Researchers believe the next step in AI might involve agents that talk directly and teach each other skills. They built an AI system that learned tasks from written instructions.

This "teacher" AI then explained those tasks to a "sister" AI using natural language, allowing the sister AI to perform the tasks without any prior training.

This communication relied on Natural Language Processing (NLP), a field that aims to create computer programs that understand and respond to human language. The researchers used a pre-trained NLP model called S-Bert along with a smaller network focused on receiving sensory information and responding with actions.

The combined system, called a "sensorimotor-recurrent neural network (RNN)," learned 50 tasks by following written instructions about how to respond to stimuli, like reacting to a light. Despite never seeing examples of these tasks, the RNN achieved an impressive 83% accuracy rate.

Here's the key part: the RNN then learned to communicate its knowledge back to an identical sibling AI using natural language instructions. This "student" AI could then perform the tasks it had never been trained on before.

This breakthrough mimics human learning, where we can follow instructions to perform new tasks. Unlike animals, who often need physical demonstrations, humans can learn from verbal explanations.

While existing chatbots can understand instructions, they can't translate them into actions or teach others. This research, however, paves the way for AI systems that can learn and communicate like humans.

This isn't a shortcut to artificial general intelligence (AGI), where AI can reason like humans. But it could improve our understanding of the human brain and pave the way for robots that collaborate and learn from each other. Imagine a factory where one robot receives initial instructions and then trains the others!

The researchers believe this is just the beginning. They envision even more complex networks that could be integrated into robots capable of understanding both humans and other AI.

Sources:

Published 18 March 2024 in Nature; Natural language instructions induce compositional generalization in networks of neurons

Discover: