The idea of "AI alignment" should not be interpreted as "does the little guy in the computer want the same thing as I do?" but as "does this software that I'm using contain bugs that introduce undesired side effects?" An LLM that produces false responses or an agent doing unexpected things is not categorically different from a shell script that has a forgotten `rm -rf ~/*` somewhere that gets triggered by some edge case condition. Scale doesn't change the fundamental nature of a problem.