How Research Resources Are Allocated: Coding, Hallucinations, and Data Labeling
Discussion of research resource allocation logic (CoT doesn't consume resources), product-oriented solutions to the hallucination problem, and the time bottleneck in data labeling.
Right now, half of our company is labeling data, and half of our core researchers are labeling data.
Model Office is a relatively certain goal, but model MIS, I think it's not certain yet-it's just a goal. Model Office, I think, is relatively certain.
CoT doesn't consume resources. Doing arbitrary research doesn't use GPUs; it only needs a few cards-it needs ideas. It doesn't consume talent resources either, because you don't need someone working on it continuously. It's not a project; it requires many people to be thinking about the problem. So there's no need to allocate resources to it, because it doesn't need resources. You need resources when you train models, build and release models, and run efficiency experiments. As I said, the consumption of both people and GPUs is minimal.
So we call this a 'lottery draw.' The barrier is low-anyone can try their luck-but who will win something, maybe it depends on talent or something else, I don't know. So we don't need to allocate resources here. The difference between us and other companies is that we spend time discussing this problem, thinking about it, and treating it as an important matter. Within the company, it's an important issue-one we spend time thinking about-but it doesn't require a lot of resources to execute.
Next, I see a question: The hallucination problem of large models affects user experience. There's a way to solve hallucinations, but it's a long-term proposition. Hallucinations can be considered a problem that can be improved through better post-training-it's solvable and improvable. It's just that not much effort has been put into it yet. Or, for me, hallucinations are a problem, but we boil it down to a product issue. We will address it, but it's not a priority problem.
Earlier, there was also a question about data labeling. Data labeling is related to our capital investment. With our capital structure, we can't afford the cost of high-quality data labeling, because it's very expensive.
The cost of data labeling in the US is no different from that in China. China has no cost advantage in labeling data, especially high-end data. That makes it hard for us to invest in labeling data like they do in the US. This path is difficult in China because labeling is too expensive-whether we outsource or do it ourselves, it's painful.
So right now, we're taking a two-pronged approach. It's not that we can't label at all, but some data labeling is cheap, some is expensive. We start with the cheap ones.
So you could say that right now, half of our company is labeling data. Half of our core researchers-the most important people-half of them are labeling data. We concentrate on labeling. Solving AI problems at this stage relies on data labeling. All you need to look at is the data.
Mr. Liang, thank you. That was very insightful and resonant-excellent. I have a few questions. First, you mentioned that Chinese models are definitely more efficient than American models. You also said there are other areas where China could surpass America in the future. Which aspects do you think we could outperform America in intelligence or other areas?
I think in many user experience aspects, we might do better than the US. Not to boast about our own experience, but I think our product experience is pretty good, and we might not be worse than the US in user experience.
In terms of product, our product capabilities might not be inferior to the US. Costs should also be lower, so China might still be competitive. As for other structural advantages, I think there may not be any. But in terms of cost and product, I think we do have some structural advantages.
Cost is easy to understand: They don't do it, so they don't develop that capability. They certainly don't place as much importance on it as we do. We can treat it as a very important matter, but for them, it's not important. As for product, native companies have decent product capabilities. So I think these two areas may have structural advantages.
Okay. Second question: You mentioned post-training, where our relative cost is high, and companies like Anthropic and OpenAI invest huge amounts. After this funding round, do you think we will increase investment in post-training?
The gap mainly lies in high-quality data labeling, and primarily in AI research. We will definitely increase investment, but high-quality data labeling is typically not a capital investment.
The bottleneck for high-quality data labeling, I think, is time-it takes time. For OpenAI, foreign companies, and Anthropic, they started earlier, have more capital, and more GPUs.
In this situation, we in China have only started in the last six months. So in terms of time, I think we need more time. This isn't strongly related to capital investment because even without more capital, the existing capital is enough to expand at the fastest possible pace.
But there's an upper limit to the speed; the bottleneck is not that we can immediately get more people. It's not constrained by money or GPUs. But it is indeed in a rapid expansion process. So I think within a year, the high-quality data problem will be relatively well solved in China-that should be expected. I think the outlook isn't that bleak, but it does take time.
Thanks. Then my third question: We see that Anthropic uses its own models to build its own products, launching many vertical applications in finance, law... Even in the future, you might move toward the medical field. Do you think at some stage we would consider such vertical applications?