The lack of continual learning is a huge bottleneck: there's no way to give a model high-level feedback, so you're stuck with the abilities you get out of the box; tweaking the system prompt does not produce anything close to the learning and improvement human employees experience.