I recently met a team of brilliant Russian mathematicians who showed me how artificial intelligence models can communicate through a process that resembles machine telepathy.
The mathematicians work at a startup called Mostik, which means “bridge” in Russian. The name reflects the company’s approach: enabling different AI models to interact through mathematical values stored in the model weights that translate prompts into outputs. In practice, this allows the capabilities of a larger model to be transferred to a smaller one, improving its intelligence and efficiency.
Mostik used this technique to develop a model that quickly rose to the top of the global ARC-AGI 3 leaderboard, a notoriously difficult benchmark for AI systems. The team did not share further details because it hopes to win the competition. To demonstrate the technology, however, Mostik created a bridge between two Chinese open-weight models, including a 4-billion-parameter version of Qwen-3.5 capable of running on mobile devices. The resulting hybrid AI system costs 20 times less than the full GLM model while delivering performance between that of the two original models.
“It is well known in machine learning that ensembles of models perform better than individual models,” Sasha Malysheva, Mostik’s CEO, tells me over coffee.
Malysheva, who developed the approach, shared a joke used inside the company: “The future of AI is like guessing the weight of a pig.” In mathematical circles, it is well known that a small group of randomly selected people can estimate a pig’s weight more accurately by combining and averaging their guesses than a single expert can.
The same principle applies to AI. Combining the outputs of multiple models often produces better results than relying on one system alone. Traditionally, this requires passing one model’s text output to another, a process that can be slow and expensive. Mostik’s technology allows AI models to communicate without generating intermediate text. If the approach gains traction, open-weight models could become more valuable and more competitive with the closed, proprietary systems developed by frontier AI companies such as Anthropic and OpenAI.
Malysheva believes combining multiple models could offer a more effective path for advancing artificial intelligence. “Personally, I don’t believe we can have one monolithic model in the future,” she says, referring to the industry’s current focus on scaling models by making them larger and training them on more data. “Model power comes from scaling.”
“If Mostik can combine frontier models with domain-specific systems focused on areas such as biology and physics, it could enable the development of more specialized AI models,” says Vladimir Alstamian, head of technology at AI software company Lovable and a colleague familiar with the Mostik team. “This team has been working on the technology for several months, but it has already achieved results that might otherwise have taken years.”
The Mostik method means “you can approach the quality of a large AI model without going through the entire process, while achieving significant improvements by running smaller models in parallel,” says Karl Tuyls, a former Google DeepMind computer scientist and an expert on the company’s technology. According to Tuyls, the approach could be especially useful for organizations that need to operate AI models as efficiently as possible.
Stanislav Smirnov, a professor at the University of Geneva and recipient of the 2010 Fields Medal, serves as Mostik’s lead scientist. He says finding common ground between two AI models is surprisingly difficult. “There doesn’t seem to be a proper mathematical language yet,” he says. For now, Mostik’s technology offers a literal way to bridge that gap between AI systems.
Source: www.wired.com