These Russian Mathematicians Taught AI Models How to Talk to Each Other Without Using Words
wired ·
A startup called Mostik, founded by Russian mathematicians, has developed a method that lets AI models communicate through their internal mathematical weights rather than text, enabling a larger model's capabilities to boost a smaller one efficiently. The approach helped create a model that topped the ARC-AGI 3 competition, and a demo bridged a 753-billion-parameter GLM model with a 4-billion-parameter Qwen model, yielding a system costing one-twentieth as much with performance halfway between the two. CEO Sasha Malysheva compares the benefits to how averaged guesses from a group can outperform an expert, and argues that combining models may be more valuable than building ever-larger monolithic ones. Outside experts say the technique could make open-weight models more competitive with proprietary ones and encourage specialized models to flourish. Chief scientist Stanislav Smirnov, a Fields Medalist, notes that no shared mathematical language yet exists for such communication, and he hopes deeper analysis might reveal connections to human reasoning. Malysheva also shared that she was motivated to prove wrong peers who doubted she could pull off the difficult approach because of her age and gender.
wired