Seems like persistent models like OpenAI's highly persistent internal model can become really effective over time. Those are the ones that drove most of the HF-OAI incident.