laitimes

Generative AI like ChatGPT: Thinking Past, Present, and Future

author:Cheng Hao How

Generative AI (tools such as ChatGPT) has now sparked a global AI boom. But how do we address the multiple challenges posed by this generative AI?

Recently, in the series of academic activities celebrating the 118th anniversary of the founding of Fudan University, Zhao Xing, deputy director of the National Intelligent Evaluation and Governance Experimental Base of Fudan University and professor of the Institute of Big Data, and Jiang Zhuoren, a researcher at the School of Public Administration of Zhejiang University, cooperated to bring a feast of thinking about the past, present and future of generative artificial intelligence.

Generative AI like ChatGPT: Thinking Past, Present, and Future

The past and present of generative AI

Jiang Zhuoren emphasized: "Generative artificial intelligence is not achieved overnight, it has a long history. "In the more than 90 years of generative artificial intelligence development, human intelligence has been shining.

In 1932, French engineer Georges Artsrouni created a device called the "Machine Brain." It completes the translation by querying a multifunctional dictionary, and the input and output are a single paper tape. Although different from today's machine translation, it fits the definition of generative AI, where a human enters a piece of content and a machine generates a new piece of content.

Later, two professors at the Massachusetts Institute of Technology in the United States created one of the first generative artificial intelligence and introduced the concept of Bayesian network causal analysis, showing how convolutional neural networks can be used to recognize images.

In 2006, Feifei Li, a Chinese-American computer scientist, set out to build a database called ImageNet. The database contains more than 14 million hand-annotated images in more than 20,000 categories. "It's with such a huge database that deep learning has emerged." Jiang Zhuoren said.

Generative AI like ChatGPT: Thinking Past, Present, and Future

Why is ChatGPT a star product?

ChatGPT has 175 billion parameters and 300 billion training words, which is its data volume. In just two months since its launch in 2022, ChatGPT has attracted more than 100 million users, making it the fastest-growing app ever.

Jiang Zhuoren believes that to understand ChatGPT, it is necessary to understand its key technologies: large model basic training, instruction fine-tuning and human feedback reinforcement learning.

"A large model refers to a large language model." Jiang Zhuoren explains, "It is a probabilistic model that can tell you the probability of a word appearing. A good language model is able to accurately predict what the next word will be. ”

On the basis of good language understanding ability, in order to make ChatGPT can have dialogue with humans, the researchers proposed instruction fine-tuning to improve the reasoning ability of large models by introducing thought chains and code generation.

"This capability is critical for large models, allowing it to excel in the open field." Jiang Zhuoren said.

Through these technologies, large models basically have the ability to answer instructions, but the quality of answers varies. In order to continuously output high-quality answers from large models, researchers design a set of reinforcement learning methods based on human feedback, which ensures the generation of high-quality answers by fine-tuning large models, training reward functions, and optimizing large-scale reinforcement learning.

"Open AI uses this approach to dramatically reduce the cost of building datasets." Jiang Zhuoren said.

Generative AI like ChatGPT: Thinking Past, Present, and Future

With new intelligence comes new challenges

Facing the challenges brought by generative artificial intelligence, Zhao Xing interpreted from four dimensions: resources, technology, application and social ethics.

From a resource perspective, generative AI requires high-quality data, and Chinese data is of poor quality than English data. Zhao Xing believes that even with powerful translation capabilities, Chinese tools similar to ChatGPT are significantly inferior to English in terms of processing results, and one of the core reasons is the poor quality of Chinese data.

"The tech world once compared the application of artificial intelligence to alchemy." Zhao joked, "People put data into the model without clear expectations, and they don't know if they can refine something valuable." ”

Obviously, on a technical level, generative AI is inherently uncertain.

"When we are ready to promote a universal tool to society without understanding its scientific principles, we run inherent risks. The core risk of AI lies in the unbearability of its results. Rarely have we been so powerless in governance. Zhao Xing said.

At the application level, the industrial development of generative artificial intelligence is certain, but there is uncertainty in risk governance. In terms of social ethics, generative AI not only carries the risk of intellectual property disputes and information leakage, but also may shape a truly information-enclosed space.

"When generative AI is by your side 24 hours a day, it subtly makes you think that everything is your own decision." Zhao Xing warned, "We are facing something that will rise rapidly, with serious consequences and unknown consequences." ”

Generative AI like ChatGPT: Thinking Past, Present, and Future

Intrinsic Security Governance: Making the Crisis "Known"

Zhao Xing believes that in the face of this new adversary, generative artificial intelligence, traditional governance cannot be adopted to "passively respond to external threats". His team is working on building a new model of generative AI governance, drawing on the "intrinsic security theory" proposed by Academician Wu Jiangxing, Dean of the Big Data Research Institute of Fudan University.

"Can we find a way to combat unknown risks before they erupt?" This is the problem that generative AI intrinsically addresses security governance. "We need to provide a new skill tree for human society before the AI risks arrive, so as to deal with the non-traditional security issues brought by AI." ”

The principle of the intrinsic security governance model is based on swarm intelligence to transform the "unknown unknown" of the individual into the "known unknown" of the group, and then further into the "known known".

"When we know what the possible risks are and where they arise, the governance of generative AI has the opportunity to translate into regular security issues, and we can try to close the loop of governance." Zhao Xing said, "However, this still requires long-term theoretical and practical exploration. ”

Zhao Xing's team also explored the application of generative artificial intelligence in scientific evaluation, and innovatively built a new paradigm of "digital sapiens" evaluation and governance, including the harmonious symbiosis of objective data, intelligent algorithms and expert review. At present, the team is conducting exploratory experiments to build an intelligent evaluation system using tools similar to ChatGPT.

Generative AI like ChatGPT: Thinking Past, Present, and Future

"Preliminary results show that although generative AI is not yet capable of serious academic evaluation at this stage, it shows interdisciplinary evaluation capabilities and emergent inference prediction potential, which deserves high attention." Zhao Xing said.

Read on