THIS ARTICLE IS SYNTHESIZED BY THE SEMICONDUCTOR INDUSTRY (ID: ICVIEWS).
Many people are paying attention, who will win the battle of AI?
I remember when Chat GPT first appeared, some people speculated that it might be another "flash in the pan" of AI.
However, those who understand the cycle of technological explosions will know that AI will definitely not stop there. In the early hours of the day before yesterday, Open AI released its latest masterpiece - GPT-4o. With its breakthrough intelligent interaction capabilities, it has completely subverted our perception of AI voice assistants. This is not only a leap forward in technology, but also a big step in the history of human-computer interaction.
At the Google I/O conference that just ended last night, Google brought the new version of Gemini AI large model and other products to try to regain the initiative on the AI track, and "roared" AI 121 times in two hours.
AI is now at the center of the tech discourse.
Many people are paying attention, who will win the battle of AI? What should AI entrepreneurs do? What are the dangers and opportunities lurking under the technological explosion?
Today, at the "AI Creation Era - 2024 Jiazi Gravity X Technology Industry New Trend" conference, Zhu Xiaohu, Managing Partner of GSR Venture Capital, Fu Sheng, Chairman and CEO of Cheetah Mobile and Chairman of Orion Star, Li Zhifei, Founder and CEO of Mobvoi, Professor and Dean of the Department of Electronic Engineering of Tsinghua University, Outstanding Young Scholar of the National Natural Science Foundation of China, IEEE Fellow, Wang Yu, Founder of Wuwen Core Dome, and Zhang Jianzhong, Founder and CEO of Moore Threads. Five unique guests, including investors, industries, and experts, gathered together to discuss the future trend of AI in China around the current status and core competitiveness of artificial intelligence.
What are the overall trends in AI in 2024?
"More calm" is the key word given by Li Zhifei. In 2023, new concepts, new vocabulary, and new knowledge are emerging one after another. For example, in March, Li Zhifei, as an industry insider, had not heard of the term AI Agent, but in April and May, everyone was discussing Agent. At the same time, the industry continues to introduce new large models, and the new knowledge and discussion are very lively. This year, one of the exciting ones is Sora, and the other is a humanoid robot. The industry feels calmer.
"Volume" is the key word that Fu Sheng thinks. In 2023, AI training models will appear again and again, in which the indicators are higher than the other. Not long ago, GPT-4o was released, and many people thought it was very shocking. However, Open AI did not release GPT5 or GPT 4.5 to greatly improve the performance of large models, but began to roll up applications, engineering, and costs. Behind the failure to continuously improve the performance of large models, Fu Sheng believes that the update of the algorithm has encountered a bottleneck.
"Infinite possibilities" is the key word written by Wang Yu. He said that more and more young people in colleges and universities are beginning to try AI. In April, Tsinghua University established the School of Artificial Intelligence. Focusing on the two aspects of AI, AI core, including the evolution of algorithms, many companies are upgrading; In terms of AI plus, all walks of life have begun to move, trying to find the power of AI in the industry.
"Imagination" is the key word in Zhang Jianzhong. "Poverty limits the imagination," he said. "Most people are poor, and for start-ups, financing hundreds of millions or billions seems like a lot, but it can't support the construction of computing centers. Even if Open AI is very rich, its computing resources are also insufficient. Therefore, the key is to solve the key problem that the company cannot experiment and iterate because of poverty and lack of resources.
"Commercial quality" is the key word of Zhu Xiaohu as an investor. He believes that starting a business in China should not pursue technical problems excessively, because technology iteration is very fast. The key is whether the company can achieve commercialization and deliver the product to the customer. There are many "daily throwaway" AI products, but it is precisely because they do not meet the commercialization requirements that customers will not continue to use them after the first login. Therefore, it is important to meet the commercialization requirements in AI startups.
Future AI genre: technology belief or market belief school
Zhang Jianzhong believes that the development of AI needs to rely on industry and technology. The industry comes first, and any technological innovation cannot be commercialized if there is no industry. Pioneer industries may be where AI is the latest to revamp and accelerate innovation. AI can have good commercialization value in many industries.
For example, he said that the accuracy of face recognition in the early days was about 60%, and it can reach 90% accuracy after deep learning. Later, after commercialization, face recognition is even more accurate than human recognition, and the twins that cannot be distinguished by the human eye can be done by AI.
At present, it is difficult to carry out R&D in China without considering commercial returns, so the industry must take the lead. However, there are innate conditions for the industry to take the lead, and cross-field and cross-professional talents are needed. If AI is to develop in a certain field, it needs people who understand both professional knowledge and AI knowledge. Behind the industry's first, it is still talent to drive the development of the industry.
