Что будет проверяться
The Interpretability team reverse-engineers the internal computations of large language models — studying features, circuits, and superposition — to make Claude's behavior understandable and steerable. Research Engineers build tooling and run experiments that turn mechanistic hypotheses into measurable results on frontier models. A technical interview would probe transformer internals (attention, residual stream, MLP layers), designing experiments to isolate and validate a learned feature or circuit, and writing efficient code to extract and analyze activations at scale.
ExoForm не аффилирован с Anthropic. Это независимая тренировочная страница.