Only by a miracle can a Level -2 organisation produce any useable software. As Level -2 organisations rarely get beyond specification they pin their hopes on automatically generating a program from that specification.

• Anthony Finkelstein, Software Process Immaturity Model

LLM’s are doing an impressive job of getting us from conversations to requirements to specifications to code, but when I look at that diagram, I have a visceral sense of what 2025’s Level -2 is.

What scares me is that we can so easily get to something that looks right, but turns to spaghetti within a year because the LLMs tend to seek out local maxima unless carefully guided toward some kind of architectural sanity. So whereas the previous attempts at automatic programming failed in obvious ways, LLMs succeed convincingly. I’ll be the first to admit that when writing UI code, I often don’t scrutinize the output as much as I do for backend code, where I actually know what I’m doing and the consequences tend to be worse (that’s not some kind of universal maxim, just the nature of my current workload).

Things are getting better in this regard. We’ve learned how to interact with the LLMs to get the results we want. But MAN does it go sideways if we don’t know the result we’re after.