PASCO’s In-Depth Discussion on GPT-like Artificial Intelligence

Since the release of the application ChatGPT, which is based on the Generative Pre-trained Transformer (GPT) model, GPT-like generative artificial intelligence tools have become integrated into various aspects of our lives over the past year. This integration has synergized human intelligence with artificial intelligence, providing numerous conveniences. While benefiting from the advantages these AI tools bring, people are also contemplating and researching the future development of this technology, its impact on individuals and the society, as well as the coexistence of humans with artificial intelligence.

(Fig.1:  Daily applications of ChatGPT – by courtesy of Dr. Shiyi Wang)

Following the presentation and the discussion on GPT-like artificial intelligence application during the 2023 annual meeting of Professional Association of Chinese Scientists and Technologists in Austria (PASCO), members of the discussion panel (the keynote speaker of the annual meeting, WANG Shiyi, Ph.D. in cluster robotics at the University of Manchester and researcher of Graz Artificial Life Laboratory; strategy consultant Ting WASNER-LIAN, MBA; and Prof. ZHANG Chi, Ph.D. in geophysics from the University of Vienna.) continued exploring this topic, aiming to present a more comprehensive picture of generative artificial intelligence.

Seven questions categorized into three groups were discussed: technic, market, and society. 

Technical Category:

1. Regarding the issue of data privacy in the use of artificial intelligence tools such as ChatGPT, how should we be cautious to protect it? How is its security ensured?

Wang: According to the user agreement of OPEN AI, ChatGPT collects and provides the conversations and data between users and GPT for further learning. Therefore, when using GPT, one should always consider the security of data, especially for companies and academic institutions dealing with core interests, which requires extra caution.

Due to this concern, many businesses are also training their own GPT-like artificial intelligence models based on the underlying architecture of ChatGPT. This allows for better application to their specific needs. Additionally, training AI language models in-house has other benefits. A self-trained AI language model can provide targeted services to businesses and organizations based on personalized training content. For example, by providing internal documents and data, a model can be trained to “understand” the information specific to the enterprise. Such a model can offer better responses in handling documents, customer service, contracts, and more.

(Fig.2: Understanding & interpretation of ChatGPT on Visual charts – by courtesy of Dr. Shiyi Wang)

Wasner-Lian: To ensure 100% data privacy, the best approach is to use open-source models and host them on your own servers. If you must use cloud service provider APIs, you need to ensure that they will not use the data you input for further training.

2. How should we understand the disclaimer at the bottom of the ChatGPT homepage: “ChatGPT can make mistakes. Consider checking important information”?

Wang: ChatGPT, as a generative AI, generates content based on a large language model (LLM). In simple terms, it generates text based on the relationships between words, as well as the probability of subsequent words, considering the input information and the massive training text. In academic terms, the content generated, which may be fabricated or not based on facts, is generally referred to as “hallucinations.” Although the generated content logically fits the expression, it can be fictional and not factual.

Regarding ChatGPT, such issues were common in its early stages because it generated content based on training data and couldn’t cross-reference the internet for the authenticity of the generated content. However, in the latest version of ChatGPT, there is an internet connectivity feature based on the Bing search engine. Additionally, users can emphasize to ChatGPT not to fabricate content when entering prompts. There are also other technical measures to address hallucination issues.

Currently, users still communicate one-on-one with ChatGPT. In more recent generative AI technologies, such as AI Agents, researchers have established multiple generative AIs. Some are responsible for content generation, some for content review, and some for content management. These AIs engage in ongoing iterations of content generation through dialogue based on certain rules to better meet user needs. The generated content is not limited to text and can include more complex outputs. For example, in software development, multiple AIs working together can generate a complete set of content, including programs, development manuals, user guides, etc. In this example, internal quality control among AIs for generated content is a more comprehensive solution to address hallucination issues and output errors. However, for current users, if they use ChatGPT as a substitute for a search engine or seek answers to complex questions, they still need to cross-check the information with the internet.

Wasner-Lian: This means that ChatGPT may produce errors or misleading information. For instance, if you want to analyze a document with ChatGPT, you need to test and verify whether the answers obtained after entering prompts are correct, ensuring the accurate extraction of data. ChatGPT is not a search engine; it’s more like a reasoning machine, making the human verification process crucial.

(Fig.3: Some ideas for generating visual charts within Midjourney – by courtesy of Dr. Shiyi Wang)

(Fig.4: Applications of different digital tools to form meta human – by courtesy of Dr. Shiyi Wang)

3.Applications of LLM-based Generative Artificial Intelligence in your Research Field?

