Text ｜ Liu Yuqi, editor ｜ Wang Yisu
In the wave of large models, the popularity of AI comes one after another.
On July 6, the 2023 World Artificial Intelligence Conference was held in Shanghai, which once rushed to the Weibo hot search list. In the 50,000-square-meter venue, the booth employees of more than 400 companies spared no effort to explain their new products, and the participants kept shuttling among them. "The feet are numb, but the mood is still very exciting." Said.
At the scene, the booths of the Light Cone Intelligent Discovery Conference are deployed according to the industry, and the software, hardware, and chips are arranged neatly in a "family". Ali, Baidu, and Tencent are three in a row, which is no different from competing on the same stage, and the audience can view the latest large-scale models in the country within 5 minutes, and compare them while watching.
As the largest AI event in China, this time is different from the previous ones in that the dawn of general artificial intelligence seems to be within reach, and the fighting spirit of the crowd has infected every witness.
Elon Musk, the founder and CEO of Tesla, opened the conference online. Yao Qizhi, winner of the Turing Award, academician of the Chinese Academy of Sciences, and president of Shanghai Qizhi Research Institute, presided over the discussion of the conference, which not only gathered Baidu Wenxin, Ali Tongyi , Huawei Pangu, Xunfei Xinghuo, Shangtang Rixin and more than 30 large models at home and abroad, as well as more than 80 academicians at home and abroad, and more than 50 domestic and foreign celebrities such as Tesla, Microsoft, Huawei, and Ali. The world of science and technology is working together to push open the heavy door of the times.
The most basic large-scale model layer and technology platform usher in a new iterative upgrade. How to empower thousands of industries and promote the implementation of industries has become the topic of discussion in the new stage.
Although the structure of AI infra is still immature, the supporting facilities for large models have caught up. The vector database adapted to the calculation and storage of large models has become a popular track, and the AI safety monitoring platform that ensures the safe operation of large models has emerged along with the trend.
The implementation of the application is ahead of the maturity of the large model, the software application has penetrated into all walks of life, and a batch of new robot products have been born on the hardware side.
China's large-scale model has run out of China's speed, and the technology layer, supporting facilities layer, and application layer have simultaneously risen in all links, instead of waiting for the underlying large-scale model technology to mature before slowly catching up. While continuing to solidify the foundation, while blooming everywhere, the whole ship is speeding up, and more Chinese large-scale model models have been derived.
Large model, walking on two legs
Since it sprung up like mushrooms in March, in just 4 months, the large-scale model is moving from "general-purpose" to "industrial landing".
The general-purpose large-scale model is a "winner takes all" battle royale. It is more important to use the large-scale model than to make it. At the World Artificial Intelligence Conference, the keyword for all large-scale model upgrades is "industry".
As providers of large-scale models, major companies such as Ant, Tencent, Baidu, Ali, Huawei, and NetEase combine their previous experience and advantages in the field to quickly explore the application of vertical scenarios.
Ali's Tongyi Qianwen large-scale model was first used in smart finance. By combining natural language understanding and generation capabilities with financial scenarios, Tongyi Qianwen cooperates with CICC Wealth Management, Bank of Hangzhou, Hang Seng Electronics, etc. to help companies identify market risks and understand customers through intelligent question and answer, intelligent outbound calls, and intelligent assistants need.
Netease, which focuses on games and education, took the lead in combining large models with these two fields. In the field of education, NetEase Youdao self-developed the first domestic large-scale education model "Ziyue", and also released the latest application results at this conference - a virtual human language coach, which can experience 1V1 oral English private education similar to real people.
And Baidu, which has accumulated many years in the field of autonomous driving, has put the large model into the field of automobile transportation. In the field of automobile manufacturing, Baidu Smart Cloud provided Changan Automobile with an artificial intelligence infrastructure platform and a digital human platform. Together with Geely Automobile, it built a large model of the automobile industry and built a factory digital brain to help reduce management and operation costs.
