Hacker News Reader

Comment by deterministic | Hacker News Reader

I'm surprised to hear that. LLMs make it practical to prove safety-critical C code correct without the cost that has traditionally made formal verification prohibitive. If you haven't explored that area, it's worth a look.

My experience with C and C++ has also been excellent. I work on very large, high-performance C++ systems used to run airlines and airports, and LLMs have dramatically improved both the quality and the performance of our code.

That said, LLMs are tools, and like any tool they take skill to use well. It took me months of experimenting and learning from others to get where I am now.

View on HN

I made mention on a reply to the other commentator but I feel the issue is due to lack of training data in the PLC/SCADA space. If an AI engine with the appropriate amount of training data was available to me I'm sure it could do the work just fine, but at the current state we have partial knowledge which is too similar to other languages. The result is agents which quickly hallucinate capabilities of the SCADA or PLC hardware and write out code that fundamentally won't work. The result of this is that I distrust AI to get it right enough that I won't kill an operator or blow up a production line. The automation world is too closed source and until the push for open standards that have begun to gain traction really takes off we're sort of stuck.

But then at the same time the code of an industrial system typically isn't the bottleneck in development. Selecting the hardware and waiting for the physical installation is a far bigger bottleneck on a project than what I do. As a result, in the end it isn't the "Correctness" of code in the sense of it is doing what it was instructed to do properly, but the correctness of matching software to hardware and that's where I've seen several failures in AI. If I tell it I have a Schneider Altavar VFD and I get code to interface with an Allen Bradley PowerFlex that's one issue. But if I get code to interface with a Altavar that expects specific parameter adjustments on the drive that haven't been set that's a separate issue and more typical for the faults I've seen within AI. If 99% of drives are configured to operate pumps and that's what AI learns for training, the moment I need it for a mill, aeration fan, or conveyor belt (which have distinct physical requirements) then we have the issues I've been seeing and it takes a trained eye to spot those deficiencies in my experience.

Thanks for the detailed reply. It’s always interesting to hear about real experiences from a different software domain.

So it sounds like the main problem is the under-specified hardware. I can definitely see how that would limit the usefulness of LLMs.

I used to work on consoles, including the N64, which are fairly low-level compared with a typical Windows, Linux, or Mac application. But you’re obviously working much closer to the hardware, which is a completely different world altogether.

Reply on news.ycombinator.com