(Report produced by: Zheshang Securities)
AI large models and industry applications present a variety of flowers
During the World Artificial Intelligence Conference, many manufacturers' AI models and application products were unveiled one after another, and the AI+ industry application ecology continued to mature.
AI + Finance: AI on the insurance side has stronger applicability
Banks and brokers: AI can empower all aspects from internal operations to external businesses, achieving the effect of reducing costs and increasing efficiency, and expanding territory. Most domestic banks and securities firms are actively integrating AI into the original structure, and the market space mainly comes from the stock demand for AI upgrades in financial institutions' own systems. Insurance: AI empowers insurance product development, marketing, underwriting, claims settlement and other links. The insurance industry has industry integration and can be embedded in the application scenarios of various industries, thereby creating rich and diversified products and services. In addition to the stock demand for AI upgrades of insurance institutions, the market space also comes from the expansion of "insurance +" application scenarios. We believe that the policy implementation of data elements will accelerate the process of data from confirmation of rights to circulation of data in various industries. In view of the high industry integration characteristics of the insurance industry, the connection between the insurance market and other markets will be closer in the future, and it is expected to achieve broad commercial value through the connection between insurance and data from various industries.
AI+Transportation: BEV+Transformer leads a new paradigm of intelligent driving technology routes
In July 2021, Tesla demonstrated a new paradigm of autonomous driving perception based on BEV+Transformer (BEVFormer). Based on the visual perception task carried out by BEV+Transformer, it can realize the integration of time and space of 360-degree surrounding, and output static (lane lines, traffic lights, road edges, etc.) and dynamic (pedestrians, two-wheelers, cars, etc.) information, so that accurate three-dimensional world information can be obtained by relying on pure vision. Since Tesla proposed the BEV Transformer technology paradigm, many mainstream car companies and autonomous driving solution providers such as Ideal, NIO, Xpeng, Pony.ai, and Baidu have proposed relevant mass production solutions, and in the realization of intelligent driving in urban scenes, "heavy perception and light map" has become the technical route agreed in the industry.
AI+Transportation: Chuangda released the Rubik "Rubik Cube" large model
On May 18, Chuangda held the Thunder World2023 press conference, releasing the "Rubik Cube" (that is, Rubik) large model, including Rubik Device, Rubik Enterprise, Rubik Studio and Rubik Auto and other products. Taking the field of smart cars as an example, the company integrates large model technology with automotive HMI design software and 3D engine Kanzi, and launches Rubik GeniusCanvas (Genius Canvas) products, which can provide automotive designers with intelligent assistance from concept creation, 3D element design, interaction and vision, special effects and scene production, application development integration, etc. With the help of Rubik GeniusCanvas, the concept creation cycle can be shortened by 70%, from the original 3-4 weeks to about 1 week; The 3D element design cycle can be shortened by 85%, from the original 4-6 weeks to about 3 days, greatly improving the efficiency of creation.
AI+ Transportation: PCITECH released a vertical large model of the transportation industry
PCITECH released PCITECH on June 29, 2023, which is a vertical model for the transportation industry. The model adopts hybrid cloud multi-mode technology, and forms a three-layer architecture of "large model infrastructure", "large model middleware" and "large model industry application" as a whole, so as to carry out reinforcement learning of "pre-training + fine tuning + feedback based on industry data". PCIDO Transportation Model already has mature application capabilities for three scenarios: intelligent customer service, intelligent O&M and emergency command for rail transit, which is expected to greatly promote the digital transformation and intelligent improvement of rail transit. Taking the intelligent rail transit scenario as an example, its essence is to use a general large model and combine industry data, experience and knowledge for further training and fine-tuning of the model. PCIDO deeply understands the characteristics, modes and laws of various rail transit systems to form a "domain knowledge base" and a "domain tool library", on this basis, it provides industry-specific application functions related to the industry, such as intelligent customer service, intelligent operation and maintenance, intelligent energy saving, integrated command, emergency drills, etc.
AI+Network Security: AI triggers multi-layered security requirements
The emergence of large AI models such as ChatGPT will be accompanied by the training and invocation of a large amount of data, exposing the security risks of AI large models themselves. During the training and invocation of large models, there may be risks of illegal data acquisition and information leakage. In the process of opening APIs to the outside world in large models, as the number of open API interfaces increases, the exposed security issues may increase. The automatic generation capability of large language models will greatly improve the efficiency of security intrusion, reduce the technical threshold of security intrusion, improve the automation capability of security intrusion, and reduce the implementation cost of advanced security intrusion. The big language model can empower both security attacks and security defense, but security attacks can achieve attack results as long as they find any breakthrough point, and security defense needs to cover all scenarios to be able to defend successfully, so the big language model in terms of automatic generation ability The bonus to security defense is much smaller than the bonus to security attacks, which is expected to further stimulate the network security protection awareness of downstream users. As large models are more widely used, problems such as information leakage and the use of AI-generated content to commit crimes have become increasingly prominent. The potential security risks of AIGC have received widespread attention, and some countries have taken certain measures to intervene. On April 11, 2023, the Cyberspace Administration of China (CAC) issued the Measures for the Administration of Generative AI Services (Draft for Comments) to promote the healthy development and standardized application of generative AI.
