Huang Tiejun, President of Zhiyuan Research Institute: China's big model is very hot but also very scattered, and we should concentrate on doing big things

21st Century Business Herald reporter Bai Yang reported from Beijing

At the recently held KLCII Conference 2023, KLCII officially released the fully open source "Wudao 3.0" series of large models and algorithms.

Founded in 2018, KLCII is a non-profit research institute that aims to promote innovation in the field of artificial intelligence development policy, academic thought, theoretical foundation, top talent and industrial ecology. KLCII started R&D on the large model in October 2020, and the release of Wudao 1.0 in March 2021 was the first and largest large model in China and the world at that time.

Huang Tiejun, president of KLCII, said that when they released Wudao 1.0, they already foresaw a new stage of artificial intelligence from "big refining model" to "big refining model". He believes that a large model needs to meet three conditions: first, the scale must be large, and the parameters should reach more than 10 billion scale; the second is emergence, which can produce unexpected new capabilities; The third is versatility, which is not limited to specialized problems or areas, and can handle a variety of different tasks.

The "Enlightenment 3.0" released this time covers basic large models such as language, vision, and multimodality. KLCII released the Aquila language model, which is the first open source language model with bilingual knowledge in Chinese and English, support for commercial licensing agreements, and domestic data compliance requirements.

It is understood that the open source of "Wudao Skyhawk" includes a series of models, including the Aquila base model, the AquilaChat conversation model and the AquilaCode (text-code) generation model.

In terms of visual large models, KLCII released the "Enlightenment Vision" visual large model series. Huang Tiejun pointed out that "Wudao Vision" includes a multimodal large model EMU that can complete everything in a multimodal sequence, a billion-level visual basic model EVA, and a general visual model Painter, which pioneers the path of contextual image learning technology. These models systematically solve a series of bottlenecks in the current computer vision field, including task unification, model scale, and data efficiency.

It is necessary to concentrate our efforts on doing great things

During the conference, Huang Tiejun was interviewed by the 21st Century Business Herald and other media, and answered the relevant hot spots of the current development of large models. He bluntly said that although tens of billions of scale will have the ability to emerge, the large models currently trained in China are still too small, and there is still a gap in the level of intelligence. "Now the domestic big model is very hot, but there is no top, this is also a problem, if everyone is repetitive efforts, it will only make resources more and more divergent, so now what China should do the most, is to concentrate on doing big things."

"The real AI innovation ecosystem should be that everyone should play their own advantages, be the strongest in the best link, and then connect all the strongest links together to form an ecology, which is the direction we should strive for." Huang Tiejun emphasized.

From a technical point of view, Huang Tiejun believes that the "big" of the current large model is far from reaching the ceiling, "the language model may take another 3 years to achieve all-encompassing, so in the next 3 years, the scale of the model will definitely be enlarged, and the ability will be stronger." ”

It also mentioned that the reason why so many large models can emerge is because of technological changes, the most important of which is self-supervised or unsupervised learning, which makes AI models not have to be the same as before - in order to solve a problem, label a type of data, train a model, make a product. "The model has become larger because there are more raw materials to feed, and the self-supervised learning method makes the raw materials no longer limited by factors such as cost and manpower."

However, although the market is now a hundred schools of thought, Huang Tiejun believes that the future big model ecology will only have a few large models, and the current dozens and hundreds of models are only intermediate products of technical iteration.

At that time, in addition to a few large models at the head, other companies need to find their place in the ecology, this position is not necessarily a large model, as long as it can be done well in a certain link after the formation of the ecology, it can also be profitable. "It's like in the early days of the Internet, they all thought that 'selling shovels' was the most profitable, but in fact, there were only a few routers sold, and everyone developed in other areas of the Internet ecosystem."

However, the development of large models still takes time. Therefore, at this stage, for entrepreneurs, it is not impossible to use a general basic model plus data in a specialized field to make a vertical model. But in the future, as the general model becomes stronger and stronger, it will become a ubiquitous service like the Internet now, it will enter another stage, just like most of the Internet application enterprises now, when "basic model + professional" will also become the most typical model.

AI challenges and governance

For ordinary people, the big model will also have two effects. On the one hand, as a technical tool, it can replace many tasks that only people can complete, for enterprises, the cost is lower, faster, better quality, for consumers, it will bring high quality and low price, not everyone can enjoy products and services in the past, now everyone can enjoy, such as some information services or images, video and other content generation.

Therefore, the big model can promote economic and social development, and can also promote the improvement of people's living standards. In Huang Tiejun's view, at least 10 years or even longer in the next 10 years will be a period of integration between humans and AI.

On the other hand, the development of large models also means that many people's jobs will be replaced. The original stable job, now AI can be achieved at a lower cost, then these personnel will face a relatively large impact.

Huang Tiejun said that to see whether an industry will be affected by AI, it mainly depends on whether intelligence elements play a key role in it, and if so, it will be affected. For example, in the manufacturing industry, if it has been automated, the impact will be relatively small, but if there are still a lot of manpower doing repetitive work, the efficiency improvement of AI will have a greater impact.

Huang Tiejun said that the big model has a good side and an impact side, but this is the normal state of human social development. "We can't fix ourselves in an era where new opportunities emerge, repetitive and unwilling jobs for AI to do, and then find new things that can better use human capabilities to do." However, it should also be noted that the next decade or two will be the honeymoon period of human-machine integration, but with the power of AI, we must also begin to pay attention to the disruptive challenges that AI may bring. ”

Finally, when talking about AI governance, Huang Tiejun believes that the ethical safety of artificial intelligence does face great challenges. In the past, when nuclear weapons appeared, they also brought great challenges, but their regulation was predictable, and the emergence of AI capabilities was unexpected, and it was very difficult to manage the unexpected.

"But from a human point of view, people are also unpredictable, often witty, flash a lot of new ideas, and even make a variety of behaviors. So since AI is an intelligent system, people are also intelligent systems, then some systems and laws of human society can also try to use AI management, in this regard, there is a lot of experience to learn from and explore, see how to manage an endless creative system. Huang Tiejun said.

For more information, please download 21 Finance APP