Cheung Kong Graduate School of Business2024-01-29 18:31
In 2024, the ship of AI innovation will enter its second year. From the continuous development of large models, to the extension of AI technology in different industries, to the battle between open source and closed-source strategies, every step of AI development is depicting the outline of future trends, and every enterprise is competing for a ticket to the future.
In the recent Cheung Kong Graduate School of Business EMBA "Sincerity Forum", Zhou Hongyi, founder of 360 Group and alumnus of Cheung Kong Graduate School of Business CEO 8, delivered a 10,000-word speech on "The Leap Forward of Enterprises in the AI+ Era", sharing his insights from the aspects of AI's impact on the industry pattern and future development trends, how large models can be implemented in enterprises, and the judgment of the top ten development trends of large models in 2024, and had an in-depth dialogue with Sun Tianshu, Distinguished Dean of Technology and Operations at Cheung Kong Graduate School of Business.
Zhou Hongyi pointed out that enterprises should put AI beliefs into action and dare to reform their own lives. The big model is the most powerful tool ever, there will always be someone to develop, and not to develop is the biggest insecurity.
Today's article shares with you the exclusive views of alumnus Zhou Hongyi, hoping to inspire you.
Share |Zhou Hongyi.
* |Cheung Kong Graduate School of Business EMBA (ID: CKemba201314).
Zhou Hongyi. Founder of 360 Group.
Alumnus of Cheung Kong Graduate School of Business CEO 8.
Perspectives Express. 01.Large models are more likely to be as ubiquitous as PCs. In the future, every person, enterprise, and individual may have their own large model, which will lead to a more distributed and privatized large model.
02.The large model learns the knowledge, laws and principles behind the data, rather than rote memorization of data and information, which is the embodiment of true intelligence. Large models will become the standard configuration of enterprise digital systems, reshaping hundreds of industries, stimulating industry innovation, and solving industry development problems.
03.We should actively embrace large-scale model technology and think about how to integrate it with our industry to create more possibilities and seize opportunities to achieve innovation. This will be a profound industry change and the key to seizing the opportunity of the fourth technological revolution and promoting social progress.
04.Only a large vertical model trained by combining "dark knowledge" can truly solve enterprise problems. Enterprises are the first driving force for the application of large models, and the combination of large models and businesses is ultimately led by the business in order to truly serve the business.
05.In order to achieve the smooth implementation of large models in enterprises, we recommend adopting the strategy of "small incision, large depth". AI transformation business is not a grand narrative, but actually a collection of several business scenarios, so it is necessary to decompose the business, find small incisions in the business, and select the business links that match the mature capabilities of the large model as the entry point.
06.On the one hand, Internet giants will continue to promote the research and development of large models in the cloud and explore new heights of human intelligence; On the other hand, the enterprise-level model will focus more on cost-effectiveness and practicality, streamlining large models into efficient, low-cost small models through techniques such as "distillation", while maintaining their performance advantages.
How did the big model come about?
In the past, what we called AI could often only complete some single and limited tasks, such as face recognition, voice assistants, etc. These technologies may be able to play a certain role in some specific scenarios, but there is still a big gap between them and the real sense of intelligence. Therefore, it is not unreasonable for some people to jokingly call it "artificial retardation".
However, with the passage of time, especially in 2023, the field of AI has ushered in a major breakthrough. This year, we've seen rapid advances in technology, and AI is starting to show more powerful capabilities. It is no longer limited to a single task or domain, but can use general algorithms to complete tasks that used to require many specialized tasks to complete. This is the prototype of what we call Artificial General Intelligence (AGI).
A key factor in the successful emergence of large models.
The successful emergence of this large-scale model can be described as a "miracle of great force", similar to the new paradigm of "violent aesthetics" advocated by John Woo. The key elements are as follows:
1. The cornerstone of the algorithm has already been laid: neural networks, deep learning, and the Transformer algorithm launched by Google have all provided the necessary algorithm foundation for the emergence of large models.
2. Historical accumulation of computing power: With the development of the Internet and the progress of GPU technology, the computing power accumulated by mankind has reached a new stage. In the past, we may have had mature algorithms and ideas, but we couldn't implement them due to computing power limitations. Nowadays, the improvement of computing power provides an important guarantee for the success of large models.
3. Integration of massive knowledge and data: The popularization of the Internet has created a large amount of knowledge and data. OpenAI and other institutions have skillfully integrated this knowledge into the training of Transformer large models, breaking the limitations of previous pragmatism and leading the new development of large models.
