Liu Qingfeng: iFlytek is not fooling Huawei's traffic. Some people are afraid that our big model is too good

Lei Diwang, Lei Jianping, October 25th

IFlytek held the 6th World Sound Expo and 2023 Global 1024 Developer Festival yesterday. Liu Qingfeng, Chairman of iFLYTEK and Liu Cong, Dean of the Research Institute, released the iFLYTEK Spark Cognitive Model V3.0, stating that the seven capabilities of the iFLYTEK Spark Cognitive Model V3.0 continue to improve, surpassing ChatGPT overall and GPT-4 in the medical field.

IFlytek also collaborated with Huawei Ascension Ecology to jointly release the "Feixing 1" large model computing platform, and launched larger parameter scale Starfire large model training for benchmarking GPT4. Liu Qingfeng said, "The curtain of the era of big models has just begun. The Starfire big model with larger parameter scales for benchmarking GPT-4 has officially started training, and will achieve benchmarking GPT-4 in the first half of 2024

Liu Qingfeng later stated in an interview with the media that the Feixing 1 is a completely domestic computing power jointly built by iFlytek and Huawei, which can train the GP4's ability base.

I also don't want everyone to think that we are always playing against Huawei's traffic and hot topics. iFlytek is a typical technological achievement of universities and is a student entrepreneurship. We have many dreams in core technology, but our commercialization ability is still weak, so we need to learn from Huawei. But it's not a miniature version of Huawei. In fact, I think there is still a big gap in our profitability and management ability compared to Huawei. But iFlytek also has dreams

At the time when iFlytek released Spark V3.0, the iFlytek learning machine caused trouble. Some parents claim to have found some distorted history and other content that goes against mainstream values in their children's iFlytek learning machines.

After the news was spread, iFlytek fell below its limit today. As of the closing bell, the company's stock price was 46.7 yuan and its market value was 108.142 billion yuan.

In response, Liu Qingfeng stated that the content had been removed from the shelves as soon as it was discovered, and penalties had been imposed on the partners responsible for ensuring content security. At the same time, the internal content review mechanism used for large models had been placed on the learning machine, striving to make the content review of the learning machine more strict.

Liu Qingfeng believes that there are behind the scenes behind this matter, some people are pushing it, and even eager to push it onto hot searches, which has an impact on the brand reputation of iFlytek. The purpose of doing this is to fear that iFlytek's big model is too good, and to fear that iFlytek will truly become a Numberone in China, and be far ahead. From a certain perspective, such suppression will make us more mature and also make us healthier

The following is part of the dialogue with Liu Qingfeng:

Question: Could you please introduce the significance of the release of the "Feixing No.1" platform for the development of domestic artificial intelligence industry?

Liu Qingfeng: Feixing No.1 is a fully domestic computing power platform jointly built by us and Huawei, which can train GP4 capabilities. This platform includes both enterprises like Huawei with domestic computing power storage network advantages and various high-quality computing libraries jointly developed by iFlytek and it.

It is precisely because of the viscosity of these operator libraries that our efficiency in training large models is higher than the original pure hardware. The same hardware, GPU, network, and bandwidth efficiency can be improved by twice or even more, which gives us the conditions and foundation to benchmark GPT-4 next year. Moreover, the Spark 3.0 we released today can be said to be the first in China to allow Huawei to not only use its own internal applications, but also, The first truly large model trained on a domestic computing power platform can benchmark GPT3.5.

We have done a lot of exploratory work in this area, and the existing achievements of many operator libraries will serve as a great reference for our domestic research institutions and enterprises in the future.

Some of the work we do with Huawei is publicly available to third-party customers, some is shared intellectual property rights between us, and some is unique to iFlytek. Therefore, we will form a comprehensive computing support for the entire industry ecosystem, which is very important in solving the bottleneck problem and also the world's second choice. I think it is a very important thing.

Of course, in the technological context of China and the United States, we did not want to announce many detailed parameters to the public. However, many people may be looking forward to this data, but I think we will wait until the first half of next year to see when we will release 4.0 and how its performance will be. I think this is the first thing.

The second thing, like China Sound Valley and our domestically produced large model, is how artificial intelligence responds to the current technological competition. I think the first one is indeed conducive to forming an industrial ecosystem, providing scenarios for artificial intelligence applications, and forming a comprehensive promotion after demonstration, which can form the entire linkage between government, industry, research, and capital. I think this is very meaningful.

