Industrial AI will also usher in the “ChatGPT moment”

The emergence of

ChatGPT may indicate that the AI ​​industry, which has been gradually considered to have reached the bottleneck of industrialization in the past few years, is still the most innovative fertile ground and contains huge opportunities.

The aftermath of the "earthquake" of ChatGPT based on single-modal GPT-3 has not subsided, and the "tsunami" of multi-modal GPT-4 has swept through the circle of friends in an instant.

"This reminds us that prediction of artificial intelligence is very difficult." OpenAI CEO Sam Altman once said this after the release of DALL-E 2. Turns out he was right. The decline of expert systems based on symbolism made people once think that artificial intelligence had come to an end. Deep learning in 2012 ignited hope again, and now it dominates the field of AI. As systems become larger, training time and capital costs continue to balloon. Just when everyone was worried that adding parameters to the model was reaching diminishing marginal benefits, GPT-3 and GPT-4 successively told the world that larger and more complex deep learning systems could indeed release more amazing capabilities, and the birth of ChatGPT , and even more so, people can see the "disruptive" application results (fake news even claims that the number of GPT4 parameters is 100 trillion).

The emergence of ChatGPT may indicate that the AI ​​industry, which has been gradually considered to have reached the bottleneck of industrialization in the past few years, is still the most innovative fertile ground and contains huge opportunities. . As new productivity begins to take shape, industries represented by industrial manufacturing may usher in deeper AI changes and usher in the industry's "ChatGPT moment". In this process, technology companies that are in line with technological trends are also expected to Be the first to get out of the circle.

1. The victory of universality

So far, the models that dominate the AI ​​field are still task-specific. Models developed by AI companies performed well within a specific range, but engineers found that their generalization capabilities were insufficient to support deployment to a wider range of scenarios. In the words of industry insiders, many models have been trained, but many more models are still needed.

This bottleneck is almost magnified by N times in the field of highly fragmented industrial manufacturing. Because there are many subdivided fields in industrial manufacturing, each field has great differences in production processes, techniques, production line configurations, raw materials and product types. Lithium battery production can be divided into more than a dozen processes, with thousands of process points, and a production line has at least 2,500 key quality control points; LCD panel production involves hundreds of processes, and there are as many as 120 types of panel defects that may occur during the production process. There are hundreds of parts for mobile phones, involving hundreds of suppliers, and each part may have dozens of defects to be tested.

Existing deep learning models have a low degree of generalization, and even in the same industry, the reusability ratio of the models is relatively low. For example, if you want to serve the entire smart production line of a world-leading mobile phone brand, you may need to create hundreds of thousands of algorithm models (excluding subsequent iterative upgrades of software and hardware).

Now, this thorny problem has become a typical scenario for the base model (big model) represented behind ChatGPT.

In 2022, a research paper [1] from institutions such as Google, Stanford University, University of North Carolina at Chapel Hill, and DeepMind introduced the "Emergent Ability" of large models, that is, Some phenomena are not present in the smaller model but are present in the larger model, and they argue that this ability of the model is emergent. Although this ability is currently mainly reflected in language models, it also inspires future research on visual models and multi-modal models.

According to the Center for Research on Fundamental Models (CRFM) at the Center for Human-Centered Artificial Intelligence (HAI) at Stanford University, “It (large models) represents a new and successful paradigm for building AI systems trained on large amounts of data. a model and adapt it to multiple applications” [2].

This kind of general capability is exactly what industrial manufacturing needs. Industrial manufacturing faces a variety of scenarios. How to create universal technical capabilities amidst highly fragmented demands through a stable technical system has become the biggest challenge for any technology company trying to show its talents here.

Jia Jiaya, the founder of Simou Technology, mentioned the concept of AI 2.0 at the beginning of the company's establishment. One of the core points that distinguishes it from the current AI companies that widely adopt AI 1.0 is the emphasis on versatility. . "We want to build a new generation of AI system architecture, use a unified architecture to solve things that others have done in a single scenario before, and make it universal in different scenarios," Jia Jiaya said. "Build a more intelligent AI system from the bottom up. Algorithms use standardized means to solve dispersed industrial scenarios and overcome key issues such as replicability and standardization."

SMore ViMo Industrial Platform, the most popular product of Simou Technology, is a typical example of universal design thinking For example, it is the first cross-industry central platform built for industrial scenarios and has multi-scenario versatility. It not only meets the needs of more than 1,000 subdivided application scenarios in multiple industries such as new energy, semiconductors, automobiles, and consumer electronics, but also flexibly supports the design needs of a variety of difficult industrial vision solutions, such as material tracking, defect location, and workpiece counting in production lines. , appearance defect detection, etc.

System architecture of SMore ViMo intelligent industrial platform.

The important feature of this path is a relatively good balance between agility, personalization and low marginal cost. With the help of the SMore ViMo platform, SMore Technology can already support hundreds of projects in different industries in the industry at the same time. In the future, it is expected to expand tenfold to support thousands of projects at the same time, bringing efficiency breakthroughs to AI industry applications.

