A group of Russian mathematicians at the startup Mostik—the Russian word for "bridge"—has developed a radical approach that allows artificial intelligence models to interact through a process akin to "machine telepathy." Instead of exchanging text, the models communicate using the mathematical values found in their weights—the parameters that determine how a prompt is transformed into an output. The startup has already used this technique to build a model that reached the top of ARC-AGI 3, a notoriously difficult competition for AI reasoning.
The Power of Ensembles
This method allows the capabilities of a large model to be fed into a smaller one, ramping up its intelligence with remarkable efficiency. According to Sasha Malysheva, CEO of Mostik, the future of AI does not lie in "monolithic" models that rely solely on scaling. Instead, she likens their approach to the mathematical anecdote of guessing a pig's weight: a group of random people can often estimate the weight more accurately than an expert when their guesses are combined and averaged.
A Demonstration of Efficiency
To demonstrate the effectiveness of the "bridge," the team connected two Chinese open-weight models: the largest version of GLM-5.2 (753 billion parameters) and a 4-billion-parameter version of Qwen-3.5 designed for mobile devices. The resulting hybrid system:
- Costs just one-twentieth of the full GLM model to run.
- Delivers performance that sits exactly halfway between the two models.
Scientific Implications
Stanislav Smirnov, Mostik’s chief scientist and a 2010 Fields Medalist, notes that finding common ground between two different AI models is surprisingly difficult because an appropriate mathematical language for such interaction does not yet exist. Mostik’s bridge is a literal interim solution. Smirnov suggests that this work could eventually reveal new insights into how AI models function compared to the human brain, potentially uncovering commonalities in how both reason through difficult problems.
