The latest article by OpenAI researcher Noam Brown tells us that the logic of AI capability growth is changing. The ability of AI is reflected in persistence, and the longer it lasts, the longer it becomes. Previously, the industry used to measure model capability through various tests. But with the development of inference models and agents, the effectiveness of this evaluation method is rapidly declining. GPT-5.5 is a typical case. When it was first released, many people felt that the improvement was limited because the benchmark numbers only seemed to be slightly ahead of GPT-5.4. But after actual experience, a large number of users found that the gap was far greater than the benchmark display level. The reason is that the capabilities demonstrated by GPT-5.5 were not obtained under the same inference budget as GPT-5.4. If comparing with the same token budget, time budget, or cost budget, the difference between the two is often significantly magnified. Stronger models seem to be better at working on longer time scales. The changes brought about by the improvement of model capabilities are not only reflected in the quality of answers, but also in the length of tasks. Search, planning, validation, rollback, retry, and self correction have become part of the inference process. The model is moving from answering questions to completing tasks. Noam believes that future model capabilities should be measured by a curve of "performance vs. inference budget" rather than a single number. More and more experiments have shown that the capabilities of modern models have not reached the plateau stage as quickly as in the past. As the inference token increases, performance continues to improve, and the stronger the model, the more pronounced this phenomenon becomes. Importantly, the biggest change brought about by the improvement of model capabilities may be the increase in task duration. The stronger the modeling ability and the fewer errors, the better it can run long-term tasks And a model that can only work for 30 seconds is essentially still a chatbot. A model that can work continuously for one hour is an assistant. A model that can work continuously for one day is an employee. A model that can work continuously for one month is more like an independent team. From this perspective, Test Time Compute is essentially purchasing working time. Providing more tokens to the model is equivalent to giving it more opportunities for search, validation, trial and error, and self correction. This is also why Sustainable Sustainable Context Work (DSCW) is becoming an increasingly important competency indicator. More importantly, longer working hours will ultimately lead to AI development. Running continuously for several hours, the model can debug code. Running continuously for several days, the model can read papers, design experiments, and analyze results. After running continuously for several weeks, the model began to have some researcher capabilities. So a new feedback loop will emerge: Stronger models → Longer tasks → More research and development work → Stronger models. This is also one of the most realistic paths for AI recursive self iteration (RSI).