DoNews news on June 21. On the 21st, Peking University’s Guanghua School of Management and Tencent announced the upgrade of the “Digital China Tower Building Plan” and jointly launched a series of courses on “Artificial Intelligence General Courses for Enterprise Managers” to help business founders and managers embrace AI. .
In the first lesson, Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industry Business Group, briefly reviewed the history of AI, and systematically explained the technological changes driven by the big language model, the challenges and responses of the industry, and the embrace of enterprises. Basic guidelines for large models.
Tang Daosheng reviewed the history of AI development, saying that the superposition of the three major factors of algorithm innovation, computing power enhancement, and open source co-creation constitutes the “growth flywheel” of AI.
Tang Daosheng said that the large model is only the starting point, and in the future, the industrial transformation of the application will be a bigger picture. In the R&D, production, sales, service and other links of enterprises in the past, there were many places that relied on people to judge, coordinate and communicate. effect.
The following is the full text of the speech:
Hello everyone! I am very happy to participate in today’s press conference to discuss the upgrading of industrial intelligence. With the development of big language models, we are entering an era reshaped by AI. From production and sales, organization of talents, to industrial innovation and social development, drastic changes will take place.
Many business managers are also thinking about how to apply large-scale model technology to their own business scenarios, such as in customer service and marketing, to bring more cost reduction and efficiency increase to business operations? How to protect the property rights and privacy of enterprise data when using large models? How to reduce the cost of using large models? These are practical issues that business managers need to consider.
Today, the “General Course on Artificial Intelligence for Entrepreneurs”, jointly established by us and the Guanghua School of Management of Peking University, hopes to help you understand cutting-edge technologies, corporate organizational changes, business model verification, and model implementation. Find new ways to solve problems together.
Here, I also share some thoughts on the integration of artificial intelligence and industry, and discuss with you how to use AI to drive industrial change.
I would like to share my observations and opinions from four parts, including the history of AI, the current state of technology, the implementation of the industry and the challenges it brings us. Let me first review the history of AI development from the perspective of technological development, which will help us better understand the current status and future development of artificial intelligence.
1. The development history of artificial intelligence
In 1950, Turing, the father of artificial intelligence, raised a key question in his paper: “Is it possible for a machine to have human intelligence?” From this the concept of “artificial intelligence” was born.
What exactly is artificial intelligence? It is a science of research, development, how to simulate and expand human intelligence. Including robotics, language recognition, image recognition, natural language processing and other directions.
To put it simply, it is to study how to make machines, like humans, able to listen, speak, see, think, and act. One of the most important aspects is to allow machines to master language, from understanding, learning, to generating expressions. This is also the “superpower” shown by large models like GPT-4 today.
Language is the most important carrier of human thinking. Yuval, the author of “A Brief History of Humanity”, even said that by mastering language, artificial intelligence has cracked the operating system of human civilization and mastered the “master key” to the future.
Over the past 40 years, AI development has continued to accelerate. There are also some well-known landmark events. For example, IBM’s special-purpose supercomputer Deep Blue won the world chess championship in 1995 by exhausting all possibilities on the chessboard.
In 2016, AlphaGo combined deep learning and reinforcement learning to defeat Li Shishi in Go. There is also AlphaFold’s contribution to biological sciences, which achieves very high precision in protein folding. Then to ChatGPT, GPT-4, Vincent graph technology Midjourney, Stable Diffusion, etc., which have recently made the public popular.
Behind these events are continued breakthroughs in underlying technologies, especially neural networks. In 1986, Geoffrey Hinton, the father of deep learning, invented the backpropagation algorithm, which laid the theoretical foundation for modern machine learning and using data to train neural networks.
The principle of the neural network computing model is to build an artificial neuron model by imitating the human brain, and abstract it layer by layer with a multi-layer architecture.
Subsequently, model architectures continued to innovate, such as convolutional neural networks, recurrent neural networks, etc., which brought great development in deep learning. The most recent important breakthrough was in 2017. Several Google researchers published a groundbreaking paper “Attention is all your need”, proposing the Transformer architecture to express the association of each word in the sequence with self-attention, including GPT today. The AI models included are all derived from this Transformer general framework.
