One thing that stood out to me about ExpRL is that it exposes a limitation of how we currently think about AI progress Most benchmarks only measure whether a model arrives at the correct answer They rarely measure how many useful ideas were explored and discarded along the way Human innovation often comes from failed attempts Not perfect ones A scientist may test hundreds of wrong hypotheses before finding the right one Yet traditional RL treats every failed path as equally worthless ExpRL suggests there may be valuable information hidden inside those failures That idea feels important beyond math benchmarks As AI agents become more capable the winners may not be the models that find answers the fastest They may be the models that learn the most from being wrong This is also why verification matters Without reliable ways to evaluate intermediate reasoning and outputs we cannot distinguish productive exploration from pure hallucination The future of AI may depend less on generating answers And more on understanding which mistakes are worth learning from @ritualnet