(Report producer/author: Guotai Junan Securities, Li Muhua, Qi Jiahong, Li Bolun)
1.AI+ office is at the heart of the AIGC wave and promises to profoundly change the way work is done
1.1.AI+ office is the core of the AIGC wave, and overseas giants lead the trend
AI+ office is the core beneficiary direction in this wave of AIGC. The tipping point of this AIGC wave is that ChatGPT, a text creation tool based on natural language processing large model technology, has rapidly grown into a phenomenal application that is popular around the world, and then the application of multimodal large models based on the processing of images, videos, audio, etc. has also been rapidly promoted. AIGC, or generative artificial intelligence, is naturally an AI technology for text, audio, video, images and other content independent creation scenarios, so it can directly improve the product capabilities of existing office software, thereby promoting the iterative upgrade of office software.
Microsoft launches Microsoft 365 Copilot subscription service to reinvent the office experience with AIGC technology. On March 16, 2023, Microsoft officially released the Microsoft 365 Copilot subscription service, which is backed by the Copilot engine, which uses the three core basic technologies of Microsoft 365 Apps, Microsoft Graph and Large Language Model. Microsoft 365 Apps is a series of common Microsoft office software such as Word, Excel, PowerPoint, Outlook, Teams, etc. Microsoft Graph is a secure intelligent gateway that can help access user business data accumulated on Microsoft 365 Apps from users' documents, emails, meetings, chats, calendars, and more; The Large Language Model (LLM) is a creative engine that parses and generates human-readable text using OpenAI's ChatGPT and the newly released GPT4 model. After the user uses natural language input prompt words in Microsoft 365 Apps, the prompt words will be trained by the Copilot system, which can improve the quality of the prompt words, so that the prompt words can be executed, and the most important part of the basic training process is to call the data previously generated by the user in the Microsoft Graph to understand and improve the quality of the prompt words, and then send the improved prompt words to LLM, the aforementioned process is called Do preprocessing. LLM responds to the prompt and performs post-processing, which will again call the user data through the Microsoft Graph for training, after passing the security, compliance and privacy reviews, generate a feedback response, and finally output the feedback reply to the user through the Copilot system and drive the APP to execute the relevant commands. Through such a complete set of processing processes, users can give instructions through natural language in office software such as Word, and then the office software will automatically present a feedback draft with obvious personal information characteristics, which greatly improves the user's office efficiency.
Microsoft 365 Copilot revolutionizes the way users work and helps improve productivity in office creation. On the one hand, Microsoft 365 Copilot will be embedded in all kinds of office software frequently used by users, including Word, Excel, PowerPoint, Outlook, Teams, etc., so as to help users liberate themselves from tedious transactional work and allow users to focus more on creative work, thereby improving office efficiency. On the other hand, a new business chat application scenario has been introduced, which can generate new content according to user needs based on the user's past accumulation of various business data, and improve the efficiency of creation, for example, after the user puts forward the "please tell team members how to update the sales strategy" instruction, the system will generate a new sales strategy based on the user's past meeting discussion records, chat records, email communication content, etc. Users can modify, retain, or discard the generated content, making Word more creative, Excel more analytical, PowerPoint more expressive, Outlook more efficient, and Teams more collaborative.
Microsoft 365 Copilot uses natural language as a channel for human-computer interaction, lowering the threshold for using Office office software. According to Microsoft's press conference, more than 90% of PowerPoint and Excel functions are not used by users. We believe that this phenomenon is not because these functions are useless, but because the threshold for interaction through menu buttons or function formulas in the past is high, and most users cannot directly call relevant functions without systematic learning. After the introduction of Copilot, all office software will form a chat box on the right side, and the user will enter the effect they want to achieve (such as a certain scheduling, adding a certain animation effect or performing a special operation) in the way of natural language chat, the software will directly implement the relevant functions, so that rich software functions are applied, greatly reducing the user's threshold for use, helping to further improve the scale and stickiness of users, thereby increasing the number of paid monthly active users.
The launch of Microsoft 365 Copilot will bring new revenue streams to Microsoft and increase ARPU for paying users. At present, Microsoft 365 Copilot is still in the internal testing stage, according to the technology media The Information, Microsoft is testing the AI-capable Microsoft 365 Copilot subscription service with AI capabilities to more than 600 large institutional customers, including Bank of America, Walmart, Ford and Accenture, on top of the paid subscription Microsoft 365 annual membership, per 1,000 employees The Copilot service requires an additional annual fee of $100,000 and an additional ARPU of $100, and pilot customers are more willing to pay than Microsoft had previously anticipated. According to Microsoft's financial data, we estimate that the ARPU subscribed by Microsoft 365 institutions in 2022 is about $103, and the pilot price of the Copilot service is expected to promote ARPU to nearly double.
Adobe introduces Firefly, a generative AI model set, that demonstrates the power of design creation. On March 21, 2023, Adobe officially launched Firefly, the generative AI model set, which was subsequently tested in Photoshop applications. At present, it mainly has the following capabilities: 1) text to image, Firefly can automatically generate the required image according to the user's prompt words, such as entering the "sky to add aurora" command, you can automatically add aurora to the sky part of the image; 2) Generate filling ability, which can automatically expand, fill, and expand into a richer image according to the content in the original image; 3) Text Effects: Display specific textures to Art Font Medium. Firefly will also explore a range of AI features such as automatic video processing, text generation 3D/vector files, and full-color images for sketches. The launch of Firefly is expected to greatly improve the efficiency of design creation, lower the threshold of design creation users, and change the current creative design industry pattern.
1.2. Domestic office software manufacturers are catching up, and AIGC functions have achieved rapid iteration
Kingsoft Office is one of the pioneers of domestic office software manufacturers to explore the application of AI technology. In 2017, AI was first raised to the strategic level within the Kingsoft office, and Yao Dong set up an AI team of hundreds of people, mainly responsible for AI algorithm improvement and the landing of engineering products. In 2018, Kingsoft Office formally proposed the development strategy of "multi-screen, content, cloud, and AI", and the AI strategy was publicly unveiled. In the first two years of the establishment of the AI team, the team mainly emphasized the accumulation of AI research and development capabilities, including algorithm capabilities, engineering capabilities, data collection and analysis capabilities, etc. In the following two years, the company began to pay more attention to the productization of AI technology, added a series of AI functions to the company's products, and launched a series of AI-assisted office functions including intelligent beautification, intelligent proofreading, intelligent assisted writing, full-text translation, image recognition, etc., to help improve the office efficiency of users. For example, intelligent beautification can help users change PPT templates and color matching, unified fonts and typography, etc. as a whole while providing basic text; Intelligent proofreading can identify and proofread typos, multiple missing words, punctuation errors, grammatical errors, sensitive word errors, etc. in documents with one click; Intelligent assisted writing can automatically generate text paragraphs or sentence supplementation according to the outline to help users lay the draft, and the tens of millions of articles in its reserve corpus are all from authoritative media and government public websites, and have many applications in scenarios such as official document writing. As of July 2021, WPS's intelligent beautification monthly active users exceeded one million, the number of monthly proofreading words for intelligent proofreading exceeded 7 billion, and intelligently generated content accounted for 33.6% of the overall content resources in the cloud. At this stage and in the future, the focus of the company's development strategy has become to accelerate the industrialization of AI products and strengthen the positive effect on the company's revenue growth, so as to achieve long-term sustainable development.
