I think degraded performance as the context window increases is one of the largest issues models are facing today. Do you think we've reached the end of any "RAM" based context performance and that external storage/ retrieval mechanisms is the only way forward?