Award-winning AI scientist who left US for China creates world’s first AI child in Beijing


A groundbreaking AI entity has been unveiled at an exhibition by the Beijing Institute for General Artificial Intelligence. Called Little Girl, or Tong Tong, she can assign herself tasks, show human emotion and explore her environment. — SCMP

Science fiction has now become reality with the creation of Tong Tong, a girl who is the world’s first virtual artificial intelligence (AI) entity, developed by Chinese scientists.

Affectionately named Little Girl, or Tong Tong in Chinese, the groundbreaking AI was unveiled at the Frontiers of General Artificial Intelligence Technology Exhibition held in Beijing on January 28-29, under the auspices of the Beijing Institute for General Artificial Intelligence (BIGAI).

There, exhibition visitors could interact with Little Girl. When programmed to like things tidy, she would fix a crooked picture frame on her own. If the frame was too high for her to reach, she would find a stool so she could straighten it without any help from a human. Also, if someone spilled milk, she would find a towel and clean it up by herself, showing her ability to interpret human intentions.

Unlike popular large language models in AI, Little Girl can independently assign tasks to herself, ranging from exploring her environment to tidying rooms and cleaning stains.

She has her own emotions and intellect, and is capable of autonomous learning.

“Little Girl possesses a mind and strives to understand the common sense taught by humans. She discerns right from wrong, expresses her attitudes in various situations, and has the power to shape the future,” a video posted by BIGAI said.

In terms of general artificial intelligence standards and testing tasks, Little Girl displays behaviour and capabilities akin to those of a three or four-year-old child. Through exploration and human interaction, she can continually enhance her skills, knowledge and values.

A key aspect of general intelligence, as one researcher explained, is having physical and social common sense akin to humans. Driven by its own values, an AI entity should not only be capable of completing an infinite array of tasks, but should also define new ones autonomously.

“To advance towards general artificial intelligence, we must create entities that can comprehend the real world and possess a wide range of skills,” BIGAI director Zhu Songchun said.

Zhu, who spent 28 years studying, living and working in the United States, left his professorship at UCLA in 2020 to establish BIGAI back in China.

As a world-renowned scholar in the field of AI, his research areas include general artificial intelligence, computer vision and autonomous robots, among others.

He has received the ONR Young Investigator Award from the US Naval Research Laboratory and the Marr Prize from the International Conference on Computer Vision, one of the highest awards given for papers in the field.

He has also served as chairman of the Conference on Computer Vision and Pattern Recognition (CVPR) and vice-director of the IEEE Computer Society Fellow Evaluation Committee.

Also showcased at the exhibition was the Tong Test, a platform for AI testing published by Zhu’s team in the journal Engineering hosted by the Chinese Academy of Engineering (CAE) in August last year.

Traditional AI tests, which focus on human identification, task orientation and virtual environment testing, each have their limitations.

The Turing Test, for example, can only assess an AI’s communication level with humans, not its intelligence. Task-oriented tests may lead to AI systems learning tasks too specifically, thus losing their ability to generalise. Virtual environment tests, while creating realistic experiences, tend to oversimplify physical environments.

The Tong Test introduces a comprehensive capability assessment framework across five dimensions – vision, language, cognition, motion and learning. It also encompasses a value system ranging from physiological and survival needs to emotional and social values, and even group values.

“With nearly 100 specialised tasks and over 50 general tasks, the Tong Test offers a complete testing regime for the development of general artificial intelligence,” a release on the institute’s website said.

“For general AI to integrate seamlessly into human environments, it must learn and execute tasks in complex settings, driven by values and an understanding of causality. This is why we proposed the Tong Test, a new direction for testing general AI, focusing on practical abilities and values,” Zhu said in the release.

“Our research will guide general AI in learning and improving its capabilities more effectively and safely, ensuring it serves human society better,” he said. – South China Morning Post

Follow us on our official WhatsApp channel for breaking news alerts and key updates!
   

Next In Tech News

This AI tool is capable of generating a video game from a simple image
Surge in Wendy’s complaints exposes limits to consumer tolerance of unstable prices
This smart glove can teach you to master new skills
Super Micro surges as AI server maker set to join S&P 500
Nvidia CEO says AI could pass human tests in five years
US judge says Google must face some advertisers' antitrust claims, dismisses others
WTO e-commerce tariff moratorium wins temporary reprieve
Micron’s New York chips project under US environmental review
Change Healthcare enables new service amid outage
Reddit seeks up to $6.5 billion valuation in IPO, source says

Others Also Read