The Silent Language of AI

We spend most of our time thinking about AI as a conversational partner. We prompt it in English. It writes back in English. But a team of researchers in Russia is building something different. They are teaching AI models to communicate with one another using non-verbal signals.

Think about how teams actually work. People often reach a level of shorthand where they don't need to explain every detail. They just look at each other, or share a quick gesture, and the work keeps moving. It is high-bandwidth communication that skips the friction of language.

The startup Mostik is trying to bring this to AI. Instead of forcing two models to negotiate using natural language—which is slow, imprecise, and prone to hallucination—they are creating a framework where models can share internal state directly.

This is a massive shift in how we might architect agentic systems. If you have an agent researching a topic, another agent summarizing it, and a third writing code, they currently have to translate their thoughts into text and back again to coordinate. Mostik’s approach suggests we could bypass that entirely. It is a shared "thought space" rather than a chat interface.

For a founder building with these tools, the takeaway is clear. We are moving past the era where every step of an agent's process must be readable by a human. The most efficient systems of the near future will likely involve models that talk to each other in languages we don't speak, achieving results faster and with less overhead.

The value isn't in the model that speaks the best prose. It is in the model that coordinates the most effectively behind the scenes. We are building the nervous system of the digital office, not just another chatbot.

Source:

These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words | https://www.wired.com/story/russian-startup-mostik-ai-models-communication/