Raj Reddy served as the founding Director of the Robotics Institute at Carnegie Mellon University from 1979 to 1991 and the Dean of the School of Computer Science at Carnegie Mellon University from 1991 to 1999. He won the Turing Award for his contribution to the research on "large-scale artificial intelligence systems" in 1994. In addition, he has fostered many well-known scholars in the field of AI. His research covers a wide range of topics in AI, including speech understanding, image understanding, robotics, multi-sensor fusion, and intelligent agents.

Raj Reddy, a pioneer in AI, has been active in the research for over five decades, which cannot be described in one simple passage. On November 13, 2023, Raj Reddy visited Zhejiang Lab and had a dialogue with the Lab's researchers, where he shared his profound insights on AI, and painted a picture of the future intelligent society together with young people engaged in AI research.

"As large language models (LLMs) are increasingly used, how can we improve their accuracy and reliability and make better use of them?" The dialogue begins with a smoking hot chatbot, ChatGPT.

"Asking the right questions and designing effective prompts will become a skill that we need to keep practicing with ChatGPT." Raj Reddy said that how to ask ChatGPT questions is critical. "Since pre-trained models and programs might make mistakes, people have to learn to judge what's right or wrong. Good students are not to toe the line, but to ask why, who, where, when, and how, that is, they should dare to ask questions, enjoy challenges, and have the spirit of questioning."

"Might GPT demonstrate superior reasoning capabilities in the future? For example, is it capable of inferring the rules of gravity from a video of apples falling?"

Regarding the reasoning capabilities of LLMs, Raj Reddy pointed out that about 90% of the tasks in the world, in fact, do not require reasoning. By imagining how to make decisions, people tend to make decisions based on the experience that has long been stored in their brains. Many people make decisions without thinking about lower-level questions like why. As far as LLM is concerned, a text corpus used for training the model is collected from the Internet in general, of which many texts are involved in reasoning, so LLMs can perform many tasks that seem to require reasoning.

"LLM uses the Web as a corpus, but those data describe past events, excluding what is going on today. I think that the power of theoretical reasoning is still needed if we want to discover new laws and 'go one step further' under the current situation." LIU Bin, a researcher at the Research Center for Applied Mathematics and Machine Intelligence, pushed the dialogue further.

For technologies such as LLMs, Raj Reddy is more concerned with how they serve the entire society. In the dialogue, he described the future pattern for interacting with LLMs, especially for semi-illiterate people. For example, in the future, digital versions of printed books will be made available, which can be downloaded to smartphones and read freely in your selected language; you can watch movies and listen to lectures without translation, and access information in your native language in real time by placing your smartphone near TV; you can translate multi-round conversations on your smartphone... The power of technology enables free and open communication among people with different languages and backgrounds, so that AI could have a positive influence on people all over the world.