Interpretability is becoming increasingly uninterpretable

What is the purpose of interpretability research? Anthropic states that the mission of their interpretability team is to "discover and understand how large language models work internally, as a foundation for AI safety and positive outcomes". I think this characterization constitutes the classical argument for studying interpretability from the lens of AI alignment. Neural networks (NNs) are black boxes—we can’t just read a model's weights to verify if it is, e.g., scheming or not—and interpreta...