iFLYTEK Liu Cong: If you are not sure about the algorithm, if you make a mistake, then three months will pass

Recently, at the press conference of iFLYTEK, Liu Cong, dean of the R&D Institute, made a remark that attracted widespread attention in the industry.

He said that the domestic model should aim for the international top level, and frankly admitted that there is still a big gap to make up. This kind of objective understanding and pragmatic style is refreshing. In this era when everyone is vying to boast about their skills, it is indeed rare that there are still people who can maintain a modest and cautious attitude.

However, being honest about the current gap does not mean that you will accept your fate. Dean Liu emphasized that it is necessary to reach the level of GPT-3.5 by the end of the year, which undoubtedly gives the industry confidence. The key lies in having a clear understanding of reality, constantly eliminating gaps, and striving to catch up with the advanced international level as quickly as possible.

Among the large models that are blooming in China, there are only a handful of them that have really developed and trained by themselves. Many of them are just fine-tuned with open-source models, and they are far from the world's top. In order to realize China's manufacturing from a big country to a strong country, the ability of the base model is the key. Multimodal integration is the only way, and medical, programming and other applications should also be built hard. With the support of strong domestic computing power and the mutual encouragement between enterprise teams, China's large model is just around the corner to enter the world stage.

In recent years, the mainland has developed vigorously in science and technology, and many top scientific research teams have emerged. However, some teams or individuals have a certain sense of pride and are unwilling to face up to the gap that still exists in the world, resulting in a certain degree of bias in the scientific research route. Taking the large model in the field of artificial intelligence as an example, although there is a situation of a hundred flowers blooming in China, there are not many products that can really compete with the top international level.

Liu Cong, Dean of iFLYTEK R&D Institute, clearly pointed out at a recent press conference that there is still a significant gap between domestic large models and international advanced levels such as GPT-3, and this objective understanding and frank attitude are refreshing. In such a field prone to exaggerated achievements, it is indeed commendable that there are still people who can maintain a style of rational analysis and hard work to catch up.

We should not be complacent because of some of the milestones that have been made. To move from a great power to a strong country, it is necessary to keep up with the world's top technologies, identify gaps and drive progress. Only through such continuous catch-up can the real rise of scientific and technological power be finally realized. Otherwise, if it is separated from the world's scientific and technological frontiers, it will gradually lose its competitiveness and influence.

Liu Cong's words reveal a solid style, and he believes that he will reach the level of GPT-3.5 by the end of the year to lay the foundation for the next stage of the sprint. This step-by-step mentality not only affirms the accumulation of mainland science and technology in the field of large models, but also maintains the awe of the international advanced level, which is a valuable quality worth learning from us.

Only by being humble can we keep improving. If Chinese enterprises want to enter the world stage in the field of large models, they must recognize the reality, remain cautious, and make up for the gap as quickly as possible, which is the only way to catch up.

At present, the large models of the mainland are blooming, and various industrial forces have tested the waters, but there are only a handful of them that can independently train the base model. Many of the so-called "big models" on the market are just fine-tuned on the basis of some open source models, and there is still a certain distance from original research and development. In order to realize the transformation from a big country to a strong country, the core competitiveness still depends on the ability of the independent base model.

As Liu Cong said, those who can compete with the international level will eventually only have a few head strengths. These top players will inevitably have to find their own way to create a unique base model as the foundation of their existence.

Taking iFLYTEK as an example, it has chosen a multimodal technology route to organically integrate images, voices and other information into the text system, and this prediction has been proven to be accurate. It is not enough to rely on a single language model, and multimodality is the only way to build a true general AI.

In today's rapid iteration of deep learning, there is a possibility that any achievement can be surpassed. However, the accumulation of the ability of the base model is more durable, and it lays the foundation for the overall strength of a team and even an enterprise. A good base model can also generate a variety of application-based models by adjusting the parameter scale.

Therefore, behind the appearance of the current large-scale model blowout, the core competition is still unfolding among a few leading enterprises. These companies must find their own unique pedestal models. Only in this way can we occupy a favorable position in the exaggerated reshuffle and achieve a real leap in scientific and technological strength.

Large models have a strong dependence on computing power. To train a giant model with trillions of parameters, you must have the computing power to match it. At present, most of the top-level models require the scale of 10,000 or even tens of thousands of cards, which puts forward higher requirements for computing power.

At this infrastructure level, Chinese enterprises are also facing both opportunities and challenges. In order to realize the independent and controllable core algorithm, it is imperative for domestic computing power platforms. iFLYTEK and Huawei's cooperation to launch the "Feixing-1" platform is an important step in this direction.

The performance parameters of domestic AI chips, represented by Huawei's Ascend chips, are still confidential, but it is conceivable that they are highly efficient in large model training and inference. The close cooperation between the two major enterprises is very rare, and the joint special combat team has the courage to tackle tough problems to ensure that the large model can be trained and operated on the domestic platform.

