Yuncong AI large model joins the battle! Support cross-modal understanding of images and texts, and solve high school entrance examination questions with GPT-4

智物

Author | ZeR0

Editor | Mo Ying

Zhidongxi reported on May 18 that today, Yuncong Technology launched the Congrong model and demonstrated on-site the functions of the Congrong model in question and answer, writing propositions, intention understanding, multiple rounds of dialogue, English writing, machine translation, programming, Ability to understand cross-modal images and text, and do real high school entrance examination questions. The large model is currently in the internal testing stage.

According to the demonstration, the calm model conducted an in-depth understanding of a pre-uploaded book "History of Western Art" with approximately 500 to 600 pages, including asking about the relevant content of the book and summarizing what kind of book it was. , find the picture of the specified requirements, etc.

After testing the ability of Convergence large model, ChatGPT, and GPT-4 to answer the real questions of last year's high school entrance examination, Yuncongcong was faster than ChatGPT. The score of answering history and biology questions was slightly lower than ChatGPT, and the score of answering geography questions was slightly lower than that of ChatGPT. The score is the same as ChatGPT, and the performance in answering political questions, ethics and legal questions is better than ChatGPT. The current level of GPT-4.0 is significantly better than other systems.

At the scene, the Big Model Information and Innovation Ecological Alliance was officially established. Its members include Youked, Xiamen Cultural Tourism, Huawei Shengteng, Nansha Public Control, CSDN, Zhongshu Information Technology, etc., aiming to promote the innovation and development of large model technology. , promote the application of artificial intelligence in various industries.

Yuncong Technology will also cooperate with CCIC, Shenzhou Information, Shenzhen News, Jiadu Technology, Jin Shiyuan, Youzu Networks, and Aiden Technology to launch quality models, financial models, entertainment models, and transportation models. Large models, large manufacturing models, large game models, large medical models, etc.

1. Demonstration of basic abilities: supports cross-modal understanding of images and text, and can answer mid-term exam questions faster than ChatGPT

The dialogue interface of Congrong Large Model is similar to ChatGPT. Ye Mao, director of the technical management department of Yuncong Technology, demonstrated the basic capabilities of the Rongrong large model at the scene.

In terms of Q&A, the reply style of the calm model is concise and practical, which can avoid the pitfalls of some trap questions.

In terms of writing, in response to a high school entrance examination essay question in Sichuan Province last year, "Write an essay of no less than 600 words with the most beautiful color as the title", I calmly wrote an essay on a large model on the spot. After putting forward more modification or limitation requirements to it, such as rewriting it in white, asking for more depth, changing it to write from the perspective of a sick child, the most beautiful color comes from the white coat of a medical staff, etc., the large model can understand the intention. and demonstrated ability to conduct multiple rounds of dialogue.

At the scene of the Congrong model, I successfully completed the creation of an English recruitment notice and modified the signature as required. After asking it to be translated into Chinese, the first reply it gave was a rather blunt literal translation, and then it rewrote the notice in accordance with the new requirement of "rewriting in Chinese habits."

In terms of programming, the calm big model first demonstrated the ability to write code and wrote a piece of code for quick sorting.

Ask it a relatively professional topic, "What is the time complexity of this code?" There is nothing wrong with the calm and large model's reply.

Requires adding code comments, it can also be completed quickly.

More difficult, let it write code that C programmers can understand. It not only completed the task, but also added comments according to the previous requirements.

In terms of reading comprehension, the leisurely large model supports the understanding of long documents or multi-document aggregation, and supports cross-modal understanding of images and text.

Yuncong technical staff uploaded a book called "The History of Western Art" with about 500 or 600 pages in advance. Congrong Model had an in-depth understanding of the content of the book, and then based on the contents of the book Interact with users in the knowledge category, including asking about the content of the book, summarizing what kind of book it is, etc.

As shown in the figure below, the middle part of the interface is the text of the book, and the right side is the dialogue interactive interface. The calm model has prepared some question samples. After clicking on a sample, it will generate a reply in real time with a reply. Some clues about the content. Clicking on these clues will link to the corresponding passage location in the book. With easy-to-use large models, users can provide image descriptions and quickly locate them in books.

Yuncong Technology tested the Congrong large model, ChatGPT and GPT-4 to determine their ability to solve real questions in the 2022 high school entrance examination in various subjects.

Judging from the test results, Yuncong Calm answers questions faster than ChatGPT. Its score in answering history and biology questions is slightly lower than that of ChatGPT. Its score in answering geography questions is the same as ChatGPT. Its score in answering political questions, ethics and legal questions is the same as that of ChatGPT. Performance is better than ChatGPT. The current level of GPT-4.0 is significantly better than other systems.

2. Industry application examples: urban transportation management, equipment maintenance, financial operations, policy solutions

To make the Congrong basic large model truly useful, it is necessary to build an industry large model.

Yao Zhiqiang, co-founder of Yuncong Technology and general manager of Guangdong Company, shared the application of the calm large model in grassroots governance scenarios, such as the One Language Intelligent Office for public services, the Intelligent Governance Elf for civil servants and grid teams, and application development Programming assistant for users and integrated command for urban transportation management center.

