I'd like to see this combined with reinforcement learning to optimize models to think computationally. Generating ideas with hypothetical results and then running them in the same thought. Their solution sounded like a lot of tokens though.