Robin Li rarely reflected on the slow internal response and backward productivity of large companies, including Baidu, and criticized China's fiery "Battle of 100 Models" as a huge waste of social resources, and that competition for running scores was "boring".
Titanium Media App learned that on the morning of December 16, Robin Li, founder, chairman and CEO of Baidu, said at a roundtable meeting that the 100-model war is a huge waste of social resources, and more resources should be Put it in super app. He also bluntly stated that large companies, including Baidu, are too slow to respond internally and their productivity lags behind. Therefore, they have repeatedly emphasized internally that Wen Xinyiyan should be used to "reconstruct" applications instead of treating them as "tools."
"I think what we are paying attention to now is the technical capabilities of large models, but this thing is really not important. What is important is whether our existing enterprises use large models to have a positive impact on the key indicators of their core business... ..It is easy to say, but not easy to do. First of all, in fact, large companies are very slow to respond. I even sometimes say that large companies are lagging behind in productivity. You must not look at what that company is doing. " Robin Li said.
The host then asked, "Is Baidu considered a big company?"
Robin Li said frankly, "Baidu is considered a big company. So I tell everyone internally every day that you must get rid of muscle memory and cannot act according to inertia. To change, we must embrace the new era. Because everyone is so accustomed to the past, they are confused about the future. Therefore, I asked Baidu to "re-do" all businesses, regardless of the past practices, but throw away things and start over. , this is a strong internal requirement. But the outside world does not do that, so when is communicated externally, I always feel that large companies should take their time. I will see some relatively medium-sized companies with technological genes embracing large companies. Model, the maturity shown when using large models to develop applications is still very admirable. "
This is a very rare occasion that Robin Li explained to more people this year after internal reflection at the end of 2022 of a big company like Baidu. The contradictions and problems existing in the company, as well as the backward challenges faced by large companies in the era of big models.
In early January this year, Robin Li bluntly stated within Baidu that it is okay for companies to "lose money because of rapid growth" in the short term, but it is not feasible in the long term. Many enterprise-level companies have no cash flow capacity and cash flow, and eventually the company's funds are broken. Will die. Robin Li also denounced the lack of business awareness among internal employees, resulting in low "income quality" of new business. (For details, see the previous article on Titanium Media App: "Robin Li's Internal Reflection: Short-term losses can occur due to high growth, but companies without cash flow will die in the long-term.")
In the approximately one-hour roundtable discussion on the morning of December 16, Robin Li answered a number of topics in the AI industry, including the gap between large models in China and the United States, how to reconstruct all applications internally at Baidu, and whether big companies like Baidu will take away all the dividends from the commercialization of large models.
Robin Li first said that in the past, AI computer vision technologies including face recognition, intelligent inspection, etc. were very popular and seemed to generate commercial value, but the scenes were very scattered, it was difficult to produce standardized products, and the value was very low. Even now, he doesn't see any particularly outstanding large companies or particularly standardized products. But in the field of large models with ChatGPT as the core, characteristics such as "intelligent emergence" and versatility have become completely different opportunities. When a set of basic technologies can be done very well and be very advanced, large models can quickly make valuable applications in various scenarios.
Robin Li pointed out that since this year, the focus of society has been on the Foundation Model itself, but he has always believed that society needs to pay attention to the millions of AI native application models on the foundation model, which is very valuable. At the same time, he also criticized the emergence of more than 100 basic large models in China's "Hundred Model War", which is a great waste of social resources. Especially when China's computing power is still limited, companies should explore the combination of applications in various industries and new App product possibilities, etc.
Robin Li also denounced that it is "quite boring" for many domestic large-scale AI model companies to compete in benchmarking, and the performance benchmarks of third-party large-scale models are not reliable.
"(We) waste a lot of resources on the training of various basic models, even running scores on the list, while relatively few resources and energy are focused on AI applications. This is what I have been feeling a little bit in the past few months. It’s normal to be anxious. When something new or a big opportunity comes, everyone will definitely rush into it, and then there will be a process of washing out the sand. This is very normal.” Robin Li said, “Train a big The model requires 10,000 GPU (graphics processing unit) card resources, and it takes a long time to train it. After training, I take a college entrance examination. What score did I get? How much money can I earn for you? Rank first on the 'list', how much bonus can the list owner give you?" Robin Li said that this may be beneficial to financing, but financing is actually an indirect result, and investors should rely on these (application scenarios). Make money.
"So sometimes I think that the evaluation of the quality of a model is the core competitiveness of a model company (itself). If you don't know what is good, rely on a third party to evaluate and give you a rating. Point, this is unreliable, even if you don’t know what you are doing. Our real model must be goal-oriented and application-oriented, and it must be able to solve practical application problems." Robin Li said .
Talking about the gap between large models between China and the United States, Robin Li said that from the perspective of technical quality, there is not much difference between the leading large model levels between China and the United States. But if it is in enterprise-level model applications or small language models, Chinese companies currently do not have the energy to do special optimization. "We (China) may indeed be lagging behind."
In terms of AI computing power, Robin Li pointed out that under the same effect, the reduction in reasoning costs brought by the model application layer should be the core competitiveness of the AI large model industry in the future. He mentioned that in March this year, Baidu’s inference cost based on the Wenxin large model dropped to 1% of the original level. "So this is a model that you only dared to adjust 10,000 times before, but now you dare to adjust it 1 million times a day. This is a completely different feeling."
Robin Li said that the process of exploring large models is "miraculous with great effort." He criticized the internal statement "What does this technology have to do with my business?" and hoped that the large model would bring efficient and growing changes to Baidu's business.
"What does your business have to do with us? Can big data help you? For example, how much has your performance increased, how much your retention rate has increased, how much your user market has grown, how much your income has increased, How much your profits grow, these are the key core indicators of your business. If you cannot make effective changes in these key indicators of your business this year (using large models), you will not have much technical level." Robin Li said.
Robin Li emphasized, "I think it's not polite to say that in the era of big models, the big manufacturers (big companies) (the industry is formed and productivity is lagging behind) will take away all the dividends." But he also admitted that Baidu is also in the field of large models. A certain "boundary" has been formed, and no company can "occupy" all the opportunities in the large model era.
"Baidu cannot believe that there are companies many times larger than Baidu. I don't think it has the ability." Robin Li Said that in the past month or two, Baidu's internal strategic planning and discussions have been taking place, and "strategy is about choices." In the end, there may be 3 or 5 startups in the large model track.
In summary, Robin Li's performance is as usual " Baidu builds Wenxin's basic large model, and other companies should build applications based on the basic model. He believes that the value of the large model may be to generate new super applications, rather than developing various basic models to compete with Wenxin. Baidu is also solving many complicated internal "inefficiency" problems.
"As a CEO, you have to make these decisions. You must do something and not do something. You must cut off what should be done. Whether it is for outside entrepreneurs or partners, in fact, you are When competing with the entire market, if you want to survive, you must be the best in the market. If you are defeated by any player in the market, you will not survive. In fact, this matter really has something to do with Baidu It’s not big, there’s no need to worry about this.” Robin Li said.
(This article was first published on Titanium Media App, author | Lin Zhijia)