AI+Network Security: Companies in the industry layout security large model research and development
1. Microsoft released Seciruty Copilot to provide productized benchmarking for the implementation of AI+ network security. On March 28, Microsoft released Security Copilot, an AI-enabled security analysis tool, which fuses OpenAI's large-language model GPT-4 with security-related models from Microsoft to provide users with ChatGPT-like assistant functions. The emergence of Microsoft Security Copilot may kick off the disruption of the global cybersecurity industry. 2. Many security vendors have increased the layout of the AI field, and the product landing can be expected. Many domestic security vendors are promoting the layout in the field of AI, and we believe that more security products with large model attributes can be expected in the future, which is expected to promote the reform and development of the network security industry: Convinced of the fact that on May 18, the self-developed security large model security GPT (security GPT) was released. Anheng Information: It is training its secure brain with massive data and multiple AI models. Qianxin: Based on ChatGPT-related technologies and accumulated massive security knowledge and data, Qianxin's proprietary ChatGPT-like security model is being trained. Venustech: Beilu Artificial Intelligence Lab focuses on the research of "artificial intelligence empowers safety" and "safe artificial intelligence" technology. NSFOCUS: The company has carried out research on ChatGPT text-generated content detection tools to avoid using AI technology to generate content phishing and fake news. Tianrongxin: It has already laid out in the field of AI and ChatGPT-like related technologies, and further increased research investment in the field of network security of large model technology, focusing on the detection and analysis of malicious samples, discovery and traceability of attack behavior, security intelligence reasoning and generation, automatic vulnerability mining and evaluation, intelligent security services and operations, etc.
AI+Cyber Security: SINFU released security GPT to benchmark against Microsoft Security Copilot
On May 18, SINFU released its self-developed security large-model security GPT (security GPT), and conducted a multi-scenario on-site demonstration of GPT technology to empower XDR in security operation work. At present, the security GPT function is integrated into the company's XDR products, which are mainly applied in three scenarios: advanced threat detection, security detection investigation, and popular vulnerability investigation. From the results of the live demonstration, the function of security GPT can be aligned with Microsoft Security Copilot to realize interactive security analysis and detection of natural language dialogue, which greatly improves the detection effect and degree of automation compared with the traditional security small model, and reduces the professional requirements for operation personnel.
AI+ public safety: Many domestic companies target generative AI landing detection products
The AI-3300 "Insight" video image authenticity workstation independently developed by Meiyapaike has two modes of intelligent identification and professional identification, covering more than 50 identification algorithms, which can identify face-changing, beautifying, generating faces, copying and tampering with the same picture or different pictures from multiple perspectives, and the recognition effect is at the international leading level, and can provide one-stop image authenticity detection and identification solutions for public security, judicial industry and related fields. With the systematization capability of the industry-based large model with multiple industries, multiple scenarios and parameters of up to 6 billion yuan as the base, Ruijian has launched different detection models for new forgery generation technologies: Software: the audio and video generated content detection tool "Ruian", the image generation content detection tool "Ruitu", and the text generation content detection tool "Ruijian Tuling", covering AIGC text, images, audio and video detection capabilities in an all-round way; Hardware: "Ruian Deep Fake Detection Special Equipment" is the first deep fake detection special equipment integrating software and hardware in China, which solves the two bottlenecks of "safe and controllable processing of sensitive tasks" and "large-scale deployment of live network traffic".
AI+Public Safety: Meyerapec Releases China's First Public Safety Large Model "Tianqing"
As a platform for SDIC Group to participate in the safe and stable development of the national digital economy, Meiyapaike released the "Tianqing" public safety model on June 28, 2023, relying on the industry knowledge system, police standards and industry data formed by more than 20 years of experience in the field of public security, committed to building an intelligent business base in the field of public security and reconstructing a new model of public security business, products and ecology. "Tianqing" Meiya Public Security Big Model is the first public safety intelligent big model in China, with rich public safety industry knowledge, strong police intent recognition, police intelligence analysis, case reasoning and other business understanding and reasoning capabilities, can continue to evolve itself from massive data, realize the closed-loop evolution of industry knowledge, business problems, and feedback solving feedback.
