Transcript

Did you see that paper from Microsoft Research last week called Intent Formalization? The author, Chengdu Lahiri, described it as the grand challenge of AI coding, and his example is so simple. Tell an LM to remove duplicates from a list. Does that mean keep one copy of each item or strip out everything that appears more than once? Now obviously those are two completely different results and the alms just going to pick one. He says that AI generated code is plausible by construction, but not correct by construction. Now clarifying intent is exactly what an experienced engineer does when confronted with new requirements or unfamiliar code. They say things like I'm just checking, you said X here, did you mean Y? At Jax, we've been working on something called Allium. It's a specification language that supports exactly this kind of reasoning. Inside ALMS, you describe what your system should do in conversation with an LLM, and it will clarify what you want before implementing any code. We've also pointed it at existing code bases to infer their behaviour and found bugs that thousands of passing tests had missed, including tests that were actively confirming the wrong answer. So this is valuable for anyone with existing code that they don't know inside out. And I define legacy code as code where the gap between how it works and what you understand about how it works is wide enough now that you can't make changes with confidence. And if we're being honest, that usually includes code that you wrote with Claude yesterday. So I hope you'll take a look at Allium. I'll post the link below. I'd love to know how you get on and if you want help with your own code base then do please get in touch.