|
|
|
|
|
|
by arty_throwaway
8 days ago
|
|
|
First thing is that "intent", according to this author, is single-author and monolithic. I'm not sure that's the way it works with any modern codebase of any size. I suppose they were perhaps just thinking about sitting down at a prompt and having a go, but I suspect it's a fiction because what is an LLM if not encoded intent of millions of us, itself running on a codebase of hundreds of authors, itself running on an OS of thousands of authors, etc. The main move here is hand execution to the machine and keep "choosing what to intend" for the human. Isn't choosing what to intend just...intent? And forming intent is what programmers always do/did. That sounds less like a new refuge, and more like a relabelled part of the job. More importantly, "intent is implementation" needs intent to be two different things at once. For the slogan to work, the intent at handoff has to be complete enough to determine the system on its own. But Section 4's own recipe is not that. It's hand off a rough idea and let the AI "expand and refine" it. Complete enough to ship, or rough enough to need the machine to finish it? "Intent" can't be both. There's no finished intent at the handoff for the slogan to be true of. Same problem with the verification argument. Section 1 says nobody needs to read the assembly (personally, I like reading assembly, but I digress). Then Section 2 says the one thing that can't be sped up is verification. The author says it's an economic reality, I would argue it's structural. Verification is expensive because there's no general procedure to confirm an implementation does what you intended! Halting problem, natch. Saying you verified the result doesn't work because that's specific: this time in that environment. Verification keeps dragging back to the code. You can't claim that code loses meaning AND claim that verification is the surviving human job, because verification is the place that code meaning won't go away. Anyway, there are frameworks that can describe how intent relates to computational process. This blog post isn't one of them. |
|
|