Post #2001915

@regehr @aburka @gwozniak In theory, yes, that sounds like a compelling argument (using LLMs for obscure stuff we don't have the time to learn). In practice, though? I teach docker and k8s, and whenever the time comes to write Dockerfiles, YAML manifests, and the like, I show students multiple techniques: editing examples found in the docs, generating them with tools, and now with LLMs. The LLMs invariably generate suboptimal results. At best they add some extraneous stuff that doesn't harm too much beyond readability and maintenance of the result; at worse the outcome is just broken in ways that are not immediately obvious. I can notice it and point it out to the students because I have a decade-ish of experience with these tools. Somebody with less experience won't know. (And no, it's not "getting better"; at least not in a game-changing way.) It's still useful as a stopgap (if some build pipeline needs a Dockerfile and you don't know how to write one, it might save you). Something I've been advocating for in that field is that you don't need the whole team to know the ins and outs of e.g. Dockerfiles, but you almost certainly need to have at least 1-2 people who are well versed with it. If you replace your in-house Dockerfile experts with Claude, the output will suffer, and you won't even be able to know how deep.