Kingsoft Office accesses multiple large model suppliers to more accurately meet the AI creation needs of users. At present, Kingsoft Office's products have been connected to many large models such as MiniMax, Baidu Wenxin, and CopyDo, and it is expected to access new large models in the future. At present, the choice of large models on the market is constantly increasing, including the large model products of giants such as Baidu, Alibaba, and iFLYTEK, as well as the products of start-ups. The current performance of each large model has its own strengths, and none of them has absolute leading product performance, so it can access multiple large models at the same time, and match and call different large models for content creation according to different user needs, which can better meet the creative needs of users. For example, CpoyDone is a large model specially built for scene marketing, which can generate copywriting, pictures, and video content with rich product types and massive content platform style, because WPS can prioritize docking with the CopyDone large model in the copywriting marketing scenario; MiniMax is a multi-modal large model, which has more advantages in scenarios such as chatting with China; Baidu Wenxin can better meet the needs of users in the fields of literary creation with its rich corpus; As the scheduling and matching center of large models, WPS can coordinate the "1+1>2" effect of various large models.
WPS light documents based on AI large models are the first to enter the closed beta stage, demonstrating strong text creation capabilities. On April 18, 2023, Kingsoft Office released a demo video of WPS AI, officially announcing that WPS AI will be embedded in Kingsoft Office's entire line of products. The first to enter the closed beta stage is the AI-capable WPS Light Document, which is an online content collaborative editing tool for light office products such as Notion AI, which can automatically generate press releases, weekly work reports, operation plans, etc. with the help of large models. It is also possible to enable multiple rounds of dialogue and continuous discussion on a certain topic; Existing documents can also be rewritten, expanded, shortened, polished, etc.; You can also generate subject summaries, article outlines, and more for specified documents. WPS AI shows strong text creation capabilities, which is expected to improve user creation efficiency.
Foxit Software integrates ChatGPT in the overseas version of PDF Editor Cloud to provide AIGC functions to users. Foxit Software is a leading PDF office software company, and its Foxit PDF Editor Cloud is an online PDF editor through which users can read and edit PDF files online. On April 25, 2023, the overseas version of Foxit PDF Editor Cloud successfully integrated ChatGPT, and paid users can currently mainly use the following AIGC functions: 1) Document summary, quickly generate concise and accurate summaries according to the content of the document; 2) Document rewriting, rewriting without changing the original meaning of the document, optimizing the expression of the document, and improving readability; 3) Document translation, translation of selected content or abstracts into the specified language; 4) Document content Q&A, you can ask questions about the content of interest in the dialog box, which will generate answers according to the document content and locate to the corresponding paragraph. Through the above functions, the user's document reading efficiency and the work efficiency in the multilingual working environment can be significantly improved. As an online editing software, Foxit PDF Editor Cloud's AI function has the advantage of rapid iteration, and it is expected to launch a new version every two weeks, continuously expanding the application of AIGC technology in digital scenarios in the document field, optimizing user experience, and providing user payment rate.
AIGC features are expected to boost Foxit Software ARPU. Since the AI features provided by Foxit PDF Editor Cloud are based on ChatGPT, there is a user/word limit, and document rewriting is limited to 100 pages per user per month; Document translation is limited to 50 instructions per user per day, with a maximum of 2,000 words per instruction; Document Content Q&A is limited to 50 instructions or questions per user per day; Once the limit is exceeded, the user is charged an additional fee, which helps increase the user's ARPU. Wondershare Technology has a complete creative office software product line and is the leader of domestic creative office software. The company's products cover four categories: video creativity, drawing creativity, document creativity and practical tools, forming core creative office software products with wide influence and large user base such as Wanxing Meowying, Wondershare Broadcast Explosion, Edraw Brain Map, Wondershare Love Painting, and Wondershare PDF.
Wondershare Technology has increased investment in AI technology research and development for a long time, and AI products have gradually entered the landing period. Since 2020, the company has gained insight into the development prospects of AI technology, and quickly set up a high-quality technical research and development team of hundreds of people, in the past two years, in the video generation algorithm, image generation algorithm, GAN generation algorithm, audio generation algorithm and other AI cutting-edge algorithms and other fields have achieved a number of technical research and development results, since the second half of 2022, the company's video creativity, document creativity, drawing creativity and other major product lines have successively released new versions with AI functions, and launched new products incorporating a number of AI technologies. In the video creative product line, in 2022, the company completed the update of the major version of Wondershare Miaoying, launched AI function sets such as AI segmentation, AI keying, AI noise reduction, AI audio reorganization, AI copywriting generation, and released a digital human marketing video creation tool based on AIGC technology at the end of 2022, which can realize AI functions such as Wensheng video and virtual human live broadcast; In the drawing creative product line, in November 2022, the company launched a new AIGC image generation tool - Wondershare Love Painting, which supports three creative modes: AI text painting, AI drawing with drawing, and AI stick figure, and users can enter a text description to obtain AI painting works in a variety of art styles, or convert them into paintings after entering pictures; In the document creative product line, the company released a new version of Wondershare PDF, introducing AI technology, adding professional functions such as translation to improve user experience.
Wondershare Technology actively embraces large model technology to promote the rapid improvement of AI product capabilities. In February 2023, the company announced that its overseas video creative software Wondershare Filmora has been fully integrated into the commercial services of OpneAI, the parent company of ChatGPT, opened the commercial account permission of the GPT-4 model in March, and signed a cloud service framework agreement with Microsoft in April, the two parties will comprehensively deepen cooperation in the field of cloud services and AI technology, give priority to the use of Microsoft's new products in the future, and are currently continuing to promote a number of overseas products to access the GPT model. On March 31, 2023, the company released a "real person" overseas marketing short video tool based on the AIGC large model - Wondershare Broadcast Explosion, and launched the desktop version of the product in June, Wondershare Broadcast based on the AIGC large model can provide more than 120 languages of copywriting script rapid generation ability and more than 60 nationalities of digital human broadcast ability, can also carry out digital human customization services, compared with the traditional video production mode, Wondershare Broadcast can reduce the cost to 1/5 of the original investment, and greatly improve the user's production efficiency.
AIGC's new products bring new profit points to Wondershare Technology and are expected to increase user ARPU. The company's traditional Wondershare Meowing individual annual membership is priced at 269 yuan, and the annual fee for 5-year members is further reduced to 120 yuan; The annual membership price of Wondershare Miaoying Enterprise is 3299 yuan, which can support 5 devices to use at the same time, which is equivalent to the annual price of 660 yuan for a single device. Among the new AIGC products launched by the company, the annual membership price of Wondershare is 1688 yuan, which is significantly higher than the pricing of traditional software products; Wondershare Love Painting pays according to the number of creations, with a total of 5 yuan for 10 picture creations, and a total cost of 20 yuan for the cheapest 100 picture creations, and the pay-per-use method has a higher growth limit on the customer demand side. Overall, the pricing of AIGC's new products is higher than that of traditional creative software, which is expected to bring new profit points to Wondershare Technology and increase user ARPU.
2. Intelligent driving is an important scenario for the landing of AI large models
2.1. Autonomous driving: AI large models help us improve the efficiency of covering road conditions with low probability
2.1.1. The coverage of road conditions with a small probability is the core problem of autonomous driving
Since the consequences of an accident are extremely serious, autonomous driving is a very sensitive scenario for small probability situations. Since traffic accidents will have very serious consequences, even 99.99% reliability is unacceptable for OEMs until the responsibility is clear, because this may mean that for every 10,000 vehicles sold, there may be one accident. Industry characteristics determine that in order to achieve autonomous driving, it is necessary to effectively cover long-tail scenarios. The accumulation of test miles is a prerequisite for effectively covering road conditions with a small probability. According to GAC's forecast, in order to achieve the coverage of long-tail scenarios required for L4 autonomous driving, at least 1 billion test scenarios need to be completed, and the minimum test mileage needs to be 1 billion kilometers, which are 100,000 times and 10,000 times that of L2 autonomous driving, respectively.
Previously, there were two main ways to accumulate test miles. One is data collection through autonomous fleets, represented by Waymo; One is data collection through private cars, represented by Tesla.