Unlike previous tech giants that focused on pragmatism, OpenAI's investment and vision are more forward-looking and disruptive. Their dedication to integrating all human knowledge into the model has led to a qualitative change that has led to epoch-making achievements such as ChatGPT.
To be clear, large models are not limited to applications in the field of natural language processing. Its strong influence and potential are gradually penetrating into various fields. By overcoming the core challenge of language understanding, large models enable AI to communicate with humans more naturally, and thus understand and depict the world's knowledge more deeply. This understanding not only improves the computer's language mastery, but also gives it a comprehensive understanding of the world structure, which provides strong support for various complex tasks.
Taking a scene in daily life as an example, the large model is able to understand and process natural language instructions such as "go to Cheung Kong Graduate School of Business in Oriental Plaza to communicate". This is due to the large model's in-depth understanding and learning ability of vocabulary and context. This is also one of the significant differences between large models and traditional search engines. Large models focus more on knowledge understanding and reasoning, rather than simple keyword matching and web page indexing.
During the training process, the large model learns a large amount of data to grasp the capabilities and paradigms behind it. This capability allows large models to take advantage of similar problems.
3. Be flexible. At the same time, the large model will also compress and optimize the data to improve processing efficiency. However, it is important to note that this compression is not simply a reduction in the amount of data, but is achieved by extracting key information and patterns.
Although large models may have some deficiencies in knowledge details, their powerful processing power and generalization capabilities make it have a wide range of application prospects in the fields of autonomous driving and robotics.
The training process of a large model.
The successful application of large models can be boiled down to four key steps.
The first is pre-training, which is similar to how we are constantly learning new knowledge and gaining experience. By absorbing massive amounts of data and information, large models gradually build up their knowledge and understanding of the world. As the old Chinese saying goes, "Read a book a hundred times, and its meaning will be seen by itself", when the large model is exposed to enough data, it will be able to integrate this knowledge and form its own understanding and judgment.
This is followed by fine-tuning, a step designed to teach the model how to better communicate with people and answer questions. Through targeted training and adjustment, large models can gradually learn to express their ideas and opinions in a language and way that is easy for humans to understand.
The third step is alignment, which is crucial. Because knowledge itself is not good or bad, but the ideas and actions of the big model must follow human morals and values. Through the alignment process, we can ensure that the large model does not deviate from the right track when answering questions, providing recommendations, and does not behave in a way that violates human ethics and morals.
The final step is application, that is, the application of large models to real-world scenarios to interact and communicate with humans. In this step, humans can guide the behavior and thinking of the large model by asking questions, providing feedback, etc. At the same time, large models are able to continuously optimize their performance according to human needs and preferences.
The intelligent performance of large models is not unfounded. Its success is inseparable from the support of massive data, the application of advanced algorithms, and continuous optimization and adjustment. In contrast, some other technologies such as web3, although high hopes have been placed, have yet to see their real explosive growth and substantial changes to life and work. However, in just one year in 2023, artificial intelligence has shown amazing development speed and application potential. This fully shows that large models, as an important representative of artificial intelligence technology, are leading a new round of scientific and technological revolution and innovation.
How to put the big model into perspective?
In the wave of artificial intelligence, large models have gradually emerged with their excellent comprehensive capabilities, not only showing their strength beyond human beings in many fields, but also being like a fish in water in various professional examinations, which is breathtaking. However, while applauding the big model, we also need to rationally examine its limitations and potential.
First of all, despite the excellent performance of the large model, we cannot overstate its capabilities. They may still have shortcomings in understanding in some areas, such as mathematics, physics, and other scientific problems. In addition, "fabrication" is a significant problem with large models, which sometimes make up information that doesn't exist. This phenomenon can have serious consequences in some key areas, such as medicine. Therefore, we need to be vigilant when applying large models and avoid blind dependency.
However, this does not mean that we should underestimate the future potential of large models. The rapid development of technology has proven that once an inflection point is reached, the rate of progress increases exponentially. As a new star in the field of artificial intelligence, large models are showing an unprecedented development trend.
Recently, Zuckerberg's rhetoric has made waves in the field of artificial intelligence. The huge computing power he shows off not only supports the huge demand for artificial intelligence training, but also indicates the possible amazing development in the future. There are concerns that human knowledge will be rapidly depleted, but in fact, as technology advances, AI is expected to break this limitation by expanding its knowledge base, such as through self-generated knowledge.
