Mexico or AI?

Professor Ras Mic (AI researcher): "Wide-open loops are currently a slop machine. The AI makes assumptions, it iterates on those assumptions, and the output compounds in the wrong direction without a human to catch it." Here is the honest picture of where loops work and where they do not: → what works: binary, fixed-feedback environments where creativity is not the primary requirement → what does not work yet: open-ended loops asked to build full applications, assumptions compound, vision misaligns, token bill climbs → the grep loop that works: push code to GitHub, greptile scores the review out of 5, loop iterates until 5/5, then merges, binary goal, measurable feedback → the limit: fails when the PR exceeds 1,000 lines, even confined loops have boundaries → The token reality: Boris and Peter have unlimited budgets; for the average builder, autonomous loops are a money burner → The current answer: human-in-the-loop remains the most effective and reliable method for building serious software Loops may be the future. They are not the present for most builders. Credit to @gregisenberg for this interview! Follow @neil_xbt for more AI engineering intelligence that shows you what the loop discourse is leaving out.