Another is to cope with international competition. At the Budapest World Athletics Games, iFlytek won the bid and we competed with the world. In fact, we have multilingual knowledge of various technologies from around the world, many of which are also the best in the local area.

But we still pull out the best local companies, even if they are similar to or slightly inferior to us, but they have local service capabilities, local certification capabilities, and integrate them into our platform. We have collaborated with more than a dozen manufacturers from various European countries and our Asian countries to launch related platforms, not to eliminate everyone, but to cooperate with everyone and give everyone more space, This has formed our industry chain and ecology with me as the main force.

Our entire research background, application background, and international public opinion atmosphere are very good, so these companies are willing to fly to Hefei. I think we should take the initiative to go out and use a more open mind, but one of the prerequisites is that you must have solid hard technology, and you can make them feel that you are indeed strong, so they are willing to cooperate with you mainly.

Question: Previously, iFlytek disclosed that the 9100b generated through cooperation with Huawei can benchmark against the NVIDIA A100. May I ask which part of the company's chip cooperation is relatively more in terms of inference chips and training chips?

Liu Qingfeng: Inference chips are relatively simple, and some in China claim that their chips can make large models, mainly in terms of inference chips. The most difficult thing is training. As I mentioned earlier, with 1000 cards and three months, A100 can train a model worth over 100 billion yuan. Being able to do training is the true domestic example.

Apart from some of Huawei's own applications, we are the only one. Huawei's latest chips are all built on a fully autonomous and controllable platform for iFlytek, which is benchmarked against ChatGpt's Spark 3.0. The next step to benchmark Gpt4 is training. Training can be done, and reasoning is a very simple thing. Of course, we can also do reasoning.

Question: The iFlytek big model has been open to the public for some time. Can you disclose the approximate number of users? What are the considerations for commercialization?

Liu Qingfeng: The number of various users of iFlytek Spark has reached 12 million, and the growth is good. And the happiest thing is that there are many deep users who have used it one or two thousand times within two to three months, and have done a lot of things. It has indeed solved user problems.

Secondly, this data is relatively compared to other data sources in China, such as input methods. I didn't include it in the analysis, did I? You cannot include all of these over a billion users here. We are talking about applications directly related to large models, with a developer team of approximately 170000 or more. From the data we can see from various companies, our development team is the largest, and it is clearly more than the second place.

So out of the 170000+, I mentioned that over 100000 are enterprise level developer teams, over 70000 are individual development teams, of which around 18000 are small assistant teams, and nearly 20000 are API based teams. You can create various applications, so we have over 170000 developer teams with a total of 12 million individual users, which is the best among the industry's developer teams, at least based on publicly available data.

Another thing it can be of great help is that after the iFlytek Spark on May 6th, our over 600 AI capabilities have been quickly utilized by consumers. Therefore, since May, the total number of new real name certified developer teams has exceeded 1.3 million, with only 170000 or so directly related to the large model, which is roughly the same data. So the specific business model of Spark is definitely free for now. In fact, we can see many business models now, but I think we need to delve deeper into these scenarios first.

In addition, for enterprises, it involves various training programs. Depending on the needs of Spark, its value varies, including our code. I would prefer to promote a model similar to contract energy management. After you use it, we will jointly build a performance management platform to form an understanding of code volume and code quality. The additional benefits generated by sharing must be incremental value to the industry, We can work together in the middle to classify, and I think this can go far.

Question: With the United States strengthening export controls on high-end chips, the cooperation between iFlytek and Huawei is currently mainly focused on free training. Will there be any further services such as computing power leasing in the future?

Liu Qingfeng: We didn't do this because of the new policies introduced by the United States today. We have been doing this for a long time, and iFlytek's localization substitution started three years ago. Last October, we further collaborated with Huawei, and the Huawei special forces were already present before the official development of the Spark. Moreover, at iFlytek, we are polishing Huawei's Ascension 9100b series platforms, which meet both iFlytek and Huawei's needs. Therefore, there is a strategic alliance between the two sides, and Ren Zhengfei attaches great importance to it.

The emergence of this policy in the United States today was already expected by us, so there is nothing new. I just feel that this matter highlights the need for us to prepare early.

To Xunfei, I think we can develop at a normal pace, but for manufacturers who previously had a lot of luck, they may have to make adjustments quickly. In addition, providing computing power online is not the business of iFlytek, and we will not do it for the time being. Our main focus is on providing external services on my cloud platform through our own applications or our developer team after training. It is possible to do this, and we will not independently develop a computing power leasing system.