Wang: From my personal experience, in engineering-related papers, the core content includes methods, experimental design, data analysis, and conclusions—essentially, a crystallization of the author’s creativity, ideas, and efforts. I’ve conducted some experiments using generative AIs like ChatGPT for this content, and the results were not ideal. However, such generative AIs can be very effective in assisting with literature reviews, analyzing research background, conceptualizing, and summarizing other aspects of a paper, significantly improving text-writing efficiency.

In academia, AI can assist in various ways, such as in conducting scientific research, defining research directions, and specifying research questions. Tools like Scite.ai can help organize relevant academic literature, present viewpoints from different researchers in the field, and brainstorm specific research problems. These are directions in which AI can help improve research efficiency.

To be honest, AI does have the capability to generate papers in certain academic disciplines. Some researchers have even tested and published papers generated by AI in journals and conferences. However, in terms of academic contributions to innovative viewpoints and advancing research progress, AI-generated papers are currently limited. Due to technical reasons, AI cannot produce truly high-quality academic papers. While some papers may pass through certain journal reviews, their academic value is generally poor.

I believe that the progress of humanity comes from generations of people continuously pursuing knowledge and truth. In this process, there are hardships, failures, but also sudden insights, and diligent efforts. Human progress stems from our own reflections and arguments about the unknown. Currently, AI cannot replace humans in this process, which is the main reason for our continuous advancement. While AI-generated papers may offer shortcuts for some individuals, such behavior is meaningless for the overall development of humanity.

Zhang: Large language models (LLMs) like ChatGPT are increasingly being used in the field of Earth sciences (source 1: https://agupubs.onlinelibrary.wiley.com/doi/full/10.1029/2023WR036288). Researchers, scientists, and educators have varying opinions on the advantages and disadvantages of using ChatGPT for research and teaching. In a forum held at our university last month titled “Who is afraid of AI,” many professors discussed the positive and negative impacts of ChatGPT on our research and teaching work. For example, ChatGPT can assist in code development, solve programming errors, or help with academic writing and editing. However, there are concerns in the academic community about risks such as data security, misinformation, bias, and plagiarism when using ChatGPT. Many in the academic community are advocating for global rules to address the ethical issues of artificial intelligence (source 2: https://rudolphina.univie.ac.at/en/governance-frameworks-for-artificial-intelligence).

It’s worth mentioning that ChatGPT cannot create entirely new concepts and cannot directly help researchers choose a novel research topic in the wide field of Earth sciences because its responses are based on existing information. While it is useful in some aspects, its role in innovative research is limited.

Undeniably, every scientist will inevitably use tools like ChatGPT. Combining other tools and research methods, we as researchers will continue to create new content and drive disciplinary progress.

Market Category:

1.Does OpenAI have competitors? What is the current status of its market development?

Wang: As of now, I believe that OpenAI is still the most advanced in terms of overall capabilities. GPT is just one of their projects, and their deep technical expertise, innovation capabilities, industry support, and platform advantages are currently leading globally. Of course, with the passage of time, the further development of the AI trend, and the tangible benefits generated by AI for businesses, more and more companies will increase their investment in related applications. On a national and regional level, whether it’s China, the EU, or many other countries and regions, AI is considered a major direction for future development. However, this requires time and technological accumulation. Regarding the development of AI in China, I am personally optimistic. From a macro perspective, China has injected a large number of resources and policies into this field. From a micro perspective, companies like Huawei, Baidu, and iFlytek have attracted many developers and researchers, investing heavily in hardware resources. The latest applications and technologies they have showcased demonstrate that China is rapidly catching up and even making remarkable and leading achievements in certain areas, such as Huawei’s recent demonstration of autonomous driving technology.

The current AI market is highly prosperous, characterized by a very close connection between academia, industry, and research. Newly proposed theories in academic papers often immediately attract a large number of developers in related industries to develop corresponding applications. Capital and policy support are also substantial, making the entire market currently very vibrant.

Wasner-Lian: OpenAI has two business segments: ChatGPT for consumers (B2C) and API services for businesses (B2B). In the B2B sector, their biggest competitor is Microsoft, as Microsoft not only provides almost all open-source models in Azure but also offers superior services in B2B APIs. For instance, Microsoft provides version management, ensuring that old versions can be used for at least three years. OpenAI does not provide this service, requiring constant updates to programs and checks on the latest outputs.

2.Specifically, which generative artificial intelligence has greater demand in the business sector?