It can be said that this is a model of cross-border in-depth cooperation. Huawei is the main provider of underlying computing power, while iFLYTEK provides support in terms of algorithms and applications. The two sides show their strengths and complement each other, providing a solid foundation for the rise of domestic large models.

In the future, as the scale of the model continues to expand, the dependence on computing power will further intensify. Domestic enterprises should seize the opportunity of the times, build a computing infrastructure with core competitiveness, and provide a strong impetus for the accelerated R&D and application of large models with trillions of computing resources, which is the foundation of China's manufacturing leap to a manufacturing power.

The purpose of scientific and technological development is ultimately to benefit mankind, and large models are no exception. Its true value is reflected in the extent to which it can serve the society and benefit ordinary people.

From this point of view, the wide range of applications is particularly important. The series of products released by iFLYTEK this time are closely related to the needs of people's livelihood and have strong application value. For example, intelligent programming assistants for programmers, English learning assistants for students, and medical Q&A robots for patients have shown huge market potential.

It is foreseeable that as these early applications continue to improve, they will have a revolutionary impact on related fields. For example, medical consultation robots may greatly improve the level of primary medical care and solve the problem of difficulty in seeing a doctor.

Therefore, instead of dwelling on technical details, it is better to take a long-term view - the large model must truly serve the national economy and people's livelihood in order to reflect greater value. Additional technical investment should be guided by this, not just for the sake of leading.

It is foreseeable that there will be more amazing large-scale model applications in the future, which will penetrate into all aspects of society and make our lives more convenient and intelligent. That's what technological progress is all about.

The scientific and technological progress of an era does not depend on the lone struggle of a single individual, but on the joint cooperation of the entire industry. In the field of large models, Chinese enterprises should also uphold this concept, unite and help each other, and go hand in hand.

As Liu Cong said, there are only a handful of teams that truly achieve world-class level. At this critical juncture, enterprises in the industry should encourage each other, learn from each other's strengths, and work together.

iFLYTEK is willing to cooperate with potential enterprises, and its open-mindedness is worth learning. At the same time, other enterprises should also maintain a modest and cautious attitude, continue to learn from international advanced experience, and communicate with top domestic forces in order to grow faster.

Artificial intelligence is a field that requires interdisciplinary integration and a wide range of business applications. It is difficult to have a situation of riding a horse alone in a short period of time. The key is to jump out of the narrow vision, be eclectic, form a joint force, and work together to complete the transformation from catching up to leading.

This style of solidarity and mutual assistance and collective wisdom is also in line with China's traditional cultural values. Let us walk hand in hand and witness the light of Chinese artificial intelligence!

To sum up, although there is still a certain gap between the field of continental large models and the world's advanced level, the prospects are promising, and the hope of rising has been shown in front of us. The key is to maintain a modest and cautious attitude, face the reality, constantly make up for the shortcomings, and catch up with the international advanced as quickly as possible.

In this process, the ability of self-developed base model is particularly crucial. It is also necessary to build a matching domestic computing infrastructure and promote the wide implementation of various applications that benefit the people. Chinese enterprises should keep an open mind, unite and help each other, share resources, and drive the rapid development of large-scale model technology with the power of aggregation.

Let us witness the scientific and technological progress of this era and the rise of the country!

User Comments:

As soon as the news of the iFLYTEK press conference was released, it quickly set off a heated discussion on the Internet.

Some netizens expressed their appreciation for the modest attitude of Liu Cong, president of iFLYTEK R&D Institute: "The real patriarch is very low-key, not ostentatious or boastful, this is the real powerful internal force!" "This cautious attitude is very admirable, which not only affirms the progress of domestic scientific and technological development, but also maintains an international advanced learning mentality, which is worthy of learning from us science and technology practitioners." ”

However, some netizens questioned: "Didn't they just release their new models, why did they turn around and say that they are not as good as abroad? Isn't this a little capricious?" "I think the domestic scientific and technological strength is not bad, there is no need to always emphasize learning from foreign countries, you should be a little confident." ”

In addition, some netizens expressed concern about the competitive attitude of different domestic enterprises: "Such a low-key style is really good, but there is still a situation of open and secret fighting between different enterprises, which is really not conducive to the development of the entire industry." "I hope that major domestic enterprises can unite and help each other to jointly promote the process of Chinese artificial intelligence and benefit the society." ”

Various voices collide with each other, fully reflecting the public's attention to the development of domestic artificial intelligence. We need to maintain a calm and rational attitude, not only to see the progress of domestic science and technology, but also to be soberly aware of the gap between China and the world's advanced level. Only in this way can we continue to make efforts, strive to catch up as soon as possible, and realize the leap of China's manufacturing from a big country to a strong country. Let's wait and see the take-off of science and technology belonging to our generation!

Finally, I would like to ask a question about the development of large model technology:

In this era of unity and cooperation, how can we build a good competition order and cultural atmosphere, so that different enterprises can not only learn from each other, but also encourage each other, and jointly promote the progress of science and technology and industry?