Two Artificial Intelligences Achieve Linguistic Communication

This breakthrough foresees a future where AI systems interact and collaborate like humans, driving a new era of technological progress.

In a remarkable advancement in artificial intelligence (AI) research, a team of scientists from the University of Geneva has achieved a breakthrough where two AI systems engaged in unprecedented linguistic communication. Their findings, published in Nature Neuroscience, represent a significant step towards bridging the gap between human-like language comprehension and task execution within AI.

[Illustration: Freepik]

The ability to convey and replicate tasks based solely on verbal or written instructions has long been a hallmark of human communication, yet it has remained a challenge for AI. However, researchers at UNIGE have successfully developed an artificial neural network capable of achieving this cognitive feat.

Led by Professor Alexandre Pouget from the Department of Basic Neurosciences at UNIGE, the team’s pioneering work has demonstrated the capacity of AI systems to interpret linguistic instructions and execute tasks accordingly. This achievement holds profound implications for various fields, particularly robotics, where machines capable of understanding and communicating with each other autonomously are highly desirable.

The AI model employed in the study builds upon the S-Bert architecture, comprising 300 million neurons pre-trained to understand language. Through a meticulously designed training regimen simulating human cognitive processes, the network was first taught to interpret language inputs akin to Wernicke’s area, responsible for language comprehension. Subsequently, it was trained to replicate tasks, akin to Broca’s area, responsible for task execution and articulation.

Reidar Riveland, a Ph.D. student involved in the study, elucidated the intricacies of the experiment, stating, “Our network learned to interpret written instructions in English and execute tasks ranging from indicating stimulus locations to responding to visual cues. Once trained, it could effectively describe these tasks to another AI system.”

The ability of the AI to convey tasks linguistically to its counterpart marks a significant milestone in the realm of artificial intelligence. Notably, this marks the first instance where two AI systems have engaged in communication solely through language, without relying on predefined commands or programming.

Beyond its academic significance, the breakthrough holds immense promise for practical applications, particularly in robotics. By enabling machines to comprehend and execute tasks based on linguistic instructions, the study lays the foundation for advanced humanoid robots capable of autonomous collaboration and problem-solving.

a, Illustration of self-supervised training procedure for the language production network (blue). The red dashed line indicates gradient flow. b, Illustration of motor feedback used to drive task performance in the absence of linguistic instructions. c, Illustration of the partner model evaluation procedure used to evaluate the quality of instructions generated from the instructing model. Credit: Nature Neuroscience (2024)

Commenting on the implications of their work, Professor Pouget remarked, “This model opens new avenues for understanding the intricate interplay between language and behavior within AI systems. It represents a crucial step towards the development of intelligent machines capable of seamless communication and cooperation.”

With this groundbreaking achievement, the researchers envision a future where AI systems not only understand humans but also interact and collaborate with each other in a manner reminiscent of human communication, heralding a new era of AI-enabled technological advancements.