As Hu Houkun, Huawei's rotating chairman, mentioned, "The key to the development of artificial intelligence is to go deeper and more realistically. The focus is on making artificial intelligence serve the production activities of thousands of industries and serve scientific research and innovation."
However, although the postures of this wave of Chinese large-scale models are different, they all began to expand outward in unison after gaining a firm foothold in their advantageous fields, competing for their industry coverage capabilities.
Take Tencent as an example. At present, Tencent Cloud has provided more than 50 large-scale industry solutions for more than 10 industries such as media and cultural tourism; Huawei's Pangu large-scale model is mainly related to finance, manufacturing, pharmaceutical research and development, coal mines, railways, etc. Combination of industries and implementation...In general, smart government affairs, smart finance, and smart transportation have become popular vertical tracks.
Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of the Cloud and Smart Industry Business Group, said, "The general large-scale model can solve 70%-80% of the problems in 100 scenarios, but it may not 100% meet the needs of a certain scenario of the enterprise. .”
In order to improve the competitiveness of large-scale models in the industry, enterprises need to go deep into the specific scenarios of vertical industries, and truly test the details and effects of technology in hand-to-hand combat.
As a technology that continues to evolve, light energy for large models is not enough. Just as the development of every technology requires two legs to walk, one is practicality and the other is advanced nature. The practicality of large models is gradually explored through the implementation of the industry side, while the advanced nature is reflected in the cross-modal The breakthrough of the problem.
Referring to the path of ChatGPT from 3.5 to 4.0, the technical advancement of generative large models is from single-modal to multi-modal. From the very beginning, Wenshengwen, which has been seriously homogenized, is moving towards Wenshengtu, Wenshengsheng, Wenshengwen Video evolution.
After Tongyi Thousand Questions, Alibaba Cloud announced the launch of a new member of the "Tongyi Large Model" AI painting "Tongyi Wanxiang". Similar to Wenxinyige and Shangtang Miaohua, the Wenshengtu track has become crowded.
After Wen Shengtu, Wen Sheng's voice and Wen Sheng's video are the next match points.
Taking video as an example, as a combination of text, sound, and graphics, Wensheng video is also the ultimate form of generative large-scale models. At present, the difficulty of Wensheng's video technology lies in the production of coherent picture materials. Based on the existing technology of digital human, Wensheng video can be realized in the digital human scene the fastest.
For example, SenseTime's Ruying large model can generate text copy, one-click knowledge sharing, brand promotion, short video delivery, training presentations, hot news and other digital human videos, as well as the 3D content generation platform "Qiongyu" (scene generation), "gewu" (object generation), based on 3D content generation technology to reproduce and interact with space and objects.
In the past, technology was often developed first and then applied. Under this wave of large-scale models, the advancement of technology and practicality are developing simultaneously, and the large-scale models that walk on two legs are accelerating.
Industry jigsaw puzzle, complete one by one
The large-scale model ran wildly for 4 months, and the jigsaw puzzle of the industry was gradually completed piece by piece.
The MaaS service platform with the large model as the core is the airport where the large model aircraft lands. How the large model lands and the speed at which it lands are closely related to the maturity of the airport platform.
At present, Baidu Wenxin Qianfan, Ali Lingjun PAI platform, Tencent Model Store, and Volcano Engine’s Volcano Ark, although there are differences in subdivisions, are all MaaS service platforms that help companies better fine-tune and deploy large-scale industry models. Apply large models to industry.
In WAIC, the upgrade of the MaaS platform is also a major attraction. At the meeting, Tencent Cloud disclosed several key industry directions, including innovative scenarios such as financial risk control, interactive translation, and digital intelligence. The next step is to lay out AI for Science, and continue to apply AI large models in fields such as astronomical exploration and cultural archaeology. .
Not long ago, Volcano Engine also released the MaaS service platform Volcano Ark, which includes two important modules: model square and model tool. The core purpose is to help enterprises better implement applications through the power of the platform.
In the process of development, security issues are gradually exposed.