Wang: Currently, the hottest direction in business is where generative AI can enhance efficiency in various tasks. The improvement of work efficiency is a top priority for all business organizations, whether they are large, medium-sized, small enterprises, or individuals.

From an enterprise perspective, many companies are researching how generative AI can enhance efficiency in daily work. Specifically, the most focused areas are where many workers primarily use Microsoft Office for tasks such as text, tables, data, and presentations. However, currently, as generative AI is a new tool, everyone is still in the exploration phase, and the information available is relatively fragmented. The market has not yet formed particularly successful cases.

It’s also worth noting that exploration of generative AI by small teams and individual freelancers is highly noteworthy. While assessing the efficiency improvement of generative AI is challenging for enterprises, for small teams and individuals, the efficiency provided by generative AI is tangible. They often have more tolerance and flexibility in their content requirements. Currently, generative AI cannot generate text and images that perfectly match human intentions. However, small teams and individuals have more flexibility to adjust their work goals, such as video creation, web content writing, image generation, etc. At this stage, generative AI may better meet their needs.

Wasner-Lian: The main areas of demand are text and document analysis, as well as knowledge management. Almost all AI startups focus on knowledge management and text/document analysis. The reason is that company’s knowledge exists all in text format. Reports, internal documents, presentations, customer feedback, etc., are all recorded in text format. Only by using LLM can you analyze this data. In the past, this technology was immature and couldn’t achieve this.

(Fig.5: Poster resources generation with DALL E by ChatGPT – by courtesy of Dr. Shiyi Wang)

Social Category:

1.In your knowledge domain, what impact has generative AI had on work or professions?

Wang: From what I understand, it mainly involves two aspects. For most ordinary people, the most immediate advantage of ChatGPT is in everyday document processing. This includes generating emails, website content, text reports, summaries, simple data organization, data analysis, and generating slide content.

For heavy graphics-related workers, such as graphic designers, 3D modelers, digital creatives, and spatial designers, generative AI in the graphics domain can provide more convenient creative ideas and material generation. In certain workflow stages, it can significantly enhance work efficiency.

Wasner-Lian: As a strategy consultant, for various types of research (such as market research, trend research, patent research, etc.), projects that used to require a small team of analysts can now be completed independently. If you provide consulting for public services or the construction industry, using image-based LLM and diffusion models, for drawing, for example, from conceptual ideas to completed renderings, what used to take a team several days can now be achieved in 30 minutes. Generative AI empowers the consulting industry, but the prerequisite is that you must understand and master all the best AI tools and know when and where to use them correctly.

(Fig.6: Restoration of old photo from the Asian Games with Stable Diffusion – by courtesy of Nenly from Bilibili, and Dr. Shiyi Wang)

2.In your knowledge domain, how is human intelligence superior to artificial intelligence?

Wang: I come from an engineering background, and when considering many issues, it often revolves around what solutions should be used to solve practical problems. Like in Marxist philosophy, the use of tools is key to distinguishing humans from animals. While this concept is not entirely accurate, it emphasizes that humans, as intelligent beings, can better create and use tools.

ChatGPT, as a generative AI, is also a tool created by humans. Human development is a continuous process of creating new tools to solve practical problems arising from contradictions and needs in human development. Artificial intelligence itself, as a tool, should not be compared to human intelligence. Instead, we should consider how using such tools can improve our work efficiency or better solve practical problems.

Furthermore, human intelligence is reflected in our exploration and research of new knowledge and the unknown world. The achievements obtained by generations of humans in this process are the crystallization and milestones of human intelligence. For research and business, such innovation and exploration are particularly valuable.

Although most people are ordinary. As an ordinary person, we may not think that we, as individuals, will make outstanding contributions to human development. But each of us is unique. As individuals, we all want to make our lives better. Our thoughts on life, followed by actions, and the expected results, are also manifestations of our individual wisdom. Therefore, I believe that the advantage of human intelligence lies in our thoughts and actions regarding individual and collective humanity, which will mutually influence each other and, unintentionally, gradually drive human development.

Wasner-Lian: Andrew Ng once said that today’s AI is like 10,000 interns at your fingertips, very good at executing relatively simple tasks but not yet at the level of writing a doctoral thesis or consulting proposals for clients.

(END)

(Text: Dr. Xiaoshan Liu from PASCO)

Be the first to comment on "PASCO’s In-Depth Discussion on GPT-like Artificial Intelligence"

Leave a comment

Your email address will not be published.


*


6 + 2 =