A Russian startup named Mostik has unveiled a breakthrough method allowing AI models to communicate via mathematical weights rather than text, potentially revolutionizing model efficiency.
- Mostik startup has developed a method for AI models to interact via 'machine telepathy' using mathematical weights.
- This approach bypasses the need for text output, saving massive amounts of time and compute.
- It allows small, efficient models to achieve performance levels close to giant, expensive models.
In a profound leap for artificial intelligence, a group of brilliant Russian mathematicians has demonstrated a way for AI models to communicate through something akin to machine telepathy. The startup, Mostik (the Russian word for 'bridge'), is pioneering a method where models interact using the mathematical values found in their weights, rather than generating and processing text.
This breakthrough addresses one of the most significant bottlenecks in modern AI: the cost and latency of model ensembles. Traditionally, to combine the intelligence of two models, one must feed the text output of the first into the second. This process is computationally expensive and slow. Mostik's approach allows the capabilities of a massive model to be directly integrated into the mathematical framework of a smaller model.
Why This Matters
BozokMedia analysis shows that this technology could fundamentally shift the power balance in the AI industry. Currently, the race is dominated by 'frontier labs' like OpenAI and Anthropic, who focus on building massive, monolithic models. Mostik's technique offers a path to high-level intelligence through efficient, specialized, and interconnected smaller models.
The future of AI may not come from simply making models larger, but from how intelligently we bridge them together.
To prove the concept, the team created a bridge between two Chinese open-weight models: the massive GLM-5.2 (753 billion parameters) and the mobile-friendly Qwen-3.5 (4 billion parameters). The resulting hybrid system functioned at one-twentieth of the cost of the full GLM model, while delivering performance that sat perfectly between the two original models.
Historical Background
The concept of 'Ensemble Learning'—using multiple models to achieve better results—has long been a staple of machine learning. It is often compared to the 'wisdom of the crowd,' where the average of many guesses is more accurate than a single expert. However, the 'bridge' Mostik is building is the first to operate at the level of raw mathematical weights rather than linguistic outputs.
| Feature | Traditional Ensemble | Mostik Approach |
|---|---|---|
| Communication Medium | Text/Data Output | Mathematical Weights |
| Operational Cost | High | Extremely Low |
| Efficiency | Moderate | Very High |
The team is led by CEO Sasha Malysheva and Chief Scientist Stanislav Smirnov, a 2010 Fields Medalist. Smirnov notes that finding a common mathematical language between different AI architectures is incredibly difficult, but Mostik's method provides a literal bridge to cross that gap.
Frequently Asked Questions
Question 1: How is this different from standard AI interaction?
Answer: Standard AI uses text to communicate; Mostik's AI uses the underlying mathematical values (weights) that define the model's intelligence.
Question 2: Can this make AI cheaper?
Answer: Yes, by allowing smaller models to act with the intelligence of larger ones, it significantly reduces the hardware and energy required.