In addition to the underlying technology, the development of AI is also limited by factors such as computing power. The training of neural networks consumes a lot of computing power. In the 1980s, computer power could only support shallow neural networks. At that time, a typical neural network had only 1960 parameters, and even the simplest text recognition could not be completed. Even in the early 2000s, computing power was still the bottleneck, and online data for training was insufficient.
In the past 20 years, hardware computing power has continued to increase. On the one hand, Moore’s Law has continuously doubled the computing power of chips; on the other hand, high-speed networks and distributed computing technologies have also continuously expanded the scale of computer clusters.
In the mid-2000s, Nvidia created CUDA to make GPUs more versatile and programmable, extending from graphics rendering to the field of scientific research and supercomputing. Based on different design concepts, GPU focuses on overcoming concurrent vector calculations. The computing power of a single GPU is a thousand times greater than that based on CPUs in the past.
Coupled with the rapid development of the Internet, the trainable data has rapidly increased, allowing the neural network to achieve larger, deeper, more parameters, and more complex model structures, thus giving birth to large models with more than 100 billion parameters.
In addition, global industry-university-research forces and open source co-creation are also important factors for the rapid breakthrough of artificial intelligence. Whether it is scientific research papers, data sets, model algorithms, or software platforms, generations of artificial intelligence scientists have selflessly opened up their research results, so that latecomers can continue to advance on the basis of their predecessors.
In terms of open source software, universities and technology companies around the world have contributed a variety of AI training and reasoning frameworks, as well as a large number of data processing tools, to developers.
Today, a large number of pre-trained open source models can be downloaded from huggingface, github and other platforms, allowing global researchers to build services on various open source models and optimize better models.
Open source co-creation, algorithm innovation, and greatly enhanced computing power, these factors superimpose to form the “growth flywheel” of AI. The advent of large models such as GPT-4, PaLM2, LLaMa, etc., let everyone see the dawn of general artificial intelligence.
2. Big models promote intelligent transformation
If it is said that the large language model has emerged a certain level of intelligence, it should be able to produce new combinations that have never appeared in the training data.
Through this painting case, we can see that the way the current AI large model handles problems is no longer by complete preservation and copying, but by recombining the trained knowledge after understanding the instructions.
It can be seen that the large model can draw pictures step by step according to the instructions, for example, draw a person with characters. The face is represented by the letter O, the torso is represented by the letter Y, and the legs are represented by the letter H. The first drawing is not ideal, but according to the feedback, you can adjust the proportions of the body and hands, and put clothes on the villain. It is a process of continuous feedback and adjustment.
On the other hand, the famous Turing test, which is whether one can identify whether the other party is a human or an AI in a conversation, is no longer sufficient to evaluate the intelligence of artificial intelligence today.
If test scores are the most direct tool for assessing human intelligence, then the big language model has surpassed the average human level in terms of language understanding and logical reasoning.
In the field of programming, GPT-4 participated in Amazon’s simulated technology test and got a full score. The test is required to last for two hours, but it only took less than 4 minutes. In the US GRE and Biology Olympiad exams, GPT-4 also exceeds 99% of humans; the results of the simulated bar exam are about the top 10%. In addition, Google’s Med-PaLM 2, also reached the expert level in the US Medical Licensing Examination.
Recently, OpenAI has added function calling capabilities to the ChatGPT API, which means that large models can also use tools. You can rely on various third-party services to try to solve the abilities that you don’t have, which greatly increases the ability of the general large model to solve problems.
The large language model represents the development of artificial intelligence and has reached a new peak. It has excellent language understanding, strong logical reasoning and communication skills, and can be brought into the role and actively think.
Models pre-trained with a large amount of data also promote new breakthroughs in AI capabilities such as machine vision, speech recognition, and robotics. Through the integration of the machine’s ability to think, hear, see, and move, AI will truly become people’s work and life assistants.
First of all, based on the multi-modal large model, computer vision has changed from “seeing” to “understanding”.
