The fourth paradigm released the text generation software for the live demonstration of AI big model ＂Shi Shuo＂
There is another entrant in the big model of artificial intelligence in China-"Formula 3.0" released by the fourth paradigm.
On April 26th, Dai Wenyuan, the founder and CEO of the Fourth Paradigm, demonstrated the various capabilities of "Shi Shuo" on the spot. The Beijing News Shell Finance reporter noticed that compared with other large models, "Shi Shuo" not only demonstrated the capabilities of text generation, picture generation and coding, but also highlighted various application scenarios of AI at the B-end in the real machine demonstration, such as automatically judging tasks after inputting text and actively asking questions to users, so that users can directly execute the "packing" goal with text.
Taking this opportunity, Dai Wenyuan put forward AIGS strategy (AI-Generated SoftwareAI Generation Software): Reconstruct enterprise software with generative AI. He said that "Style Theory" will be positioned as a new development platform based on multi-modal large model, which will improve the experience and development efficiency of enterprise software and realize AIGS. "C-end products are approaching the upper limit of user experience, while enterprise-level software at B-end is often a very complex execution system, and it is not too much to pile up more than ten layers of menus and thousands of functions. At present, the extremely complex interactive experience of these B-end software and the extremely low development efficiency brought by complexity leave enough reconstruction for generative AI.
What is the ability of "expression"? Writing, drawing, programming and then combining the three to "container"
The Beijing News Shell Finance reporter saw at the scene that the fourth paradigm has prepared several scenarios, including AI dialogue, AI group chat summary, AI photo, AI scheduled meeting schedule, and AI applications in finance, medical care, aviation and other fields.
"Style Theory" first shows the daily functions of copywriting, such as generating travel plans, writing and developing large language models, etc. It also shows the ability of continuous dialogue. In the demonstration of script writing, Shi Shuo first wrote a script of Wandering Earth 3. When Dai Wenyuan asked Shi Shuo to write a script of Wandering Earth 4 on this basis, and added the elements of Fourth Paradigm Company, Shi Shuo also fulfilled the requirements. "It can make the artificial intelligence of Fourth Paradigm Technology Company apply to movies, for example, make artificial intelligence become a movie.
In addition, "Shi Shuo" also shows the functions of drawing pictures and writing codes, such as "drawing a basketball shoe with bright colors" and "writing a code for multiplying two numbers with VBA".
It is worth noting that the fourth paradigm ingeniously "integrates" the above three capabilities, and demonstrates the process of "packing" the container on the spot. In the real machine demonstration, Dai Wenyuan gave the instruction "Help me to carry out a packing task", and then the "formula" showed its "thinking" process in the interactive interface, indicating "I think this is a packing task", giving the "task goal", and actively asking Dai Wenyuan to input the container size, quantity and other constraints. Finally, the animation of the packing demonstration was generated, which took about 1 minute, which was undoubtedly faster than manual writing.
"In the past, it was difficult to call the functions of enterprise software through human language (natural language). Now, when we have stronger semantic understanding and generation capabilities, coupled with the ability of GPT task translation, task distribution and reasoning, we can call the functions through better dialog interaction, and we no longer need to find a function under a menu directory of more than ten levels." Dai Wenyuan said.
In addition, for the B-end application scenario, the fourth paradigm also shows the ability of "expression" to understand pictures, such as making them "find the same" after inputting pictures.
In Dai Wenyuan’s view, to achieve AIGS, a big model does not necessarily need to be a generalist with extensive knowledge and a decathlon champion. What is more important is that the model has the ability of Copilot and CoT(chain of thoughts).
Shell Finance reporter learned that, in fact, "Style Theory" added multimodal and Copilot in the 2.0 stage, because the data in many enterprise software is multimodal, and Copilot can translate human instructions into which API to call in the background. In the previously released Demo of "Shi Shuo" 2.0, store employees sent instructions to "Shi Shuo" through interactive means such as voice and text. After the "Shi Shuo" was understood, the networked store monitoring software called out the pictures of the kitchen without a mask, and directly output the pictures to the employees in the form of a dialog box.
Dai Wenyuan said that it has been of great value for the big model to call the built-in functions and data of the software to complete the task in a dialog box. However, employees will also face complex tasks when using enterprise software, requiring people to perform functions in sequence. Therefore, "Formula Theory" 3.0 emphasizes Copilot and thinking chain CoT, which has stronger reasoning ability. After learning a lot of data and "Raiders", it can form intermediate logical reasoning steps, so as to split and perform complex work.
How to choose the development direction of the incoming big model? The fourth paradigm AI should take the "AIGS strategy"
The fourth paradigm told the Beijing News Shell Finance reporter that when BERT (the natural language processing framework released by Google in the early years) came out, the Paradigm Research Institute had begun to pay attention to and put into this technical field, and it was more clear that GPT3 was going in this direction after it came out. ChatGPT craze has helped the company the most, that is, the confidence of the whole market has been adjusted from zero, and the investment in certainty has become greater, and then it is to promote products and commercialization.
The company also revealed the iterative process of "Shi Shuo"-"Shi Shuo 1.0" is the first generation product launched after the explosion of ChatGPT, which has the ability to generate languages; On the basis of language ability, "Shi Shuo 2.0" adds multimodal input and output capabilities such as text, voice, image, table and video, and increases the Copilot capability of enterprises. In order to connect with the internal application library and private data of enterprises, analyze information and data, answer employees’ inquiries or perform related tasks, and become business assistants from knowledge assistants; On the basis of generative and language ability, "Formula 3.0" exerts Copilot and COT (multi-step reasoning, complex task splitting and data flywheel) to transform the experience and development efficiency of traditional B-end enterprise software, so it is called AIGS, which reconstructs enterprise software with generative AI.
Compared with the domestic big model "peers", the fourth paradigm indicates that there is no company in China that is absolutely leading in big models like OpenAI, and there will be more big models. The big model is a new productive force, and everyone has to have a big model as a base, so the threshold for entering the game has become higher, but after reaching this threshold, the focus is on how to choose the direction.
The fourth paradigm believes that the greater opportunity lies in transforming the entire enterprise software industry, that is, AIGS. The technical direction of the big model is that Copilot can be controlled (execution can be controlled, mistakes can be corrected) and COT (chain of ideas, multi-step reasoning, complex task splitting) can be used to form a data flywheel (for example, putting data and processes in a vertical field into a big model can soon form a thinking chain of the model in this field).
Dai Wenyuan said that the AIGS strategy of the fourth paradigm refers to the transformation of enterprise software into a new interactive paradigm based on the Copilot+COT capability behind the formulaic model, and the continuous learning of the use process of software in the new interaction, thus forming a "thinking chain" of domain software. Finally, due to the emergence of new forms of interaction, the development efficiency of enterprise software becomes higher.
Dai Wenyuan’s live demonstration of the "style" big model photo by the Beijing News reporter
Reporter contact email: luoyidan@xjbnews.com
Beijing News Shell Financial Reporter Luo Yidan
Editor Yue Caizhou
Proofread Liu Baoqing
Reporting/feedback