In 1956, the concept of Artificial Intelligence (AI) was first proposed, and it has been more than sixty years since then. In the past 60 years, AI has gone through a process from outbreak to cold winter and then to barbaric growth. With the improvement of technologies such as human-computer interaction and machine learning, AI has become a new trend in the technological era.
In 2022, the AI industry will once again usher in a new node, AI Generated Content (AIGC, AI Generated Content) will come from behind and become a major event in the history of technological revolution at a speed beyond people's expectations. Whether it is the "AI painter" DALL-E2 or the "universal chatting" chat robot ChatGPT, generative AI is rapidly giving birth to a new technological revolution system, pattern and ecology.
Turning the clock to 2023, the enthusiasm caused by AIGC has not diminished but increased, and the new era of intelligent creation will not only bring about profound changes in productivity, but will also further change the evolution of human thinking. In this regard, the 21st Century Business Herald's digital economy research group planned a series of reports on "Chasing the Waves AIGC" to interpret the technical possibilities and business prospects brought by AIGC in multiple dimensions.
21st Century Business Herald reporter Bai Yang reports from Beijing
Under the new wave of AI, a global arms race around AI has also kicked off. Right now, although ChatGPT is leading the way, it is actually only the tip of the iceberg. Next, AI applications based on large models will continue to emerge. Just like the advent of the mobile Internet ten years ago, a new era of change is quietly unfolding to.
Faced with the opportunities of the times, people will always be excited, and technology giants at home and abroad are gearing up and ready to go. Zhou Ming, the founder and CEO of Lanzhou Technology, said in an interview with the 21st Century Business Herald recently that Chinese companies should not rest on their laurels and learn from others when they build large-scale models. , because in the past two decades, China has made great progress, and it has also been able to walk out of Chinese characteristics in the field of AI.
Zhou Ming gave an example, "For example, making each function of the large model more controllable, or taking the lead in the implementation of To B, these will become Chinese characteristics, and with these things, a 'Chinese faction' in the martial arts can be formed. , It can also let colleagues see China's power."
In fact, in the past ten years, the entire AI industry has been in a period of rapid development, and many Chinese companies have also invested huge resources in this field, which has also made China a global leader in some AI segments. Among many Chinese technology companies, Tencent has an early layout of AI and has rich practices in AI applications. Therefore, this article will use Tencent as a sample, hoping to observe its AI development path, which can bring benefits to the future development of the industry. Some enlightenment.
Layout sixteen years ago
China's AI initially emerged around the needs of products. For example, the starting point of Tencent AI was in 2007. In that year, Tencent invested 100 million yuan to build the Tencent Research Institute.
Wu Yongjian, who is currently the vice president of Tencent Cloud and the head of Tencent Cloud Intelligent Research and Development, joined Tencent in 2008. The first department was Tencent Research Institute. He told the 21st Century Business Herald reporter that the research of Tencent Research Institute was very application-oriented at the beginning. For example, one of the jobs he was doing at the time was to develop image processing technology around QQ images.
"Later, with the help of our technology, the processing time of QQ video was reduced to about 60% of the original, and the effect was very obvious. Then this technology was applied to other departments such as games," Wu Yongjian said. It was also from then on that Tencent Research Institute discovered It is more appropriate to make technical reserves by yourself, so the whole team began to transform, from a product-oriented team to a technical support team.
Subsequently, Tencent Research Institute has made many achievements in pattern recognition, multimedia communication, data mining, image processing, and word segmentation. By 2011, Tencent had applied for more than 4,000 patents, which was more than the sum of other domestic Internet companies, of which Tencent Research Institute contributed more than half.
Originated from Tencent Research Institute, Wu Yunsheng, Wu Yongjian and others later formed the Youtu Lab team, becoming the industry's top computer vision laboratory. Later, Tencent also successively established a number of technical research teams, such as the WeChat Zhiling voice team established in 2011, which mainly develops voice artificial intelligence technology.
