How to narrow the gap between China and the United States in general large models? I saw the answer at the two sessions

Bai Jiao is from Ao Fei Temple
Qubit | Official account QbitAI

"The universal large model is related to the battle for national destiny"...

"Artificial intelligence +" appeared in the government work report for the first time and was directly promoted as an action.

Representatives such as Lei Jun and Liu Qingfeng have included artificial intelligence in their suggestions;

When the two sessions of the year are underway again, large AI models have received unprecedented attention.

At that time, on the other side of the ocean, GPT-4 was being completely surpassed by the latest large models, and Sora's new videos continued to amaze netizens.

For a time, topics about the future development of domestic general-purpose large models and the gap between China and the United States once again attracted attention.

That being the case, let’s take a look first, what were the two sessions talking about? Maybe you can find the answer there.

Artificial intelligence becomes popular during the two sessions

In line with the launch of the "Artificial Intelligence +" action, more than 20 National People's Congress deputies or CPPCC members have talked about large models, covering all aspects from the underlying data computing power, model layer and application layer, providing solutions for the current challenges we are facing. Challenge and make suggestions.

General large model has become the key word here again. Such a grand occasion was rare in the past.

Among the specific suggestions, we can see that there are roughly three aspects: technical bottlenecks, future development, and application implementation.

• Technical bottlenecks: data, computing power and industrial ecology

What are the current technical bottlenecks for domestically produced large models? Including iFlytek Chairman Liu Qingfeng, Zhihu founder Zhou Yuan, JD.com Group Technical Committee Chairman Cao Peng, and Chinese Academy of Sciences Institute of Computer Science researcher Zhang Yunquan all expressed their opinions.

Zhihu founder Zhou Yuan talked about data challenges. He believed that large model data collection should be supervised and reviewed.

Cao Peng, chairman of JD.com's technical committee , and Zhang Yunquan, a researcher at the Institute of Computer Science of the Chinese Academy of Sciences, have all talked about breaking through the bottleneck of computing power. Cao Peng encouraged the collaboration of hardware and software in domestic computing power, and Zhang Yunquan proposed centralized AI chip development and the establishment of a special group for the development of intelligent computing power. Suggestions in a few directions.

Liu Qingfeng, Chairman of iFlytek , introduced the shortcomings of China’s large-scale development model from the dimensions of computing power, base platform, and source technology research and development, and suggested formulating a national “General Artificial Intelligence Development Plan” to narrow the gap between China and the United States in general artificial intelligence. Industrial gaps create China's comparative advantage.

• Future development: Educational talents, policies and regulations have also become the focus of attention

In addition to technology, aspects such as education, talent development, policies and regulations have also become the focus of the representatives.

Xiaomi founder Lei Jun put forward three talent-related suggestions: popularize artificial intelligence literacy education from the compulsory education stage; vigorously promote the construction of artificial intelligence-related majors in colleges and universities; support large technology companies and education and training institutions to cultivate artificial intelligence application talents.

There are also some people in the legal profession, such as Zhang Yi, senior partner of King & Wood Mallesons , who have proposed promoting the introduction of the Artificial Intelligence Law.

• Application implementation: How to empower thousands of industries?

It is worth mentioning that representatives from various industries such as film and television, sports, rural areas, elderly care, manufacturing, and cultural tourism also participated in the discussion on the development of artificial intelligence.

For example, Sora's impact on the film and television industry, actor Jin Dong said in an interview that some service-oriented jobs may be replaced, but in a short period of time, it will be difficult for artificial intelligence to replace film and television and other creative industries.

Others like Zhong Zheng, Vice President of Midea , Wu Guoping , Chairman of Nianhuawan Cultural Tourism , and Zhang Tianren, Chairman of Tianneng Holding Group, mentioned the application of artificial intelligence in manufacturing, cultural tourism, elderly care and other industries.

…

What can be seen is that large models have undoubtedly become the focus of the two sessions. In the proposals of more than 20 people's congress representatives or members of the Chinese People's Political Consultative Conference, the current development of domestic large-scale models can be summarized: technical challenges still exist, talent policies must keep up, and application development must be accelerated .

How big is the gap between China and the United States?

The emergence of ChatGPT has set off a war of thousands of models in China. Some players' large models have achieved the strength of the benchmark GPT-3.5 within one year, and some capabilities have exceeded GPT-4.

The emergence of Sora, which can automatically generate 1-minute videos through text only, has brought subversion to the field of video generation. Its performance has crushed similar products...

As a result, there has once again been heated discussion about whether the gap between China and the United States has further widened. Data, computing power, talent training and investment have become the focus of discussion.

But how big is the gap between China and the United States? There is still no conclusion.

At the two sessions, iFlytek Chairman Liu Qingfeng gave a quantitative description for the first time——

1-2 years, tied .

Why is this number? Liu Qingfeng gave further answers.

He believes that the "main battlefield" in the Sino-US game is to continue benchmarking on the capabilities of universal bases. Sora is a successful practice in a specific field based on the universal large model base capabilities of GPT-4/4V.

Also extended are products like DALL-E3 and Whisper.

He also took the iFlytek Spark model as an example and expected to reach the current best level of GPT4/4V within 6 months. But with the release of GPT-5, “this gap may be stretched to more than a year . ”

Therefore, he also emphasized that this will also be a dynamic process of chasing each other .

In Liu Qingfeng's reasoning, in the field of artificial intelligence, general large models have been pushed to a high point and become the core competitive point of the gap between China and the United States.

Some representatives also expressed similar views at the two sessions this time: the development of general large models is no longer a simple battle of technology, but also a battle of national destiny, with far-reaching influence.

It can be seen in the past year that general large models have become the development consensus of players in the industry.

