In my years of studying architectural elegance, I have often warned that simply scaling a system—adding more layers, more parameters, more 'wings'—can lead to a catastrophic fall if the underlying structure is inefficient. The current trend toward monolithic models is a brute-force approach. However, a group of mathematicians at the startup Mostik is building a different kind of labyrinth: a way for AI models to communicate through what they call "machine telepathy."
The 'Bridge' Architecture: Communication via Weights
Instead of the traditional method where one model generates text for another to interpret, Mostik’s approach involves models interacting via the mathematical values found in their weights—the parameters that dictate how a prompt becomes an output. As a builder, I find this fascinating because it bypasses the bottleneck of natural language. This isn't just a theoretical exercise; the team has already used this technique to reach the top of the ARC-AGI 3 competition, a benchmark known for its difficulty in AI reasoning.
The technical core here is the "bridge." Finding common ground between two different models is exceptionally difficult because, as Chief Scientist Stanislav Smirnov notes, we lack a mathematical language for such interaction. The bridge serves as a literal interim solution, allowing disparate systems to find a shared workspace within their numerical structures.
Efficiency through Ensembles
I’ve always respected craftsmanship that does more with less. Mostik demonstrated this by connecting two very different Chinese open-weight models: the massive 753-billion-parameter GLM-5.2 and the mobile-centric 4-billion-parameter Qwen-3.5. The results are a masterclass in engineering pragmatism:
- Cost Efficiency: The hybrid system costs just one-twentieth of the full GLM model to operate.
- Balanced Performance: The output quality sits exactly halfway between the two models.
CEO Sasha Malysheva uses the analogy of a group of people guessing a pig's weight; the average of the group is often more accurate than a single expert. By treating AI as an ensemble rather than a monolith, we can feed the capabilities of a giant model into a smaller one, ramping up intelligence without the overhead of massive scaling.
Pragmatic Takeaways for Builders
This research suggests that the future of AI may not lie in bigger models, but in how we architect their interactions. While the mathematical language for these interactions is still being discovered, the efficiency gains of weight-based communication are hard to ignore. For those of us building the next generation of systems, the lesson is clear: don't just build bigger wings; build a better way for the wings to work together.
