Chinese AI entrepreneurs in conversation: "Next time starting a business, no AI, choose food and beverage without U.S. sanctions

Recently, the fourth BEYOND International Science and Technology Innovation Expo was held in Macau, China.

During the opening ceremony "What's Next? AI Panel" roundtable discussion, He Jiandong, President of the Macau Technology Association （澳门科技总会会长贺建东）, Xu Bing, Co-founder of SenseTime（商汤科技联合创始人徐冰）, Liu Qingfeng, Chairman of iFlytek（科大讯飞董事长刘庆峰）, and Zhang Wen（壁仞科技创始人张文）, Founder of Birun Technology, discussed topics such as AI's advantages in the Chinese and American fields, the reasonable boundaries of AI development, and entrepreneurial opportunities in AI.

Xu Bing mentioned that the biggest gap between China and the United States now lies in computing power, which is roughly ten times the gap between them. However, the computing power gap between China and the United States can be bridged with significant investment.

During the discussion, Zhang Wen said he doesn't want to do AI again and might choose the restaurant industry for his next venture. "I don't know if this business will affect McDonald's in the United States," but he hopes to avoid sanctions.

(Note: SenseTime, iFlytek, and Birun are currently on the U.S. Department of Commerce's "Entity List," and SenseTime is also included in the U.S. Department of the Treasury's "Non-SDN Chinese Military-Industrial Complex Company List," as mentioned at the beginning of the dialogue.)

Liu Qingfeng stated that large Chinese companies have financial and market strength, but they "cleanse" the market, leading to a situation similar to "bad money drives out good," which is actually very detrimental to innovation in small and medium-sized enterprises. Therefore, Chinese companies need to promote a more benign ecosystem, which is the biggest gap between China and the United States.

Here is a brief summary of the core information of the 25-minute roundtable discussion provided by Titanium Media AGI for readers to refer to:

Question: What are the differences (gaps) in the AI industry between China and the United States?

Liu Qingfeng:

Talking about GPT-4o, we see that Jiandong smoothly switches between English and Chinese and is very fluent. If there are others who can switch like him and interact so vividly with each speaker, I feel that currently, AI still can't do that, so, in fact, humans are still smarter and wiser now.

So, how do we view AI between China and the US?

Many people were surprised when ChatGPT came out, saying they originally thought China and the US were very close in the field of artificial intelligence, but suddenly the gap widened. In fact, ChatGPT is a deep pre-training method that came out in 2018, and in 2020, we had GPT3.0 and ChatGPT-3.5.

This method is something that we, including the AI research fields of several companies present here, including those in China, have always been paying attention to. We even used the GPT method for the first time to surpass human performance in English reading comprehension, which was done by Chinese technology. But why did the others not roll out (similar models)? It actually represents an important leap from quantitative to qualitative changes. It scaled up the model to a larger size, up to 175 billion, and later, GPT-4 reached over 1 trillion floating-point parameters, with substantial data training, and then organized a massive number of people for human-machine collaborative reinforcement training, which ultimately led us to catch up quickly from quantitative to qualitative changes.

I think, from this example, one can see that when it comes to the development of general artificial intelligence between China and the US, the computing power foundation in the US is certainly more advanced than ours, and the flexibility of fund allocation in the US financial market, as well as the overall societal atmosphere for original technological innovation, is still stronger than in China. This is something we should acknowledge. However, in this current round, in the field of general artificial intelligence represented by large models, the gap between China and the US is not as large as many might think. China is the only country in the world that could potentially achieve the emergence of secondary wisdom. Today, the smartest and best large model on the market is still GPT-4, which has now evolved into GPT-4v based on the intelligence of GPT-4, including its Sora video, speech recognition, and today's voice technologies.

Actually, in these technologies, the underlying logic is still GPT-4. Our gap with GPT-4 is only about half a year. By June or July this year, we believe we'll be able to catch up to his current level, at least in English if not in Chinese. However, once GPT-5 is released, the gap might widen to over a year. At that time, we will catch up again because of stronger computing power, more data, and earlier training.

The gap between China and the US in the general large model base is a dynamic catch-up of about half a year to a year and a half. We won't be left far behind; that's a basic assessment.

So, based on this, when we are able to keep up with the base capabilities, in various vertical tracks, we might be able to do even better than them. Because in the development of China's internet and the original digital economy, we actually have advantages in some infrastructure and data-driven business ecosystems and innovation models compared to the US in this field.

For example, in education, we've been collaborating with Macau in education. At the end of last year, Japan wanted to promote its 2025 World Expo and Osaka wanted all public schools to use large educational models to enhance their English teaching. In a global tender, despite direct competition with GPT, we eventually won the bid as a single source, and the speech synthesis effect of our GPT-4o speech is no worse than theirs.

Secondly, in fact, we set international standards. In 2020, in Korea, the standard called "full duplex communication" was defined by us. Thirdly, when it comes to getting results in 200 millimetres or 300 millimetres, that's end-to-end training requiring computational support to unify video, text, and speech training. It takes about three months to half a year to complete. As long as the computing power is there, we can do it.

So today, I've been talking a lot about the gap between China and the US in general large models. I think it's about half a year to a year and a half, and China might reach the second tier in the world in terms of AI intelligence without falling behind.