In the banking business, a lot of receipts, invoices, applications, business emails and other data have to be processed. For example, a commercial bank we cooperate with needs to process more than 10,000 emails and faxes in the asset custody business every day. From different business systems such as investment, insurance, financing, etc., the content includes bills, ID photos, etc., in various styles. Relying on manual processing and inputting into the system is time-consuming and labor-intensive, and more intelligent machine recognition is needed.
Under the traditional algorithm model, 2,000 receipts need to be input before the machine can recognize one kind of receipt, and there is no ability to organize them into tables or labels.
Now, based on the capability of the large model, our TI-OCR only needs 50 marked documents to quickly identify a type of document. At the same time, according to the analysis ability, the core tags can be automatically extracted, and electronic data files can be generated for subsequent business analysis.
The large language model not only understands multiple human languages, but also masters multiple programming languages, and can also help programmers write code.
We have also created Tencent Cloud’s new generation AI code assistant, which realizes AI’s understanding of code, assists programmers in writing, troubleshooting and testing, assists the entire process of software development, and improves development efficiency and code quality.
This is a new video released last week. The robot dog Max of Tencent Robotics X Robotics Laboratory has been upgraded again. As you can see, two robot dogs are running an obstacle course, put them randomly in the field, one chases, one hides, and there is a random flag. The robot dog who is hiding must try to touch the flag without being caught. After touching the flag, the role is reversed.
In this process, the two robot dogs have to judge their own behavior based on each other’s actions in real time, and at the same time keep in mind the target, that is, touch the flag or grab the opponent. At the same time, you must immediately correct your strategy after encountering the flag.
Through this video, we can see that the action of the robot dog has better flexibility and autonomous decision-making ability due to the addition of pre-trained AI models and reinforcement learning.
The large language model can not only communicate with people, but more importantly, through the fine tuning of the model, a series of execution steps can be generated according to the needs, such as the ability to call different plug-ins online, and multi-modality allows AI to understand the graph at the same time. Understand words, plan, and act, so that more powerful applications can be made, making AI more like a real smart assistant to complete more advanced tasks.
For example, online ad placement staff need to refresh a large number of advertising materials every day to ensure the ROI of advertising. If combined with advertising effect data and Wenshengtu’s ability, they can continuously generate delivery strategies based on data analysis, adjust delivery channels, and target Generate delivery materials, the degree of automation and efficiency will be higher.
3. Ways and paths for enterprises to embrace large-scale models
The convergence of so many changes also means that massive innovations are about to explode. The big model is just the starting point. In the future, the industrial transformation of the application will be a bigger picture.
In fact, no matter which industry you are in, you should actively embrace AI. In the past R&D, production, sales, service and other links, there were many places that relied on people to judge, coordinate and communicate. Today, we should look at which links, The productivity of AI can be superimposed to improve quality, reduce costs and increase efficiency.
At present, some large overseas companies have begun to invest in and adopt large-scale model technology. For example, Morgan Stanley directly accesses GPT-4 and uses it to integrate and analyze massive investment strategies and market research reports to provide direct reference for investment consultants.
According to the analysis of the self-media “Qubit”, we can see the influence and acceptance of generative AI (AIGC) on different industries. As can be seen in the figure, the content and e-commerce industries are most affected. Technologies like Vincent’s graph will greatly change the process and cost of content production.
Since the big model is so important, the entrepreneurs and managers here may also ask, how can we use it in the enterprise and seize the dividends of technological change?
I can give some advice to business managers:
First, focus on the enterprise’s own business, select specific scenarios, and make AI an incremental service.
Second, ensure the quality of training data, sort out test cases, and establish an online evaluation process.
Third, ensure service compliance while paying attention to data property rights and privacy.
Fourth, use cloud vendor tools to build an integrated model service, which is more efficient and saves training, operation and maintenance costs and time.
In the specific implementation, the model, data and computing power are the three points that everyone needs to pay special attention to.
First is the model. Although everyone has high expectations for a chatbot with a general large language model, it is not the only large model service method, nor is it necessarily the optimal solution to meet the needs of industry scenarios.
