HAL Update: Data Plane Specification

Time for some Friday fun. A couple of weeks back I posted an idea for a little experiment to see just how good AI first coding can be. Or not. As a quick recap the idea was to see how well it would cope with building a next generation HPC scheduler / workload manager. The idea here is first and foremost to see how good AI coding is without the hype and to be fully transparent about how everything was done.

This is far from perfect. It’s a little too manual, error prone and labour intensive but for now I’ve added prompt history markdown files to the repo. Any time code is generated using AI the details are captured into the prompt history file, and the commit should include that file as well as stating in the commit comment that AI was used. I’ve also segregated commits such that any human written code is not in the same commit as AI generated code. Folks at Anthropic/ Cursor etc if you’re watching something like would be pretty handy to enable easier code reviews too as well as understand change history if it could be captured automatically.

So far, I’ve only used VS Code with Copilot (I now prefer this to Cursor curiously) but that’s because I’m still at a stage where I’m specifying what needs to be built. Yes even that is getting done using AI. The intention is to let Claude Code / Open Code / Codex loose on this once there is a good specification. Firstly, to create the tests and then later the implementation. 

I’m starting with what I hope will be something easy, controlling the data that the compute will need.

Back over to you again for a bit. What’s missing? What else would you like an HPC data plane to do?

GitHub - hmxlabs/hal

Contribute to hmxlabs/hal development by creating an account on GitHub.

GitHubhmxlabs