The whole algorithm logic, including the pace they're leading now, is catching up closely. But as long as we're not left behind, we might surpass in various vertical tracks like medical education, speech interaction, and many aspects related to smart hardware.

In the entrepreneurial ecosystem, it seems like by 2026, 80% of companies worldwide will be using large models. Currently, it's around 5-10%, and US startups are more active in applying large models compared to China. This is because their capital is more inclusive, and big companies have more respect for small startups than in the Chinese market.

Chinese large companies have financial and market strength, often dominating the market, which I think is not favourable to innovation in small and medium-sized enterprises. So, we need to advocate for a more benign ecosystem, and this is where the gap between China and the US lies.

I believe in the future, in the commercialization of various vertical tracks globally, China may eventually have the upper hand in this round of AI.

Xu Bing:

First of all, a big thanks for inviting us to this stage, and it's great to connect with everyone here. I remember vividly, ten years ago when starting my entrepreneurial journey, I visited Qingfeng, Mr. Tang, and iFlytek for learning. Zhang Wen, the former president of SenseTime, also had a successful entrepreneurial journey after his time at SenseTime.

So, to begin with, China indeed has a group of AI pioneers. Over the past decade, in the development of AI in China, we've been diligently working on solidifying these foundational technologies step by step, which has led to the current situation where, as Qingfeng just mentioned, the gap isn't that wide, right?

We know the three key elements of AI: computing power, data, and talent. Presently, the biggest gap between China and the US lies in computing power, with about a tenfold difference. The US predominantly owns the most cutting-edge high-performance GPUs, with many countries and companies globally purchasing GPUs from them. In the past decade, the value created by AI is roughly $2 trillion, equivalent to NVIDIA's market value.

We are one of the largest purchasers of NVIDIA GPUs in Chinese history, with about 45,000 of them. However, when we compare it to the US currently, there is a significant gap, roughly over ten times. But I believe this gap can be bridged. On the one hand, our domestically produced chips are rapidly developing, and on the other hand, computing power is essentially a commodity, right?

Computing power is a commodity; as long as we are willing to invest in it and it has its own strong financing attributes, it has already exhibited investment properties similar to real estate. It can increase investment scale through high turnovers, fund circulation, leveraging, and amplification of investment size.

Currently, we see rapid progress in the construction of GPU computing centres in the Asian market. In this context, I believe there is an opportunity to narrow the gap in computing power between China and the US.

Talent has liquidity under this background. Talent is where there is abundant computing power. Currently, the best AI talents are indeed in the US, primarily because the US possesses the most computing power. However, once we can bridge the gap in computing power, many talents will also have mobility and flow into the Asian market, flowing into China, Hong Kong, and even Macau to innovate in large models.

Speaking of data, I believe China has a significant advantage in data because the breakthrough of large models originally came from the aggregation of internet data, subsequently abstracting the physical world, right? We can use language to express what the physical world looks like, and large models possess this ability.

Now, incorporating vision and as Qingfeng just mentioned, adding speech, it has evolved into a multimodal form. So, in terms of data, China boasts a rich reservoir. Additionally, another noticeable advantage in data for China is our ability to finely sculpt data.

We can work with data more efficiently than in the US, so during this period, the level of large models in China has improved significantly, moving from being out of reach compared to ChatGPT to approaching GPT-4 and some companies have already surpassed GPT-4 in some public evaluations.

A significant part of the reason behind this is that we have devoted a lot of effort to meticulously refining the data. With this meticulous refinement alongside massive data support, under more computing power, we have the opportunity to catch up to GPT-5 and GPT-6.

So, in this setting, I believe that in the long run, we need not worry about a significant gap between AI in China and the US. This, I believe, is crucial for every industry because the world needs an alternate choice. In the exportation of AI, having an alternative choice is vital.

It signifies that this technology is going global, being able to create enormous value in various dimensions of the global economy. In this process, I believe numerous new trillion-dollar and billion-dollar companies will emerge and bring about disruptive original innovations in education, healthcare, social interactions, games, and entertainment, right?

So, although currently, we may lag in the Chinese market, we have a more energetic environment, fiercer competition, which is advantageous for China's long-term AI innovation. It's similar to the electric vehicle industry; initially, we were behind, right? Tesla dominated the electric vehicle market, but looking at it now, the Chinese electric vehicle market is more robust and vibrant. We can export overseas; I believe AI will follow a similar path in the next 5-10 years.

Zhang Wen:

Thanks to Jiandong and the organizers. Liu, from iFlytek, is my fellow townsfolk; we're both from Hefei. As for Xu Bing, we started a business together back then. Unfortunately, SenseTime had it worse than Bitmain; it got sanctioned during a visit by the Transport Minister from both sides. Yes, I was involved in receiving them at SenseTime. We never expected a company visited by a Transport Minister to get hit with sanctions.

So, in the end, I started my own venture and founded Bitmain. But three years later, Bitmain got sanctioned by the US, making it the second time for me. For my next venture, I might delve into the restaurant business. I'm not sure if it'll affect American McDonald's, and I hope to avoid any sanctions.

Regarding the AI gap between China and the US, I'll answer from various angles. In the short to medium term, especially in large models, we might lag due to limitations in computing power and technology. But looking ahead in the long term, I believe this gap will gradually narrow.