At present, general-purpose large models are generally trained based on extensive public literature and network information. The above information may contain errors, rumors, and biases. Many professional knowledge and industry data are insufficiently accumulated, resulting in industry-specific and accurate models. The degree is not enough, and the data “noise” is too large.
However, in many industrial scenarios, users have high requirements for the professional services provided by enterprises and have low fault tolerance. Once enterprises provide wrong information, it may cause huge legal liability or public relations crisis. Therefore, the large-scale models used by enterprises must be controllable, traceable, and correctable, and must be tested repeatedly and fully before they can be launched.
We believe that customers need more industry-specific industry models, coupled with the company’s own data for training or fine-tuning, in order to create highly practical intelligent services. What enterprises need is to really solve a certain problem in actual scenarios, rather than solving 70%-80% of the problems in 100 scenarios.
In addition, the more training data, the larger the model, and the higher the cost of training and reasoning. In fact, most enterprise scenarios may not require general AI to meet the needs. Therefore, how to choose a suitable model at a reasonable cost is what enterprise customers need to think about and make decisions about.
Secondly, data is the raw material of large models. For specific scenarios, the coverage and quality of relevant data are crucial. The management of labeled data is also an important task in model iteration.
The model must eventually be implemented in real scenarios, and to achieve the ideal service effect, it is often necessary to use the company’s own data. In the process of model development, we must not only pay attention to the protection and security compliance of sensitive data, but also need to manage a large amount of data and labels, and constantly test and iterate the model.
Thirdly, computing power is the basis for the continuous operation of the model. High performance, high flexibility, and high stability of computing power require professional cloud services.
In the process of training and using large models, a large amount of heterogeneous computing power is required, and the requirements for network speed and stability are also high. In addition, GPU servers are less stable than ordinary servers. The more frequent the investigation, the difficulty and workload of the overall operation and maintenance will be much higher.
In the training cluster, once the network fluctuates, the training speed will be greatly affected; as long as one server overheats and goes down, the entire cluster may stop, and then the training task will be restarted. These problems will greatly increase the training time. Increase, the cost of investing in large models will also soar.
Based on the consideration of the practical problems and needs of these enterprises, just two days ago, Tencent also officially announced the panorama of Tencent Cloud MaaS services.
Based on the Tencent Cloud TI platform, the selected large-scale industry model store will cover 10 major industries such as finance, cultural tourism, government affairs, medical care, media, and education, and provide more than 50 solutions. Based on these capability models, partners only need to add their own unique scene data to quickly generate their own “exclusive models”.
We also launched a fine-tuning solution for large-scale industry models based on the Tencent Cloud TI platform. Help model developers and algorithm engineers to solve tasks such as model calling, data and annotation management, model fine-tuning, evaluation testing and deployment in one stop, and reduce the pressure of creating large models. We can also use the TI platform to implement privatized deployment of models, authority control, and data encryption, so that enterprise users can feel more at ease when using models.
For example, based on the “Cultural Tourism Large Model”, we and the top domestic online travel companies have created a robot customer service that can automatically judge user intentions and automatically call the corresponding API to complete user consultation and services with high quality.
If a user asks, “Dragon Boat Festival will be out of Jiangsu, Zhejiang and Shanghai for three days, what is the itinerary recommendation, and how should it be arranged?” If it is a customer service robot based on a general large model, it can only give some simple introductions to scenic spots and route planning.
But when we added industry data and fine-tuned the model, the customer service robot’s answers became more detailed, and it was able to plan daily traffic and scenic spot arrangements, including hotel recommendations and introductions of different grades, and even directly provide booking links and platform discounts. coupon information. The intelligent customer service system can not only realize a humanized service experience, but also has a stronger sales transformation ability.
On computing services. The stable computing, high-speed network and professional operation and maintenance provided by Tencent Cloud can greatly reduce the pressure on equipment operation and maintenance for algorithm engineers, allowing them to focus on model construction and algorithm optimization.
