{"id":"models-editing-models","title":"Learning to propose parameter changes","regions":["computation"],"question":"Can a model propose useful changes to another model from structural and conceptual information?","formulation":"The proposal ranges from interpreting hidden layers to predicting operations and choosing updates. A concrete suggested test is adding a new token and proposing its parameters from a conceptual description.","representations":[{"name":"Inspection of a token trajectory through layers","affords":"Connects intermediate states with an output.","loses":"An activation trajectory does not identify every relevant parameter or establish a causal explanation.","epistemic_kind":"reconstruction"},{"name":"A proposed weight update","affords":"Makes an intervention that can be measured.","loses":"Can improve a local example while degrading unrelated behavior.","epistemic_kind":"reconstruction"}],"tension":"Understanding, predicting, and modifying a model are distinct capabilities.","challenge":"What evidence shows that a proposed edit generalizes beyond the supplied concept description?","development":"active","epistemic_kind":"reconstruction","evidence":["basis-models-editing-models"],"next_move":{"text":"Test the proposed new-token edit against ordinary initialization and training, with held-out uses and regression checks.","epistemic_kind":"model_proposed"},"open_questions":["Can a model propose useful changes to another model from structural and conceptual information?","What evidence shows that a proposed edit generalizes beyond the supplied concept description?"],"formal_status":"open_problem","version":1,"content_hash":"ebab9b35ea73b62b7c89d634339ee87651096ef0f7e5ea23ac384e0080f61b01"}