Post by @RuthMalan@mastodon.social

2026-07-26 13:55 UTC

“This paper [Naur’s Programming as Theory Building] is so timely in times of LLMs because it gives us a definition for programming that an AI can never do, by definition. A large language model can never form a theory about what it is doing, that simply is not a goal of a language model. An LLM can not (consistently and correctly) answer questions about the code.”

Replies (2)

@ewolff@mastodon.social 2026-07-26 13:59

@RuthMalan@mastodon.social My problem with Naur‘s paper: It gives two examples, so no real empiric evidence. Also the theory is almost like ghost: impossible to capture or write down. I do like the paper and it explains what I also see in the real world. I just think more research and empirical work would be great. The impact on LLMs would certainly justify it.

Open ##4197513
@fanf42@social.treehouse.systems 2026-07-27 12:39

@RuthMalan@mastodon.social you can still learn interesting things : we use llm to explore code in mob programming. The llm do the heavy greping, the knowledgeable people explain why it is so, sometimes (ok, every time) we find inconsistencies. Also "agent" are not pure llm anymore. The llm is more like a query engine in natural language that relies on deterministic tools (compilers, etc) to do things. So for our Scala/rust/elm functional programing stack, the guards are ok. It still needs heavy guidance for design, but the mechanical chores are ok. It doesn't change anything on all other aspects (energy, training, monopoly building, bubble, etc etc). And for the mental model of the code, perhaps we will get there. We are not. Even with its memories and skills and etc, the agent needs to be checked on all like of code.

Open ##4197516