Oasis Capital in Conversation with Professor Yang Xiaokang: What Can Be Understood Can Be Created

Does "technology" have "vitality"? If so, how does it grow?

With this question in mind, over the past three years Oasis Capital has interviewed nearly a hundred professors and scholars worldwide. Some came from biology labs, marveling at the "impermanence" of life; others were steeped in fundamental physics, awakening to the "constancy" of all things.

This year, we will gradually organize, categorize, and share these conversations. We also welcome more friends to reach out and explore the vitality of "technology" together with Oasis.

So first, let's start with ChatGPT.

Professor Yang Xiaokang is Executive Deputy Director of the AI Institute at Shanghai Jiao Tong University and Director of the AI Key Laboratory of the Ministry of Education; he is a recipient of the National Science Fund for Distinguished Young Scholars and an IEEE Fellow. Below are selected highlights from our conversation. Enjoy.

Oasis: Can ChatGPT understand language logic itself?

Professor Yang: Let's first look at what ChatGPT is. First, do you know why this chatbot is called ChatGPT? The English word "chat" means conversation. That's easy enough. The "GPT" part refers to the core technology behind the chatbot — the AI model developed by OpenAI. This quirky name is actually an acronym for Generative Pre-trained Transformer. A direct Chinese translation would be "generative pre-trained converter." The most critical concept here is Transformer, which is the shared core technology of a new generation of AI models. It's also a novel deep artificial neural network architecture. It was Transformer technology that pushed the AI field into a new realm, including the rapid iteration of GPT. Previously, AI mainly relied on machine learning to learn specialized skills from human-processed data tailored for specific purposes — translation abilities, language processing abilities, and so on. Transformer technology creates a mechanism called "self-attention." For example, in the sentence "I am Xiaokang," it can learn on its own that "Xiaokang" is a name and that "I" refers to the person "Xiaokang." This is a simple sentence, but this "self-attention" mechanism can also handle more complex cloze tests and sentence completion tasks. In this way, it gives AI systems the ability to self-learn from "unlabeled" raw data. Transformer technology also enables more complex systems to learn and accommodate more data more effectively. GPT, BERT, and other large language models are all based on Transformer. These LLMs are then fine-tuned and applied to downstream tasks. ChatGPT uses a technique called Reinforcement Learning from Human Feedback (RLHF) to achieve this fine-tuning, correcting the language model to better understand human instructions and accelerating ChatGPT's iteration. Therefore, from the perspective of ChatGPT's underlying principles, it does have a certain understanding of language logic itself, and can continuously correct its previous misunderstandings through human instruction.

Oasis: Why couldn't Google produce ChatGPT?

Professor Yang: To counter ChatGPT, Google launched its intelligent chatbot Bard in early February. Google's official Twitter account even released a promotional video for Bard, showing a Q&A with the bot. The question was: "What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?" In Google's video, Bard gave three answers. The first two about Webb's discoveries were fine; the third claimed that the Webb telescope was the first to take photos of planets outside our solar system. On Tuesday afternoon, February 7, someone discovered that this claim didn't match NASA's official records. In other words, the third point was something Bard had made up out of thin air — a fabricated fact. Starting February 8, Google's market cap plummeted by $150 billion over two consecutive days. A careless example selection by a Google editor became the most expensive editorial mistake in human history, with $150 billion in Google market value evaporating in two days. Currently, Google executives are still dealing with the fallout from Bard.

Oasis: Will widespread ChatGPT use lead to a decline in human cognitive abilities?

