"Run the LLM in a loop and lay off the employee" only works if the loop closes on its own. The entire rest of the post is an argument that it doesn't - that the value is in catching the out-of-distribution moment, the "this part wasn't even needed" moment, the overfit-to-the-test moment, and steering around it. That catching is a person. So the meat proxy does get replaced, but not by the loop by the colleague who didn't turn their brain off. The moat was never the typing; it's the taste and the QA judgment, which is precisely what the meat proxy has opted out of.
So "there's no point at which this methodology works for the employee" reads as true to me, but the thing being punished is the brain-off part, not the using-an-LLM part. The person supervising agents and cutting the diff to 25% (the Gary Bernhardt example) is doing more valuable work than they were pre-LLM, not less.
So "there's no point at which this methodology works for the employee" reads as true to me, but the thing being punished is the brain-off part, not the using-an-LLM part. The person supervising agents and cutting the diff to 25% (the Gary Bernhardt example) is doing more valuable work than they were pre-LLM, not less.