minimax founder and CEO Yan Junjie
html From July 4th to 6th, the 2024 World Artificial Intelligence Conference and the High-level Conference on Artificial Intelligence Global Governance (waic 2024) was held in Shanghai. At the main industrial development forum of the conference held on the afternoon of html on the 4th, Yan Junjie, founder and CEO of minimax, a domestic AI large model "unicorn" company, participated in a roundtable discussion. He said that although many domestic companies are catching up with chatgpt, the current model error rate is 60%-70%, far less than the 30%-40% of gpt-4.Therefore, in Yan Junjie’s view, reducing the overall error rate of large models is the most core issue at present, and solutions such as continuously improving training efficiency, researching new network structures, and doing better alignment are all aimed at promoting errors in large AI models. The rate can be reduced to single digits after half a year or a year, thus truly assisting human work.
Talking about the "price war" phenomenon of domestic large AI models, Yan Junjie believes that, overall, the lower prices of large models is a positive thing. For most AI companies, the lower cost and price of large models means there will be more users and longer usage time, which in itself generates greater value. Based on the increased traffic and value, we will find a better business model.
After the meeting, Yan Junjie told TMTpost agi and others that a "killer" AI super application (killer app) similar to WeChat, Douyin, and Toutiao will not be realized until at least three years later.
"The largest application in the AI era, I think it still has a chance, but (currently) this (super) app is definitely not that big. I think it may take three years (maybe) before there will be something popular, but It doesn’t matter. When you can do the first one, and then your ability becomes stronger, your resources increase, and your technical skills become better, it is possible that you can do it. At the same time, this thing will be done step by step, and then I think (killer app). At least three years later." Yan Junjie said.
It is reported that minimax was established in December 2021. The person behind the company, Yan Junjie, graduated from the Institute of Automation of the Chinese Academy of Sciences and served as vice president of SenseTime. During this period, he was responsible for building the tool chain and underlying algorithm of deep learning, as well as the development of general intelligent technology. At the
technical product level, minimax has provided end-to-end self-developed multi-modal large language models and abab and minimax open platforms, as well as c-side Hoshino, Conch AI, etc. Regarding the
open platform, minimax revealed this year that the company has served more than 30,000 customers, including many leading Internet companies including Kingsoft Office, Xiaohongshu, Tencent, Xiaomi and China Literature. In terms of financing for
, in the two years since the establishment of , minimax has completed investments supported by institutions including Alibaba, Tencent, Sequoia China, Hillhouse Ventures, idg Capital and other institutions. The company’s valuation has exceeded US$2.5 billion (approximately RMB 18.169 billion). , becoming one of the most highly valued large AI model companies in China, and also the leading AI unicorn company in China.
At present, minimax, Baichuan Intelligence, Zhipu AI, Dark Side of the Moon, and Zero One Thing are known as the "Five Big Model Tigers".
When talking about the topic of whether going overseas and whether it is "volatile" in China, Yan Junjie said that overseas is also "volatile" in the same way as domestically. The core is that the fundamentals of China and the United States are different. For example, the best models in the United States are obviously better than domestic models. This is a fact, but overseas models also have shortcomings. In fact, they do not have an engineer bonus. For companies with the same value, there are definitely many more engineers in Chinese companies than in Silicon Valley, and may even be better than the average. The main thing is talent. Therefore, there is still a bonus for engineers in China’s AI field, and it may even be a 3-5 times gap. In addition, in the mobile Internet era, China also has many excellent commercial products. "This is actually a 'Tian Ji horse racing'."
According to Yan Junjie, in the first half of this year, the company began to realize that in some productivity-oriented scenarios, "we began to have local advantages."
“Frankly speaking, I think most (domestic) companies have not differentiated themselves yet. They are all about the same, probably with similar model levels and similar products, and then they will start to compete on price.I don’t think this is a bad thing. In fact, it forces everyone to do better technological innovation. " Yan Junjie said frankly.
Yan Junjie pointed out that the current business development of minimax is to penetrate more users. At the same time, the company's technology needs to continue to become stronger in the process. "The more important thing is that our text model (must) become stronger. "
Regarding what is the clearest goal of minimax in the short term, Yan Junjie responded that the most important thing is to solve the "error rate" problem, because the error rate of large language models is actually still very high, such as the current error rate of 30% , leading to a significant decrease in users.
When talking about the integration and contradiction between technology and commercialization, Yan Junjie said bluntly, “I think whether technology is online depends on whether you are a qualified company. If the technology is not good enough, maybe he should be eliminated. This is the first one; but good technology is useless, because there are still AI filters. For example, if you look a year from now, it must be It may not be entirely from a technical perspective, but may be viewed from a commercial perspective. Therefore, I think the most important thing in the short term is actually the advancement of AI technology. "
Yan Junjie emphasized that , one year later, the commercialization of large-scale AI models must be very important. For the enterprises themselves, they still need to continue to explore cutting-edge technologies (no man's land).
Titanium Media agi learned , Minimax is also expected to demonstrate its cutting-edge technical capabilities such as Vincent Audio and Vincent Video.
"Whether a company does it independently or in other ways, the core still depends on how many users you create and how much your technology can do." How good it is, and how well you can cooperate with users, this is the most essential thing. "Yan Junjie told TMTpost agi.
(This article was first published on TMTpost app, author | Lin Zhijia, editor | Hu Runfeng)