After being the first to use Transformer technology in large-scale industrial scenarios to greatly improve the efficiency of intelligent manufacturing, Simou has once again embraced large models for the first time. The Simou team is the first team to conduct research and industrialization of the Emergent Ability of large models in the industrial field. Its industrial large models use a small number of defect samples for in-context learning, so that the basic model can quickly adapt to specific industrial scenarios and complete specific tasks.

In the view of some industry insiders, the success of ChatGPT and the more versatile technology behind it will push AI applications into a new stage. Among all walks of life represented by industrial manufacturing, companies that have taken root in the industry in the past, embraced this trend, and completed the closed-loop implementation of data and technology have more advantages and will be more favored in the future explosion of applications.

2. Accelerate the inclusive development of AI

In the field of industrial manufacturing, there are also profound gaps between different "languages". Industry insiders said that the industrial manufacturing industry has accumulated a lot of data, but manufacturing engineers (such as mechanical engineers and materials engineers) still rarely write programs to utilize this data, and AI developers also face the challenge of understanding industrial problems. , which restricts the implementation of technology to a large extent.

Algorithm engineers from Simou Technology said that the technology behind ChatGPT, such as RLHF (Reinforcement Learning from Human Feedback, reinforcement learning based on human feedback), allows them to see that they can go further on existing work.

RLHF is an extension of reinforcement learning that incorporates human feedback into the process of training large models, providing a natural and humane interactive learning process for machines, just like humans learning from another professional The same way you learn professional knowledge. By building a bridge between AI and humans, RLHF enables AI to quickly master human experience.

They said that industrial AI can give birth to an active learning AIaaS (AI As a Service, artificial intelligence as a service) platform in the future. Through the cooperation of algorithm engineers and annotation experts, RLHF technology can be used to train large models and human Knowledge allows AI to understand industrial problems and meet the requirements of specific industrial tasks, allowing industrial experts who cannot program to train AI models.

Currently, Simou Technology is already exploring application scenarios that combine RLHF with industry.

In addition, the simple interaction mode of ChatGPT is very similar to the strategy of implementing AI in industrial manufacturing. The scenarios in the industrial field are complex, and good products must be simple and easy to use. For example, through concise interactions and one-click deployment solutions, training costs and learning burdens during the delivery process can be reduced.

Many programmers say that ChatGPT is equivalent to rebuilding a magnificent Tower of Babel. Communication with computers is no longer the exclusive preserve of programmers. It can already understand some requirements and produce simple code solutions. . But now, we can foresee that in the near future, practitioners in the manufacturing field can also implement self-programming on the AI ​​platform and develop models based on production line needs. This can also help solve the shortage of AI talents in the manufacturing industry.

"Only when computer systems can break through several major problems in industrial implementation and realize automatic algorithm combination and deployment, and humans only need to participate in a small amount of customized algorithm design, can the cross-domain large-scale industrialization of AI be realized? Possibly." Jiajiaya once said.

In fact, Simou Technology has been imagining building a development platform that can achieve rapid technology iteration since very early. Just upload the image, you can automatically mark defects, and get a product-level model or SDK for one-click testing. Reduce algorithm costs that are heavily invested in projects.

With the iteration of the project, Simou Technology gradually integrated more mature industry solutions and practical experience into its products, and then launched a complete product type, allowing customers to do it themselves without the help of Simou Technology employees. Experience and use, thus forming the earliest commercial application of the product.

With the advancement of technology, whether it is for consumers or for industries such as industrial manufacturing, we have seen more inclusive technology applications, which are bringing huge opportunities.

3. ChatGPT is just a starting point

In the past ten years, the commercialization of AI technology has been questioned by many. This time, the technological breakthrough behind ChatGTP heralds the arrival of a revolution, and AI may truly become a universal productivity infrastructure.

"GPT (generative pre-trained transformer) can also be the abbreviation of general-purpose technology (general purpose technology)", an article in "The Economist" wrote, "A ground-breaking innovation can be like Like steam engines, electricity, and computers, productivity has improved in all walks of life” [3].

The personal computer revolution, which began in the 1980s, began to really boost productivity in the late 1990s as these machines became cheaper, more powerful, and connected to the Internet. The turning point of deep learning occurred in 2012, when the AlexNet neural network won the championship in the ImageNet competition. Since then, a large amount of research began to be carried out, inspiring people to apply it in various fields. In more than ten years, deep learning technology is crossing the threshold of large-scale enabling industries.

Looking back at the development history of intelligent industrial manufacturing, there have always been challenges such as the inability of technical capabilities and algorithms to meet actual application needs, poor replicability and difficulty in implementing solutions, and high communication costs between new technology companies and manufacturing companies. The current basic model (large model) shows the general ability to multi-domain and multi-task. It is breaking down the "barriers" of these industries and "sweeping" industrial applications with extremely low fault tolerance and cost-sensitive in a low-cost and inclusive way.

Using AI to solve industrial problems contains opportunities, and ChatGPT is a starting point. With the continued deepening of some technology companies rooted in the industry, more and more industries are ushering in the "ChatGPT moment" of AI applications.

The above is the detailed content of Industrial AI will also usher in the “ChatGPT moment”. For more information, please follow other related articles on the PHP Chinese website!