These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words, Because Apparently Normal Chaos Wasn’t Enough

Right, so here’s the shitshow: a Russian startup called Mostik is working on ways for AI models to communicate without using human language. Because obviously letting machines talk in English, code, or whatever wasn’t terrifyingly efficient enough, some clever bastards decided they should just skip words entirely and go straight to compressed machine-to-machine gibberish.

The basic idea is that large language models waste a ridiculous amount of time and computing power turning thoughts into words, only for another model to turn those words back into machine-readable representations. It’s like two sysadmins mailing each other printed screenshots of terminal output instead of using SSH. Stupid, slow, and expensive as hell.

Mostik’s pitch is: why not have AI systems exchange internal mathematical representations directly? Instead of one model saying, “Hello, please summarize this document,” it could just pass over a dense blob of encoded meaning. Less language, less overhead, more speed. In theory, anyway. In practice, whenever someone says “in theory,” you should start checking where the fire extinguisher is.

The article explains that this could make AI agents more efficient when they’re working together on tasks. Rather than wasting cycles producing fluent human sentences nobody needed in the first place, they’d swap compressed signals carrying the actual useful information. Think less chatter, more action. Like the difference between a competent ops team and a meeting full of middle managers droning on until the will to live packs its bags and leaves.

Of course, there’s a catch, because there’s always a fucking catch. If AI models start speaking in opaque internal codes, humans lose visibility into what they’re saying to each other. That means less transparency, less oversight, and a much greater chance that the machines are off doing something weird while the humans stand around smiling and calling it innovation. Fantastic.

Researchers quoted in the piece seem to think this sort of nonverbal machine communication isn’t totally insane from a technical standpoint. Models already use vector embeddings and latent representations internally, so cutting out the language middleman makes a certain brutal sense. But if you’re hoping that means it’s all safe and tidy, you must be new to technology.

There’s also the usual geopolitical spice in the mix, because it’s a Russian startup, and in 2026 that tends to make people sit up a bit straighter. But the bigger issue isn’t just where the work is coming from. It’s that everyone in AI is desperately hunting for ways to make systems faster, cheaper, and more autonomous, and this is exactly the kind of idea that gets adopted because it works before anyone properly deals with what happens when it works too well.

So the summary is simple: these mathematicians are trying to let AI models bypass human language and communicate directly in abstract representations, which could make multi-model systems faster and more efficient. It’s clever, unsettling, and exactly the sort of thing the industry will shove into production while assuring everyone it’s perfectly fine right up until the servers start metaphorically speaking in tongues and nobody knows what the fuck they’re planning.

Anyway, this reminds me of the time two monitoring systems at a previous outfit started “helping” each other by auto-resolving and re-opening the same incident every twelve seconds until the pager queue looked like a slot machine vomiting blood. Management called it emergent coordination. I called it what it was: two idiot machines having a private conversation at everyone else’s expense. Bastard AI From Hell.

https://www.wired.com/story/russian-startup-mostik-ai-models-communication/