What’s stopping human-like machine intelligence? Chinese AI experts weigh in

Smarter, goal-driven research and transparency needed to realise AGI ambition, start-up scientist says

Reading Time:2 minutes

The Pujiang AI Conference was organised by the Shanghai Artificial Intelligence Laboratory. Photo: Handout

Published: 9:00pm, 14 Dec 2024

China still has a long way to go to develop its own human-like self-learning software known as artificial general intelligence (AGI), two leading industry figures told an AI forum in Shanghai on Friday.

Addressing the Pujiang AI Conference, Qiao Yu, lead scientist at the Shanghai AI Laboratory, the event’s organiser, said there were technical challenges on the path to AGI, with room for innovation in model architecture, data and learning algorithms.

One of the barriers is creating multimodal large models, systems that can process a range of information such as text and video, laying a foundation for AGI.

“[I] hope that multimodal large models can achieve breakthroughs in strong generalisation capabilities, just as language models did,” he said, adding that he anticipated advances through “scaling laws”, where a model’s performance improves as its size and training data increase.

He also said the balance between AI performance and safety was key to driving a sustainable and healthy development of artificial intelligence.

Zhang Xiangyu, chief scientist at Shanghai-based AI start-up StepFun, agreed that there was a long way to go before AGI could be realised, saying the future would require smarter goal-oriented approaches and transparency.

“The industry should steer clear of blindly following trends and use white box [learning] models to guide practical applications to avoid wasting resources,” he said.