The development of the artificial intelligence industry is advancing by leaps and bounds, and the large-scale model technology has exploded in an all-round way, driving the demand for computing power to soar. As a fortress of historical civilization and the hinterland of transportation in the Central Plains, Henan bears the function of a computing hub in central China.
In 2019, APUS established its second global headquarters in Zhengzhou, Henan Province, and then successively established the "Global Digital Gene Bank" and "Intelligent Computing Center" to help Henan build a highland of data and computing power. In August 2023, the APUS model will be launched in Henan, providing an innovative foundation for the development of Henan's AI industry and helping the transformation and upgrading of Henan's thousands of industries.
On December 2, Li Tao, Chairman and CEO of APUS, was invited to attend the "Fifth High-end Training Course of the Central Plains Private Entrepreneur Training Program" of the School of Economics and Management of Tsinghua University, sharing the theme of industrial application and value creation of artificial intelligence.
In this lecture, Li Tao analyzed the evolution of artificial intelligence and AI large model technology, the opportunities and risks of the current development of the global artificial intelligence industry, and made a detailed interpretation from the aspects of the application of large models, the value of large model industry and future intelligent governance.
The first thing to be clear about is that AI is not the same as digital technology.
In the process of transformation, many traditional industries tend to see artificial intelligence, digital economy and digital technology as the same concept, but in fact, there are big differences between them. Artificial intelligence is a technical means to process information, solve problems, and achieve automatic decision-making and intelligent control by simulating human intelligence and thinking; The digital economy refers to an economic form based on digital technology, which assists the digital transformation of society and enterprises, and promotes the development of enterprises and social progress.
However, China's artificial intelligence development does lag behind the world's leading technologies, and it starts later than others in chip research and development and algorithm innovation. In addition, there are not enough chips to support the evolution of computing power, and the "acceleration" of development will inevitably slow down.
However, I believe that we must objectively recognize the current gap and face up to the gap where we are lagging behind, so that we can know what direction we should strive for, so that we can really identify the shortcomings and continue to catch up.
If I can't compete with others in computing power, how can I keep my AI model iterative and evolve when the computing power is less than others? For APUS, we would choose to train long text by updating the data quickly in real time through incremental pre-training.
In fact, APUS did just that, and the APUS model trained 128K text very early.
As a global business, APUS is also facing various challenges such as international barriers and cultural shock. From 2018 to 2022, we have been exploring the application of AI to 200+ products around the world, but from a business perspective, the actual results have not been great. Why?
For the past 4 to 5 years, APUS has used small models. Small models do not have scale effect and replication effect, and the marginal cost of using small models for application is constantly increasing. The real role that algorithms play in building a business is very small. In different industries and enterprises, building business applications with small models almost requires "starting all over again".
In contrast, the one-time fixed cost of a large model is high; But when various industries and scenarios begin to use it, its marginal cost begins to slowly approach zero, which is also the time when the value of the large model base really begins to appear.
In November last year, ChatGPT was released, and APUS quickly followed up the large model as soon as possible, and in December, it decided to shift all the technical routes to general artificial intelligence, began to reserve talents in the field of computing power and AI, and invested in the research and development of its own large model; In April 2023, the entire company is All in AI, and launched the 100-billion-level multi-modal large model - APUS large model, which is self-developed in China.
In the first half of the year, everyone was talking about artificial intelligence, everyone was talking about large models and computing power, in the second half of the year, everyone was talking about artificial intelligence applications and agents, and now everyone was talking about the industrial application of artificial intelligence and what is the application value of large models in the industry.
I believe that one of the great values of large models is to use artificial intelligence to help the industry think and calculate the overall situation.
In the past, manual decision-making lacked global thinking and the ability to do global calculations. Taking agriculture as an example, in the early stage of production, it is difficult for us to judge the global sowing amount, fertilizer amount, etc., judge the growth of crops, harvesting and transportation, including market transactions, to judge the input-output ratio; At the same time, it is also difficult to forecast and calculate variables such as future weather that affect output. But today, in the era of large models of artificial intelligence, these things can be done.
For example, in the field of "autonomous driving", which we are hotly discussing today, the artificial intelligence currently used is still a "small model", so L35 of the smart cars can only solve the problem of "artificial assisted driving"; If a large model is added to the autonomous driving technology, it may solve the problem of full autonomous driving, and only then can the intelligent car be truly pushed to the level of L4 or L5.
The "large" model needs massive data to be "fed" for its Xi, and it can also quickly calculate the results and generate many new things, including new applications such as intelligent driving and intelligent assistants. Slowly, we will find that in many fields, including many traditional fields that are difficult to digitize, artificial intelligence is needed to stimulate industrial value.
The industry generally believes that artificial intelligence is the fourth industrial revolution, so in the fourth industrial revolution, there is a very important supporting force is the "large model". I often say that the large model is the "underlying operating system" of artificial intelligence, and the underlying capabilities used in various fields in the future will be provided by the large model.
However, with the development of large models, especially the latest release of OpenAI on November 7, many practitioners have also seen potential global AI ecological risks.
So let's first take a look at what openai's upgrade has released, and what challenges and risks the entire artificial intelligence ecosystem may be facing from this upgrade.
In fact, the core released three things: the first is the release of GPT Turbo; The second is GPT Store, and it unifies the entire GPT entrance; The third releases GPTS.
GPT Turbo is essentially OpenAI's open empowerment of the world's leading artificial intelligence technology and provides a base for large models for developers around the world. This seems to be a good thing from the perspective of artificial intelligence developers, but many people ignore that this world-leading large model base can create a "developer monopoly", on which OpenAI will build its own "unique" artificial intelligence ecosystem.
The essence of GPT Store release is also to build an ecosystem as the core purpose, through the application market, developers' products and technologies can be better used by more users. But this time, the GPT Store is particularly different – it's not just a store, it's a unified entry point for all users. The developer's application is more in the form of plug-ins, which are integrated into GPT, and the developer's personal brand is obscured; Compared with the Apple Store and Google Store, the commission rate charged by GPT Store to platform users may only be high.
GPTS is actually an agent, it will be like a digital assistant and a digital doppelganger, and it will solve problems according to your needs very secretly. However, through the agent, OpenAI can even easily obtain all kinds of private and confidential data of developers and users, thus forming a "triple monopoly" in technology, traffic and data, and truly achieving the business purpose of building an "AI empire".
As a practitioner of Chinese artificial intelligence, I also put forward some ideas on intelligent risk and governance.
First of all, large economies should have their own AI general model base. For example, large independent economies such as the United States, China, and even India, the European Union, etc., should have their own general model base.
Second, every country, or even every independent civilization, should have a set of AI frameworks that align with the values of its own country or civilization, and use such a framework to accommodate and constrain large models.
In addition, it is necessary to build a healthy artificial intelligence industry ecology, so that every link in the artificial intelligence industry chain can have its own living space, so that every link and every chain can develop its own potential.
At the beginning of APUS's investment in the research and development of large models, it was established that it would "customize AI large models for China". After the tempering and practice of the market, the APUS model has also clarified its strategic direction of "growth": adhere to intelligence for good, constrain the model with a red framework, align values, and integrate the application of the model with value creation.
The development of technology requires the joint construction and correction of every link under the wave. In the process of developing artificial intelligence and large model bases, we should follow the tenet of "intelligence for good", and establish a positive and binding framework based on this tenet to truly guide the AI industry to be born for good, so as to better serve enterprise production, people's lives, and social and economic development, and better create AI application value.