So did they close the feedback loop with an AI, or did they just use some ML hackery to create a model of the system? Does anyone know how this compares to a classical control system?