Mostik, a startup founded by Russian mathematicians, has created a novel approach for AI models to communicate without using text. This method allows smaller models to leverage the capabilities of larger ones more efficiently, potentially enhancing their performance significantly. Mostik's hybrid model recently excelled in the ARC-AGI 3 competition, demonstrating its effectiveness by combining two distinct AI models with varying parameter sizes.
The technique could reshape how AI models are developed and utilized, particularly in making open-weight models more competitive against proprietary systems. By enabling different models to interact directly through their mathematical structures, Mostik aims to advance AI capabilities without relying solely on larger models, which could lead to more specialized applications across various domains.