Whether it is AI fraud or Samsung employee data leakage, model security has once become the focus of society. At this stage, the premise of solving the problem is to find the problem and turn on the safety red light to solve it in a targeted manner.
To this end, Ant launched the trustworthiness security evaluation system "Ant Jian 2.0". It can explain and evaluate capabilities, and can serve more than 10 large-scale and complex business scenarios in the fields of finance, education, culture, medical care, and e-commerce.
In addition, Ants will also release the 1.0 version of the open source framework of lingo with privacy computing technology as the core, the popular science interactive game "Privacy Computing Security Battle", and the digital copyright protection platform "Que Chisel", etc.
Ant’s technicians told Lightcone Intelligence: “When faced with multi-model decision-making, companies can open large models to the AntJian platform for scoring. There will be 4 different scoring standards, just like the plagiarism checking system for papers, and the detection model. The safety index and score to provide guidance for enterprises."
However, according to Light Cone Intelligence, at present, Ant Jian can only detect problems, and does not provide targeted solutions to data security problems. It also needs other manufacturers to jointly improve the industrial chain.
On the other hand, the development of data technology is also breaking through the limitations of the large model itself.
As mentioned above, the large model is undergoing a change from general to industry. This process is limited by two aspects: First, the general large model has only "short-term memory". Whether it is domestic or foreign, OpenAI has recently released the open source version of GPT4 Zhong also emphasized that there is a Token limit of 8k; the second is to train industry-specific large models. Enterprises need to combine private data with general-purpose large models. How can private data be safely connected to large models? How to process dynamically changing data more efficiently and at low cost?
This is why, after the big model became popular, the vector database became popular.
Sun Yuanhao, the founder of Transwarp Technology, explained to Guangcone Intelligence: "The vector database is like the external brain of the large model, which can help the large model have long-term memory, and private data can be stored in the vector database through vectorization, which can satisfy Instantaneous access and output also protects corporate private data from being leaked."
To this end, Tencent Cloud has released Tencent Cloud VectorDB, a cloud-native vector database, which supports a vector retrieval scale of 1 billion and controls latency at the millisecond level. Compared with the traditional stand-alone plug-in database, the retrieval scale is increased by 10 times, and it also has a peak capability of millions of queries per second (QPS).
Transwarp released the distributed vector database Transwarp Hippo. It supports storage, indexing, and management of massive vector data sets, and provides capabilities such as vector similarity retrieval and high-density vector clustering, effectively solving the problems of low knowledge timeliness, limited input capability, and low accuracy of large models.
Behind the popularity of vector databases is the large model driving the development of the entire data industry chain.
Data preparation (cleaning, filtering, purification) and data processing (selection, quality, labeling) before training a large model can only be "fed" to the large model after going through several processes, and China's current data industry is relatively scattered, resulting in In the process of data circulation, it is easy to encounter security problems, and the efficiency is not high.
For example, complex unstructured data needs to be vectorized (embedding) and unified into coordinate values in a multi-dimensional space before it can be used in large models. However, vector databases such as Zilliz do not help companies complete vector process.
This requires a one-stop data processing platform. Sun Yuanhao, who has been dealing with data for more than ten years, has condensed his experience into Sophon LLMOps of Xinghuan. Through this platform, users can complete a complete closed loop of data collection, knowledge accumulation, and iterative improvement of large models. At the same time, the combination of the vector database and the graph database can better help the large model to learn and optimize cross-domain knowledge, so that the large language model can better understand the professional terms, abbreviations, common vocabulary and grammar in different fields, and bear the unified responsibility Semantic understanding function to solve business domain problems.
The continuous completion of the industrial chain is the cornerstone of the development of large models, and it will also bring about many opportunities in the industrial chain to build a complete model ecology.
Software, hardware, both hands
In the entire exhibition hall this year, the robot known as the "King of Popularity" was robbed of the limelight by digital humans.
Crowds gathered outside the digital human exhibition hall, circle after circle. SenseTime built the Ruying digital human generation platform to demonstrate the process of 3D digital human generation on the spot; NetEase Youdao displayed the virtual human language coach on the vertical screen, which attracted exhibitors to come forward and interact; Three episodes in a row.