AI+ Office: Kingsoft Office released WPS AI, leading office intelligence
As a domestic office software leader, Kingsoft Office has been accumulating strength in the field of AI for a long time: 1) AI is the company's core strategic direction, Kingsoft has been exploring in the field of AI for a long time, its Chinese image recognition and understanding, document conversion technology has reached the domestic leading level, OCR and machine translation technology level ranks the first echelon in China. 2) Office scenarios involve content generation, collaboration and cloud storage, and office products and services have a good integration with AI. Since April 2023, Kingsoft Office has released WPS AI series products three times. On April 18, Kingsoft Office officially released WPS AI, a generative artificial intelligence application with large language model capabilities, which applied AI functions to light document products for the first time. On May 16, WPS AI launched a new product again, WPS AI embeds large model capabilities into four components: tables, text, presentations, and PDFs, and currently AI capabilities achieve full coverage of the main scenarios of content generation. Recently, the WPSAI product line has been further upgraded, the original light document and light form upgrade has been renamed as smart document and smart form, and AI function has been launched on the form tool on the mobile terminal to launch intelligent form tools.
AI+ office: iFLYTEK, large model + office application, combination of software and hardware
Based on its own core technology, iFLYTEK has actively deployed the smart office track, and has now launched products needed in various office scenarios such as conferences and daily offices, and has achieved outstanding performance at the software and hardware levels. Based on the achievements of Xinghuo's cognitive big model, product-level applications in the office field are launched: smart office books, smart microphones, voice recorders and other product applications are released, meeting minutes, one-click drafting, discourse rounding and other capabilities, the accuracy of speech recognition reaches 98%, and the record can reach 400 words in one minute.
AI+ office: Wondershare Technology, AI helps document editing and drawing and other office scenarios
The company vigorously promotes the research and development of AI technology, and has achieved major version updates of all core products, completed the quality improvement and scenario expansion of video, drawing and other products, and continuously improved product competitiveness and user experience. Wondershare has been following the international technology trend and has been in the field of AI for a long time, and PDFelement, Filmora, EdrawMax and Ink Knife have introduced AI-related functions to cover office scenarios such as drawing and document editing.
AI+ remote sensing: Aerospace grand map, Tianquan big model empowers the application of remote sensing data elements
Aiming to solve the limitations of sample annotation and model generalization under the existing "AI + remote sensing" business model, and strive to build an integrated intelligent remote sensing ecosystem of "segmentation, detection and generation", and empower multiple application fields such as national defense security, land resources, transportation and water conservancy. Segmentation of everything: Break the limitation of sample labeling, realize large-scale multi-source data panoramic fine-grained segmentation, introduce interactive semantic extraction mode, reduce the demand for specific scene modeling knowledge, training calculation, and data labeling, and complete the image segmentation task under a unified framework. Detect everything: break the traditional target detection and solidification method, liberate the model's dependence on labeling, improve the generalization of AI detectors, and flexibly migrate to different tasks; It has self-supervised learning ability, significantly reduces the cost of training and research and development, and can conduct full-image retrieval of objects of interest with "document examination". Generate everything: AI generative models apply new scenarios, simulate and simulate massive remote sensing sample data, and by storing knowledge understanding into a large number of parameters and fine-tuning specific tasks, "generation-segmentation-detection" improves each other, promotes the model to "train itself", and greatly expands the generalization ability of the model.
AI+ remote sensing: Zhongke Star Map, Star Map Intelligent Brain Engine to create a computable digital earth
AI big models are empowering industry applications. Taking the Star Map of China Science and Technology as an example, in August 2022, the Aerospace Academy released the first biometric pre-training large model for cross-modal remote sensing data based on Huawei's MindSpore. Spirit Eye", effectively fills the gap of cross-modal generative pre-training model in the professional field of remote sensing, has been tested in many fields such as national defense and security, real-world 3D, etc., and the measured results in target detection and recognition, feature element classification and other aspects have been significantly improved compared with the general visual model, and is expected to be further promoted to more industries such as land resources, housing and transportation, water conservancy and environmental protection in the future, providing a set of solutions for the integrated application of space, space and ground, and achieving cost reduction and efficiency increase;
Star Map Brain Engine creates a calculable digital earth. The star map intelligent brain engine released by the company is based on the big model of Lingyan, deeply integrates earth big data, analysis and interpretation algorithms, and supercomputers to build a calculable digital earth core engine. Through intensive "intelligent computing", it provides users with intelligent processing of earth data, intelligent perception of earth information, intelligent reconstruction of earth scenes, and opens various computing power interfaces to Internet users, so as to quickly build cloud applications and provide high-quality spatiotemporal content services for earth science research, remote sensing industry applications, and public daily life.
