Comment by seydor

Perhaps people feel that the problem of modeling long-range token relationships has been solved. The problem is now how to get this model to produce tokens that are valid and ingenious solutions to problems, with RL or otherwise.