Russian Startup Mostik Enables AI Models to Communicate via Latent Representations
By bypassing human language, Mostik allows AI models to exchange complex internal states more efficiently.
A Russian startup founded by mathematicians is attempting to remove the human language bottleneck from artificial intelligence. Mostik has developed a communication protocol that allows different AI models to collaborate and exchange information without relying on traditional text-based prompting.
Rather than using words, the system utilizes latent representations—hidden mathematical vectors and values—to facilitate interoperability. According to Wired, this non-verbal approach allows models to transfer high-dimensional internal states directly. By bypassing the need to translate complex data into human-readable tokens, the system aims to eliminate the information loss and precision errors that typically occur when one model writes a prompt for another.
The Limits of Text-Based Orchestration
Most current AI orchestration relies on "chaining" or agent-based workflows. In these setups, a primary model acts as a coordinator, generating a text-based instruction that is then passed to a specialized secondary model. This process is inherently limited by the constraints of human language and token limits, which can lead to "hallucinations" or a degradation of nuance during the hand-off between different systems.
Implications for AI Architecture
Moving toward a non-verbal protocol could fundamentally change how AI systems are built. If models can communicate via high-dimensional vectors, they can theoretically transfer nuanced data far more accurately than text allows. This shift suggests a future where a "swarm" of small, highly specialized models can be linked together to create a cohesive system. Such an architecture could potentially outperform a single, massive large language model (LLM) by combining the best of various specialized capabilities without the friction of linguistic translation.
The Path Forward
As Mostik continues to refine this interoperability layer, the industry will be watching to see if this mathematical approach can scale across different model architectures. While the potential for a "super-model" effect is significant, the primary challenge remains ensuring that these latent representations remain compatible across diverse AI frameworks. For now, the project represents a pivot away from the linguistic nature of current generative AI toward a more raw, mathematical form of machine-to-machine collaboration.