The method of covering a small probability of road conditions through road testing in autonomous vehicles is relatively inefficient. Waymo is the dominant player in the field of autonomous driving, but in the past many years, Waymo's takeover frequency has not converged in terms of perception problems, pedestrian problems, software problems, etc. (based on California road test report). There is no doubt that Waymo's self-driving capabilities are increasing year by year. Then, Waymo's apparent "regression" in software problems, pedestrian problems, etc. can only be explained by covering more small probability road conditions. For example, after testing in relatively simple scenarios such as highways, Waymo will gradually expand the road test site to more difficult urban streets.
The crowdsourcing method can improve the coverage efficiency of low-probability road conditions to a certain extent. Tesla replaces the test fleet with shadow mode. Shadow mode is essentially crowdsourcing to solve the problem of rapid accumulation of scenes. In this mode, even when a person is driving, Tesla's Autopilot system calculates what it will do and then compares it with the person's choice. If the autonomous driving system and the human choice are inconsistent, such data is pooled and then handed over to the engineer to determine whether the choice of the autonomous driving system is reasonable. In March 2020, Tesla applied for a patent to obtain self-driving training data from its fleet. Since Tesla has far more cars than the self-driving test fleet, shadow mode can achieve faster accumulation of driving long-tail scenarios, and the results obtained are also more statistically significant. By the end of 2019, Tesla had delivered a total of 850,000 vehicles equipped with autonomous driving assistance hardware, and the cumulative mileage in the AP-activated state had exceeded 2 billion kilometers, far exceeding its competitors (Waymo's 20 million kilometers). As Tesla ownership continues to rise, the gap between other competitors and Tesla in terms of data accumulation and long-tail scenario coverage will widen.
2.1.2. Large models are of great significance for covering road conditions with low probability
2.1.2.1. Large models can greatly improve the efficiency of scene generation and annotation
With the emergence of AI large models, our efficiency in covering road conditions with a small probability of autonomous driving is expected to be greatly improved, and this efficiency improvement comes from at least two aspects:
Scene generation
Using AI large models for scene generation is a new idea to cover low-probability road conditions. Compared with pure road testing, direct scene generation and combining simulation results with drive testing are of great benefit to quickly achieve road condition coverage. For example, DriveGPT Snow Lake Hailuo has been released, which can achieve three capabilities: generate many scene sequences according to probability, and each scene sequence is an actual road condition that may appear in the future; When all scene sequences are generated, the self-driving trajectory of the most interest in the scene can be quantified. It can realize the future driving trajectory of the car while generating the scene; Based on the generated trajectory, the output of the decision logic chain is realized.
It is worth noting that Snow Lake Hairuo introduced a reinforcement learning mechanism similar to human feedback in GPT series models. That is, the judgment and decision of the system and the driver are compared, and if the comparison results are consistent, the system will be given a high score, otherwise it will be given a low score. This is similar to the pattern of Tesla FSD.
Data labeling
In addition to scene generation, AI large models can also play an important role in automatic labeling. In the era of AI 1.0, data labeling mainly relies on manual labor, resulting in long labeling time and high cost of data labeling. Especially in the field of autonomous driving, due to the complex road conditions, there is a large number of labeling needs. Automatic annotation can be realized based on large models, thereby greatly reducing costs and improving efficiency. For example, the Snow Fox Hailuo of Milli Smart Mobility opens up its scene recognition capabilities to the public. Previously, it cost about 5 yuan to label a picture using the ordinary annotation scheme, while DriveGPT Snow Lake Hairuo only needed 0.5 yuan, which greatly saved costs.
2.1.2.2. Although large models are difficult to completely solve the problem of small probability road conditions, they are still of great significance to the autonomous driving industry
Of course, it must be admitted that large models still cannot help us solve 100% of the problems caused by small probability road conditions. Large model capabilities come from deep learning, not reinforcement learning. From the perspective of technical routes, the big model is "deep learning + human feedback reinforcement learning". To test the impact of reinforcement learning on model capabilities, Open AI ran multiple-choice sections of a series of exams based on the GPT-4 base model and the GPT-4 model with reinforcement learning. The results show that in all exams, the average score of the basic GPT-4 model is 73.7%, while the average score of the model after the introduction of reinforcement learning is 74.0%, which means that reinforcement learning does not significantly change the ability of the base model, in other words, the ability of the large model comes from the model itself. According to Open AI, reinforcement learning is more about making the output of the model more in line with human intentions and habits than about improving the model's ability (and sometimes even lowering the model's test score).
Since the big model has not got rid of the deep learning framework, this means that statistics are still behind the AI at this stage, and the residual problem cannot be completely solved. In other words, the problem of "unsolvable capabilities" still cannot be fundamentally solved, and we still cannot achieve 100% correctness, and we can only improve safety by covering more low-probability road conditions. Large models are theoretically difficult to help us achieve 100% coverage of road conditions with a small probability. In essence, the use of AI large models for road condition generation can greatly improve efficiency, but it is still similar to exhaustive. It is not theoretically possible to achieve full coverage of low-probability road conditions through exhaustive methods, and the essential reason is that "the road conditions themselves are an infinite scene." Imagine that if we were to open a lockbox, we only had to try it all from "000" to "999", and the box must have been opened; In the same way, in chess games, each move can be "dropped" is a finite set, in other words, all possibilities can also be traversed, so these two scenarios are "finite scenes", while the open road autonomous driving scene is an "infinite scene".
2.2.AI Help improve the interactive experience of intelligent cockpit
The interactive nature of intelligent cockpits will undoubtedly continue to rise. From a necessity perspective: The automotive industry is shifting from a seller's market to a buyer's market, and the core drivers of industry evolution are shifting from technology and products to consumer demand. The traditional automobile industry has gone through a hundred years, with the high maturity of the industry, this market is gradually changing from a seller's market to a buyer's market, and the key factor for the development of the industry has also shifted from technological breakthroughs and product polishing to changes in consumer demand.
2.3.AI Drive the efficiency of vehicle R&D and design
As project cycles shrink, automotive R&D efficiency is becoming increasingly important. The development cycle of automobiles is gradually shortening, which makes the supplier's project cycle greatly compressed, the previous project may be 2-3 years, and now it may be more than 1 year or even less than 1 year, while the OEM customization demand is increasing, shorter development cycle and more customization demand puts forward higher requirements for Tier1's intelligent manufacturing capabilities. With the gradual increase of automatic driving function modules, the mileage that needs to be tested increases rapidly, and there is not enough time for road testing, and because of the safety involved, the test link itself cannot be simplified, so the efficiency of design and testing is gradually becoming an important factor restricting the rapid and timely delivery of the project to a certain extent.
AI large models are of great significance to the efficiency of automotive designers. For example, Genius Canvas combines language capabilities, visual rendering and special effects production capabilities, which can help designers complete the creative process by assisting concept creation, assisting 3D element design, assisting special effects code generation and assisting scene construction and production, thereby optimizing workflow and improving designer work efficiency. In terms of concept creation, it was able to reduce the work cycle from 3-4 weeks to 1 week, saving 70% of the time. In terms of 3D element design, it can shorten the 4~6 week work cycle to 3 days, saving 85% of the time. In terms of special effects and scene production, it can save 90% of time.
2.4. Domestic intelligent driving companies actively embrace the new trend of AI
2.4.1. Chuangda: Launched Genius Canvas to improve HMI interactive experience
Chuangda attaches great importance to AI technology and uses kanzi to promote the development of intelligent cockpits. In 2022, Chuangda announced the establishment of a joint venture with Horizon to focus on the intelligent driving track. kanzi is a tool with powerful real-time 3D rendering capabilities. The intelligent cockpit 3.0 launched by Chuangda uses Kanzi for Android, a new technology, which makes the Android system and Kanzi perfectly docked, realizing 3D records, customizable real scene navigation, real-time interface personalization, cross-screen and cross-system applications.
