Qiming Venture Partners: GPT4 will be born in China by 2024 at the latest, and long text is the key to large language models

Author: Zhao Ying

Image source: Generated by Unbounded AI tool

On Friday, July 7th, the second day of the World Artificial Intelligence Conference, Zhou Zhifeng, a partner of Qiming Venture Partners, shared about the AI model and the development trend of the entire industry.

Qiming Venture Capital is one of the earliest venture capital institutions in China that pays attention to and is most active in the field of AI. Its investment companies are located throughout the entire AI ecological chain. The upstream, midstream, and downstream actually have designs, such as chips, basic software, and Like large models and downstream vertical applications, etc.

The first principle of large models is to compress digital knowledge, information and data

Zhou Zhifeng emphasized that large-scale computing power and large-scale data are very important for the future development of AI. **

**The first principle of large-scale training is to compress all digitized knowledge information and data in the human world on a large scale. ** You can see from this first principle that large-scale computing power and large-scale data are very important to the future development of AI, and indeed large-scale computing power and large-scale data are the chips of human beings in the past 40 years These technologies are continuously accumulated in various fields of Internet big data, and then promote the explosion of AI. Large-scale pre-training learning generation, the underlying logic of the whole trend is actually very clear, and it will continue to develop at a high speed in the future. Deep learning is called AI1.0 by the venture capital industry, mainly because it is aimed at a specific small task. A small model trained from a specific small data set, and it is mainly pattern recognition, such as face recognition and speech recognition. Today's large-scale speech field learning is more commonly known as AI2.0, mainly because it is general The large model trained by the data has the ability to generate decision-making.

The wave of AI2.0 will rebuild the entire industrial structure

Zhou Zhifeng further pointed out that the new wave of AI2.0 will rebuild the entire industrial structure. The entire ecological architecture will be divided into three layers, the bottom layer is the infrastructure layer, the right side provides computing power, such as aws volcano engine, Alibaba Cloud, etc. are providing this kind of computing center platform, the right side is the tool chain, and it is mainly for large models Optimize training, reasoning, and deployment.

The middle layer is the most important. The first layer is the model layer. The model layer also has several modes. The first one is to provide the basic model base model, and then export the model to the outside world, and customize the model through the API.

Then there is another type of self-built large model. After making his own large model, he will optimize it for a specific industry-specific scenario, and then provide an end-to-end solution from model to application.

In the third layer of application layer, the one is to directly implement vertical application through self-built models. Maybe 80~90% of the companies on the left are those who use the capabilities of third-party models to build familiar scenarios or industries An application, **This is the three-tier architecture we understand. Indeed, this new architecture has also brought about great changes in the way the entire world builds products. ** The left side is actually the past few decades. Whether it is a car or an Internet social software, it is actually such a structure, that is, the product manager obtains the needs from the user, and the developer obtains the design from the product manager, and then the user uses it. Developed products to use.

In the past, Tencent did a good job, Alibaba did a good job, and any company did a good job. In a sense, it said that he turned the flywheel most effectively and was able to continuously iterate the flywheel.

There are two types of future enterprises: +AI, AI+

When it comes to industrial development, Zhou Zhifeng pointed out that future enterprises will be divided into two types: +AI and AI+.

In the future, it will be more about putting the capabilities of the new generation of AI into the workflow, **it is actually an enhancement of the old scene. ** There is another category where he will use this ability to build a new product, which is actually an application of the so-called AI native, which I call the reshaping of old scenes, or the creation of new scenes. **

At present, there are still a few very leading AI companies, mainly because AI has not realized such a real situation of empowering all industries, and its industrialization is not satisfactory. Zhou Zhifeng said:

ChatGTP has rekindled the wave of AI2.0. You can see that the entire global financing amount has undergone tremendous development. This time it will be a bubble that lasts for two years and then goes down, or will it really move forward and develop? To a general artificial intelligence, I think this is a very worthwhile question.

The large model will definitely become more and more powerful. The CEO of OpenAI also said that they may get involved in products like Microsoft and make an office productivity product. They are also yearning for a broader space. Are we Can you find a golden channel to start your own business, and finally go to a vast world?

Maybe the reality is like this, the road we have to go may be a death canyon, both sides will be constantly squeezed, their technology is changing dynamically, how can we go through this death canyon, I think this is entrepreneurship Sometimes, we have to use our rational thinking and work hard to think.

Every wave of technology will definitely give birth to new kings and new great companies.

Prospects for Ten AI Development Trends

Looking forward to the future, Qiming Venture Partners teamed up with Unfinished Research to jointly release a blockbuster report "Generative AI" | State of Generative AI 2023 and summarized ten development trends:

First, based on the information we have seen about companies invested by Qiming, we know that in 2024 or even earlier, China will definitely have a multilingual model comparable to GTP4. We have clearly seen the progress of several companies in this direction .

Second, long context will definitely be a key point in the development of the next generation of large-scale language models. We will see that you can actually have a contextual communication with a large model for several days and months, instead of just talking for 3 rounds and 5 rounds today.

Third, we think there are several ways to make a vertical large model. In fact, we have summarized 5 methods.

Fourth, we believe that although today's stabele-diffusion is a very good diffusion model architecture, we believe that whether it is stability or other companies, a new model will emerge in the next two years.

Fifth, the text-to-image model will be more controllable in the future, and we have seen that many top teams in the industry have made some scientific breakthroughs in this area.

**Sixth, the third and fourth quarters of this year will be a breakthrough point for music generation. **We believe that there will be a major breakthrough in the generation of v6 and 3D next year.

Seventh, there will be a major development in the intelligence of how to combine large language models with real physical space control robots and humanoid robots.

Eighth, although transform is now the mainstream, as I said, the ultimate goal of AI is to use the best method to compress the digital information of all human beings. **transform is definitely not the end, and more advanced architectures will appear. ** Ninth, from a business perspective, we believe that in the next three years, model capabilities and applications cannot be decoupled, and truly disruptive applications must come from those companies that have mastered the core underlying model R&D capabilities, rather than a pure application companies, I mean disruptive applications, because we don't see this possibility of decoupling in three years.

Tenth, it is still a golden period that can produce platform companies. We believe that some start-up companies established in the next three years may become a company with a market value of 100 billion trillion.