AI+Medical: Weining Health, released WiNEX Copilot in October
Weining Health carried out the R&D and training of WiNGPT, a large-language model in the medical vertical field, in January 2023, and the number of model training parameters reached or will reach 6 billion, 15.6 billion and 65 billion as of April, June and September, and is expected to be officially released in October. WiNGPT will be integrated into WiNEX products using Copilot-assisted diagnostic mode. The first scenarios include medical knowledge Q&A, Internet consultation, medical report summary, etc.
AI+Healthcare: Entrepreneurial Wellcome, BosftGPT will become the AI middle platform
Wellcome signed a tripartite strategic agreement with the Institute of Computer Innovation Technology of Zhejiang University and the Zhejiang Smart Medical Innovation Center to build a BSoftGPT convergence platform to promote the research and development of AI large models in clinical decision support, public health and personal health scenarios. API call + local deployment, aggregate general GPT. At the same time, through the local deployment of embedding vector database as well as the company's own domain knowledge base. Output AI intelligent services to internal and external application scenarios, such as medical services and personal health.
iFLYTEK: Spark Cognitive Big Model empowers educational apps and hardware products
Spark Cognitive Model Empowers Essay Correction: The language ability of the Spark Model combined with the learning machine can provide in-depth understanding of essays, and automatically generate comments and sample essays, so that AI can correct essays like a teacher. The Spark Cognitive Big Model realizes human-like dialogue and empowers the speaking training evaluation scene: The AI learning machine equipped with the Spark Cognitive Big Model can realize free dialogue like a human, and the translation evaluation can be automatically carried out during the dialogue, and it covers all primary and secondary school course standard topics, and the topic can be completely open. Xinghuo Big Model continues to iterate, and upgrades the AI Speaking Assistant to the Spark Language Companion APP: At the Spark Cognitive Big Model V1.5 results conference on June 9, the AI Speaking Assistant was upgraded to the Spark Language Companion APP, which can be opened to all English learning enthusiasts, including primary and secondary school students, college students, and business people, and can simulate oral teachers to provide oral dialogue services for better interaction.
iFLYTEK: The performance of learning machines continues to improve
Whale reference data shows that from December 2022 to May 2023, the brand share of iFLYTEK in the learning machine/learning tablet category of JD.com's e-commerce platform increased from 11.80% in December 22 to 23.60% in May 23, and rose to the first place in the monthly brand market share in May. T20 and T20 Pro equipped with the Spark cognitive large model have become the main products of the learning machine category. Since March 2023, the Xinghuo model learning machine on the Jingdong platform has contributed more than 60% of the revenue of iFLYTEK learning machine category, becoming the main product of C-end learning machine, and is expected to further open up the consumer market space in the future.
Huayu Software released the "Vientiane" legal model to help legal services in the new era
On July 10, 2023, the 2023 Legal Science and Technology Conference with the theme of "Digital Intelligence Integration and Vientiane Update" opened at the China National Convention Center in Beijing, hosted by Huayu Software. At the meeting, the company officially released the "Huayu Vientiane" legal model, which integrates the industry accumulation and technical accumulation of Huayu Software for more than 20 years of legal technology, and is polished based on the leading domestic base model.
The model has eight capabilities that can be empowered in 100+ application scenarios: The Vientiane model has eight capabilities, including legal knowledge retrieval, legal knowledge Q&A, legal element analysis, legal content summary, case fact combing, legal text generation, model compliance control, legal knowledge update, etc., providing a very good experience for judges, prosecutors, lawyers, corporate counsel and other legal professionals in more than 100 business scenarios.
Based on years of industry experience, the company has explored the road of building a large legal model that can be implemented: including research and demonstration, fine training, knowledge enhancement, user and scenario verification. Firstly, the general large model is studied and verified, and the appropriate pedestal model and business scenario are selected through instruction learning, legal concept learning, and expert feedback reinforcement learning. Then the general large model is used for legal fine training, and the knowledge of the large legal model is enhanced by integrating knowledge retrieval, legal map, knowledge vectorization, application plug-ins, etc., and finally the user verification of the legal scene can truly meet the needs of users.
Excerpts from the report:
(This article is for informational purposes only and does not represent any investment advice from us.) For information, please refer to the original report. ）
Selected report source: [Future Think Tank]. 「Link」