Post #2769029

2026-03-18 14:17 UTC

@kornel@mastodon.social The reason we have programming languages is because natural language is such an inadequate way of specifying requirements. Otherwise we'd have compilers interpret them. That's essentially what an LLM coding model attempts to do, making stochastic guesses at what will satisfy the description, based on ingested codebases. Not enough people understand why this is a ridiculous approach, and that's why so many are choosing such poor tools, that will ultimately fail them. @iainvdw@mastodon.social

Replies (2)

@iainvdw@mastodon.social 2026-03-18 14:41

@happyborg@fosstodon.org @kornel@mastodon.social Very much agreed! Aside from natural language not being adequate enough to specify requirements clearly, we as humans are notoriously bad at describing things we want, let alone *need*. If we remove human interaction from the whole process, I'm afraid we also lose a large chunk of critical thinking in the sense of "should we actually do this?". LLMs can't do that critical thinking part for us.

Open ##2769030
@kornel@mastodon.social 2026-03-18 14:41

@happyborg@fosstodon.org It's not entirely the same problem as earlier attempts at making natural-language compilers. The "large" part of LLM gives it more context and a bit of "common sense". It's still a guess, but drawn from a more likely distribution. The learned context is so strong they can sometimes surprise Babbage and give right answers to wrong questions. Yeah, they routinely fail at software engineering we have today, but I wonder how much the world will adapt to "cheap but inaccurate" code.

Open ##2769035