@phaedr0s The last code I wrote is a special case. The code itself is dwarfed by the documentation; the source is included as an appendix. Most of the documentation was a very detailed explanation of the physics and engineering the code implements, comparisons with other codes/reports, and all the equations, data, and citations (with specific page and equation numbers) required for another person to replicate the computer results by hand. The code itself is far less valuable than the documentation of what it does and how it does it; the whole purpose of my code is to validate a critical piece of a larger production code that is IMO not adequately tested or documented.

A pobabilistic text generator will not help with any of this and will almost certainly make the result worse and slower to produce (recall how many lawyers and researchers have been called out for fabricating citations from using GenAI...). These codes estimate how much radioactive aerosol makes it from damaged reactor fuel to a place where it could be transported to the environment - this is an application where it's more important to be right than fast. Writing code is an incredibly small part of that.

Using a chatbot is an unacceptable risk here and would degrade our team's ability to understand and defend our work against technical and regulatory scrutiny. But you don't have to be working on reactor safety and public health to be doing something important where you depend on the results being high quality or ensure that people are being responsibly cared for.