In addition to the improvement of computing power, the field of artificial intelligence is also actively exploring new training methods. Using the knowledge generated by AI to train AI itself has become a cutting-edge and promising research direction. This approach not only enables the continuous generation of new training data, driving the continuous learning and progress of artificial intelligence, but also potentially unlocks more unknown possibilities for us.
At the same time, the rise of multimodal learning has also injected new vitality into the development of artificial intelligence. With the increasing abundance of multi-content, artificial intelligence has been able to process text, images, and other types of data at the same time, so as to obtain more comprehensive and in-depth knowledge and information. One can imagine how amazing the growth curve of AI will be in the future if cameras around the world can provide learning materials for AI.
Development and security of large models.
With the rise of GPT and domestic large models, artificial intelligence has begun to attract widespread attention. By simulating the way humans type and communicate, the concept of artificial intelligence has quickly become popular all over the world. However, many people may only understand the large model as a chatbot, but in reality, its potential is much more than that.
Three application directions for large models.
1. Robots.
Large-scale model technology has injected new vitality into the robot industry. In the past, training a robot required tedious rule training for each task, but now large models allow the robot to understand the world more deeply and act accordingly. This technological advancement heralds a major breakthrough in the field of robotics, and robots are expected to play a wider role in more fields.
2. Autonomous driving.
Large-scale model technology is also expected to drive the development of autonomous driving. Current autonomous driving systems rely heavily on technologies at the perception level, such as radar and cameras. However, the introduction of large-scale model technology will enable vehicles to think as deeply as a human driver, accurately judging obstacles on the road and making decisions. This will significantly improve the safety and reliability of autonomous driving systems, and promote their development to a higher level.
3. Scientific research assistant.
Large model technology can not only be used for language communication and personal assistants, but also is expected to be a powerful assistant for scientific research. In the fields of biology and disease research, as well as physics and mathematics, large models can assist scientists in complex data analysis and theoretical derivation. The application of this technology will accelerate the progress of scientific research and promote the progress of human society.
With the maturity of computing power, data and training methods, the application threshold of large model technology is gradually reduced. The key lies in how to skillfully apply large model technology to real-world scenarios and maximize its potential. Now, both at home and abroad, it is recognized that large-scale model technology is leading a new round of industrial revolution and will reshape various industries. Therefore, we need to think about how we can make the most of big model technology in our own industry to bring about real change.
Unlike comparing large models to operating systems, large models are more likely to be as ubiquitous as PCs. In the future, every person, enterprise, and individual may have their own large model, which will lead to a more distributed and privatized large model. At the same time, with the development of large-scale model technology, cutting-edge applications such as digital twins will also become possible. This will provide us with a completely new digital experience and even some degree of digital immortality.
In addition, large models will also be widely used in various terminal devices, such as mobile phones, home devices, etc. This will provide users with more intelligent and convenient services. Especially in the automotive industry, cars without large models will soon be phased out. With the big model, the car can have a full dialogue with us, managing the entire autonomous cockpit and entertainment system, providing a more advanced experience than it already has.
Therefore, we should actively embrace large-scale model technology and think about how to integrate it with our industry to create more possibilities and seize opportunities to achieve innovation. This will be a profound industry change and the key to seizing the opportunity of the fourth technological revolution and promoting social progress.
Development of large models and security issues.
Regarding the development of large models and their accompanying security issues, there are two opposing views in the world.
On the one hand, the "security faction" advocates a cautious or even restrictive attitude towards the development of large models. They believe that the potential risks of large models could pose an irreversible threat to humanity. These concerns are particularly prominent in the application of the military field, such as OpenAI's recent lifting of restrictions on the entry of artificial intelligence into the military field, and the US Department of Defense has begun to experiment with the use of large models for various applications, which has undoubtedly intensified the concern about the security of large models.
On the other hand, the "developmentists" firmly believe that the big model is an important tool for social progress. They believe that the application of large models will provide individuals and enterprises with powerful knowledge assistants and support forces, help solve various problems, and promote the rapid development of social economy. This optimism is supported by the widespread use of large models in a wide range of industries, and these application examples demonstrate their tremendous impetus for the development of the industry.
However, regardless of the view, we cannot ignore the security issues in the development of large models. In order to ensure the security and controllability of large models, it is necessary to strengthen technical research and regulatory measures to prevent them from being maliciously used or producing uncontrollable consequences. At the same time, it is also necessary to solve content security issues to ensure that the content generated by large models is credible and reliable.
