First, the stronger AI becomes, the more valuable judgment and creativity are. Today I want to talk about what truly matters in the AI era, what TikTok is doing, and how I personally use AI.
Regarding that audience poll about what matters most in the age of artificial intelligence, the answers included human judgment, interpersonal relationships, human creativity, human purpose, and so on. I think all of these things are very important. But interestingly, we don't actually need to get stuck on how exactly to define "human" 鈥?that word is too broad. Everyone in the room has a different judgment, and people from different fields judge differently too. But we can quickly reach a consensus: there really is something called "judgment." Interpersonal relationships and human creativity stand in direct contrast to the problems AI is currently facing. These things are not the same as our innate biological attributes, and the way we do things may be completely different. So I basically agree with all of the options above.
Second, the most urgent issue is not what AI can do, but how to make it do the obviously correct things. There is a lot of discussion about AI alignment (the research direction of keeping AI's goals consistent with human intentions), and there are some very extreme examples. The paperclip example is famous: you tell a machine to produce more paperclips, and it just focuses on producing paperclips regardless of all other costs. This is the "Paperclip Maximizer," a well-known thought experiment in the field of AI alignment. So the challenge we face is how to handle something as powerful as AI. It has already shown the ability to solve very complex math problems. I chatted with some entrepreneurs over the weekend, and my judgment is that within the next two to three years, because of artificial intelligence, we will see a real scientific breakthrough in fields like physics, biotechnology, chemistry, and materials science. Since capabilities may arrive this quickly, the question then falls to the next step: how to align the models, how to align the machines, so that they do the things that are obviously correct for us humans. Which boundaries should you hold on moral and ethical grounds, and which things you should not do. Frankly, I cannot answer this question right now. It is indeed very complex, but in the AI era, it will become crucial.
Third, AI can generate everything, but people still prefer to see what is real. For TikTok, the topic of creativity is especially relevant. We already allow users to make a large number of very realistic AI videos with models, and I have made a few myself. But there is still an imbalance: the proportion of creators making AI-generated videos is far higher than its share of views, and AI content currently accounts for only a single-digit percentage of views. In other words, user demand for real human content remains very strong. Because view counts are also the result of AI recommendation algorithms running their course, this is a very democratic process, truly reflecting what people actually want to watch. From this, we can see that people have tried all kinds of ways to use AI, and AI can indeed create more content, much of which is more creative than you might imagine. But at the same time, people's acceptance of AI-generated content is still much lower than expected. People still value what is authentic and original. So human creativity will continue to play an important role. Authenticity remains very important in content creation. People increasingly believe that this should be a combination of humans and AI, rather than letting machines take over all decision-making. As technology advances, we are continuously improving the infrastructure so that agents can do more for us, while also building up trust and guardrails (safety constraint mechanisms that prevent AI from overstepping). In the long run, we may allow agents to autonomously complete more and more things. But in the short term, at least over the next two to three years, whether you can use AI well will become very important.
Fourth, leave routine and low-risk tasks to agents, and keep high-risk, nonlinear tasks for humans. Regarding what AI can do, internally we try to simplify it into a 2x2 matrix. One axis looks at whether the task is routine or requires judgment and is highly differentiated; the other axis looks at whether the risk is high or low. The conclusion comes quickly: routine and low-risk tasks can be automated by agents and executed very quickly; high-risk but routine tasks need more caution, but most frontier models can also do them; high-risk, highly differentiated, and very detail-oriented tasks still require a lot of human involvement. For example, TikTok has a huge number of users globally. Suppose we want to know what a 14-year-old French-speaking Nigerian child sees on TikTok. In the past, we had to have people from the trust and safety team register accounts, test, watch 100 videos, and write reports. Now we can deploy a thousand agents and automatically generate reports overnight, telling you how many of them violated our guidelines, with the entire process fully automated. Another example is content moderation, and livestreaming is the hardest. At any moment, tens of millions of livestreams are running, and in the past we could only randomly sample, making it hard to be both precise and accurate. But now with VLM (vision-language models, models that can turn visual content into text descriptions), real-time video streams can be converted into text narratives very quickly and then compared with community guidelines. For example, if someone is driving, AI can recognize that the speedometer shows 160 kilometers per hour. The human eye may not catch it, but AI knows this is very dangerous. This greatly improves our ability to get a lot of work done. Of course, we also see new risks, such as cybersecurity risks. But on the other hand, the productivity gains AI can bring to the company also make us very excited.
Fifth, don't just read materials; only by trying it yourself will you know where the boundaries are. 1. If you don't use it, you won't know what it can and cannot do. Inside the company, one thing we value most is giving everyone, not just engineers but also lawyers and people in various functional departments, firsthand experience. If you don't personally use large models and AI products and only read materials, you cannot truly know what it can and cannot do. So what will really matter in the next few years is getting hands-on and trying it yourself. We encourage everyone to stay curious and get hands-on, while of course also being clear about where the boundaries are. We have created many projects for internal employees to expose them to the most cutting-edge models, with the goal of encouraging everyone to emphasize practice, truly put these new tools to use, and at the same time pay close attention to changes in the market. 2. I treat AI as someone to talk to. I often treat AI as someone to confide in. For example, I recently became a U.S. taxpayer, and the U.S. tax system is simply incomprehensible; I could not handle it at all without AI's help. I also use AI as a personal nutritionist. Take a photo of food, and it will tell me not to eat too much of this and that I can eat more of that. It feels like having a particularly smart person by your side to ask at any time. Earlier this year I had a shoulder problem, and I probably asked the surgeon 200 questions, about 90% of which AI answered first, saving the doctor a huge amount of time repeatedly explaining things. Of course, AI will remind me, but it will not physically stop me from eating what I shouldn't eat. It is not at that stage yet. AI reminds, and the person stops themselves. This is human-machine cooperation, not an either-or choice.
Sixth, the risks are real, but don't draw conclusions too early. 1. Users may not know what they have authorized. Cybersecurity risks are real. Many users give agents a lot of permissions but do not truly understand what they have authorized. For example, if you ask an agent to use your credit card to buy things for you, it may want so much to complete the task that it ends up buying a bunch of things you don't need at all. When I was in elementary school, the internet was not yet widespread, and my parents were especially afraid of buying things online. Later, e-commerce gradually matured, and now there are very powerful payment functions. I believe these things will eventually be realized, but there is still a lot of educational work to do. For enterprises, cybersecurity risks are equally real. Spam, platform abuse, and similar behaviors used to be very costly to carry out, but now the cost has dropped significantly, so platforms need corresponding tools to defend against them. But building these tools takes time. Employees also need training: do not tell public chatbots business secrets, and chatbots are not real lawyers and cannot provide legal assistance. Clarifying what can and cannot be used is a very important boundary. Precisely because AI is so useful, people sometimes overuse it. 2. It is too early to draw conclusions now. There is a lot of discussion about whether AI will threaten humanity, replace humans, or push humans aside, but it may still be too early to draw conclusions now. On the one hand, today's technology is far from that point. The example people often cite is that the human brain uses only 20 watts, while data centers are measured in gigawatts. But there are many things the brain can do that these models still cannot. For example, retrieving long-term memories 鈥?I can recall things from childhood, while today's large models still find that very difficult. People's visual learning ability is also very strong; with one look, they can learn a lot, and models are still catching up. But on the other hand, technology keeps developing. You can absolutely see a world where silicon is superior to carbon: longer lifespan, easier to scale, more convenient to connect. So AI may indeed be stronger than humans in some respects. But technology has never been anything new. Even if technology keeps moving forward, humans will still always remain at the center, enhanced, able to do greater things, and do them faster.
Comments