On April 16, the 2024 Uniview Partner Conference was held in Wuzhen, the millennium. At the conference, Mr. Zhang Pengguo, President of Uniview Technology, said: Large model technology will definitely bring many new possibilities to this era and the industry in which we live, and is the biggest technical variable in the entire AIoT industry.
The following is the full text of the speech delivered by Pengguo Zhang, President of Uniview, at the 2024 Partner Conference:
Dear partners, good morning!
Recently, I chatted with several channel bosses, and everyone generally has technical and industrial anxiety about large models, especially Musk, Huang Jenxun, and Ultraman's wheel war, and the invisible pressure is very great: Wensheng video for a while, Wensheng song for a while, and Wensheng 3D again in the second half of the year. It's over, our company is going to be finished, we're going to lay off employees... Every night before going to bed, I feel that there are thousands of roads, and when I wake up in the morning, I still continue to walk the old way.
This time, we're going to focus on exploring our path.
There is no doubt that large models are the most dazzling technology in the field of artificial intelligence and represent the highest achievements of human beings in the field of information processing. It is also the most exciting technological breakthrough in the science and technology industry in more than 20 years since 3G wireless communication technology.
Regarding whether large-scale model technology is a technological revolution or a technological upgrade, whether it is the "Galileo era" or "the eve of the Newton era"?
Closed-source and technocratic believers are more inclined to this is a technological revolution, because the large model runs through the Turing machine, has a breakthrough technical framework, and also has amazing application effects. For example, Dora of Wensheng Video and Suno of Wensheng Song;
In the fifties of the last century, almost everyone believed that as long as all the dictionaries were entered into the computer, all natural language understanding and translation could be done. However, this fact is too difficult, and today, nearly 70 years later, natural language processing is like a point, and the computing power has increased by 10 trillion times.
The fact is that there is a gradual process of technological innovation, and large-scale model technology is no exception. From the perspective of our industry, everything is just beginning.
What is certain is that large model technology will definitely bring many new possibilities to this era and the industry in which we live, and it is the biggest technical variable in the entire AIoT industry.
The general model is the battlefield of the world's top companies, which requires hundreds of billions of dollars. And these general large models have their own merits, and you will catch up with me, and they will be open source successively, which is beyond doubt.
On the basis of the open-source general large model, the training data, subdivision scenario requirements, management mode and operation mode requirements of various subdivided industries can be superimposed, and the industry large model can be built. Today, I would like to tell you that Uniview has done it and started to iterate.
At the channel conference last year, we released the Sycamore industry model. In the past year, we have continuously upgraded the technology of Sycamore, and there have been many application cases. At present, although the publicity of the application effect of the general large model (in the Internet scene on the C-side) is more eye-catching, the commercialization prospect of the industry large model is more certain and clearer.
How can the industry model be monetized with real money, and how can it be better commercialized? This is a question we have been thinking about.
The industry's point of view is two directions: large-scale equipment and large-scale equipment. The former is similar to +AI, and the latter is AI+.
In the last wave of deep learning boom, there was also a debate in the industry about +AI and AI+. In terms of practical results (at least in the 2B/2G field), +AI has won by a wide margin, and there are several definitive conclusions:
First, the business logic of selling algorithms (including large models) is definitely not compatible, and there are countless big pitfalls;
Second, it is very difficult for algorithm/software companies to transform into hardware products, and the underlying system capabilities are very different, and the challenges are too great.
That is, the road to equipment/tooling of large models is very bumpy. Yesterday on the way to Tongxiang, I heard that the price of a large model has fallen to less than 200,000, and at this time last year, it was still several million. Cliff-like impairment curves. Why? Because the software ecology and the value of software have not been well solved, and they are getting worse and worse. Don't expand on it, see Chen Guo George's article "China's Enterprise Software Industry Has Reached the Most Dangerous Time".
In the era of large models, there is also the problem of route selection for commercialization. From the perspective of the entire industry, tools/equipment products may be the first commercial landing point of large model technology, and office software/APP such as mobile phones/PCs, office/adobe, etc., are rapidly integrating large model technology and becoming big winners.
Uniview's AIoT products are both equipment and tools. Moreover, Uniview has a mature product system capability, so for Uniview, the route for the commercialization of large model technology must be: equipment large model, referred to as equipment.
Scene fragmentation is a feature of AI technology, and large model technology does not improve this, because the exponential rise of parameters has greatly amplified fragmentation, which is a big pity. Challenge the company's channel marketing capabilities/brand marketing capabilities.
In addition, the difficulty of landing subdivision scenarios is not only the difficulty in the technical dimension, but also the difficulty in the closed-loop of business logic.
Too many parameters and too many subdivision scenarios are not effective requirements, and it is not easy to find a balance between user value, TCO cost, and the price that users are willing to pay. We can only prioritize the subdivision scenarios that are most likely to find a balance to meet the real needs of users.
