Good reinforcement learning and other'reasoning' benchmarks to measure progress, some set of increasingly harder tasks that can measurably show the different strengths of various models. My thoughts are that it wasn't just the data, but everything around image-net that really pushed the field forward, the yearly competition, the talks and progress graphs the anticipation and excitement to see how far the teams pushed the limit this time. Reinforcement learning still needs its'image-net moment', ideally some annual competition that can gain traction over time, have the big teams invest resource to push the limits. The field lends itself well to simply adding more complex tasks as the models get stronger and stronger. I merely answered this question as in'what would I as an outsider like to see', so feel free to disregard', but I think there is something in the human nature about competition which drives progress.