In the process of driving the development of large models, it is important to have a deep understanding of their principles and mechanisms. We need to explore how to limit the capabilities of large models to manageable limits so that they can be useful tools to serve humanity rather than potential threats. In addition, using the advantages of large models to monitor and correct their behavior is also an effective means to ensure their security.
The state has also put forward a slogan "digital transformation and intelligent reform". "Digital transformation" refers to digital transformation, emphasizing the use of large models as the basis to promote the digitalization process of all walks of life. Without digital transformation, you can't collect enough data to extract valuable knowledge to train your own large models. "Intelligent transformation" refers to intelligent transformation or intelligent upgrading.
How to implement large models in enterprises?
Problems encountered by enterprises in using large models.
Enterprises do encounter multiple challenges when applying large models. Although large models are trained with public knowledge and data, valuable "dark knowledge" within the enterprise, such as proprietary drawings, exclusive recipes, and unique management processes, is the core of its competition. These unique information assets are often underutilized in general-purpose models.
The crux of the problem is twofold:
The depth of the industry is insufficient. General models may be comfortable answering basic, general questions, but when they touch on the deeper questions of the professional field, they are stretched thin and may even give misleading information.
Poor adaptability of enterprises. It is difficult for the general large model to fully understand and adapt to the specific environment and needs of enterprises, which limits its effective application in enterprise scenarios.
In addition, data security issues cannot be ignored. Entering sensitive internal information into an external model carries the risk of leakage or abuse. Therefore, private deployment becomes the key to ensuring data security.
The problem of "gibberish" in large models is especially acute in enterprise environments. Individual use may tolerate occasional inaccuracies, but in an enterprise scenario, such missteps can lead to serious business consequences, such as confusion caused by false alerts.
Cost is also a challenge for small and medium-sized enterprises. The high cost of computing power and resources required for the training of general-purpose large models has discouraged many enterprises. Therefore, in the future, the development of large models should pay more attention to industrialization, verticalization, enterprise and industrial customization to meet the actual needs of different enterprises.
This trend will bring unprecedented opportunities to the industry. Unlike the Internet field, the industry has a wide range of application scenarios and untapped markets. Combined with the "dark knowledge" of enterprises to train large models, it will effectively solve various business problems and promote the sustainable development of the industry.
It is expected that in the future, large models will show a dual-track development trend: on the one hand, Internet giants will continue to promote the research and development of large models in the cloud and explore new heights of human intelligence; On the other hand, the enterprise-level model will focus more on cost-effectiveness and practicality, streamlining large models into efficient, low-cost small models through techniques such as "distillation", while maintaining their performance advantages.
For many enterprises, building a 100 billion model is costly and impractical. Conversely, the 10 billion model is optimized to run on a single computer, is low-cost, and is easy to scale. Although smaller, the 10 billion model excels in specific areas.
With the popularization of large model technology, the key is to apply it to specific scenarios and combine it with enterprise knowledge for training. Aligning big models with enterprise business requires technology and business insights. In the future, AI will become the core competitiveness of enterprises, and talents who master AI technology will have a competitive advantage. Therefore, enterprises should actively promote the popularization and application of AI, and cultivate employees' awareness and trust in AI to cope with the new industrial revolution.
AI content" assessment.
In today's enterprise application of AI, we observe two core levels: first, the use of AI general products to grasp its working principles and limitations, and optimize internal business processes accordingly; The second is to explore how to use AI to enhance the products and services provided to the outside world. Although it is often asked how to effectively use AI in enterprises, the reality is that there are still not many cases where AI can be used to create new products on its own.
In 2023, we face the reality that while large model technologies continue to advance, they are still not perfect, and their potential is enormous but not yet fully realized. As a result, we have yet to witness the kind of disruptive applications that AI brings. However, that doesn't mean we can't get real value out of AI. The success stories of Microsoft and Salesforce are proof of this. Rather than trying to create entirely new products, they cleverly integrate AI technology into existing product lines to enhance the user experience and product value.
This raises questions about the nature of innovation. Innovation is not the same as invention, and sometimes it is also an innovation to present an existing product in a new way. For people with non-IT backgrounds, it may not be practical to ask them to create entirely new products using AI. Instead, we can learn from the strategies of Microsoft and Salesforce and prudently incorporate AI technology into existing products to innovate and improve over time.
