The LLM productivity debate refuses to close
A piece titled 'The LLM Critics Are Right. I Use LLMs Anyway' generated a thread that circled the same tension for several hundred comments: LLMs are unreliable enough that critics are technically correct, but the productivity gains for people who already know what they are doing are real enough that dismissing them is also wrong. The sharpest comment in the thread: 'LLMs amplify what you already have.' People without strong opinions or structure get noise back; people with both get speed.
The trust problem is the crux. One commenter framed it well: a small, reproducible code change with clear tests is reviewable regardless of whether a human or an LLM wrote it. A large opaque diff is not reviewable regardless of source. The tool is not the risk; the workflow is.
Separately, the thread on whether LLMs can perform deep technical comprehension of computer architecture papers landed skeptically. The abstract of the paper was itself identified as AI-generated and poorly written, which commenters found difficult to overlook in a paper grading AI output.
So what?
Founders shipping AI-assisted code into production need to enforce the same review discipline they would for any large diff, regardless of who or what generated it. The failure mode is not that LLMs write bad code; it is that they generate plausible-looking large changes that are hard to review. The fix is workflow design, not model quality.