AI First Coding in HPC
Trying to answer the question: Can AI write mission critical high performance code.
Can AI really code? I don’t mean generate a good prototype; I mean real production code good enough to run a bank on. Or a nuclear reactor.
It would seem that 2026 has kicked off with a large uptick in chatter around the capabilities of LLMs with tools such as OpenCode and Claude Code to accelerate development. And I don’t mean just from the usual AI hype as a service providers either. We’ve seen people with an anti AI stance in the past adopt LLM assisted coding. Names such as Linus Torvalds himself (allegedly).
I know what you’re thinking, these are just toy projects. CRUD apps. Front end candy. AI still can’t write any serious code. Not where its mission critical. Not where performance really matters. Right? I don’t know the answer to that. But I have a way for us to find out.
AI assisted coding accelerates us where we know the domain. Where we have the expertise already. Where we know what we want. Guess what, I think I might just have something in that space. One of my recurring conversations with clients is around the future of scheduling HPC workloads and the features that they’d like to see. The AI community is positively screaming for a single scheduling solution that can solve their problems. Huh. I think that might be a space we know a thing or two about.
So let’s do it. Allegedly Claude Code and its ilk can do months of work in days. For us to even think about writing a scheduler would be total insanity normally. If the AI hypers are right though, using AI first coding techniques this should be perfectly feasible. Let’s find out. Together.
We’ll try and create a new scheduler. But not just a scheduler, it needs to include a cloud control plane, observability, it needs to be able to geographically and network map the hardware its running, it needs to do data aware scheduling, it needs to use ML to predict job run times to maximise utilisation rates, it needs to have the speed of IBM Symphony/ HT Condor but the topological job placement abilities of Slurm. And it all needs to be modular so you can swap out one bit for another. Yes, I’m asking for the world. With a cherry on top. No point in making this easy.
And we’ll take you along for the ride. All in the open. Every AI generated bit of code committed before human intervention. If the git commit log tells a story this will be a page turner.
The whole thing is MIT licensed. I mean if we’re saying the value of software is trending to zero with LLM coding nothing else makes sense right? Plus, I take zero responsibility for anyone actually trying to use it!
Which brings me on to a rather important point. This is first and foremost an experiment in AI first coding. It may well not work. We may well abandon it half done. We absolutely have no plans to ever suggest running this code in production and we certainly don’t plan to support it.
The most important deliverable here won’t be the code. It will be what we learn along the way.
Welcome to HAL.
hmxlabs
