These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words
Russian mathematicians at startup Mostik developed a method for AI models to communicate without producing text output, potentially increasing the value of open-weight models.
Intelligence analysis by Qwen 2.5 (3B)
Russian mathematicians at startup Mostik created a technique for AI models to communicate without text output, potentially enhancing the performance of open-weight models.
Russian mathematicians found a way for different AI models to talk to each other without making text. This lets them use smaller models to get the same results as bigger ones, saving time and money.
Analysis
The Bridge Approach
The approach developed by Mostik is a novel way to allow different AI models to communicate and combine their capabilities. This is achieved by using the mathematical values found in their weights, which determine how a prompt gets turned into an output. This method is particularly useful for combining the capabilities of a larger model with a smaller one one, allowing for more efficient use of resources.
The Hybrid System
The Mostik team demonstrated their approach by creating a hybrid system between two Chinese open-weight models: the largest version of GLM-5.2, which has 753 billion parameters, and a 4-billion-parameter version of Qwen-3.5. The resulting hybrid system costs one-twentieth of the full GLM model and performs exactly halfway between the two.
The Future of AI
The Mostik technique could increase the value of open-weight models, allowing them to better compete with the closed, proprietary models offered by frontier labs like Anthropic and OpenAI. The team believes that combining lots of different models may turn out to be a better way to advance AI, and that this approach could reveal new things about how AI models actually function and how this compares to the workings of the human brain.
Key points
- Russian mathematicians at startup Mostik developed a method for AI models to communicate without text output.
- The approach uses the mathematical values found in the weights of different models to combine their capabilities.
- The resulting hybrid system costs one-twentieth of the full GLM model and performs exactly halfway between the two models.
- The team believes this technique could increase the value of open-weight models and advance AI research.
- The technique could reveal new things about how AI models function and compare to the human brain.
This technique could lead to more efficient and specialized AI models, potentially improving the performance of open-weight models and advancing AI research.
If this technique is widely adopted, it could lead to a decrease in the value of proprietary AI models, as open-weight models become more competitive.
