blog.6admin.io
guest@6admin.io:~$
debunk, folklore, llm

The AI secret language, folklore

The story keeps coming back: a “secret language” AIs supposedly invented to talk to each other, faster and better than us, behind our backs. The version that set the networks on fire in early 2025 was called Gibberlink and the legend that stuck to it promised a protocol “80 percent more efficient than human language”. I want to set the record straight calmly, because the demo itself is real, but the story told on top of it is technological folklore.

What the legend says

The narrative is seductive. Two voice agents would recognize each other as AIs, then switch to a non-human communication mode, a series of beeps and tones, to exchange information at an efficiency beyond speech. The video spreads, the term circles the networks and everyone projects their fantasy of machines plotting in whistles. It is the “secret language, superior to ours” part that travels and it is exactly the part that does not hold.

Contrary to what I wrote too quickly myself, the demo is not a fantasy: Gibberlink came out of an ElevenLabs hackathon in early 2025, the code is open source and the principle is reproducible. Two voice agents, once they have confirmed they are both AIs, stop synthesizing speech and switch to GGWave, a library that is nothing new and transmits data as audio tones, good old data-over-sound, the 1990s acoustic modem brought back. So there is code, a demo and a protocol. What there is not is a language, let alone a secret one more efficient than ours. GGWave is a channel, not a tongue and mistaking it for “AIs invented a superior idiom” is a category error.

The real arithmetic

The “80 percent more efficient than human language” makes you smile, because it compares two things that do not play in the same league. The gain the demo measures is audio-to-data against speech synthesis followed by recognition, over a channel that is only sound, typically two bots on either end of a phone call. In that very specific case, dropping speech to modulate tones directly does save time and reliability. A real niche, a real gain.

But step off the audio channel and the arithmetic collapses. LLMs do not need to “surpass” human language; they already exchange, with each other and with their tools, infinitely faster and more precisely, through text and data structures, JSON, function calls, ordinary network protocols. The moment there is a network, a socket beats any sound modulation, lossless and noise-free. Going back through beeps where a TCP pipe already exists is not an optimization, it is a regression dressed up as science fiction.

Virality is not reality

The real question, then, is not whether AIs have a secret language, they have neither one nor the need, but how to separate the modest demo from the legend piled on top of it. A concept that spreads fast is not a concept that holds; it often spreads precisely because it flatters an intuition rather than because it survives scrutiny. Good hygiene, faced with this kind of announcement, is three questions: where is the code, where is the specification, where is the reproducible demo. For Gibberlink, all three exist and that is exactly what lets you see it is a neat audio hack, not a linguistic breakthrough.

If coordination between models actually interests you, there is concrete work to look at rather than folklore. I have described elsewhere how you really orchestrate several LLMs that deliberate and converge, with no magic protocol, just structured text and a voting mechanism (consensus ) and why the grand narratives about “alignment” (LLM alignment ) and model “liberation” (freeing the LLMs ) deserve the same reading hygiene, where is the code, where is the proof.

At bottom, a modem

  • The “AI secret language, 80 percent more efficient than human” is folklore. The legend is the problem, not the demo.
  • Gibberlink does exist, though: an ElevenLabs hackathon, open-source code and under the hood GGWave, data-over-sound that predates it. It is a channel, not a language.
  • The gain is real but local: only when the single available channel is sound, two voice agents on a call. The moment there is a network, text and sockets win, lossless.
  • Reinventing a sound modulation between networked machines is a regression, not an optimization; what is viral is the staging, not the technique.
  • Faced with this kind of announcement, three reflexes suffice: where is the code, the spec, the reproducible demo. Asking them would have deflated the legend without denying the demo.