Innovation of a technology for communication without any speech between AI models!
In a development that could reshape the future of artificial intelligence, a team of Russian mathematicians working at a startup called Mostik has developed a new method that allows different AI models to communicate with each other without the need to generate intermediate text. According to a Wired report, this approach resembles a form of “machine telepathy.”
In a demonstration, the team connected two different models: GLM-5.2, which has 753 billion parameters, and Qwen-3.5, which has 4 billion parameters. According to Mostik, this hybrid system operates at a cost approximately 20 times lower than running the larger GLM-5.2 model alone, while achieving quality that falls between the results of the two original models.
The core idea is to bypass the traditional method in which large language models interact via text, where one model generates a response, which is then received and processed anew by another model. Mostik proposes transferring information directly through the internal mathematical representations of the models, entirely skipping this intermediate stage. This approach aims to address the challenge of efficiently integrating multiple specialized models while avoiding the computational burden of repeated text conversion, which leads to information loss and resource waste.
However, the technology remains shrouded in mystery so far. The team has not published a scientific paper, open-source code, or a cost-calculation methodology, nor has it released a specific benchmark for comparing GLM and Qwen. The Wired report notes that the company has not yet provided precise technical details, and there is no independent verification of its claims, raising doubts about the novelty of this technology. It remains unclear whether Mostik’s technology represents a genuine innovation or merely a rebranding of existing concepts such as knowledge distillation or communication through embedding spaces.
The scientific research at Mostik is led by Professor Stanislav Smirnov, a Fields Medalist in 2010 and a professor at the University of Geneva. Smirnov states that one of the main tasks is to find a common mathematical language that allows different models to understand each other’s internal representations.
Observers note that releasing standardized data, opening part of the system, or securing significant funding could help validate these claims. In any case, this development highlights the intense effort to find more efficient ways to leverage multiple AI models, a challenge faced by major technology companies worldwide.