If we say that before 2012, Tencent’s technology research and development team was more to serve its own business, then since the establishment of AI Lab in 2016, Tencent has started to walk on “two legs” of basic research and industrial practice. Therefore, Tencent's AI path is to continuously extend from the service business to the upstream cutting-edge technology research.
In 2019, at the World Artificial Intelligence Conference held that year, Ma Huateng, chairman and CEO of Tencent, stated that Tencent has established four AI laboratories, covering AI from comprehensive basic research to various application development, and has also established cutting-edge technology. Explore the matrix of laboratories, covering robotics, quantum computing, 5G, edge computing, IoT, etc.
According to the data, in 2019, the number of patent applications of Tencent in major countries around the world has exceeded 30,000, and the number of authorized patents has exceeded 10,000. At that time, this number ranked first among domestic Internet companies and second among global Internet companies, second only to Google.
Explore cutting-edge technology
In Tencent's laboratory matrix, there are many seemingly "didn't do business" research, which is actually Tencent's research on future basic technologies.
For example, many people know that in 2016, Google's AlphaGo defeated the human Go champion. In fact, after Tencent AI Lab's Go AI "Fine Art" was released in 2016, it also won the world's top tournament championships four times, and has since Since 2018, he has been working as a dedicated AI for the training of the Chinese National Go Team for free.
Another example is that in 2017, Tencent applied artificial intelligence technology to the medical field and released the AI product "Tencent Miying" that can assist doctors in medical imaging screening and medical diagnosis. In November 2017, the Ministry of Science and Technology announced the list of the first batch of national new-generation artificial intelligence open innovation platforms, including relying on Tencent to build a national new-generation artificial intelligence open innovation platform for medical imaging.
In 2021, Tencent released the first multi-modal quadruped robot Max with self-developed software and hardware. At that time, Max relied on the integrated design of the foot wheel to realize standing and moving from quadruped to biped, and can complete backflips, fall self-recovery and other actions.
Max was born from Tencent Robotics X Laboratory, which was established in 2018. The core research direction of this laboratory is robots, including the perception ability as the basic technology of robots, and the three pillar technologies of sensitive movement, dexterous manipulation, and intelligent body. At present, in addition to Max, the laboratory has also released products such as the robot dog Jamoca and the wheel-legged robot Ollie.
In addition, Tencent also has a long-term plan for the large-scale AI model that has attracted much attention recently. In April last year, Tencent disclosed for the first time the development progress of its "Hunyuan" AI large model. It is reported that the Hunyuan AI large model completely covers basic models such as NLP (Natural Language Processing), CV (Computer Vision), multi-modality and many other industry models. VCR, MSR-VTT, MSVD and other authoritative multi-modal data sets have reached the top of the list.
Recently, the Hunyuan AI large-scale model team also launched the NLP trillion large-scale model, which not only once again broke the record of the three major lists of CLUE, but also benefited from the characteristics of low cost and inclusiveness, the model has also successfully landed in Tencent Advertising, Search, chat and other internal products and serve external customers through Tencent Cloud.
The Tencent Hunyuan AI large model team stated that since larger neural network models often mean stronger model performance, the Hunyuan NLP large model will focus on exploring larger model parameter scales in the future on the one hand, and on the other hand. Combine audio, image, video and other multi-modal information to further create a more powerful multi-modal AI large model. In addition, with the hot rise of the AIGC direction, the Hunyuan AI large model will continue to promote continuous upgrading in the fields of text content generation and Vincent graphs in the future.
Focus on scene application
On the other side of basic research is industrial practice. Ma Huateng has repeatedly stated: "Tencent's AI layout focuses on scene applications, not research for research."
Just like in the early days, Tencent AI started from user scenarios and used AI technology to solve internal product needs. In the mid-term, it promoted the development of general artificial intelligence with research + scenarios, emphasizing that "academics have influence and industry has output." Now, Tencent is using AI to solve problems in vertical industry scenarios, incubating customized solutions into standardized AI platform tools.
