Solve the problem of domestic large-scale models: avoid "passive water", the last mile needs to connect algorithms and chips

Source: The Paper

Author: Hu Xiner, an intern, Shao Wen, a reporter from The Paper

Image source: Generated by Unbounded AI

Dai Qionghai, an academician of the Chinese Academy of Engineering, said: "Our country should deepen the cultivation of artificial intelligence personnel and basic research in terms of policies, mechanisms and investment, strengthen original innovation, and avoid falling into the dilemma of 'water without a source'."

Wang Yu, a tenured professor of the Department of Electronic Engineering of Tsinghua University, pointed out: "There are already many chip companies in Shanghai, and there are also many algorithms in Shanghai. How to achieve efficient and unified deployment and run such algorithms on chips is a very important issue. question."

On July 7, at the 2023 World Artificial Intelligence Conference "General Artificial Intelligence Industry Development Opportunities and Risks in the Era of Large-scale Models" forum, a number of experts in the field of general artificial intelligence focused on large-scale models, respectively from basic innovation, application technology, and future prospects. In-depth discussion of artificial intelligence at other levels.

"Our country should deepen AI talent training and basic research in terms of policies, mechanisms, and investment, strengthen original innovation, and avoid falling into the dilemma of 'water without a source'." Dai Qionghai, counselor of the State Council and academician of the Chinese Academy of Engineering, emphasized in his keynote speech.

Wang Yu, a tenured professor and head of the Department of Electronic Engineering at Tsinghua University, said from the perspective of landing applications that it is currently difficult to deploy large-scale models in the vertical field, and domestic large-scale models face three challenges: high field deployment costs, a large gap in model computing power, and domestic Chip replacement is difficult. "In the last mile of the large model landing, we need to connect the algorithm with the chip." Wang Yu said.

"Brain intelligence is the new direction of the future"

Dai Qionghai believes that in the innovation of the large model "0 to 1", domestic disruptive achievements in the field of basic research are weak. "From the perspective of the intelligent development industry, we are both optimistic and not optimistic." In his view, most of China's artificial intelligence talents are concentrated in the application layer, so there is a lot of room for application scenarios and technology layers. However, China is obviously at a disadvantage in terms of talents at the basic level, and lacks original innovation.

Dai Qionghai said that the innovation and development of artificial intelligence requires three pillars, namely algorithms, data, and computing power. Algorithms determine the level of intelligence, data determine the scope of intelligence, and computing power determines the efficiency of intelligence. At the algorithm level, large models are expected to become a key basic platform in artificial intelligence applications in about five years.

Dai Qionghai also pointed out that brain intelligence is a new direction in the future. The new artificial intelligence algorithm that integrates the brain and cognition is at the forefront of the industry layout and will lead a new generation of intelligence. He suggested at the forum that the government should encourage enterprises to lead the construction of large-scale models, explore the combination of biological mechanisms and machine characteristics, further create new paradigms of artificial intelligence, and promote basic research and application expansion simultaneously. He predicts that artificial intelligence with cognitive intelligence as its core will begin to be applied ten years later.

In addition, Dai Qionghai believes that it is necessary to be vigilant about the security issues of large-scale model applications. Large models are not yet capable of authenticating outputs such as generating deceptive content. "This means that once there is a problem with the application of the large model, it is not as simple as the current computer network virus, just kill and kill the virus, which will have a subversive impact. Therefore, when the large model is applied, security and reliability should be combined. Believability was discussed clearly.”

Domestic large-scale models should focus on solving the four pain points

Wang Yu said in the forum, "Shanghai is very concerned about artificial intelligence and chips, but from another perspective, our most advanced models and relatively important computing power are actually subject to certain restrictions. Our computing power should be more Which direction to go, how to better make up for domestic computing power, and how to support the country's development in large model training and reasoning with such a space, these issues have become extremely important."

Wang Yu also mentioned that currently in foreign countries, only Nvidia and AMD can choose chips with large computing power. Nvidia dominates the market, and its software ecosystem is relatively good. "So various foreign models, such as OpenAI, Microsoft, and Google, are purchasing Nvidia chips in large quantities, and then develop them on top of Nvidia's software framework. The foreign ecology is very simple. Enterprises do a good job of algorithms. In this field, the deployment is supported by Nvidia's software system."

"However, the development of large computing power chips in China is still in its infancy." Wang Yu believes, "There are already many chip companies in Shanghai, such as Tianshu Zhixin, Suiyuan Technology, Cambrian, Biren Technology, etc. There are also many in Shanghai. Algorithms, how to achieve efficient and unified deployment, and how to run such algorithms on chips is a very important issue."

At the same time, Wang Yu emphasized that it is very difficult to deploy large-scale models in the vertical field at present, and domestic large-scale models face three major challenges: high field deployment costs, a large gap in model computing power, and difficulty in replacing domestic chips.

"Different from a model in the past AI 1.0 era targeting a specific task, today's AI 2.0 era is a model that solves multiple tasks, and the application layer, algorithm layer, and system layer need to be optimized collaboratively." Wang Yu said that at the end of the large model landing One kilometer, domestic large-scale models should focus on solving four pain points.

"We first need to deal with the problem of long text, that is, to use it well." Wang Yu said that the current trend of algorithms is to increase the length of text supported by large models, but long text will also bring about a surge in load. Transformer (A deep learning model developed by Google, on which OpenAI developed GPT) The load on the architecture increases dramatically as the input becomes longer. Therefore, adjusting long text is an extremely important requirement.

Another requirement for large models is the improvement of cost performance. "If Google uses a large model in its search engine, it will increase costs by US$36 billion, and its profits may be lost by 65%." Wang Yu said that if the company can reduce the cost of a click, the overall loss of profits may be reduced. . Moving in this direction, everyone is expected to be able to afford large models.

In addition, large models need to empower a variety of vertical fields. For all walks of life, there are not many large models that have a lot of knowledge. Especially in fields such as medical care and finance, the acquisition of corpus data is expensive and very scarce. "If you can add a general-purpose base model and fine-tune it, the basic performance of various industries is expected to be further improved." But Wang Yu also pointed out that if it develops in the vertical field, the general-purpose large model must be fine-tuned, and the larger the model , the cost of fine-tuning also increases significantly. Therefore, how to design an efficient fine-tuning algorithm is a topic that needs to be discussed.

At the same time, large models also bring new requirements for one-stop deployment. In the optimization of software and hardware, if operator optimization, compilation optimization, and hardware deployment are deployed in layers, a total of 100 manpower is required per day, while for one-stop automated deployment, only 10 manpower is required per day. Wang Yu pointed out that one-stop deployment can optimize labor costs, further increase the scale of compilation optimization space, and is expected to promote the development of the entire industrial chain.

View Original
The content is for reference only, not a solicitation or offer. No investment, tax, or legal advice provided. See Disclaimer for more risks disclosure.
  • Reward
  • Comment
  • Share
Comment
0/400
No comments
Trade Crypto Anywhere Anytime
qrCode
Scan to download Gate app
Community
English
  • 简体中文
  • English
  • Tiếng Việt
  • 繁體中文
  • Español
  • Русский
  • Français (Afrique)
  • Português (Portugal)
  • Bahasa Indonesia
  • 日本語
  • بالعربية
  • Українська
  • Português (Brasil)