Question: What is the most difficult level behind the official release of the Feixin No.1 platform based on the Shengteng ecosystem for the cooperation between iFlytek and Huawei?

Liu Qingfeng: I also don't want everyone to think that we are constantly fooling around with Huawei's traffic and hot topics. I think we need to learn from him. Of course, Huawei found a particularly demanding scenario, using a powerful switch, and then succeeded in the communication field. We are envious of making mobile phones, and there are also many things worth learning from.

IFlytek is a typical technological achievement of universities and a student entrepreneurship. We have many dreams in core technology, but our commercialization ability is still weak, so we need to learn from them. But it's not a miniaturized version of Huawei. In fact, I think there is a significant gap between our profitability and management capabilities.

But we may all have dreams, and perhaps we have more dreams and ideas, so I don't think we need to compete. The biggest pressure on the flight platform we just mentioned is, of course, that we are doing this with our own mindset. If you were to use NVIDIA to accomplish something that could have been done in a month, you would have to use Huawei. There are various uncertainties in your calculations or porting, and it may take you three months to complete it. There is a war going on in the market there.

So if we were not using a domestic platform now, and I were using an already formed NVIDIA platform, my Spark effect today might be even better. But you have to take this step, and I also think Huawei is very proud of us. I am also proud of my own team, so his biggest difficulty at that time was that you were the most luxurious weapon used by others while running quickly, while you were fighting while honing your weapons and moving forward.

So, in fact, the test of mindset and the impact on one's future long-term strategy are both very important, and in this process, you need to invest more. You need to make money and invest on a large scale, both software and hardware. So, I think in the middle, do you truly believe in the future of artificial intelligence from the bottom of your heart, believe that you can step out and lead the industry, Otherwise, no one would have put in so much extra effort in this process than others.

Question: There were some inappropriate content in the content review of iFlytek AI's learning machine, and the company quickly responded by removing it from the shelves. Does iFlytek have a better prevention mechanism?

Liu Qingfeng: We have always attached great importance to information security in education. This time, it was indeed an accident, but we have already dealt with it internally. It happened on the 16th, and it was only a problem discovered by a parent. They are a cooperative partner and have a clear agreement with us to ensure content security. They have just been put into trial, and we are also reviewing various aspects.

The scope that everyone can adjust is one of a large number of articles. When this problem occurred, we quickly removed them from the shelves and punished our partners. We also took corresponding measures internally. At the same time, we quickly put the content review mechanism we used on the large model on the learning machine for stricter review.

In the past few years, our learning machine has become the most well-known learning machine in the industry. We have been under extreme pressure from the United States several times, and are rooted in China's own hard technology. We hope to help every child better teach according to their aptitude, develop their mental and physical health, and promote five education simultaneously.

So this is the main line, and there was an accident in the middle. It was indeed due to the excessive content mentioned by our partners. Before we finished the review process, the relevant personnel hurriedly asked everyone to try it out, and the problems caused have now been completely eliminated.

But from another perspective, I also think it's not easy to innovate in China. Today, when Spark 3.0 is released, if you see that we still have a lot of work to do on site, it's still quite good. However, just after we released it, we suddenly realized that the overwhelming public opinion was pushed behind the scenes.

Many media friends have told us that someone specifically feeds them and wants them to send it, and then the investment institution receives it, hoping to affect our stock price and our social image.

What is it after all? It's because we're afraid that our big model is doing too well, and we're afraid that iFlytek will truly become a Numberone in China, and be far ahead. So I think the more this kind of suppression, the more it inspires our fighting spirit, but at the same time, the more we see education. Indeed, we need to be more strict in ensuring the content, and make our sunshine and health education more practical.

From a certain perspective, such suppression will make us more mature and also make us healthier.

In fact, we have already been taken off the shelves and are still spreading. We have also reported to relevant departments, reported to our public security and internet information departments, and we are also collecting evidence, but there is a cycle in the process of collecting evidence.

Some people are pushing it, and even wish to push it to hot search. Of course, it has an impact on our brand reputation, but I think these are all waves in the process. What are our core fundamentals? Is it possible for us to truly create products that can make China's education better in the future? I think that is the fundamental.

Lei Di was founded by media person Lei Jianping. If reprinted, please specify the source.