Mostik Startup Bridges AI Systems for Efficient Communication
A Russian startup, Mostik, has developed a method enabling different artificial intelligence systems to communicate directly via their internal mathematical weights, bypassing text generation.
A Russian startup called Mostik has created a technique that allows artificial intelligence systems to communicate directly without generating text. The method, developed by mathematicians, enables these systems to interact using the mathematical values in their weights, acting as a bridge between different ones.
Sasha Malysheva, Mostik's CEO, developed the approach. She shared a company joke comparing the future of AI to guessing a pig's weight, where combining many estimates yields a better result than relying on a single expert. Similarly, combining outputs from multiple AI systems often improves performance, but traditionally this requires feeding one system's text output into another, which is time-consuming and expensive. Mostik's method bypasses this step entirely.
The Bridge in Practice
To demonstrate, the team built a bridge between two Chinese open-weight systems. They connected the largest version of GLM-5.2 with a much smaller Qwen-3.5 system capable of running on a mobile device.
The resulting hybrid system's performance and cost were compared to the original ones.
| System | Parameter Count | Relative Cost | Performance Note |
|---|---|---|---|
| GLM-5.2 (largest) | 753 billion | Full cost | Baseline |
| Qwen-3.5 (mobile) | 4 billion | Low cost | Baseline |
| Mostik Hybrid | N/A | A fraction of GLM cost | Exactly halfway between GLM and Qwen |
The startup also used its approach to build a system that reached first place on the ARC-AGI 3 leaderboard, a notoriously difficult AI competition. The team declined to provide further details to protect their competitive advantage.
Potential Impact on AI Development
Malysheva argues that combining diverse systems may be a better path for AI advancement than simply scaling up single, monolithic ones. She told the source she personally does not think we will have a monolithic system in the future or that capabilities will come from scaling.
If successful, the technique could increase the value of open-weight systems by letting them compete more effectively with closed, proprietary ones from companies like OpenAI and Anthropic. Karl Tuyls, a former Google DeepMind computer scientist familiar with the tech, says it allows approaching large-system quality without the large system handling the entire loop.
Vladimir Arustamian, tech lead at AI software company Lovable, noted the team's rapid progress. He said this team has been at it for a matter of months and already has something running that he would have guessed was years out.
Mathematical Challenges and Insights
Stanislav Smirnov, a Fields Medalist and Mostik's chief scientist, explained the core difficulty. He said there seems to be no appropriate mathematical language yet for finding common ground between different AI systems. Mostik's approach serves as a practical bridge in this interim.
Smirnov suggested the work might also reveal new insights into how AI systems function and how that compares to human reasoning. A deeper mathematical analysis could uncover commonalities in how both tackle difficult problems.
Malysheva recounted being told the bridge approach would be too difficult, with some peers warning it might be too hard for a young girl. Her response was decisive: she decided she needed to prove them wrong.