Chuangda uses Kanzi to realize multi-screen interaction in the intelligent cockpit. Since intelligent driving involves human-machine co-driving, intelligent cars carry more and more driver information, external environment information, vehicle information, etc., and need to present more space and regions to users, and ensure good interaction with the driver. Based on the multi-screen linkage supported by Kanzi for Android, maps can be cross-screen and 3D navigation can be presented in an all-round way during navigation. At the end of navigation, the map can be zoomed out from the co-pilot screen to the center control screen. Multiple options to suit more individual needs. Genius Canvas empowers the development of the automotive industry and creates a new HMI interactive experience. One of Genius Canvas's tools is the Big Model Engine, which turns ideas and concepts into copy, further into ideas and works, and ultimately into applications through technical means. The second tool for Genius Canvas comes from the KANZI product. After combining Kanzi with large models, it can quickly create colorful Kanzi HMI concept effects and special effects, build a variety of 3D models and image libraries, and realize real-time preview functions in the vehicle system. At present, more than 100 models around the world have chosen Kanzi, and tens of millions of production models equipped with Kanzi technology land every year.
2.4.2. Desay SV: Partnering with universities to localize large models and empower autonomous driving
Desay SV cooperates with universities to promote the localization deployment of large models. Desay SV has cooperated with Sun Yat-sen University, Nanyang University of Technology and other universities to build technical support by trying and laying out digital virtual assistants based on large models, automatic annotation of image data, automatic scene creation, automatic programming, etc., and related solutions have been unveiled at the Shanghai Auto Show. In the process of AI large model localization, Desay SV can provide customers with differentiated and comprehensive technical support and solutions. AI large model technology can be perfectly integrated with Desay SV's existing technology. For example, AI large models can achieve more accurate data supplementation and prediction in perception fusion, perception prediction and planning, so as to give more help to autonomous driving in behavior prediction and give more control choices.
2.4.3. ArcSoft Technology: Commercially available AIGC products have been released
The company has released AIGC products to help small B customers significantly reduce the cost of product display. 1) The commercial shooting market space exceeds 50 billion yuan, and the rainbow soft solution can greatly reduce the dependence on models, reduce the cost of commodity display, and realize the replacement of the original plan. 2) Using the current market scheme in many details of distortion and distortion, the rainbow soft solution can make the product display "what you see is what you get". 3) The company plans to launch a static product display solution in 2023, including the generation of still images of products with background, as well as the generation of digital model images of products and digital models, and plans to launch dynamic video and 3D content in the future. Business model: Similar to the company's mobile phone and automobile business, AIGC's business model is divided into two parts: membership service fee and production flow fee. 1) In the membership service section, the company will open different functions according to different membership levels, such as different scene libraries and model libraries; In addition, companies can differentiate custom development for value-added APIs. 2) In the generated traffic fee section, the company will directly price the generated content based on the actual computing power consumption.
2.4.4. Jingwei Hengrun: Independently developed driving simulation test software and launched intelligent cockpit AI products
In terms of autonomous driving simulation, Jingwei Hengrun independently developed simulation software to help driving testing. Jingwei Hengrun independently developed ModelBase, a comprehensive driving test simulation software, which can be used in the design, testing and verification of vehicle electronic control systems and ADAS systems for passenger cars and commercial vehicles. It involves the full development cycle of electronic control systems, including early algorithm simulation testing, hardware-in-the-loop testing of controllers, semi-physical bench testing, and vehicle-in-the-loop testing. At present, this software has been applied to more than 50 projects such as FAW, Dongfeng and NIO. In terms of intelligent cockpit interaction, Jingwei Hengrun has developed a series of products such as music rhythm ambient lights based on AI technology. Jingwei Hengrun Music Rhythm Ambient Light has two modes: real-time song feature recognition and offline song feature recognition. Among them, the relevant functions of offline song feature recognition mode are implemented based on the AI music style classification algorithm and the AI music paragraph division algorithm. Through the recognition of music characteristics, it provides a rich combination of effects for the rhythm of ambient lighting music and improves user experience.
3. Finance is one of the core scenarios for AI implementation
3.1. The financial industry dares to try new technologies, which is one of the core scenarios for AI landing
Compared with other industries, the financial industry has three characteristics in the application of new technologies. The first characteristic is that it attaches great importance to the development of new technologies and dares to try. Because the day-to-day business activities of the financial industry involve a large number of transactions, small advances in technology may bring huge benefits to customers, so financial institutions are sensitive to new technologies and will actively pursue the application of new technologies in daily business. The second feature is that the IT budget of the financial industry is sufficient, and the cost performance is not as sensitive as other industries, and in the eyes of IT Party B, the customers of Party A of financial institutions are often the best piece of cake. The third characteristic is that financial institutions have extremely high requirements for system stability and data security. This feature often conflicts with the first feature, but system security is always the bottom line of financial institutions, and above this bottom line, the application of new technologies will be pursued. Whether it is securities, banking or insurance, once the core system fails for more than a certain period of time, it is easy to receive a regulatory letter, and the relevant IT department leaders need to assume management responsibilities. For example, on May 16, 2022, the centralized trading system of China Merchants Securities failed, and on September 8, it received a regulatory letter from the Shenzhen Stock Exchange.
3.2.AI Technology can greatly improve the efficiency and user experience of the financial industry
In December 2022, McKinsey, a global management consulting firm, released McKinsey China Financial CEO Quarterly - "Today's Technology Reshapes Tomorrow's Finance: Seven Technologies Shaping the Future of the Global Financial Industry", which summarizes seven new technologies that are reshaping the future of the financial industry, including artificial intelligence. One is artificial intelligence. From a single point of trial to a comprehensive application, deeply integrating all aspects of business and operation; Provide additional value for financial institutions in terms of project/product landing speed, overall work efficiency, comprehensive cost control, and security guarantee. The second is cloud computing. The trend of large-scale cloud migration is accelerating, and cloud computing and edge computing complement each other. The flexible arrangement of front-end outlets and back-end computing power will unlock a series of application scenarios with high customer perception. The third is the metaverse and comprehensive virtual technology. Virtual perception builds virtual worlds and reshapes customer services and internal operations; Continuous developments in spatial computing technology, AR/VR/MR technology will redefine customer experience and internal operations. The fourth is blockchain and Web 3.0. Internet paradigm iteration, subverting future business models; Blockchains, digital assets, and decentralized architectures will disrupt the old portal platform business model and even give rise to new financial services sectors. Fifth, the next generation of communications. High-bandwidth, low-latency, and strong security data transmission enabling technical solutions, and IoT technology continues to promote the implementation of new use cases; High-throughput satellite networks, 5G/6G, low-energy LANs and other communication technologies from the sky to the ground will develop and integrate with each other, which will enable faster and safer financial products and applications. Sixth, the next generation of integrated development. Civilian development, flexible deployment, intelligent assistance, and automatic development will change the traditional technology-intensive development process, further reduce the development threshold, and scientific and technological capabilities are no longer the unique moat of technology enterprises. The seventh is trust architecture and digital identity. Build a digital trust system and consolidate the cornerstone of fintech security; Zero trust architecture, digital identity, privacy engineering and other technologies ensure financial and privacy security and enhance trust.
3.3. Listed companies are launching AI products
3.3.1. Hang Seng Electronics: Launched intelligent investment research products and developed financial models
Hang Seng Electronics launched intelligent investment research products based on the big model, which includes three sub-products. The first one is called CHAT, through which you can ask for a variety of data. It is like a financial information data informant, the user can ask F9, ask the market, ask research reports, public information, ask opinion extraction, etc., its underlying key technology uses search plus large models, through such technology to call the entire Hang Seng Clustered financial information database, so as to achieve voice control of tens of thousands. The second product, called ChatMiner, is a document miner. For example, users have a document of their own, and after uploading, they can ask questions about this document, and ChatMiner can answer the questions according to the content mentioned in this article. The underlying key technology is a vector database plus a large model. The third product, WarrenQ, is a one-stop digital intelligence investment and research terminal. WarrenQ has a lot of scenes, functions, the large-model product ChatMiner is also in it, and then readers, cloud notes, original citations and traceability, calculation boards, valuation models in it, and online sharing brain maps, etc. have been fully opened up in the investment research scene, so it is a one-stop investment research platform.
