or platform compatibility into C, CUDA, Kotlin, or Swift as needed?

I'd love to learn more about where the world is going here. I found a few papers (e.g. https://arxiv.org/abs/2406.03003) and I'd love to learn more about what compiler folks think, what kinds of benchmarks exist and what directions people are going in. I can only assume that PyTorch -> performant fused CUDA kernels is a thing that any number of folks are looking into.