Chinese AI vs. the US: When asking questions counts

Kai-Fu LeeCEO of Sinovation (venture capital), former president of Google China and author of AI Superpowers (2018), a book in which he predicted that China will become a world leader in AI, has a basic explanation that may prove crucial: it is not the programming ability or the essential huge data processing centers, but the education of users. Especially when it comes to asking questions to generative AI. “In the era of big models, the most powerful person is not the one who can write the best content, but the one who can ask the best questions. Pairing a skilled questioner with an assistant [una IA] “The first-class software is much more powerful than the best content generator.” Anyone who has used this type of application knows this. It requires learning and practice by the people who use it. Ultimately, it is about convincing, or stimulating, the machine.

Advertising

Knowing how to ask good questions has always been essential for the advancement of science. Now, with this AI, it is also the same. In fact, through questions, questions with natural language, through prompts or instructions like DAN (“Do anything now”), the way in which Generative AI responds can be changed and it can be forced to respond to questions without restrictions, without having to change its code. Some experts believe that even in this way, without coding, its codes can be changed. The Generative AIs themselves that were consulted deny this. But these generative AIs learn and improve with the questions, answers and re-questions of the users. With conversations, in short.

In addition to the fact that such verbal interactions are limited in China, the much lower diffusion of Generative AI may be due to the weaker publicity or media impact of this technology, part of the reason for the explosion of its use by users in the US and Europe. It may also be due to the fears of Chinese authorities that this technology will lead to bypassing censorship for general users, at least in some matters. In general, learning to ask questions (and receive answers) requires greater doses of freedom (which many Chinese researchers demand), not democracy (which is not a significant demand).

It also has to do with the capacity for originality. To quote a Chinese source, according to Zheng Yongnianof the Chinese University of Hong Kong (Shenzhen) and director of the Qianhai International Affairs Research Institute, “the future of originality is likely to continue to be led by the United States.” This is not because the Chinese are less intelligent. Many advanced chip design teams in the US, he explains, are led by Chinese. But because of a lack of “institutional reforms,” a complaint often heard at international conferences, such as the one we quote herealong with the need for a business culture that “generates returns.”

This is what the Big Four are trying to do. startups Chinese generative AI companies: Zhipu AI, Moonshot AI, MiniMax and 01.ai. The Chinese themselves, faced with this challenge, look to Europe. And what do they see? A negative example: excessive regulation and conservatism, a lack of originality, are what have led to a lack of development of the industry on the Old Continent. In the field of large language models, the United States currently occupies a monopoly position.

However, the idea that took hold a few years ago in the US that China did not know how to innovate or lacked creativity is also a thing of the past. For example, China is ahead in practical advances such as mobile payments without credit cards, digital money, or short videos, as started by Tik Tok, a Chinese company, and they began to use influencers Chinese to advertise clothing and brands at a incredible speedMajor American platforms have had to adapt. In fact, experienced Chinese staff in live shopping are training Americans to sell on the TikTok Shop.

While the implementation of Generative AI (GPT Chat, Gemini, etc.) has been very fast among American users, it is much slower in China. More than the restrictions imposed by the US on high-tech exports to China, the reasons, as noted above, may be cultural and political. Economic, technological and geopolitical priorities also matter in the AI ​​race.

«One of the great historical battles of China was to design typewriters, when they emerged, without abandoning their model or falling into literacy»

But there are other aspects that are less talked about. The large language models, the basis of this generative AI, among other things, operate by dividing words, among other techniques, into syllables, and looking for the next syllable that is most appropriate for the answer, based on their enormous databases. Mandarin Chinese operates on ideograms. In fact, one of the great historical battles of China was to design typewriters, when they emerged, without giving up their model, without falling into a literacy that they considered degrading like the Japanese or Koreans. Thomas S. Mullaney explained it in a fascinating way in his book The Chinese Typewriter: A History (2018).

These were years of enormous efforts with various systems of gigantic maps of ideograms that were selected by the operators, since women were better at these tasks. These efforts ended with the emergence of digitalization that solved the problem. In the face of generative AI, Chinese models have managed to use the characters or ideograms themselves, which have problems with polysemy (their meaning changes depending on the context), or break them down into their “radicals,” signs that form them or are repeated, and thus generate generative AI from Chinese texts.

Chinese texts, really? Yes, some large Chinese language models are trained only in Chinese, for example, for AIs focused on specific fields such as literature or legal issues. Others with little English text or with multilingual training. But English texts are dominant in the world’s large language models. They are the first “training” tools for generative AIs. Certainly in the US, but also in China where between 50% and 80% of the first texts that feed their generative AIs are in English. Only later do Chinese texts come in.

It is worth considering whether these texts, at least for some purposes, introduce the values ​​and biases of the cultures in which they were written, whether papers academics, the press, or other sources, not to mention the vast field of codification. This introduces the prejudices of each culture, among others, because language reflects cultural values.

That is to say: Is liberal culture, especially American, permeating Chinese Confucian culture through this development of generative AI? Can the opposite happen with, for example, the AI ​​contained in Tik Tok, a Chinese company that the American government wants to ban in the US? Can we reach a clash of cultures to the Huntington, now in generative AI? We will have to answer another time. Chat GPTs and other proven brands are not yet clear, but they admit the possibility. Important issue when generative AI, large language models, as Salvi, Ribeiro, Gallotti and West point out in a investigationthey can be more persuasive towards people than other human beings.

For all this, one of the most important issues with these AIs is knowing how to ask. That is learned. Some in the West thought that the arrival of the Internet was going to liberalize China. It was not knowing the system. The regime has universalized the network in its society, but it has turned it into an instrument, along with others, of greater control, of techno-surveillance. It can also happen with this generative AI, as with the Internet. Not only in China. Also in the West, as Shoshana Zuboffven has explained in The era of surveillance capitalismand Edward Snowden himself in his memoirs, both written before the immense possibilities, positive and negative, that this new technology brings. One that has yet to find its business model, very different from that of Google (Alphabet) or Meta. Ask, ask. But questions are also monitored. And used.