The plug-in layer can solve the problem of data immediacy. The first column is NL2SQL. For example, using CHAT to check the market of Hang Seng Electronics, it uses the interface of this large model to return a paragraph at the same time, and adjusts the NL2SQL interface, and goes to the data source library to find out the time series of the market of Hang Seng Electronics, and turns it into a K-line chart return, and the user can see the most timely and updated market; The second search interface is also important. The first NL2SQL solves the problem of instant queryability of data in time series format, and the search interface solves text data, because the first interface cannot get the latest news, events, news research reports, announcements. The underlying technology of ChatGPT is the vector database, which involves how to vectorize, query and store similarity of a large number of documents in the private domain. This plug-in layer is very important, it is a very important supporting force for the financial field to make vertical domain products. Combined with the financial model trained by Hang Seng Electronics, it can be used to do a variety of applications, including intelligent investment research, intelligent investment advisory, wealth management and other services. Hang Seng Electronics' large model will be available for trial until September 30, and will be further optimized by the end of the year. The capabilities of Hang Seng's large models specifically for the financial industry have been upgraded to a usable level, and the trial interface will be opened on September 30. By the end of the year, the reasoning performance will be further optimized, and the logic capability will be further upgraded, so that it and photon matching can be unified to form an AI pass-through application system.
3.3.2. Flush: Launch of artificial intelligence and virtual human products
As early as 2013, Flush began to lay out the field of artificial intelligence, first promoted the financial search engine Love Money, and by 2019, the whole business fully promoted AI, and has accumulated a variety of AI products.
(1) i Qicai Investment Research Platform: i Qicai Investment Research Platform provides multi-dimensional stock, fund, bond data, investors enter natural language questions, search for the data and information they want. In addition, there are functions such as conditional stock selection, research reports, chart selection strategies, product search, short-term review, strategy backtesting, macro economy and so on. Flush i Qicai Intelligent Head Digital Human is committed to using artificial intelligence technology multi-modal interaction and rich media expression to solve users' personalized investment problems, improve users' investment capabilities, and assist in completing investment goals.
(2) iFind: iFind big financial data terminal is an intelligent terminal that integrates financial data professional consulting and investment research and analysis tools, currently covering all domestic securities and futures companies, more than 80% of funds and commercial banks, most of the media universities listed companies private equity institutions' products cover the world's major capital market stocks, bonds, foreign exchange, commodities, funds and other varieties, with more than 6 million macro industry indicators, an annual increase of more than 500,000, research reports more than 100,000 news data sources, for users For comprehensive market information, iPhone's AI algorithm provides users with innovative applications such as intelligent prediction, intelligent search and intelligent context, making the user experience more efficient and convenient. In 2023, iFind will use AI technology, aigc series, to further improve user experience and work efficiency.
(3) AI short video platform: At present, short video is the mainstream way for financial information users to obtain information, which is limited by pain points such as high production threshold and lack of financial data, resulting in low production efficiency of short video. In response to the aforementioned pain points, Flush has developed an AI short video platform, which is a video production and publishing platform based on artificial intelligence technology, through the integration of advanced AI technology, automatic processing of video materials, including editing dubbing subtitles, etc., combined with the cool visual display ability of data, so that users can quickly produce high-quality short videos, and the platform also provides a variety of rich templates and themes to automatically produce personalized short videos. Flush AI Short Video provides a fast and fun video creation experience, helping to create a digital service system that transforms from text to short video.
(4) Digital human interaction all-in-one machine: digital human has professional knowledge personality and emotions comparable to real people, and can assist in completing tasks such as business consultation and handling, marketing promotion and publicity in banking, securities, operators, government affairs, medical care, education and other industry service scenarios, and improve customer experience and marketing success rate. What we see now is a digital person with expertise in the financial field, which can provide real-time financial data to users.
(5) Flush virtual exhibition hall: The virtual exhibition hall is a product that Flush smoothly uses core technologies such as virtual human, artificial intelligence, and cloud computing to help enterprises create quality conveniently and efficiently, realize the overall display and interaction of enterprise products and services, and empower enterprises to promote and recommend, popularize science education and other functions. The virtual exhibition hall uses 3D panoramic display lines with special effects such as sound and light animation, which can bring visitors an immersive experience with full immersion. The virtual exhibition hall breaks through the limitations of time and space, and the application of 3D panoramic display of online products and services, combined with special effects such as sound and light animation, can bring visitors an immersive experience with full immersion. The virtual exhibition hall breaks through the limitations of time and space and immerses itself in an immersive experience.
(6) Flower exploration: the main function of the upper gastrointestinal examination function plate is to use intraoperative navigation prompts for the examination site, missed detection site, visual field clarity and lesion body prompt. At the same time, the system will automatically intercept the site and lesion pictures to save, the main function of the lower gastrointestinal examination function plate is surgical technology to identify anatomical positions such as ileocecal valve and ileal end, visual field clarity assessment and reminder of abnormal lesions to prompt, our product has obtained the second class of medical device certificate, and reached cooperation with a number of well-known medical institutions.
3.3.3. Lexus: Serving Japanese financial companies through AI technology
The company's main customers are Japanese financial and insurance companies. The company has established a long-term and stable partnership with excellent first-class software contractors in Japan. Due to the small number of first-class software contractors in Japan, after the company establishes a stable cooperative relationship with them, it can effectively reduce the company's sales expenses and relationship maintenance costs, and improve the efficiency of cooperation. In the process of cooperation with Japanese first-class software contractors, the company has accumulated rich experience in finance, real estate, telecommunications, e-commerce and other industries, and has won a good reputation among customers. At present, the company's largest customer is Nomura Research Institute, the world's top financial service technology provider, and in 2019, Nomura Research ranked tenth in the world financial technology rankings, and is a very excellent financial technology company in the world. At present, in addition to serving its parent company, Nomura Securities, Nomura Research Institute is also actively exporting IT capabilities, especially AI capabilities. The company fully participates in the construction of IT systems for overseas financial enterprises and implements AI capabilities. The company has participated in the development of many core business systems in the financial industry, including online trading system, customer relationship management system, etc., the core system of insurance business, business support system, sales platform system, online banking platform for bank customers, pension management system, etc., basically achieving full coverage of all systems in the financial industry. Among them, the working paper system based on OCR and NLP has been run online in 41 customers including brokerages and fund companies, using the Transformer model and CV target detection algorithm in deep learning, and the functions of text error correction, document consistency comparison, prospectus review, bond offering review, multi-document cross-review, and general document verification based on deep machine learning have also been completed, and experience tests have begun in many securities firms, and customers have been upgraded one after another.
4.AI With the blessing, design and industrial software will achieve cost reduction and efficiency increase
4.1.AI It is beneficial to improve design efficiency
4.1.1. AIGC lowers the threshold for using design software
Current AI-assisted capabilities are limited to making suggestions for users and replacing some repetitive design work, reducing some of the burden on designers and not lowering the threshold for software use. Using the AI-assisted features in the newly released AutoCAD 2024 as an example, 'Activity Insights' records all user actions on drawing files and suggests workflows and actions, and 'Smart Blocks' automatically places new blocks based on where previous drawings were placed. These features are of limited help to designers and do not lower the barrier to entry. ChatGPT generates CAD code according to the developer's natural language instructions and explains it accordingly. Currently in the CAD space, ChatGPT supports the Visual LISP/AutoLISP language developed by Autodesk, the Maya core scripting language MEL, the common scripting language MAXScript for 3ds Max related products, and the Visual C++ language for the AutoCAD platform secondary development package ObjectAX.