In the scene of the city operation management center, a demonstration of the city operation intelligent large-screen AI assistant was played on the big screen. If a commander commands "Move the fifth screen to the middle," the AI ​​will understand the command and execute it quickly.

The commander then asked, "Are there many people and cars around here?" The large AI model then used its multi-modal capabilities to automatically analyze the flow of people and vehicles, select the video footage of the routes with the most people and cars around it, and tell "Based on the video content." "The traffic situation around the lake is good, but there are vehicles occupying the road and parking illegally. At the same time, some places are crowded with people. It is recommended to arrange security personnel to maintain order."

The commander continued to ask: "What suggestions do you have for citizens to travel today?" The large AI model responded to the weather conditions and gave suggestions on "suitable for travel" and "good sun protection".

An example for equipment maintenance management scenarios is the intelligent maintenance accompanying system developed by Yuncong Technology based on industry large models.

The maintenance plan comes from two sources: one is based on equipment classification, equipment historical maintenance records and recent production plans, etc. will automatically form a maintenance plan; second, the inspection robot or probe will be used during the robot inspection process. It detects a problem somewhere, which may be temporary, and it knows what the fault is, what the likelihood is, and whether it needs to send a repair request.

In the process of preparing for maintenance, by studying maintenance manuals, maintenance records and expert advice, large machines gradually developed into "master master" level maintenance skills. Before maintenance, the system will provide a complete preview course including pictures, texts, audio and video, and key points of maintenance, so that maintenance engineers can preview the possible faults of the equipment to be repaired, what causes the faults, and what tools should be brought and what methods should be used. Repair, replacement of spare parts, etc.

During the maintenance process, Yuncong Intelligent Maintenance Accompanying System provides two typical solutions: one is "guidance", the engineer can ask the large model how to repair the equipment failure; the other is "supervision", the large model can monitor the engineer Whether the maintenance actions are standardized and whether any important repairs are missed.

When the maintenance work order is closed, there is no need for workers to write it themselves. The system will automatically analyze the entire video record to form a graphic maintenance record, accompanied by the core video content of the maintenance, for maintenance and repair. and provide a source of knowledge for subsequent repairs.

For financial scenarios, large models can improve the efficiency of bank internal business operations. The large-scale model can transform banks' massive data resources into more valuable information, breaking through the supply bottleneck of professional knowledge and helping to improve banks' capabilities in inclusive finance, bank operations, and serving the real economy. The performance of the AI ​​virtual account manager in answering professional questions related to financial management was demonstrated on site.

Compared with general large models, industry large models optimized with local knowledge base can provide more professional and rigorous answers and avoid random fabrication.

Take the Customs Policy AI Elf as an example. The Customs Policy AI Elf is based on more than 2,000 materials from the General Administration of Customs to form a local knowledge base. Through semantic segmentation, semantic retrieval, prompt learning and other technologies, it constructs accurate information for a large-scale model. Prompt words can then give full play to the capabilities of large models and provide users with accurate policy answers.

Yuncong Technology has also incubated a number of large-scale model application entrepreneurial projects internally. For example, the Damai Digital Human Live Broadcast Platform can realize functions such as intelligent construction of live broadcast rooms and provision of live broadcast pre-heating corpus.

For educational scenarios, the intelligent education AI wizard can form a self-generated question bank based on existing course syllabus, question bank and other basic models, generate customized exercises and study plans based on students' daily performance, and can automatically generate corresponding comprehensive evaluations based on student performance. Analysis to reduce teachers’ daily workload.

Conclusion: In the next few years, technology will continue to unlock scenarios

After demonstrating basic general abilities such as language, mathematics, and reasoning, large models are moving towards the industry, showing their application potential in professional knowledge fields such as finance, law, medicine, and policy.

Zhou Xi, chairman and general manager of Yuncong Technology, believes that large-scale models will subvert traditional interaction methods and are mainly displayed in three forms: question and answer, companionship and hosting. Among them, "question and answer" refers to the current GPT; "accompaniment" means that AI will be like a friend, accompanying you to perform many things; "hosting" means that one thing is mainly left to AI to do, similar to "on-hook training" in online games ". Once the “hosting” stage is reached, people are freed up to do more meaningful and interesting things.

He said that with the entire platform framework and the basic capabilities constructed through the basic large model, the skill package of the industry large model can be continuously added, and a more powerful industry system can be constructed. This system can serve all walks of life including To G, To B, and To C.

In Zhou Xi’s view, without a strong basic large model, directly building an industry large model will not have long-term sustainable vitality, because if you want to make the industry large model practical enough, you need to train the basic large model in turn. If you want In order for large industry models to be used in mass production in the industry, efficiency and cost control must be achieved to the extreme, and ultimate optimization requires mastering the basic large models.

The development of future industries depends on emerging technological breakthroughs, and industry applications do not simply rely on creativity, but require technical support. In the next few years, technology will continue to open up new application scenarios, and scenario parties will also continue to try to reconstruct industry efficiency and experience.

The above is the detailed content of Yuncong AI large model joins the battle! Support cross-modal understanding of images and texts, and solve high school entrance examination questions with GPT-4. For more information, please follow other related articles on the PHP Chinese website!