A person from Tencent said that the Tencent AI team is different from the traditional research team. It is a systematic construction. From algorithms, engineering, quality, data, products, to the entire commercialization model, there may be first and last, such as research. Go first, and commercialization comes in last, but the whole is a car building and moving forward.
Wu Yongjian pointed out, "If the goal is difficult enough and the scene is complex enough, it will lead us to make a world-class algorithm. Similarly, when your algorithm research solves a world-class problem, the algorithm is more valuable. , not purely for publishing papers".
In order to accelerate the industrial implementation of AI technology, in November 2021, Tencent officially released the "Tencent Cloud Smart" brand, through the aggregation of the products and technical capabilities of AI laboratories such as Tencent Youtu Lab and Tencent AI Lab, as well as years of industrial practice Experience, external output from the underlying computing power support to the AI development platform, AI product solutions, and top-level digital intelligent transformation methods of the whole chain of services.
For example, at the underlying computing power level, Tencent uses "one cloud with multiple cores" as the basis to accelerate computing power performance with the help of self-developed AI chips; at the AI development level, Tencent uses the "Tencent Cloud TI Platform" as the core to help customers quickly create and deploy AI applications.
Zixiao is Tencent's self-developed chip for AI reasoning scenarios. It has been adapted to the Tencent Cloud TI platform, which has improved the performance of a single card by 200%, reduced the cost of unit computing power optimization by 50%, and saved energy consumption of green computing power. 60%. Song Dandan, head of Tencent Cloud's heterogeneous computing products, told the 21st Century Business Herald that these chips will first be deployed on Tencent's self-developed business, and in the future it is hoped that it will serve external services in the form of PaaS services.
Around the TI platform, Tencent has also built a product matrix, including TI-DataTruth labeling platform, TI-ONE training platform, TI-Matrix application platform, TI-ACC acceleration tool, and also includes TI-OCR training platform, TI-AOI industrial quality inspection training platform, etc. These products have also been applied in pan-interaction, finance, industry, media, pan-government, medical and other industries, helping to realize many subdivided fields such as intelligent industrial quality inspection, financial AI middle platform, smart city operation management, and disease auxiliary diagnosis. Development of AI applications.
Li Xuechao, vice president of Tencent Cloud and head of Tencent Cloud Intelligent Platform, told the 21st Century Business Herald that the entire AI has indeed entered the deep water area in terms of implementation. "In the past, customers only needed you to provide some AI capabilities, but now, what customers propose are all scenario applications, and you need to integrate AI into business scenarios."
In Li Xuechao's view, through the current hot "pre-training large model + downstream task fine-tuning" model, AI applications will definitely become more generalized. On this basis, the original AI application scenarios will be made deeper. At the same time , AI will also penetrate into more scenes.
However, it also pointed out that the priority of doing AI applications is to solve problems, so in many scenarios, the original AI model can solve the problem, so there is no need to catch up with the heat. After all, the use of large models will also bring additional benefits to customers. the cost of. But for some scenarios, such as intelligent customer service, if the use of large models can bring direct effect improvement, you can try it while weighing the cost performance.
In this global AI competition, we need to pay attention to and compete with the most cutting-edge technology research. At the same time, we need to do some down-to-earth things according to market conditions. Zhou Ming told the 21st Century Business Herald that the services of To B enterprises in China are very different from those in foreign countries. The SaaS ecology in foreign countries is very mature, and small and medium-sized enterprises have become accustomed to receiving services through SaaS, but many enterprises in China do not accept SaaS. deploy.
This means that it takes more effort to serve To B customers, such as understanding customer needs, doing a good job of "last mile" business process and system connection, and also considering delivery and maintenance costs. "If your model is fragile, you may lose one project for one project.
Therefore, you must do a good job in the foundation, and you must also understand customers and have the ability to iterate quickly. This is what Chinese companies must face when they make so-called large models. In reality, from this point of view, if you only want to quickly copy a ChatGPT and then make quick money, it is very naive," Zhou Ming said.