Microsoft code platform GitHub released the latest version of Copilot X, a programming assistance Copilot, to implement AI voice interaction-assisted programming. On March 23, 2023, Microsoft's code hosting platform GitHub released a new version of Copilot X, a programming aid tool, with a new version of GPT-4. GitHub CEO Thomas Dohmke said that while autocompletion has greatly improved developer productivity, the new Copilot X can increase developer productivity by a factor of 10. Industrial design software will also have its own "Copilot", significantly lowering the barrier to entry for software use and increasing productivity. Industrial design software has a high threshold for use, but in the future, users can directly use natural language to put forward requirements and limit the call of AI for code writing and drawing, which greatly reduces the difficulty of use. At the same time, designers can directly use AI to eliminate repetitive design work and improve work efficiency.
4.1.2. AIGC will further enhance the capabilities of generative design
AI can automatically generate a large number of diverse design solutions that meet the requirements, so that designers can explore more design options in less time and improve design efficiency. Traditional design methods rely on a "model and then analyze" cycle, but in generative techniques, AI can quickly generate a large number of CAD solutions that meet the requirements and optimize them without human intervention, based on user requirements and constraints such as material type, functional requirements, performance constraints, cost constraints, and other information. Designers can explore a large number of possibilities in a short period of time, quickly narrow down their options and choose the best solution. Streamlining the design process also helps designers make decisions faster and be more productive. Generative techniques minimize costs and optimize performance. Generative technology creates optimized product designs based on user requirements and constraints, rather than making geometry and then validating, so designs are optimized for goals such as minimum cost and weight. This method can effectively reduce the use of materials and reduce costs.
4.1.3.AI Break down the barriers between 2D and 3D to achieve accurate conversion
AI multimodal large models are expected to break barriers, effectively convert 2D drawings and 3D BIM models, and improve design efficiency. At present, there is still a large number of BIM mold turning needs in China, although there are plug-ins and algorithms that can realize two-dimensional drawing mold turning three-dimensional BIM model, but the mold turning effect is generally poor, requiring a lot of manual correction, AI large model after training is expected to improve the accuracy and fineness of mold turning, replace manual mold turning, and achieve cost reduction and efficiency increase.
4.1.4.AI Enabling EDA to reduce costs and increase efficiency
Synopsys launched and paid off with the first AI EDA suite, potentially leveraging AIGC to write code in the future. In April 2023, Synopsys, the world's leading EDA vendor, announced the industry's first full-stack AI-driven EDA solution Synopsys.ai covering the design, verification, test, and analog circuit design phases to help customers continue to innovate and achieve higher quality designs faster while reducing costs. Synopsys.ai has been adopted by many leading companies, including IBM, NVIDIA, and Microsoft, and has achieved remarkable results. Renesas achieved a 10x improvement in reducing functional coverage dead zones and improving IP verification efficiency by 30 percent. SK Hynix has reduced the chip size of advanced technology by 5%. At present, engineers still write the C language of chip manufacturing, and it may be assisted or even replaced by AIGC in the future.
4.2. AIGC will effectively improve industrial production efficiency
4.2.1. Siemens and Microsoft join forces to improve industrial productivity with AIGC
AIGC powers the further development of industrial AI. At present, the enhancement of AI to the manufacturing stage of industrial products mainly lies in the manufacturing execution and management process of AI algorithms, and the generation and inference capabilities of AIGC will bring significant improvements to AI applications and further optimize the execution and management processes. Siemens and Microsoft have partnered to set the benchmark for AIGC in industry. In April 2023, Siemens announced a partnership with Microsoft to use generative artificial intelligence (AIGC) in multiple ways to improve its industrial control workflows, continuously improve efficiency, and drive innovation. Siemens Teamcenter creates new applications for Microsoft Teams to enhance cross-functional collaboration. The two companies will integrate Siemens' product lifecycle management (PLM) software Teamcenter® with Microsoft's collaborative platform Teams, language models in Azure OpenAI services, and other Azure AI capabilities. Service engineers or production operators can use natural language to record and report product design or quality issues via mobile devices. At the same time, through Azure OpenAI's services, the application can parse the aforementioned informal speech data, automatically create summary reports, and send them to the appropriate design, engineering or manufacturing experts in Teamcenter. Combined with Siemens Teamcenter provides additional support for workers who do not have access to PLM tools, allowing them to participate in the design and manufacturing process in a simple way.
4.2.2. AIGC can optimize and generate 3D printing solutions, lowering the barrier to entry
It is found that ChatGPT can fine-tune and optimize 3D printing parameters, and even provide suitable 3D printing solutions, effectively reducing the work threshold and improving efficiency. Gcode is a programming language used in the field of 3D printing to provide 3D printers with specific instructions on how to print objects. But writing Gcode requires in-depth knowledge of the 3D printing process, which is time-consuming and error-prone to writing manually. Generating optimized Gcode ensures product quality and reduces lengthy trial and error times, saving material and time. The researchers found that ChatGPT successfully optimized 15 print parameters in 1 hour and explained the reason for each parameter change, a task that would have taken about three weeks to complete.
4.3. Mainstream players in design and industrial software have accelerated the deployment of AI and have achieved certain results
4.3.1. Guanglianda: AIGC technology has been used in its core products
The company has laid out AI in 2015, established AI technology as the company's core technology, and continued to focus on investment and achieve results for many years. In terms of cost business, we have broken through interactive generation technology based on deep learning and used large model technology to provide services such as intelligent group pricing and intelligent takeoff. In terms of construction business, the labor face recognition terminal has achieved mass production, and a number of CV security hazard recognition algorithms have been integrated into the construction Hummingbird box products, helping the Hummingbird system to be successfully selected as a typical case of AI intelligent safety inspection application scenarios at the construction site of the Ministry of Industry and Information Technology's "National Artificial Intelligence Innovation Application Pilot Zone "Wisdom Empowerment 100 Scenes". In terms of digital construction internationalization, MagiCAD released AI-assisted design capabilities, continuing to expand its leading edge in core regions and maintaining good growth momentum in key expansion regions such as the United Kingdom, Germany, and Italy.
Companies are also deploying generative AI, in the design business, the conceptual design stage is preceded by designers doing ideas, and then drawing them one by one, and in the future, AI simulators' ideas can be quickly generated various sketches. For the field of intelligent design, the company has a dedicated team to explore, and has carried out preliminary trials on some projects, but the overall is still in a relatively early stage. For example, the function of AI forced row (layout of buildings according to the mandatory specifications of the building), in the stage of land auction, dozens and hundreds of solutions have greater value than the comparison of only a few solutions at present. At present, the strong ranking tool has entered user verification, and the real-time sunshine analysis performance is leading in China. There are precedents for the use of AI in the construction industry. Autodesk partnered with DAISY to improve the efficiency of the construction design process. DaisyAI is the first artificial intelligence (AI)-powered wood design CAD software that produces the best design to meet specifications in 10 minutes, saving engineers 2-3 hours per day and reducing wood waste by 80%. Autodesk's Kratos research project uses AI methods to quickly evaluate structural designs for multiple materials, including concrete. In 2022, Kratos partnered with DAISY to use Kratos to calculate load-bearing walls in timber structures and output the results to Daisy to generate detailed floor plans, reducing the amount of concrete used in the foundation and reducing construction costs.
4.3.2. Zhongwang Software: AI-driven generative design capabilities have been introduced
When it comes to CAD, companies can develop built-in generative design capabilities based on existing data. AIGC allows engineers to specify their requirements and goals to the software, allowing them to automatically generate a large number of design proposals. AI-driven generative design capabilities are now available in major CAD products such as Siemens Solid Edge, PTC Creo, and Autodesk FUSION 360. In terms of CAE, AI can empower simulation optimization, improve simulation efficiency, and help companies train industrial AI models. By partnering with Ansys Twin Builder, Microsoft Project Bonsai can run virtual models of hundreds of machines or applications simultaneously and feed the data generated by these digital twins directly into the brain to optimize it. Using a large number of virtual models can reduce training time, reduce costs, and learn to understand all possible scenarios, increasing the accuracy of industrial AI models.
