TMTPOST--Yan Junjie, the founder and CEO of leading Chinese AI large model company MiniMax, said that while many domestic companies are striving to catch up with ChatGPT, the current error rates of these models are around 60-70%, significantly higher than GPT-4's 30-40%.

He made the remarks in a roundtable discussion during the 2024 World Artificial Intelligence Conference (WAIC 2024) held from July 4 to July 6 in Shanghai.

Yan noted that reducing the overall error rate of large models is the most critical issue. Efforts to improve training efficiency, explore new network structures, and enhance alignment are all aimed at lowering the error rate to single digits within six months to a year, enabling AI models to truly assist human tasks, he added.

Regarding the ongoing "price war" in China’s AI large model industry, Yan views the overall reduction in large model prices as a positive development. Lower costs mean more users and longer usage durations for AI enterprises, generating greater value, he said, adding that this increased traffic and value will help in discovering better business models.

In a post-forum conversation with TMTPost, Yan said that it will likely take at least three years to create a "killer AI app" like WeChat, TikTok, or Toutiao.

"The biggest application in the AI era has great potential, but the current app isn’t there yet. It may take around three years to create something for the masses. It’s a step-by-step process. I believe we will see a 'killer app' at least three years from now," Yan remarked.

Yan founded MiniMax in December 2021. He had served as Vice President at SenseTime. At SenseTime, he was responsible for building deep learning toolchains, underlying algorithms, and the development of general intelligence technologies.

MiniMax offers end-to-end self-developed multimodal large language models and open platforms such as abab and MiniMax, as well as consumer-facing products like Xingye and Hailuo AI. The company serves over 30,000 clients, including major internet companies like Kingsoft Office, Xiaohongshu, Tencent, Xiaomi, and Yuewen.

MiniMax has raised investments from Alibaba, Tencent, Sequoia China, Hillhouse Capital, and IDG Capital, among others, achieving a valuation exceeding $2.5 billion.

This makes it one of the highest-valued AI large model companies in China and a leading AI unicorn in China.

Yan admitted that while both domestic and international markets are competitive, the fundamental differences between China and the U.S. are notable. For instance, the best models in the U.S. are stronger than those in China, but the U.S. lacks the engineering talent pool that China enjoys. This talent pool provides a significant advantage, potentially offering three to five times the workforce compared to Silicon Valley.

Reflecting on the domestic market, Yan observed that many companies have yet to differentiate themselves significantly. Most models and products are similar, leading to competition based on prices. However, he believes this competition drives technological innovation.

Currently, MiniMax aims to penetrate more users while continuously strengthening its technical capabilities. "The primary goal is to improve our text models," Yan noted.

He identified reducing the error rate as a short-term priority since high error rates affect user engagement.

Yan also addressed the balance between technical advancement and commercialization. He asserted that a company's technical prowess determines its qualification to compete, but technical excellence alone is insufficient. Commercial viability, alongside technical progress, is crucial for long-term success.

Looking ahead, MiniMax plans to demonstrate capabilities in generating audio and video content from text, among other cutting-edge technologies.

"Whether a company builds its models independently or not, the core focus is on how much value it creates for its users, how good its technology is, and how well it tangoes with users. This is the essence of the business," Yan added.