"Long-term technology, short-term market" is the answer given by Fu Sheng. Let's start with technology, which is changing the productivity, structure, and efficiency of society. The impact of AI on the industry model definitely exists, so Fu Sheng is still optimistic about technology. But in the short term, it's important to note that technology is not a linear process. Technology is often a breakthrough brings a wave of market application, technology is phased, when it really lands, don't believe that technology is the same every year as last year.
In the 80s, when the cost of robotics was higher than the cost of manual labor, the automation industry was transformed into flexible production with manual assembly, which greatly improved production efficiency. Therefore, entrepreneurs must closely integrate with the market when looking at things for 1-2 years.
In addition, Fu Sheng believes that under the condition of limited resources, it is more able to innovate according to the needs of the market. The meaning of the "resource trap" is that sometimes excessive superstition in the explosive power brought by the technology itself, regardless of the cost of investment, will lead to the bursting of the industry bubble. Therefore, AI entrepreneurs should pay more attention to the application and how the market can get returns from the market.
Li Zhifei believes that technology and pragmatism can enable enterprises to develop better in the trend of AI. For technology entrepreneurs, belief in technology is an instinct hidden in the genes, but many technology entrepreneurs in the early days will have a misunderstanding: it is too unpragmatic. Avoiding real problems under the banner of ideals, such as making a lot of products, always thinking that cutting-edge technology is used, and no one uses it is the user's problem. In fact, it is more important for entrepreneurs to pay more attention to business, face more competition, and understand more about whether technology is a user need.
Wang Yu also gave two paths, technology and business. In universities, the obsession with technology is natural. He shared more about business. The industry is always waiting for technological breakthroughs and looking forward to a batch of commercialization advances. Therefore, "transformation of scientific research achievements in universities" has attracted considerable attention in the industry, but "social information input" has been rarely mentioned. The key is whether university professors and researchers can realize what is lacking in the world at an early stage. Research out a systematic technical solution to solve the problem in theory. Now from university and business communication, and back to technology, the closed loop of this route is not fast enough.
In a company, it is not easy to do research for 1-3 years. However, for universities and research institutions, they can do research for 5-10 years, so if some technical problems are put into universities to form a better ecology, the industry still needs to work together.
At this point, what worries Huang the most?
Wang Yu said that for large models, the emergence of the ecosystem makes it possible to need 2,000 operators now, but now only 200 operators may be needed, so the demand for GPUs will be affected. It turns out that the performance of the GPU is good, and everyone is talking about the need for a GPU. Slowly, the market can offer only hardware without using CUDA. In the way of plug-ins, so what chips are used at the bottom may not necessarily have such clear requirements.
Li Zhifei said that China and the United States have many differences, from entrepreneurship, choice to path. For NVIDIA, 80% of the largest AI revenue still comes from various giants, so the biggest worry should be that the giants do it themselves. For example, Mata, Microsoft, and Google are all developing their own chips. When all the applications and prices of the model are uniformly converged, large companies will make a hardware with similar performance, or even cheaper, and can control themselves. In essence, whether the investment in the large model can match the needs of users determines whether the algorithm fee is sustainable.
Zhang said Huang should be worried about what will be the next application of AI.
A chicken-and-egg or egg-and-egg question: is it better to accelerate Transform in the GPU, or is it only possible to have Transform with the GPU? Ecology is important, and everything needs ecology. He believes that with GPUs, Transform can develop better; At the same time, with Transform, it is also promoting the improvement of GPU architecture. Now that Transform has gone through various changes, what's next.
In the process of accelerating large model languages, every small algorithm change is a new improvement in the technical architecture. The communication requirements and scaling requirements are the best places to improve the GPU. So, if someone asks if GPUs can adapt to the next generation of future technology? The answer is: OF course. In addition, Zhang Jianzhong said that the iteration speed of GPU products is not slower than that of large models, but the iteration of GPU architecture is also accelerating.
Huang may believe that GPUs will never fail and will not be replaced by others, but Fu Sheng denies this. He said that the entire AI industry is moving in the direction of clipping parameters. Google released the first large model with a local mobile phone, and as a demonstration, Google identified the scam call on the spot. That is, the anti-fraud APP is installed on the mobile phone, which does not need to consume any server-side GPU. If this road is 10%, then 10% of the computing power demand will be reduced. At present, Google and Apple are doing small parameter models.
On the other hand, Open AI is also streamlining the architecture of Chat GPT 4 and optimizing inference to save more GPUs. Therefore, when the arms race has come to the present, a huge trend is to use less computing power, lower cost, and provide better services.
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.