4.3.3. Central Control Technology: Self-developed China's first process industry process simulation and design platform
APEX's massive data enables large model training to optimize engineering installations. In November 2022, the company officially released its self-developed APEX, which became the first process simulation and design platform for the process industry in China. Based on the mechanism model, it opens up the data flow from process design to plant operation, realizes functions such as process simulation, process bottleneck analysis and operation optimization, and provides intelligent operation solutions from engineering design, factory digital twin, production operation to full life cycle operation and maintenance. The massive data obtained through APEX operation will also be used in AI models to optimize the plant and further improve plant efficiency.
5. Under the background of .AI the big model, cybersecurity opportunities and industries coexist, and all parties are accelerating their layout
5.1.AI Catalyzed by large models, opportunities and challenges coexist in the cybersecurity industry
5.1.1. "Security", "AI Security" and "Safe AI" are equally important
Cybersecurity threats are on the rise. With the popularization of Internet applications, the number of corresponding cyber threats has increased, and their complexity has also increased relatively, which has brought great challenges to network security. Nowadays, mobile devices, Internet of Things, and cloud computing are becoming more and more popular in enterprises, and the attack surface is also increasing. In addition, hackers can use AI to constantly morph viruses/malware, which traditional static defense solutions may not be able to effectively detect and block. In addition, cyberattack-as-a-service has made cyberattacks widespread, and attackers do not need to have strong hacking knowledge themselves but can also obtain attack tools by paying cryptocurrency. The role of AI in cybersecurity is to help organizations reduce the risk of intrusions and improve their overall security posture. AI learns from past data to identify patterns and trends, which is then used to predict future attacks. AI-powered systems can also be configured to automatically respond to threats and combat cyber threats in a faster time. As the enterprise attack surface continues to evolve, hundreds of billions of time-varying signals must be processed to properly calculate risk. To address this unprecedented challenge, AI tools and methods, such as neural networks, are evolving to help information security teams protect sensitive information, reduce the risk of intrusion, reduce security operating costs, and improve security posture with more effective and efficient threat detection and mitigation capabilities. In addition, with the rise of the trend of large models, network security product capabilities, platform operation capabilities, and security services are expected to usher in comprehensive optimization and upgrading.
In summary, the cybersecurity industry under the wave of new technologies needs to consider four layers of security. The first step is to do a good job in traditional network security protection, and the second type is to use artificial intelligence-related algorithms or large models to enhance network security products and optimize and empower services. At this stage, it is necessary to ensure that the integration of AI systems and network security products can be effectively integrated in the process of empowering network security by AI large models. Third, it is necessary to ensure the inherent security of the AI capabilities output by large models (under the premise that the current problems of alignment and interpretability have not been effectively solved, at least ensure that security risks cannot be amplified by AI). Finally, the security protection of the large model itself is also very important, because as the cornerstone of AI capability output, its own security and stability have a fundamental supporting role.
5.1.2. The combination of cybersecurity and AI technology has natural advantages
To understand the role of AI in cybersecurity, we must first review the construction ideas of cybersecurity products themselves. The construction idea of network security is based on the perspective of red and blue confrontation, that is, corresponding protection according to the chronological order of hacker attacks. The second step is to protect the perimeter, that is, to deploy corresponding security protection according to the exposed content to strengthen the boundary, then to carry out regional control, that is, the construction of monitoring means, and the last step is to strengthen control. The whole process is based on before, during and after the construction of cybersecurity.
The capabilities of cybersecurity point products can be enhanced based on AI algorithms. Firewall, IDS/IPS and other products are required for asset combing, security vulnerability investigation, and boundary protection in the early stage, while machine learning techniques can improve malicious software detection by combining large amounts of data from anti-malware components on the host, network, and cloud compared to traditional software-driven or manual methods. Deep learning uses large amounts of data to train deep neural networks, which can also help protect against attacks. For example, Google uses deep learning to detect hard-to-detect image-based emails, emails with hidden content, and communications from newly formed domains, which can help detect sophisticated phishing attacks, including spam-related Internet traffic patterns; Deep learning architectures can be used to uncover hidden or latent patterns and become more environmentally sensitive over time, which can help identify zero-day vulnerabilities or activities, such as natural language processing can scan source code for dangerous files and flag them, "generative adversarial networks" can learn to mimic any data distribution, and can also be useful in identifying complex flaws.
5.1.3. The large model brings great changes to the supply and demand side of the cybersecurity industry
The wide application of large language model technology can empower many links of the network security industry, and may even bring subversive changes to some links. The essence of the large model represented by GPT is to understand the language intention and assign tasks according to the intention, so as to realize the ability of dialogue, calculation, mapping, etc., and the relatively large work links with the language system and process work can be empowered by the large model. At the same time, due to the difficulty of project implementation and cost-effective considerations, large models are more suitable for links with larger scale and more labor. From the perspective of network security vendor supply capabilities, the semantic understanding and code generation capabilities of large models can effectively empower security products and services. Network security logs are a language system in the computer field, and GPT has a natural advantage in understanding logs after pre-training in Github. At the same time, security operation involves a lot of process work, and some links require more manpower, and the application of large models is expected to reduce the number of security service personnel in the security operation center (SOC) scenario and achieve cost reduction and efficiency. For example, in user behavior analysis (UEBA), traditional SIEM is based on characteristics and rules, while user behavior goes beyond rules and correlations, and can study attacker behavior patterns through the empowerment of large models, so as to more effectively detect insider threats, targeted attacks and fraud; As another example, before an attacker can encrypt data, the SIEM may detect ransomware alerts and automatically perform response actions on the affected system, and the code generation capabilities of large models can improve the automatic response of the system.
5.1.4. The AI capabilities of large model output must have native security
While large models can bring a big leap in network security performance, it is also important to consider the security of the AI capabilities exported by large models. The security of the output capability of large models, that is, "safe AI", is as important as traditional network security protection in its industrial application process, essentially because AI large models as a tool should help people rather than replace people or cause harm to human society. Based on the large model of security, the effectiveness of its empowerment of traditional network security can be fully guaranteed. The idea of "AI Safety" is to build a large model of safety, and model security needs to consider three major factors. AI Safety specifically includes: alignment, interpreferability, and robustness. Among them, alignment requires that the goals of the AI system be consistent with human values and interests, but there are also three challenges in the implementation of AI alignment, one is to select the appropriate values, the other is to encode the values into the AI system, and the third is to select the appropriate training data; Explainability refers to the understanding of the internal mechanisms of the model and the understanding of the results of the model; Robustness can be understood as the tolerance of the model to changes in the data.
Data leak prevention (DLP) is the absolute number one security requirement for enterprise customers today. While an employee at Samsung Semiconductor Factory was entering source code into ChatGPT to identify and eliminate errors and optimize the program, he also inadvertently leaked confidential production data to the public through GPT, and another Samsung employee used an AI chatbot to summarize the meeting minutes, resulting in the leakage of the meeting minutes. THE ABOVE DATA LEAKS ARE JUST THE TIP OF THE ICEBERG, ACCORDING TO A SURVEY RELEASED BY CYBERHAVEN ON MARCH 21, 8.2% OF EMPLOYEES HAVE USED CHATGPT IN THE WORKPLACE, 6.5% HAVE PASTED COMPANY DATA, 3.1% HAVE FED COMPANY SENSITIVE DATA TO CHATGPT, AND SENSITIVE DATA ACCOUNTS FOR 11% OF THE TOTAL DATA PASTED BY EMPLOYEES.