Tencent Cloud has also built a new generation of HCC (High-Performance Computing Cluster) high-performance computing clusters for model training, equipped with the latest next-generation GPU, combined with multi-layer accelerated high-performance storage systems, coupled with high-bandwidth, low-latency network transmission, The overall performance has been improved by 3 times compared with the past, and has been highly recognized by many customers. Several large-scale unicorns have launched computing power cooperation with us.
In addition to the “hard power” of computing clusters, we have recently launched a “soft power” that is more suitable for AI computing – the vector database, which can more efficiently process unstructured data such as images, audio and text, and supports a single index The scale of 1 billion is 10 times higher than the scale of stand-alone plug-in retrieval, and the efficiency of data access to AI is also 10 times higher than that of traditional solutions.
4. Challenges and Countermeasures of AI Development
The value of AI is huge and the speed of development is astonishing, but from the social level, we must also pay attention to the risks and challenges it brings.
Recently, Geoffrey Hinton (Geoffrey Hinton) left Google because he was worried that AI could not be controlled. He mentioned in an interview that the information architecture of artificial intelligence may be more powerful than the information architecture of the human brain.
There are approximately 86 billion neurons in the human brain, and approximately 100 trillion connections are formed between these neurons. Although the parameters of GPT4 are not disclosed, it is generally estimated that the number of parameters is only one percent of the neuron connections in the human brain, which is about 500 billion to 1 trillion.
However, the knowledge loaded by GPT4 is thousands of times that of ordinary people, and the learning efficiency is also higher. This shows that, to a certain extent, the current neural network may have a “better” information processing architecture and learning algorithm than the human brain. Once sufficient computing power is obtained for training, it can learn a large amount of information and knowledge more quickly.
Another point worth mentioning is that AI can download and copy the model through an online connection. In a relatively short period of time, a new machine can copy a large amount of knowledge, and each can synchronize with each other after learning different knowledge.
However, human knowledge and wisdom must be passed on through complex, changeable and imprecise language. As a medium for transmitting information, the cultivation of each person almost starts from zero (except for the memory of the hard code in the gene), and needs to start from a young age. Decades to learn, decades to gain experience.
breaking latest news’s powerful and continuous generalization capabilities have made many people very worried that humans will gradually lose control of AI. In particular, breaking latest news can be connected to the Internet, programmed, manipulate other systems (because APIs of other systems can be called), understand people (because a large number of books are imported into the model, understand the history of the development of human civilization for thousands of years, understand people’s way of thinking and weaknesses, and understand people every day. Interact with many people, even communicate emotions), it masters language (so it can affect people’s thinking and behavior), it can generate pictures and videos (so it can make people have optical illusions), and there may be more abilities we have not yet discovered .
Therefore, Hinton also proposed that AI poses four threats to human beings. He did not hesitate to leave Google, where he had worked for 10 years, to promote public attention to the potential risks of artificial intelligence and establish norms for the safe use of breaking latest news.
In the face of various problems brought about by artificial intelligence, there are still many things worth thinking about. Includes human development, ethics, education, and more.
These questions, I believe that each of us here has his own thinking. But one thing I want to say is that the development, evolution, and changes of technology are always beyond human imagination, and the courage of human beings to embrace change, the wisdom of innovation, and the ability to turn challenges into opportunities are often beyond our own imagination.
Just like in the early days of the Industrial Revolution, there were also concerns about the collapse of the rural economy and the worrying living conditions of workers, etc. However, in the end, we came over in a uniquely human way, and made the production efficiency and quality of life of all human beings geometrically exponential. soaring.
There is no doubt that AI’s changes to the world must also be achieved through integration with industries. A series of changes such as machine decision-making, autonomous generation, and natural interaction have promoted the industry to achieve higher efficiency, lower costs, better experience, and greater innovation. Enterprises in the future will also evolve towards intelligent natives.
Facing the future, Tencent is also willing to continue to contribute its own capabilities. With an open mind and endless curiosity, we will work with experts, scholars, and business managers to jointly explore, innovate, and embrace new opportunities in the intelligent era.