“For continual learning in AI models, direct weight updates are expected to be more effective than in-context learning over longer time horizons.”