Over the past decade, the way to measure progress in artificial intelligence has been simple and straightforward: give the model a question, compare its answer to the standard, and score it. But now, this approach is becoming obsolete.

The identity of AI has changed. It is no longer a passive question-answering machine but an active agent capable of taking initiative. Planning its own itinerary, calling various tools, making continuous judgments in complex tasks—this new generation of AI is gradually taking over the work once done by humans.

The question that follows is: since AI is doing more than just spitting out a sentence, but completing entire tasks, can we still evaluate it using the traditional "right or wrong" testing standards?

Imagine a task with no single correct solution. AI uses an unexpected but more effective method to solve it. Under traditional evaluation methods, this would be considered a failure. But what is the reality? The goal is achieved. This is not just a technical detail but a systemic challenge—how you evaluate AI determines whether it has truly learned to solve problems or merely learned to please the rules.

Therefore, the AI research community has now reached a consensus: don’t just look at the results, focus on the process. The latest research and practical experience point in the same direction—evaluation should not focus on a single answer but on the entire chain of actions. How AI understands the task, how it breaks down steps, when it should call tools, whether it can adjust strategies based on environmental changes—these are the truly important aspects to observe.