Continual learning—where model weights are updated based on what the model learns from deployment—may not be necessary if in-context learning can get so good across longer and longer time horizons that you don't need to distill deployment learnings back into weights.