Zhou Hongyi, founder of Chinese internet security company 360 Security Technology, which is participating in China’s race to launch its own ChatGPT-style large-scale language model, said Sora’s introduction was “like cold water being poured over China’s head.” “It’s like,” he said. Media Yisai reported on Friday. “It cools many people’s heads and makes us realize the gap with foreign leaders,” he added.
This week, the Chinese government asked Sola to give its most trusted state-owned enterprises the lead in AI. The State Council’s State-Owned Assets Supervision and Administration Commission on Monday called on enterprises under the central government to “embrace the profound changes brought about by AI.” Ten of these companies were named as his AI champions, but the watchdog did not reveal the names of the selected companies.
Caining Xie, an assistant professor of computer science at New York University’s Courant Institute for Mathematical Sciences, denied having any involvement in Sora’s development, emphasizing the importance of talent, data, and computing power. In a widely publicized social media post, Mr. Xie asked whether China was ready to deploy Sola, saying the technology “will not be misused as a tool for profiteering or manipulation by some people or groups. “We need to do that,” he said.
Sora access is currently restricted. Unlike some of OpenAI’s previous models, this one is not open source, so only a few people have access to a trial version of the model.
Chinese entrepreneurs express awe and fear over OpenAI’s Sora video tool
Chinese entrepreneurs express awe and fear over OpenAI’s Sora video tool
However, few companies can match Sora. One reason for this is that new diffusion transformer (DiT) architectures are not yet in use.
Meanwhile, London-based unicorn Stability AI is following the popularity of Sora with its text-to-image model Stable Diffusion 3, which also uses DiT, as this architecture could become mainstream for building generative AI. Released. A likely path for Chinese AI engineers is to “first decode Sora and train it with their own data to mass produce similar products,” said one Chinese developer, who asked not to be named.
Xu Liang, an AI entrepreneur based in Hangzhou in eastern Zhejiang province, said it won’t be long until similar services appear in China. “We’ll see models like Sora coming out of the Chinese market as early as the next month or two, and we’ll see a lot more models coming out within the next six months,” he said. However, Xu pointed out that there may still be significant differences between Chinese products and Sora.
05:03
How will China’s AI stack up against ChatGPT?
How will China’s AI stack up against ChatGPT?
Wang Shuyi, a professor specializing in AI and machine learning at Tianjin Normal University (TJNU), said that the past year’s experience in LLM development will help Chinese big tech companies accumulate and accumulate know-how in this field. said. The necessary hardware will be provided and he will be able to produce products like Sora within the next six months.
Sora’s launch led to speculation about the secret to its incredible power output. “Data is probably the most important element to Sora’s success,” tweeted New York University’s Xie, one of DiT’s two developers. He estimated that Sora may have about 3 billion parameters.
“If true, this is not an unreasonable model size,” he wrote. “This may suggest that training Sora models may not require as many GPUs as expected. We expect very fast iterations going forward.”
A few months before Sora was announced, a group of researchers launched VBench, a video generation model benchmarking tool designed to evaluate the performance of Runway’s Gen-2 and Pika. Among the 16 dimensions, Gen-2 stands out in areas such as image quality and aesthetic quality, but was weaker in dynamic range and appearance style. Pika, co-founded by Guo Wenjing, a Chinese doctoral candidate at Stanford University, has good background consistency and temporal flicker, but the image quality needs improvement.
The VBench team, made up of researchers from Nanyang Technological University in Singapore and Shanghai Institute of Artificial Intelligence in China, found that Sora offers superior overall video quality compared to other models, based on a demo provided by OpenAI. I discovered that there is. There is limited information about how the model converts text prompts into videos.
Lu Yanxia, research director for emerging technology research at IDC China, said tech giants such as Baidu, Alibaba and Tencent will be among the first to roll out similar services in the country. Local AI companies iFlyTek, SenseTime and Hikvision (all licensed by Washington) also plan to join the race, he said.
But analysts say China still faces an uphill battle, with its tech market increasingly isolated from the rest of the world in terms of capital, hardware, data and even talent.
The gap in market value between China’s top tech companies and US companies such as Microsoft, Google, and Nvidia has widened since the Chinese government decided to kneel down to big tech companies in the name of curbing “irrational capital expansion.” has expanded significantly in recent years.
And while China was once seen as having an advantage in data volume, it now faces a lack of the quality data needed to train these new models, leading to increased demand for advanced chips. Limited access exacerbates the challenge, Lu said. Lu said the talent shortage is also a concern, as the country’s best talent in the AI field often shines when they work for leading American companies.
For example, OpenAI has a core group of Chinese-educated technical experts. Of his 1,677 associated OpenAI members on LinkedIn, 23 studied at Tsinghua University in China. Tsinghua University is her ninth-most ranked higher education institution among startup employees, ahead of Cambridge University and Yale University.
OpenAI’s top three schools by number of employees are Stanford University, University of California, Berkeley, and Massachusetts Institute of Technology, with 88, 80, and 59 employees, respectively, and have these schools listed on its LinkedIn profile.
But even if the necessary talent is available, experts question how far China’s homegrown generative AI can go in the face of existing constraints from U.S.-China trade tensions.
Ping An Securities said in a report that continued restrictions on semiconductor exports from the United States “may accelerate the maturation of the domestic AI chip industry,” but warns that “domestic substitutes may fall short of expectations.” did.
The U.S. government prevents Chinese companies from accessing the world’s most advanced semiconductor tools through restrictions on related products containing U.S.-originated technology. In October, the US tightened these restrictions again, blocking mainland access to GPUs designed by Nvidia specifically for Chinese customers in response to previous restrictions.
Alexander Hallowell, principal analyst for advanced computing at technology research and advisory group Omdia, noted that China has options other than GPUs for LLM training. “I wish I could use Google’s TPU” [Tensor Processing Unit]Huawei’s Ascend, AWS’ Trainium, or one of quite a few startup products,” he said.
However, replacing a GPU is costly. “The further you move away from the GPU route, the more effort it takes for software development and system administration,” he says Harrowell.
There are also opportunities specifically for the Chinese market, said Hangzhou-based entrepreneur Xu. “The publication of a technical report on Sora and the upcoming open source video model will set the foundation for Chinese players to learn,” he said. He added that the local video model will have better Chinese language support.
TJNU’s Wang noted that one of Sora’s demo videos included a scene of a dancing Chinese dragon, which he felt was a classic depiction of the activity. He said China’s large number of ethnic groups, folk traditions, customs and geographical diversity provide rich material for leveraging local video models to better cater to local users. .
Wang also balked at the idea that there is an “insurmountable gulf” between Chinese and American AI.
“Will Chinese companies simply follow suit and churn out plagiarism every time their U.S. peers develop a new product, or will they set larger goals to work toward secure artificial general intelligence?” Will you?” asked Mr. Wang.