There needs to be break through papers or hardware that can expand context size in exponential way. Or a new model that can address long term learning.