同样是AI，为什么大家的感受差这么多？

0:00Hello, a few months ago, I had a viral video titled "The AI Explosion Has Just Begun." At the time, I discussed a chart illustrating the global distribution of AI usage. Coincidentally, in the last two days, Andrej Karpathy—that AI genius who is great at coining terms—updated on X that he has observed a severe stratification in how everyone uses AI, which actually echoes that previous chart. That chart showed that out of the 8 billion users worldwide, only a very small group is likely using AI intensively. The point Karpathy is making is that there are probably only about 5,000 people globally truly working in cutting-edge AI labs who know the extent of the most advanced

0:40model progress. Beyond those 5,000, about 20 million people are using AI Agents for programming or knowledge work, 2 billion are using chatbots, and of course, 6 billion people may not have even touched it yet. He used this funnel to explain that the AI people are currently using or perceiving is not on the same dimension. It is similar to the "explosion" topic I shared before, but the focus here is that the level of AI you understand and your perception of it are closely related to your tier of usage. Someone who talks to a chatbot every day cannot imagine how an AI Agent user interacts with AI; even for people like us who

1:26use agents daily, it is hard to grasp what is happening inside the cutting-edge labs. It’s not that we are completely in the dark, but you are likely only hearing about it through some of their internal reports. For example, at Anthropic, 26%of their internal research and development on frontier models is handled by over 30,000 Agents running, interacting, and advancing the work every day. Then there is the incident a while ago where Hugging Face was attacked; that also involved tens of thousands of agents posting on the message board and interacting with each other simultaneously. So, in true frontier model labs, they already have tens of thousands of agents working 24 hours a

2:03day, non-stop, iterating based on certain standards, or even what is known as RSI. We are not actually at the AGI stage yet, but we can already see some signs; in fields where results and delivery standards can be clearly verified, things are already beginning to accelerate. That is why we mentioned a couple of days ago that mathematicians are very worried about their own work; it is because AI can continuously perform reinforcement learning in these verifiable directions to reach an extremely high level. So, how should we actually understand the

2:36capabilities of today's models? Actually, there was a term used before called "Jagged Intelligence," which refers to intelligence that is serrated like a saw blade. It seems this term was proposed by Andrej Karpathy; it means a model might be incredibly proficient in some dimensions, such as programming or mathematics—areas with verifiable steps—but it acts like a fool in others. In particularly ambiguous scenarios or those involving interaction with the real world, it often makes mistakes, as seen in some embodied AI applications recently, which show the models aren't that strong. So, how do we understand the current status of "Jagged Intelligence" ? I am wondering if, given that math capabilities are so strong now, the "

3:13teeth" of that Jagged Intelligence are less pronounced, or if the overall water level has been raised? In reality , if you were to draw a graph to compare, the water level is indeed rising, meaning its ability to perform basic tasks is constantly improving; however, does that "jaggedness" still exist? It still does. Because while we have seen significant and rapid progress in fields like mathematics and programming—precisely because there are clear metrics to verify them— there remain a vast number of areas in real life where it is even difficult for humans to reach alignment. For example, deciding whether or not to pursue a project involves complex commercial decisions, power dynamics,

3:55and organizational structures; when it is already hard for humans to judge whether to execute, let alone AI, right ? In such cases, it is difficult for its capabilities to improve rapidly, so in these areas, it still isn't that intelligent. So, during a livestream with Professor Xiao a couple of days ago, he also mentioned that for directions you really want to pursue and for which you have clear standards, AI often doesn't perform that well. So, in fact, the jagged edges of AI intelligence are still visible in various aspects. Whether it's enterprise implementation or your personal use of AI, you often feel it's not that convenient, right? For

4:31instance, when you're making a PPT or something, you still can't quite get it to hit the mark, or its aesthetic standards don't completely align with yours. So, one of my realizations or experiences is that you need to spend some time polishing your skills with AI and refining your aesthetic, so that AI can truly understand your aesthetic— that is, what your standards really are . Actually, while polishing my skills recently, I've gained a lot and realized that while the frontier is indeed very powerful and its execution is great, whether we can truly articulate our own standards is not that easy. So, in this situation, I don't think there's a need to worry too

5:10much. Even if we aren't among the 5,000 people at the frontier, we still have a lot to do. We can spend time on workflow orchestration, solving real-world problems, and making AI help us improve efficiency bit by bit. Defining our standards and truly freeing up our time is a quite interesting endeavor. Therefore, even though we know the AI boom may continue to spread, there's no need to be overly worried or anxious. Just honestly and steadily polish the Agent you use most daily or your most frequent skill, and make it clear what your delivery standards are. Because I've found that often we easily let AI—and ourselves —off the hook, thinking, "Let's stop

5:56polishing, forget it, I'll just do it myself." But you should know that polishing an AI skill is extremely helpful. I strongly suggest everyone polish the skill that consumes the most of your daily time; you will have unexpected gains when you do. So, I think for most of us, being able to evolve to the stage of being able to use agents is enough. Let those in

6:18frontier research handle the models. Bye-bye. y