A compelling distinction may be this:
Current AI is mostly trained to possess capabilities before deployment.
A more general intelligence must remain capable of creating, repairing, and reorganizing its capabilities through ongoing experience.
The next question is not only how learning continues at runtime, but what exactly gets preserved from experience: weights, predictions, policies—or reusable structures, constraints, failure patterns, and validated paths?
Runtime learning may be the beginning. Runtime structural writeback may be what turns experience into lasting intelligence.