New Anthropic research: Forecasting rare language model behaviors. We forecast whether risks will occur after a model is deployed—using even very limited sets of test data.

→ View original post on X — @anthropicai