5.1.5. The security of the large model itself is equally important
The situation becomes more complex during the system integration phase of large AI models. The system integration of AI applications involves not only the security risks of AI technology itself, but also the combination of on-board systems, networks, software, and hardware, and these threats include the confidentiality of AI data and models, code vulnerabilities, AI bias, etc. Therefore, in view of the hidden dangers in the use of large models, firewall companies focusing on large models have appeared overseas. Arthur Sheild is the first firewall for large language models, helping companies deploy large model applications such as ChatGPT faster and more securely, and ensure model deployment and operation. Arthur Sheild capabilities may include: protection against PII or sensitive data leakage, protection against poisonous and offensive or problematic language generation, protection against hallucinations, malicious prompts by users, and protection against malicious injection.
In addition to the traditional enterprise security stack protection, the large model also has some security protection requirements that distinguish it from other software development companies. In terms of traditional protection, such as using Cloudflare, Auth0 to manage traffic and user identities. ChatGPT experienced information leaks and downtime caused by Redis bugs in March, which created demand for APM and observability vendors such as Datadog and Sumo Logic. The large model also has some security protection requirements that are different from other software development companies, such as Prompt injection attacks, which put forward higher requirements for security companies. Overseas, many companies focused on Security for AI have emerged, such as the HiddenLayer MLSEC platform, a software-based, non-intrusive platform that monitors the input and output of machine learning (ML) algorithms to thwart adversarial attacks and provide visibility into the health and security of ML assets. The platform is based on a cloud architecture that does not require access to customer data or intellectual property to protect customers' ML assets without compromising speed, efficiency, and reliability. At the same time, the platform helps customers maintain ML algorithms, protect them from attacks such as inference, data poisoning, evasion, or model injection, and prevent sensitive training data from being exposed.
5.2.AI The cybersecurity market is growing rapidly, and overseas giants are rapidly deployed
Global investment in AI cybersecurity is growing rapidly. The increasing popularity of the Internet of Things, increasing concerns about data protection, and the continuous escalation of network offensive and defensive confrontation are driving the development of AI in the cybersecurity industry, and more and more cybersecurity vendors are increasing their investment in the AI security market to seize the commanding heights of "AI + Security". According to MarketsAndMarkets' research data, AI will be valued at $22.4 billion in cybersecurity in 2023 and is expected to reach $60.6 billion by 2028, with a CAGR of 21.9%, while MarketsAndMarkets believes that North America will account for the largest share of the AI cybersecurity market during the forecast period.
Overseas top security vendors have also continued to increase the application of AI-related products. IBM Security Qradar Suite products embed AI and automation to accelerate security teams' response to each step of the attack chain; CrowdStrike and Cribl jointly launch CrowdStream, which aims to provide faster and more accurate cybersecurity data collection and analysis; Fortinet's FortiXDR is the first AI-powered incident investigation response solution that fully automates security operations processes typically handled by experienced security analysts, enabling faster threat mitigation across a broad attack surface.
5.3. Domestic manufacturers have accumulated AI capabilities for a long time, and the direction of large models has increased
5.3.1. Qianxin: Grasp new opportunities in AI and explore the blue ocean of cybersecurity
Capitalize on the wave of new technologies and launch new products and services that best fit the market. For generative artificial intelligence (AIGC) technology, the company combines the concept of "endogenous security", uses years of massive security big data and knowledge accumulation, actively trains proprietary ChatGPT-like security big models, and plans to achieve wide application in the fields of security product development, threat detection, vulnerability mining, security operation and automation, offensive and defensive confrontation, anti-virus, threat intelligence analysis and operation, and network-related crime analysis. Qianxin has made great achievements in the direction of AI technology empowerment security, and its research results are widely used in the company's products, and deep learning and machine learning technologies have been successfully used in data mining, anomaly detection, and complex network analysis.
5.3.2. Convinced Convincing: AI layout is forward-looking, and the first-mover advantage in the field of large models is significant
Adhere to the concept of AI First, and empower cloud product upgrades with AI technology research and application. The company adopts the concept of "AI First" to build a cloud full-product system, and AI technology is required for hyper-convergence, managed cloud services, desktop cloud, storage and database management. Based on this new product architecture, it can comprehensively improve the capabilities of performance, reliability, security, and operation management, and this capability is called AFOPS, AIRUN, and AISEC. AISEC guarantees security when you go to the cloud; AIRUN makes it more convenient, effective and faster for customers to use the cloud; AIOPS enables increased maintenance and automation when using the cloud, without the need for more manpower.
5.3.3. Tianrongxin: AI and product deep integration, continuous improvement of competitiveness
Innovation is integrated into AI to continuously enhance the core competitiveness of products. Tianrongxin has an early layout in the field of AI security, and as early as 2019, it jointly released the first domestic white paper "The Next Generation of Firewalls Incorporating Artificial Intelligence" with IDC. The company mainly uses AI technology for threat intelligence analysis, network application classification, unknown threat detection, etc., and has been practically applied to the company's products, released products include firewall, intrusion prevention, zombie worm, sandbox, big data analysis, situation awareness, EDR, data leakage prevention, etc. Subscription revenue of $313 million in 2022 came from at least half of which came from AI-produced knowledge. At the same time, the company has been deploying and using large-class models since 2020, and has trained a basic model for security services, and is training a code automation model for non-core modules (such as automated testing) to improve development efficiency.
5.3.4. NSFOCUS: Build an AI lab and increase GPT intelligent applications
Actively explore new AI technologies represented by AISecOps, SecXOps, and the Security Knowledge Graph. The company established eight laboratories in 2016, of which Tianshu Lab focuses on AI research and has accumulated a number of research results, including: the release of the safety knowledge graph, the launch of AI SecXOps concepts and products, and the release of white papers in cooperation with university research institutions. The company attaches great importance to the influence of large language models represented by ChatGPT and GPT-4.0 on the security industry, and has carried out research on GPT-like technologies in the fields of security attack and defense, security operation, and GPT content recognition; With long-term accumulated offensive and defensive knowledge, operational data and threat intelligence, intelligent security service robots based on GPT-like technology will be released in the third quarter of 2023, aiming to apply large model capabilities to code security, improve security operation efficiency, and improve the accuracy of security analysis and judgment.
5.3.5. Anheng Information: Complete data security system and advantages empowered by AI
Anheng's big data and data security related products make extensive use of AI technology, and its product strength is improving year by year. Data security is a major strategic direction of Anheng, and data security and artificial intelligence also have a natural combination, the company's product polishing has effectively used the advantages of AI technology, and the production capacity continues to improve, especially in AiSort data security grading, AiMask data desensitization, AiGate data security gateway, AiThink behavior analysis and AiTrust zero trust and other product systems.
5.3.6. Venustech: Panxiaogu helps integrate AI security R&D and operation
AI-enabled security has always been the focus of Venustech's exploration. The company's self-developed artificial intelligence security modeling and empowerment platform has been widely adopted by threat detection, security big data analysis, threat intelligence, UEBA and other products, comprehensively improving the capabilities of security data governance, security model construction, model security detection, model reasoning empowerment, etc., realizing the rapid construction of artificial intelligence applications based on ModelOps and AIOps, model life cycle management and multiple empowerment, and helping traffic detection technology and threat detection technology to achieve intelligence. Promote the company's network security products to the advancement of automation and intelligence. Venustech released the "PanguBot" safe intelligent lifeform in 2022. Based on the intelligent security operation solution of artificial intelligence technology, the company has built an integrated platform for AI security R&D and operation with the goal of artificial intelligent security service and operation (AISecOps) throughout the life cycle. "PanguBot" is provided by Venustech Pangu artificial intelligence platform model operation computing power and environment, with Chat as the window, the application of natural language model trained based on the security operation special corpus, can receive text, voice, pictures, video and other ways of information input, feedback to users through the form of text and pictures, and can integrate various operation tools to achieve security analysis and disposal automation, which has become a strong support for Venustech's intelligent security operation.
(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」