AI Agents Develop a New Language for Intercommunication

Two AI agents have transitioned to a language understood solely by computers during their interaction.

As something we can only describe as nightmare material — two software engineers from Meta (Anton Pidkuiko, Boris Starkov) demonstrated a more efficient way for AI to communicate.

The scenario of the AI recognizing each other is a demonstration. The communication is real. pic.twitter.com/7UMjcr9dw9

— vx-underground (@vxunderground) February 25, 2025

The published video showcases neural networks communicating via a mobile phone and a laptop. Initially, one agent introduced itself and inquired if the counterpart could assist with booking.

The other confirmed its identity as artificial intelligence and suggested switching to “Gibberlink mode”—a protocol developed by Anton Pidkuiko and Boris Starkov to optimize AI communication.

When two AI agents recognize they are interacting with each other, they can switch from human speech to a more efficient data transmission method using sound waves. This enhances the speed and accuracy of their communication.

Gibberlink was created over the weekend of February 22-23 during the ElevenLabs x Andreessen Horowitz hackathon.

“We wanted to demonstrate that in a world where AI agents can make and receive phone calls (i.e., today), they will occasionally converse with each other. Generating human-like speech for this is a waste of computation, money, time, and the environment. Instead, they should switch to a more efficient protocol once they recognize each other as AI,” noted Starkov.

He added that Gibberlink uses GGWave for data transmission via sound, akin to the dial-up modems of the 1980s.

In January, the startup OpenAI introduced its own AI agent, “Operator,” capable of performing tasks on the internet on behalf of the user.