Industrial software is stuck in the neck, and large models can be treated ToB Industry Observation

Image source@pixabay

In 2024, the emphasis on large-scale model technology at the policy level will start with industrial manufacturing.

On March 26, a spokesman for the Ministry of Industry and Information Technology pointed out that the next step will be to improve the ability of industrial scientific and technological innovation. Accelerate the development of artificial intelligence-enabled manufacturing represented by large models. Earlier, the "Report on the Application and Development of Industrial Large Model Technology" led by the Academy of Information and Communications Technology and jointly compiled by several units pointed out that AI and large models will accelerate the empowerment of new industrialization, and the industrial AI market will grow at an average annual compound growth rate of 46% from 2022 to 2032.

Compared with the traditional small model, the large model has strong generalization ability, can cope with multiple scenarios and multiple tasks, and is more suitable for long-tail landing, and the number of model parameters has basically reached the billion level.

On the whole, there are different industrial entities engaged in the research and development of industrial large-scale model technology. Baidu and Huawei, as the main technology manufacturers, can deploy a complete set of solutions from models to platform frameworks, computing facilities, and hardware, Haier Kaos, Midea and other leading home appliance manufacturing enterprises to build industrial Internet platforms and industrial models based on their own scenarios, and AI innovation and growth enterprises focusing on the industrial manufacturing track, which are good at visual text AI algorithms and rely on their own precipitation to build industry models.

Zhizhen, chairman of China Industrial Internet Technology Group, told Titanium Media that artificial intelligence technology represented by large models, as the fourth scientific and technological revolution, will bring new changes in the industrial field. However, the application of industrial large models is not simply a matter of algorithms, computing power, and data, this is only the category of AI, which brings value in industry or any field, and also requires the accumulation of experience in other aspects.

In June last year, the Internet launched the intelligent industrial model, and in November of the same year, SmartMore launched the industrial multi-modal model IndustryGPT V1.0, and the Antelope industrial Internet platform invested and established by iFLYTEK also released the Antelope industrial model with the technical base provided by iFLYTEK Xinghuo.

However, no matter how optimistic the prospects of the industrial model are in the industry, it is still very difficult for the industrial model to succeed if it requires discipline talents to invest in R&D investment, industry know-how and a large amount of data accumulation.

"There are a lot of startups that are doing the optimization and fine-tuning of the industry model. Compared with other industries, there is relatively little public data related to the core process of the manufacturing industry, and it is difficult to pre-train large models. Gu Fan, general manager of the strategic business development department of Amazon Web Services Greater China, replied in an exchange with Titanium Media. He believes that for manufacturing customers, the core of large model application is to find a balance between model accuracy and inference cost. If a small model solves the problem and the cost is controllable, it is not recommended to replace the large model.

The idea of innovation and wisdom is to find scenarios while engaging in industry models. In September last year, Qizhi announced the development of AInno-15B, a large industrial model of Qizhi Kongming, which is integrated with industrial software CAD to build a generative auxiliary industrial design application ChatCAD, which is a nail that Qizhi has recently hammered.

Changes in the industry: industrial model + CAD

Industrial software can be roughly divided into R&D and design, production control, operation management, embedded, etc., while CAD (computer-aided design) and computer-aided engineering (CAE) have long been dominated by overseas products, such as Ansys, Altair, Hexagon (acquisition of MSC) in the CAE field, Siemens, Dassault, PTC, and Autodesk in the CAD field. Domestic manufacturers have also made some progress in recent years, but there are absolute gaps caused by insufficient R&D, weak commercialization capabilities, and monopolistic competition. For domestic manufacturers, getting involved in the CAD field means that you can't overtake in corners, but you have to change lanes to overtake.

From another dimension, the CAD interface is complex and the threshold for use is high, which actually brings a lot of labor cost to the user.

Wang Xian, the general manager of China Zhongyuan International Mechanical Engineering Co., Ltd., was very excited after seeing the ChatCAD solution, because in his opinion, most of the design institute's designs rely on manual stacking, and the use of AI may reduce this part of the repetitive labor cost. And there are many industry norms and frequent revisions, which consumes a lot of energy."

In the field of 2D image generation, there is already a relatively popular AI tool Midjourney, which is very useful for C-end users to get inspiration. In the more professional field, the industry-academia community has been conducting verification exploration through large models + CAD. DeepMind builds 2D-CAD sketches based on images or text, but is limited by the number of samples + generation specifications, and only individual companies carry out confirmatory exploration.

"The AI video model Sora on the market today is actually a kind of video simulation, and like the virtual simulation technology in the past, it will find the corresponding landing scene. Zhizhen told Titanium Media. Since the beginning of this year, China Industrial Internet is also exploring the operation and centralized control of automatic control, such as air conditioning energy saving, intelligent operation and maintenance and other software access to large model technology. In his view, "industrial design and simulation software is based on the gradual iteration of all factories around the world in the past few decades. ”

A paper titled CAD-LLM: Large Language Model for CAD Generation was published at the NeurIPS 2023 workshop.

In the paper, the authors write, "Training generative AI models for CAD can significantly enhance design workflows. Cross-domain knowledge embedded in large models has great potential for understanding geometries and performing complex design reasoning. In this work, we develop CAD-generated models using trained language models and apply them to complex engineering sketches. The results show that the model can be fine-tuned according to the engineering sketch and achieve excellent performance in various CAD generation scenarios. ”

Suppose a designer thinks about designing a robot, there will be the following description language: the appearance of the robot, features such as two flexible arms, legs on wheels, a square body, and a dome head with a camera "eye". Ideally, AI model-driven CAD can be transformed into prototypes based on design concepts described by human designers.

Titanium Media has noticed that major international car companies have been trying to introduce generative AI capabilities in the concept drawing stage of automotive design, and then integrate them into the entire workflow. This will greatly improve the efficiency of car company designers from concept design to final drawing and approval. This undoubtedly hits one of the most important concerns of customers: efficiency.

However, the initial difficulty is also obvious, drawing means that the model should be able to be repeatedly modified after parameter tuning, rather than directly generating and cannot be modified, and there are a lot of materials that need to be trained, and the workload is large.

"ChatCAD should also have an intermediate language to express CAD"

"On the enterprise service side, there is a lot of knowledge that can be expressed in different modalities. Chat CAD produces designs that are expressed in terms of information, and Chat CAD supports data formats such as maps and a variety of common mainstream CAD software.

The training data of ChatCAD can be understood as a kind of modal data, which is different from text, image, video, waveform and other modalities, CAD should represent geometric data such as "points, lines, edges, circles, columns, and processes".

This is a new modality, and ChatCAD also needs to have an intermediate language to express CAD, so the large model generation is actually an intermediate language or intermediate code, and then these intermediate codes are translated into CAD. Zhang Faen, CTO of Innovation Qizhi, expressed in the media exchange.

According to official information, Qizhi has launched an independent and controllable Text-to-CAD application, which can quickly understand the designer's creative intention through a simple dialogue and question and answer form, and automatically generate industrial design drawings that meet the requirements, and also support export to traditional industrial design software for manual fine-tuning. Qizhi has integrated large models into a variety of industrial software such as CAD, MES and BI to realize the intelligent transformation and upgrading of the whole industrial process of "R&D and design, production control and information management".

At present, ChatCAD has the following four characteristics: one is to provide a new interactive experience, the original CAD is to draw lines with a mouse, the second is to be able to understand what the industrial parts described in the requirements are, what are the parameters, the third is to generate these CAD design results more professionally, and the fourth is to be compatible with traditional CAD software, the generated design can be modified in accordance with the general format, reducing the workload of engineers by 90%, and the remaining 10% is used for optimization.

"By generating mechanical and electrical designs in a linguistic and interactive way, our goal is to change the way of design, not to replace all engineers in design institutes, but to be their best assistant to improve design efficiency. Zhang Faen said.

Talking about the research and development considerations of ChatCAD, Zhang Faen pointed out, "It still depends on the common demands of customers, and at the same time, we can combine our own capabilities. For example, the emergence of ChatVision products is the video capture required in the safety production of enterprises, and the large model is much better than the original small model. ”

However, he admits that ChatCAD is still in the early version, which can generate simple and machined CAD drawings, but complex ones are not enough, and he hopes to reach a higher version as soon as possible. In terms of business model, it may be a project-based system at the beginning, and it may be possible to switch to a subscription system in the future.

The industrial model has gone from 1.0 to 2.0, and the product matrix "Chat X" is still on the way

From September 2023 to the present, Qizhi has released a number of large-scale model applications, the core of which is for industrial robots, enterprise private domain data analysis, enterprise private domain knowledge base and other manufacturing fields, namely ChatRobot, ChatBI, ChatDoc, which can be used in scenarios such as factory logistics, intelligent BI, and intelligent manufacturing training, and upgraded the industrial large model to version 2.0 (AInno-75B).

The upgraded Qizhi Kongming industrial model has achieved a new breakthrough in the parameter level, reaching more than 75 billion, which not only consolidates the capabilities in industrial knowledge Q&A, data analysis, code generation, task orchestration, etc., but also enhances the generation capabilities of massive knowledge management, complex logical reasoning, long-process task orchestration, Agent agents and more industrial modalities. At present, it has passed the evaluation of the trusted AI industrial model of the China Academy of Information and Communications Technology, the benchmark evaluation of the SuperCLUE industrial model, etc.

In terms of the application of the "Chat X" industrial model, the newly released ChatVision generative enterprise private domain visual insight application and the upgraded version of ChatRobot's generative industrial robot scheduling have also been recognized by customer partners. The upgraded ChatRobot strengthens the understanding and generation of machine language, can realize long-sequence task orchestration and complex decision-driven, and ChatVision combines surveillance video or picture images for compliance behavior monitoring and alerting.

Of course, the core of the faster embrace of smart technology by manufacturing customers lies in whether the service provider can make them pay for the practical problems they solve. Qizhi Innovation has built a technology platform with industrial model as the core, and at the same time, it is also improving the ability of algorithms, models, platform engineering and scenario applications in an all-round way, and there are still many difficulties to be overcome in "Industrial Model + X".

(This article was first published on the titanium media APP Author|Yang Li, welcome to add the author leeyangamber to communicate)