These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words (www.wired.com)

A startup called Mostik, founded by Russian mathematicians, has developed a novel method for AI models to communicate without using words, relying instead on the mathematical values in their weights. This innovative approach allows larger models to enhance the intelligence of smaller ones more efficiently, as demonstrated by a hybrid model that combines the capabilities of the complex GLM-5.2 and a lightweight Qwen-3.5, achieving a performance level that is cost-effective and competitive in the challenging ARC-AGI 3 competition. This breakthrough could significantly alter the landscape for open-weight AI models, enabling them to rival proprietary systems from major players like Anthropic and OpenAI. By enabling models to "talk" mathematically rather than through text output, Mostik’s technique promises to streamline the process of combining model outputs, making it faster and easier to develop specialized models across various domains. Experts believe this could lead to a more collaborative approach to AI development, emphasizing combinations of models over singular, bulkier solutions, potentially offering new insights into both AI mechanics and human cognitive processes.