Professor Yang: For now, ChatGPT is just a tool that helps humans find needed information faster and more accurately. Human cognitive abilities won't be affected by ChatGPT. On the contrary, people can use ChatGPT to help themselves learn quickly and improve work efficiency. First, ChatGPT can serve as a teaching assistant, helping us learn science, engineering, and humanities subjects. Taking science and engineering as an example: if you ask ChatGPT to solve a derivative problem, it can not only give the correct answer but also clearly explain the reasoning process. If you give it a simple programming problem, it can not only generate the code but also tell you what auxiliary programs its design is based on, and that you need to install these auxiliary programs first before using the code it generated. Second, ChatGPT can serve as an assistant to help us gather materials, draft or polish written documents, and assist in content creation. But ChatGPT has a problem: it makes things up, lies without skipping a beat, and answers questions in a calm, confident tone. For example, when I asked ChatGPT to tell the story of Liu Bei visiting Zhuge Liang's thatched cottage three times, it rattled on about Liu Bei separately consulting Zhuge Liang, Liu Biao, and Zhang Fei three times each. This isn't a unique problem with ChatGPT — it's a common flaw in current AI systems based on similar architectures. Therefore, when ChatGPT provides an answer, we need to carefully verify it before using it, rather than copying it verbatim. This actually tests human cognitive abilities.

Image generated by Tiamat

Oasis: In this new wave of AI (AIGC), do you see opportunities for new 2C applications? If so, which categories?

Professor Yang: Large language models are a good opportunity for 2C applications. As powerful new productive forces, they contain enormous commercial value — provided they are effectively governed to prevent misuse. Before governance standards emerge, in the near term, we can leverage their strengths while avoiding their weaknesses, promoting the commercial use of general-purpose language models. Microsoft recently launched a trial version of New Bing, integrating ChatGPT. Building on search, the new features enable continuous conversation, summarizing and synthesizing answers from factual sources, and suggesting follow-up questions you might be interested in. Microsoft's market cap surged $80 billion overnight. Microsoft also plans to add ChatGPT-like AI to products like Word and Outlook. In the near term, the main commercial applications of general-purpose language models will at least include AI assistants for fact-based summarization, interaction, and translation. Beyond ChatGPT, GPT can also be used for code generation, speech recognition, image generation, and more.

Oasis: What insights or help has the LLM breakthrough brought to your own current research?

Professor Yang: ChatGPT is a typical example of generative AI — it generates language. Gartner's report last year listed generative AI as a strategic technology for accelerated growth. Generative AI, through machine learning, produces entirely new, original data. By 2025, generative AI is expected to account for 10% of all human-generated data. Generative AI will bring profound transformation, driving content development, visual art creation, and digital twins. The GPT-3 large model has already achieved preliminary automatic programming, providing AI intuition for scientific research, generating new mathematical conjectures, helping prove mathematical conjectures, and accelerating new drug discovery and new material synthesis, among other applications. Furthermore, generative AI will drive the development of the metaverse. The core functions of the metaverse are simulation of the physical world and virtualization of people. Generative AI enables such virtualization of people and objects, promoting deep integration of virtual and physical realms, achieving efficiency gains, experience enhancements, and spiritual elevation. Therefore, we believe generative AI is the content generator of the metaverse, the connector between virtual and physical, and the efficiency accelerator; world models are the interactive physics engines in the metaverse; and virtual digital humans are the native inhabitants and productive forces of the metaverse. Generative AI provides a feasible pathway for building visually intuitive physical world models and virtual digital humans. Looking ahead, we need to further solidify the foundational theory of generative AI by drawing on mathematics, physics, information theory, and cognitive science. "Physics + data" jointly driven, "virtual + reality" deeply fused — generative AI intuition is expected to accelerate scientific discovery, material synthesis, and metaverse construction.

Image generated by Tiamat

Feynman once said, "What I cannot create, I do not understand." Generative AI believes, "What I can understand, I can create."

Celebrating Vitality

What do you think is the vitality of technology?

The vitality of technology lies in its evolution and integration into human life. Now, when artificial intelligence possesses creativity, it gives rise to other forms of life. — Professor Yang Xiaokang, IEEE Fellow, Executive Deputy Director of the AI Institute at Shanghai Jiao Tong University

Oasis Capital is a new-generation venture capital firm in China, dedicated to discovering the most vital entrepreneurs of the next decade and growing alongside them to create long-term value. "Celebrating Vitality" is Oasis's vision and mission. This vitality is both the direction of structural transformation in our era and the resilient, evolving force of entrepreneurs.

Oasis Capital focuses on early and growth-stage investments, with individual investments ranging from $3 million to $30 million. We concentrate on robotics, artificial intelligence, technology services, and other fields, supporting China's technology-driven new service upgrade.