There's no point at which turning your brain off will work

Argues that turning your brain off while using large language models is a flawed strategy: letting an LLM take actions (write code, summarize, fix bugs) and assuming correctness produces fragile, often incorrect results. While “meat proxy” workflows - where a human acts as a loose loop around an agent - have become more effective as models improve, there is no point where the human is meaningfully safe from being replaced, because a company can automate the loop. High-value tasks expose the limits: agents overfit to tests and metrics, mis-handle out-of-distribution situations (obscure languages, modern board games), and produce plausible but wrong guidance that harms novices. Concrete failures include bots weaker than simple heuristics, software that breaks user flows or creates escape-proof loops, and large diffs full of hacked-in workarounds that experienced engineers reject.

The practical conclusion is that humans must stay engaged as QA, architect, and decision-maker: constant supervision, careful prompts, and insistence on edge-case handling are required. Some personal projects and data-analysis loops benefit from iterative agent use, but outsourcing thought yields poor commercial-quality results and makes jobs vulnerable as models run unattended. Converting complex codebases or build systems took months even with agents, demonstrating too many intangible requirements and unknowns for a “brain-off” approach to be reliably production-ready.