Although AI anchors and AI digital humans have poured into the short video platform on a large scale recently, causing a wave of discussion, but after shopping around, no one exhibited 2D digital humans, which is obviously more cool and fun than its practicality. , Interactive 3D digital human can grab people's attention even more.
Walking into the venue of the AI painting platform is still dazzling. Ali built a giant screen to display the effect of its Wensheng diagram, and SenseTime used a super-large scrolling screen to demonstrate its second-drawing function. Digital human, AI painting, AIGC has become the most concerned thing of the public, and it is also the fastest running direction in the application of large models.
The popularity of AIGC remains high, and it has become a topic of conversation after dinner, but more importantly, it is necessary to promote the implementation of the industry.
Zhao Zeng, the person in charge of Netease Fuxi pre-training and generative artificial intelligence platform, told Lightcone Intelligence that games are currently one of the core application scenarios for image generation. "Game scenes require the creation of a large amount of content, such as characters of various dynasties and landscapes around the world. The text and image function can quickly generate manuscripts, stimulate creativity, and improve game production efficiency."
During the light cone intelligent shopping exhibition, it was found that compared with the beginning of the year, the types of content generation on the market have become more abundant. Vincent graphs are almost the standard configuration of every company. In addition, the multi-modal capabilities of more dimensions are becoming more and more perfect. For example, Tencent has launched the XMusic generative general composition framework in the field of music creation by using Wensheng audio technology.
From the initial Wenshengtu to the current Wensheng Audio and Tushengtu, the types of generated content have become more and more diverse, showing a trend from single modality to multimodality. But this is far from enough. With reference to the development of software such as Midjourney abroad, Wensheng Video will be the direction of the next stage, while domestic companies are still in the early stages of exploration in this area.
"Left-handed software, right-handed hardware", the AIGC application is showing a trend of "grasping with both hands".
Just after the opening of the conference, Musk proposed in his speech: "In the future, humans will have more robot products, and it is expected that the number of robots on the earth will exceed the number of humans.
It is understood that a total of more than 20 robots were unveiled at the scene this year, and many of them were first released, including the biped robot equipped with large-scale dialogue capabilities, the Tesla humanoid robot Optimus Prime, the Netease Fuxi embodied intelligent engineering robot, and the minimally invasive robot. Surgical robot, Meituan UAV V4, Yushu quadruped robot, HKUST aircraft robot dog, Yunshen Jueying Lite3 quadruped robot, etc.
The large model allows the software to be redone again, but there is still a long way to go before it is actually applied to the robotics industry.
For a long time, robots have been facing problems such as insufficient front-end demand, high unit cost, and inability to scale production. The person in charge of the market of a certain robot manufacturer revealed to Lightcone Intelligent that in a few years, the cost of our robots has been reduced to one-third of the previous one, but as far as the industry is concerned, there are still many unknown problems in practical applications that need to be worked out.
The above-mentioned person in charge said that ideally, the large model can make the robot grow a soul, let it make decisions independently, and interact with the actual environment.
But unfortunately, it is still difficult to do it at present. The edge resources of the robot are limited, and the existing computing power cannot support the deployment of large models, and the deployment of expensive GPUs will continue to increase the unit price of the robot.
In this regard, some entrepreneurs said that the AI technology route has not yet been cleared, and it may be difficult to see the application of large-scale humanoid robots within five years.
However, many entrepreneurs, including Musk, believe that robots are the "embodied intelligence" of artificial intelligence. In the longer term, the maturity of robots will bring the application of artificial intelligence from the information industry to the deeper and larger physical world.
epilogue
China's large model is in a thriving situation.
However, amidst the excitement of the industry, we have also seen that the large-scale technology path, industrial landing, supporting facilities construction, and application ecology are still in the early stages.
As Zhou Jingren of Alibaba Cloud said, "The process of catching up with each other in the field of large-scale models has just begun, and everyone still needs to be patient."