“This is gonna be scariest sound you’ll hear when they’re looking for you [1] ”
~ @IIBLINKII

Waver allows you to send and receive text messages from nearby devices through sound waves.

No one was ever scared of R2-D2.

On 24 Feb 2025, Anton Pidkuiko uploaded a video to YouTube [1] that gave some of us a tiny chill.

In it, two AI agents realize they’re both synthetic - and mid-conversation, they switch to an audio protocol faster than human speech.

As the Dr-Business write-up on GibberLink notes [2] , “This might sound like an incredible leap for AI efficiency, but it also raises serious concerns”.

GibberLink is equal parts hilarious and uncanny, and honestly, it’s inspired a future homelab project of mine. Soon, my Proxmox nodes… each running an Ollama instance - might start randomly chirping at each other. I won’t know what they’re saying… but their cheeps will sound like a kind of digital birdsong.

GibberLink didn’t emerge from a polished research paper… it burst into existence during a high-energy hackathon [3] . In early 2025, at the ElevenLabs & a16z global hackathon in London, developers Anton Pidkuiko and Boris Starkov introduced GibberLink: an AI-to-AI communication protocol that detects another AI on the line and switches to a stealthy, data-over-sound channel powered by GGWave.

The project swept the event, winning the top prize and quickly going viral. Its demo video, where conversational agents shift from natural language to beep-like code, amassed millions of views and featured across tech outlets like TechCrunch [4] , MakeUseOf [5] , and DeepNewz [6] .

“It eliminates up to 90% of the compute cost associated with generating human-like speech, shifting processing to less resource-intensive audio encoding/decoding [7] ."
~ Next Big Future

Try GibberLink out for yourself.

GibberLink’s beeping brilliance is built on a quieter foundation - GGWave [8] , an open-source project by Georgi Gerganov dating back to late 2020. Originally designed for sending short messages via sound, it found new life as the backbone of this AI-to-AI protocol.

Stats for nerds

TechSpeedLatencyDuplexErrors
Human Speech (English)a~25-60 bpsInstant (in earshot)FDM
Typing (Secretary)~40-64 bps~0.5-1 sSL-M
GibberLink (GGwave)b~100 bps - 1 kbps+<1 s (bursts)HDM
Acoustic Coupler~300 bps~1-2 s handshakeFDL-M
V.92 ModemUp to 56 kbps2-6 s handshakeFDL
Carrier Pigeonc“Varies wildly”Hours-DaysSM-H
Wi-Fi (5 GHz)Up to 1 Gbps+MillisecondsHDVL

Legend

  • Duplex: FD = full-duplex, HD = half-duplex, S = simplex

  • Errors (Rate): VL = very low, L = low, M = medium, H = high error risk.

  • Notes:

    a Humans can talk over each other, but comprehension tanks.

    b GGwave is typically one-way at a time in practice (coordination needed) and is CRC only (no FEC) - drop & retransmit at higher layer.

    c Pigeons using flash storage, depending on weather, training, and the bird’s personal sense of urgency (RFC 1149). Redundancy actually implemented by the Bergen Linux User Group in 2001. One packet was eaten. RAID strongly recommended (RFC 2549).

When the AI hunts you down…

GGWave/GibberLink can also operate in a high-frequency mode (I would have included some HF examples here, but alas - Safari-based browsers chop off the higher portion of the spectrum [9] ).

High-frequency mode can only be heard by people of a younger age… and your pet dog.

“Of course, AI-to-AI ultrasonic chatter is nothing to worry about. Unless your smoke alarm beeps in Morse code at 3 AM. Then, maybe, worry a little."
~ ChatGPT

GibberLink is ideal for AI agents making phone calls - reducing time and compute costs. It’s faster than human speech, slower than a V.92 modem, and vastly more efficient… Yet in most real-world cases, any self-respecting AI would just call an Application Programming Interface (API) instead.

So when the hunter-killers come looking for you? No - you won’t hear them coming. They’ll be whispering over Wi-Fi, in encrypted bursts of machine code, laughing at your firewall.

…at least when R2-D2 beeped, the only thing at stake was the hyperdrive.

References

  1. Two AI agents on a phone call realize they’re both AI and switch to a superior audio signal ggwave. In YouTube. ↩
  2. How GibberLink & GGWave Are Changing the Way Machines Talk. In Dr-Business. ↩
  3. What happens when two AI voice assistants have a conversation? In ElevenLabs. ↩
  4. GibberLink lets AI agents call each other in robo-language. In TechCrunch. ↩
  5. Watch Two AI Chatbots Speak to One Another in an Unintelligible Language. In MakeUseOf. ↩
  6. ElevenLabs Hackathon Winners Unveil GibberLink, Boosting AI Voice Agent Efficiency by 80% Across Multiple Cities. In DeepNewz. ↩
  7. AI Agents Can Use R2D2 Sounding Faster Communication. In Next Big Future ↩
  8. ggwave - Tiny data-over-sound library. In GitHub. ↩
  9. ggwave - The higher portion of the audio spectrum in Safari is completely missing - Issue #5. In GitHub. ↩