March 2027 -> Neuralese recurrence and memory

> For example, perhaps models will be trained to think in artificial languages that are more efficient than natural language but difficult for humans to interpret.