TMTPOST--The paid closed large model is a sort of “stupid tax,” as the real task for artificial intelligence (AI) is to ensure practical applications, and large models must be effectively implemented in enterprises for them to truly benefit from AI, said Fu Sheng, the Chairman and CEO of Cheetah Mobile.

Fu, who is also the Chairman of OrionStar, made the comments during the 2024 World Artificial Intelligence Conference (WAIC 2024) held from July 4 to July 6. 

He mentioned that the open-source and closed models should be developed together, instead of falling into a scenario where one dominates and the other fails to develop.

"I am not entirely inclined towards the open-source camp. In some sense, closed models might be slightly better because of the significant investment of money and manpower. However, open-source is often sufficient and develops rapidly. I believe they will not be a situation where one is far ahead, and the other cannot develop, " he elaborated, adding “The development history of AI shows that the open-source ecosystem is not something new; it has been strong in areas like speech-to-text recognition. Even if models are not open-source code, it still follows the principle of 'many hands make light work.' It allows more people, research institutions, and small companies to use open-source model products, creating a massive feedback network. It's like an army of ants; the power it generates is considerable. Of course, in scenarios like the overall capability of GPT-4, open-source models have not surpassed closed-source ones, which is a reality. If you look at the applications, in many scenarios, the capability of open-source models is sufficient. In our current practice with many clients, the effect produced in specific scenarios is enough. Besides, closed-source models are very costly, requiring a lot of computing power, with high costs and data security concerns."

"In fact, open-source large models already perform quite well, and many enterprises use them without paying fees. If a company uses a paid closed-source large model, that's a 'fraudulent deal,' especially when high model licensing fees and API costs are involved, spending millions a year, only to have it as a showpiece that employees can't even use. Therefore, to effectively use large models in enterprises, it is essential to integrate them with actual applications. Regardless of the model chosen, the ultimate goal is to combine it with the enterprise's real-world scenarios to strengthen applications, allowing enterprises to truly benefit from AI," Fu further explained.

In 2009, Fu became CEO and Chairman of Keniu Software. On November 10, 2010, Kingsoft Security and Keniu Software merged to form an independent company, with Fu Sheng as CEO of Kingsoft Network. By March 2014, Kingsoft Network rebranded to Cheetah Mobile, becoming a leading Chinese internet company focusing on cybersecurity, web browsers, and mobile applications.

On May 8, 2014, Cheetah Mobile was successfully listed on the New York Stock Exchange (NYSE), with its Clean Master app exceeding one billion downloads worldwide, setting a benchmark for Chinese internet companies going global.

Around 2016, Cheetah Mobile ventured into AI and robotics, establishing OrionStar.

In December 2023, Cheetah Mobile announced it had increased its stake in OrionStar, founded by Fu, through two wholly-owned subsidiaries to 35.17%. In January 2024, OrionStar announced a capital increase of approximately 369 million yuan.

In Cheetah Mobile's Q1 2024 financial report released on June 7, Fu announced that the company is transitioning from a consumer-focused company to an enterprise-focused one. The strategic focus will be on developing custom applications based on large language models (LLMs) for enterprises and leveraging these applications to enhance its service robots for enterprises, aiming to commercialize large models with strong AI capabilities and successful product development experience.

Currently, Cheetah Mobile's revenue comes from two major segments: internet business and AI and others.

More detailed business segments include: app business (BeoFun Technology), international advertising business (Cheetah Overseas Marketing), cloud management business (Juyun Technology), and AI business (OrionStar).

In his speech delivered on July 6, Fu said that Cheetah Mobile aims to become a leading provider of new productivity tools in the AGI era. He pointed out that Cheetah Mobile's large model application products primarily focus on global enterprise AI applications, achieving significant leaps in enterprise data security, accuracy, and efficiency through private large models, private data, and deeply customized applications.

However, Fu also noted that humanoid robots based on AI and "embodied intelligence" are challenging to commercialize, while wheeled service robots are easier to deploy and scale profitably.

"Robots will still require many years of investment. I don't believe that creating a humanoid robot that can do everything will immediately sell worldwide. Looking back at the invention of the automobile, it took many years to replace horse-drawn carriages. The earliest cars had various faults and issues. Similarly, robots are a vast industry. If we talk about real-world deployment in the service sector, wheeled service robots are already showing significant growth this year. For instance, OrionStar's service robots have moved from leasing to overseas sales, especially in developed markets where income continues to grow. We've also discovered more scenarios where robots are needed, such as in Japanese nursing homes and retail stores, indicating increased market acceptance. It will take much longer time for Bipedal humanoid robots to become viable," Fu told TMTPost.

Fu believes that future robots should not necessarily resemble humans but should be seen as tools to assist humans with various tasks. "Cars did not need to look like horses to emerge. Biological and mechanical designs do not have to match exactly, even though there are bionic technologies."

Fu acknowledged that the scaling law of large models has somewhat slowed, especially with GPT-5's continuous delays and the emergence of intelligent phenomena remaining in a "gray box" state. However, this slowdown has provided more opportunities for the deployment and development of open-source edge-side small models.

"This phenomenon is quite apparent. GPT-5's delay until next year indicates some difficulties with the scaling law. The scaling law requires connecting 100,000 computing units, and the U.S. is experiencing severe energy shortages, with cities running out of electricity, creating many physical limitations. Data is also insufficient today, and the entire system consumes excessive resources, so the scaling law may not be the optimal solution. For us, this slowdown is beneficial. Some people say that it could take only one year for the application of AGI, I disagree with it, but a major technological revolution is undeniable. For example, combining robots with large models has greatly improved response capabilities, making it a good time for application developers to thrive," Fu remarked.