In order to promote the widespread application of AI in enterprises, I proposed the concept of "AI content". This means that employees need to be trained to master the basic principles and application scenarios of AI, and include them in the assessment system. Only in this way can we cultivate pioneers with an in-depth understanding of AI and lead the company to deeply integrate with AI. If a company doesn't know anything about AI, it can't make a real difference just by buying a big model. Therefore, "AI content" is not only a requirement for employees' skills, but also an important indicator for companies to promote digital transformation and enhance competitiveness.
Unlike cloud computing and big data technologies, the application of AI is highly dependent on specific business scenarios. After purchasing a GPT account or connecting to an external large model, many enterprises still find it difficult to effectively integrate it with their own business. This is because AI needs to be deeply coupled with the business and driven by the business. AI application scenarios may vary greatly between different industries and enterprises, so enterprises need to define and guide AI applications based on their own business characteristics.
In order to achieve the smooth implementation of large models in enterprises, we recommend adopting the strategy of "small incision, large depth". Enterprises should start by engaging business experts to break down their business needs into a series of specific, well-defined tasks and scenarios. These tasks are then matched to scenarios and capabilities that the large model has already demonstrated.
In the matching process, there are three key elements to focus on: whether the capabilities of the large model can play a role in the scenario, whether there is enough data and knowledge to support the application of the large model, and how fault-tolerant the scenario is.
Taking the field of human resources as an example, enterprises can select several scenarios with the most prominent pain points, the highest cost, or the worst current performance from multiple potential scenarios, and try to introduce AI technology to solve them. At the same time, in order to ensure the effective training and evaluation of large models, enterprises must ensure that they have sufficient historical data and knowledge accumulation. In terms of fault tolerance, enterprises also need to be cautious about the application of AI technology in key business links to ensure that there is no irreparable impact on the business.
Floor-to-ceiling frame. The application of large models in enterprises requires a comprehensive framework to ensure effectiveness and security. This framework consists of the following three main parts:
First of all, the concept of "data factory and model factory" is essential for every enterprise with its own large model. This means that enterprises need to integrate a part of local knowledge into the model, and through the support of data factories and model factories, they can train large models that meet the characteristics of the enterprise to better adapt to and solve actual business problems.
Secondly, the construction of an external knowledge base is also key. The knowledge within the enterprise is constantly updated, so in addition to integrating knowledge into the large model, it is also necessary to manage and correct possible errors in the model through an external knowledge base. By combining the accuracy of knowledge search with the intelligence of large models, it is possible to ensure that large models do not make mistakes when processing enterprise knowledge, thereby improving the reliability of their applications.
Finally, the introduction of the agent framework gives the large model more powerful execution capabilities. An agent can be understood as an extension of a large model, enabling it to connect with the internal CRM, ERP, OA and other systems of the enterprise, and even interact through the Internet. In this way, the large model is not just a dialogue tool, but can actually participate in the business process of the enterprise, improve the efficiency and intelligence level. However, the introduction of agents also brings security risks, so security issues must be addressed to ensure that their double-edged nature does not cause harm to the enterprise.
Judgment of the top ten development trends of large models in 2024.
Popularization of large models: Large models will become the standard configuration of every enterprise and are ubiquitous.
The outbreak of open source large models: With the increasing maturity of technology and the vigorous development of open source communities, the technical barriers of large models will gradually disappear, and they will enter thousands of households with a more people-friendly model.
Small model intelligent terminal application: In order to meet the needs of more intelligent terminals, small models will come into being, injecting new vitality into the Internet of Things, edge computing and other fields.
The rise of the enterprise-level market: With the increasing demand for intelligence from enterprises, the application of large models in the enterprise-level market will usher in explosive growth.
Agent: The large model is like a powerful engine, and the agent is the chassis and wheels that combine it with the actual application.
Killer applications emerge in TOC scenarios: In consumer-facing scenarios, a number of killer applications are expected to emerge, leading a new round of consumption upgrades and lifestyle changes.
Multi-modality: With the advancement of technology and the improvement of application requirements, multi-modality will gradually become popular in enterprises and become the standard technology for processing multi-modal information such as audio and audio.
AIGC: Artificial Intelligence Generated Content (AIGC) technology will make breakthroughs, gradually expanding from the initial image generation to the first generation and other fields, providing powerful auxiliary tools for marketing, design and other fields.
Robots: With the continuous penetration and application of large-scale model technology, the robot industry will usher in new development opportunities and is expected to achieve explosive growth in the next few years.
Promote basic scientific research: Large models will gradually become a powerful tool for basic scientific research.