The release of ChatGPT by OpenAI shook the world. In just two days from its release, ChatGPT was being used by one million people. In two months, that number rose to 100 million. The use of ChatGPT is spreading at a rate unprecedented in history. Everyone -- not only AI researchers and specialists, but also companies, politicians, experts, and members of the general public -- is surprised and fascinated by generative AI such as ChatGPT and is watching closely to see what impact it will have on society. How will generative AI change the IT industry? How will it affect businesses and social systems? And how will our own lives be changed by generative AI? In this running feature, we will focus on large language models (LLMs), the foundation of generative AI. We will learn about their key technical points, how they can be used in business, and their future prospects.

In part one, we presented the nature and position of generative AI with respect to AI technology as a whole. In part two, we looked at Toshiba Digital Solutions’ own efforts from the perspective of using generative AI in actual business operations. In part three, the final part of this series, we will envision the society of the future—a society in which generative AI is part of our everyday lives. In the first half of this issue, we will focus on initiatives for applying AI to multimodal application and to improving design and development operation efficiency, two fields where we are currently carrying out verification testing. In the second half of the issue we will explain the ideal evolutionary path for AI and how it connects to people, as envisioned by Toshiba Digital Solutions.

Skip to the first half of the issue.

In the late 1980s, there was a popular American television series about a man who teamed up with a sports car with its own AI to fight crime. The show was also broadcast in Japan. In the show, the AI and the main character communicate with each other, sometimes even joking around. Various sensors in the car read and interpret information from the surrounding world, so the car understands the situation around it, warning the main character of dangers or operating entirely autonomously. A look at recent advances in generative AI research will show that the creation of AI like that shown in the show appears to be just around the corner.

Let’s look at our own vision of the ideal evolutionary path for AI and its connections to society (Fig. 5).

Until around the 1990s, AI was nothing more than a tool that provided a single function. However, with the development of discriminative AI over the past few decades, for example, AI has come to be used as an assistant that supports operations such as defect detection (“Stage 1: Communication” in Fig. 5).

Now, with the arrival of generative AI, the AI usage stage has advanced to the “associate” stage, in which AI are entrusted with entire roles (“Stage 2: Collaboration” in Fig. 5). For example, in customer service operations, until now, human operators have predicted what customers would like to know, have searched for the information that best fits those customer questions, and have provided that information. Generative AI can now identify the meaning of customer questions from the questions customers submit, generate new text based on this customer intent, and answer customer questions through this newly generated text. In other words, generative AI is becoming increasingly capable of handling part of the work that has been performed by humans until now.

What will the world look like during the next stage, then? We envision a future in which AI mixes naturally with people and blends seamlessly into society as a partner (“Stage 3: Coexistence” in Fig. 5). When humans operate systems, they predict what situations will arise next, thinking about which future changes would be correct and which would be incorrect. They intentionally control systems. As multimodal application technologies advance, generative AI will likely become capable of combining various external information, such as sensor data and images from cameras, with document information such as system instruction manuals, reports, and problem resolution records, to assess a system’s current status and predict potential future conditions. If this happens, it will lead to a world in which generative AI autonomously controls (operates) systems.

As we have discussed, generative AI has the potential to understand the meanings associated with various multimodal data and datasets. The key point is that people control systems with specific intentions, such as changing situations such that they benefit people and society. By accurately teaching generative AI these human intentions, we can have generative AI autonomously control systems in ways that match those intentions. In a future where generative AI such as these exist, we see people as naturally coexisting with AI, enriching people’s lives and society.

So how will generative AI have to evolve in order to create this future of natural coexistence between humankind and AI? We will look at this question from three perspectives.

The first perspective is the resource problem, specifically the power and fresh water problems. According to agencies such as the Japan Science and Technology Agency, the total amount of power consumed around the world by data centers, which are growing in line with increases in data processing volume, will, in a worst case scenario, be equal to the whole world’s power output in 2040^[6]. The amount of fresh water used to cool data centers in 2040 is also forecast to be equivalent to the amount of total fresh water consumed by Japan in a single year. Scaling laws show that transformers cannot achieve higher levels of accuracy without dramatically increasing the number of dimensions used in calculations, so if things continue on their current course, there will be a tremendous negative impact on global power and fresh water consumption. However, expectations are high for measures that aim to compress calculation space while maintaining output performance, along with the invention of new AI algorithms with structures that are optimized for computers with non-von Neumann architectures, totally different from modern transformers.

The second perspective is the issue of the proliferation of generative AI engines. Currently, implementing high-accuracy, high-performance generative AI requires massive computational resources. Because of this, it also requires the ability to pay for the tremendous operational costs involved. The reality is that only a limited number of research organizations have the resources to develop generative AI. If, however, it becomes possible to develop high-performance generative AI engines whose operation costs far less than current AI engines, it is highly likely that all kinds of countries, companies, associations, and organizations could have their own proprietary generative AI engines. Generative AI not only has the ability to operate autonomously, but it can also faithfully reproduce the intentions of those who train them. The world must collaborate to create mechanisms that will prevent the unfettered proliferation of generative AI engines that are detrimental to society.

The third perspective is the quality and ethics of AI. Humans cannot be expected to achieve 100% perfect, completely error-free levels of quality. To prevent errors, humans use ethics and exercise self-control. For example, if a person obtains information, they can determine what kinds of effects the distribution of that information would have and make the ethical decision to take steps to prevent negative effects, such as fact-checking the information or choosing not to distribute the information. Humans have spent long years building ethical standards and behavioral standards from perspectives such as “how should we, as living beings, exist?” “what should humans do to exist as members of society?” and “what should humans do to live sustainably as part of the global environment?” Generative AI has yet to develop these autonomous ethics. Humans must be actively involved in generative AI, steadily and consistently building ethically-based computational limitations into generative AI from the perspective of coexisting with it. Instead of simply leaving everything up to generative AI, in the future, humans will need to actively fulfill these responsibilities.

The essence of generative AI lies in understanding the intentions of people who ask questions. By training generative AI to understand language, we enable generative AI to interpret and understand the intentions behind questions asked using words, and to use words to provide answers to those questions. Similarly, by training generative AI to understand multimodal data, we can have generative AI provide answers about system conditions based on their interpretations of the datasets provided to them.

However, generative AI is not all-powerful. The people who use generative AI must control the AI by defining, in advance, what it will be trained on and what it will output—in other words, how it will be used. This is one of major roles and responsibilities of humans. It is only when people can make the fullest use of generative AI that they will be able to coexist with it and help create a truly rich digital society.

Through this three-part series, we have explained generative AI, looked at Toshiba’s own AI initiatives, and discussed the future potential of generative AI. With the release of ChatGPT, various companies and organizations have begun deploying “ChatGPT-equivalent” dialogue and search tools. The use of generative AI in operations will continue to grow in the future. We hope this running feature has provided you with a better understanding of the features and essential nature of generative AI, and that it will assist you in leveraging generative AI in the industrial field.

• Toshiba AI Technology Catalog