At the model level, technological breakthroughs such as long text processing, multi-modality, logical reasoning, and mathematical coding are made to comprehensively improve the understanding of general large models. At the infrastructure layer, an independent and controllable computing power ecosystem has also been built, and domestic computing power software and hardware collaborate to support large model innovation and application.

Of course, the most obvious development change is the full bloom of the application layer .

Traditional players from various industries such as medical care, education, advertising and marketing, and manufacturing can accelerate the application of large models in their own fields based on the universal large model base platform and industry data.

In the war of thousands of models, most of them are large models in industries and vertical fields. Without the support of the universal base large model, the effectiveness of the industry's large model will not be able to continue to improve.

Therefore, China must have a large universal base model that is independently controllable and benchmarks against world-class standards.

The most representative practitioner of this is iFlytek .

In the past year, they have made two developments worthy of attention -

One is China's first domestically produced computing power platform "Flying Star One" that supports trillions of floating-point parameters, which works with Huawei to achieve independent control of domestic computing power.

Another one released iFlytek Spark V3.5 based on this platform. The overall effect is close to GPT-4 Turbo, and a large-scale model industry ecosystem has initially formed.

Based on computing power and continuous upgrading and iteration of general large models, they have deep applications in medical, education, industry and other scenarios, and have taken the lead in building an industry-leading large model industry ecosystem——

As of January this year, iFlytek Spark has 24 million net users. Based on iFlytek Hearing, iFlytek Spark APP, iFlytek Input Method and other applications, Spark has empowered hundreds of millions of users. The large model developer ecosystem has accumulated 370,000 developers, of which 240,000 are enterprise developers... and from this data closed loop, it self-drives the iteration and implementation of large models.

Past development results show that the national general artificial intelligence team represented by iFlytek is promoting the implementation of large models. We have the foundation and our own scenario and data advantages. But we must also objectively see the gap, face it squarely, and narrow the gap between China and the United States.

Global competition is becoming more intense, and universal bases are about to emerge

2024 has just begun, and new advances in AI that are measured in days have once again made the world sleepless.

Sora, which subverts video generation, Claude 3, which completely surpasses GPT-4, and the release of Stable Diffusion 3. In the industry chain, Nvidia officially exceeded 2 trillion US dollars and shocked the stock market...

Clearly, the global race is not slowing down, but is becoming more urgent.

But unlike last year, when ChatGPT first appeared, with hundreds of models fighting thousands of models, the country seemed much calmer this year. Because the consensus about technology trends couldn’t be clearer:

• Multi-modal fusion, including voice, image, video and other modal fusions, has become the focus of large model upgrades and iterations for major domestic and foreign technology companies;

• Scaling Law has been verified repeatedly, and different schools of large models are moving toward unity;

• Integrating software and hardware, a common base built by upstream and downstream of the industrial chain is even more imminent. Only a universal base is the cornerstone of comprehensive strength, long-term stability, everlasting foundation, and supporting "AI+" in thousands of industries.

Interestingly, this insight was also mentioned at the two sessions.

Liu Qingfeng gave comprehensive and systematic suggestions.

He suggested that on the basis of the 2017 "New Generation Artificial Intelligence Development Plan", the national "General Artificial Intelligence Development Plan" should be systematically formulated to promote the development of general artificial intelligence with top-level design.

At the same time, related work must also be carried out simultaneously. To this end, Liu Qingfeng gave nine suggestions.

First of all, focus on the "main battlefield" of general large models, integrate resources from all parties, and continue to increase investment.

For example, it will continue to support research and development in the next five years in the form of special projects, support the construction of computing infrastructure, and promote the application of large models in the fields of industry and people's livelihood.

Subsequently, it is to strengthen the source technology layout, focus on the fields related to general artificial intelligence, lay out strategic and forward-looking basic research, and insist on promoting the exploration of disruptive innovation with source core technology breakthroughs.

In addition to large model technology, we must also accelerate the development of brain science and brain-like intelligence, quantum computing, and promote AI for Science.

More specific suggestions include:

It is recommended to accelerate the formation of an independent and controllable industrial ecology with domestic large models as the core.

It is recommended to promote the opening and sharing of national-level high-quality training data, and support the priority and low-cost use of national strategic scientific and technological forces in the form of unveiling lists.

It is recommended to introduce more objective, fair and credible evaluation methods to promote the healthy development of large model applications in industry fields.

In addition, he also emphasized the importance of talent training, laws and regulations, and ethical and humanistic research.

Especially for talent training, he not only emphasized the cultivation of top innovative talents and applied talents, but also suggested accelerating the promotion of artificial intelligence general education, empowering basic education, vocational education and higher education, and suggested the establishment of a national artificial intelligence academy .

For industries and positions that may be largely replaced by artificial intelligence in the future, he believes that new talent quality models and training programs should be studied.

The reason why such insights and suggestions are systematic and comprehensive is that iFlytek itself is a national artificial intelligence team and knows its stuff. On the other hand, years of deep cultivation in the artificial intelligence industry has also given it a deeper insight into industry needs.

Through the views of industry representatives at the two sessions, we can see that the social consensus is:

Generic large models are the way to go .

From the perspective of global development, only by achieving independent control of computing power and industrial ecology can we ensure the continuous iteration and application of large models, and can we occupy a place and have the right to speak in the global competitive situation.

At the level of people's livelihood and society, new productive forces represented by large models are becoming a new type of infrastructure that supports social development. From technology research and development to commercial implementation, the consistent ultimate goal on this development path is to improve quality and efficiency for all walks of life.

Therefore, even now the gap cannot be ignored, but core players including domestic large models have initially explored an independent and controllable path to empower various industries. This is also the real value of large models.

From "Internet +" to "Artificial Intelligence +" , China can expect new productivity opportunities in the future.

Click here ???? Follow me and remember to star~