Large model technology is the biggest technical variable in the AIoT industry, which will definitely bring us new market opportunities. A big opportunity is the update and iteration of a large number of online products.
With the help of industry large-scale model technology, edge and end products are expected to achieve better application results at a lower cost in some segmentation scenarios (which still need to be explored).
Of course, technology upgrades also put forward new requirements for product and solution architectures, such as requiring that chips at the cloud edge support the Transformer architecture.
Yes, the upgrading of chips means the most intuitive need: all products will be redone.
Technological progress will give rise to more non-security scenarios, such as AI sports and physical testing, at the site of this summit, you can see a lot of physical testing screen products, welcome to taste and give opinions and suggestions.
Why AI in the cultural, educational, and sports industries? First, the user base is huge, and the market space is large enough. Second, political correctness is in line with the direction of social progress (rescuing the elderly from overtreated hospitals, dragging middle-aged people out of the sea of Wenshan Hui, and fishing children out of the sea of questions). Third, there are not so demanding accuracy requirements (it doesn't have to be 100%), leaving room for technological advancement (including the illusion inherent in large models).
Uniview has always been positioned as a product and solution provider, and our core strategy can be summed up in one sentence, "software hardware, hardware equipment, equipment serialization". Algorithms are also software. We firmly believe that this strategy is in line with the development trend of the industry, and is highly compatible with Uniview's existing capabilities and capabilities.
We have been following this path for so many years, we have enough technical accumulation, and we also have absolute organizational confidence. This is in the genes of Uniview, strong and reliable, experienced in a hundred battles, no one is afraid. Please rest assured, and look forward to your continued support for Uniview.
The improvement of the efficiency of the use of information and energy has always been the two main lines to promote the progress of human society. In the era of large models, the relationship between the two has become more and more intimate, like a pair of double helixes.
The existing large-scale model technology follows the idea of "heavenly parameters + unlimited computing power" = "brute force cracking with miracles", and some people ridicule it as "a silly boy sleeping on a cool kang, all by fire". Such training and inference processes require a lot of computing resources and consume a lot of power.
This is different from the way the human brain thinks about problems, the human brain is too low in power consumption, according to the Buddhist point of view, there are 36,000 images behind a thought, and the power consumption is only a few watts? Because of this, the reasoning school does not recognize the statistics school, but the statistics school has won for the time being, and we will not expand on it here. The carbon-based base, which has evolved for hundreds of millions of years, is still very good and will not be defeated so easily.
High energy consumption will pose great challenges to the supply capacity of the power grid, the operating costs of enterprises, and the global climate and environment.
There are two ways to solve the problem, one is to reduce the energy consumption level of large model training and inference, and the other is to change the energy supply structure, and use more clean energy, such as wind, solar and hydropower generation, most of which are scattered, so it is necessary to build a distributed energy system. Distributed energy resources are also an extremely important measure for national energy security, which you can't experience in a peaceful country, but I think the countries of Northwest Europe have experienced it.
Based on this understanding, we decisively entered the new energy business field, first cutting into the market from charging pile products and home energy storage systems, and iteratively making progress together with AIGC technology.
This sentence was mentioned for the first time last year in order to promote the transformation of the entire company to real channel marketing, and to promote the company's various organizations to treat channels as customers and effectively serve channels.
Today, we understand this sentence in more depth.
I paid attention to a sentence by Miao Wei, former minister of the Ministry of Industry and Information Technology, which is particularly insightful: we should learn to use the method of playing Go in China to further plan, and how to empower the manufacturing industry and empower various specific fields through our large model under the condition of relatively backward technology.
This is the most objective and rational view I have seen in recent times, and it is neither arrogant nor arrogant, and it is very wise.
Yes, Uniview will empower the manufacturing industry through a large model, and then empower the channel, so that the channel can empower various subdivisions, so as to complete the business closed loop of each subdivision scenario.
Therefore, we decided to fully open our capabilities to the channel, from R&D, manufacturing to quality management, brand marketing, as long as the channel wants, we must quickly empower and give quickly.
"Accurately identify the environment, fully meet the challenges, focus on the core proposition, and actively promote change", several changes in Uniview over the years, especially the major changes promoted by myself, are all closely focused on these four sentences.
Determine our judgment on environmental changes, meet our challenges and propositions, take the initiative to seek change, take the initiative to revolution, and believe that we will be able to achieve victory after victory.
Please always remember these three sentences:
1. The big model is just a technological upgrade, not a revolution
2. Hardware-based software, hardware-based equipment, and serialized equipment
3. All equipment will be reworked in the future
Thank you for your support, and I wish you good health, all the best, and a long-lasting foundation!
Zhang Pengguo, April 16, 2024, Wuzhen, Tongxiang.