Imitation won’t produce continual learners

Training large language models by imitating static corpora (and relying on longer context windows, scratchpads, or retrieval) cannot substitute for true within‑lifetime continual learning that changes an agent’s inductive machinery. True continual learning requires update mechanisms that permanently alter model weights/algorithms, enabling the discovery of new concepts and ways of thinking not present in the training data. — If true, this limits how quickly society can expect LLMs to autonomously innovate, self‑improve, or replace human long‑term learning, shaping regulation, deployment risk assessments, and industrial strategy.