The landscape of artificial intelligence is currently dominated by a "bigger is better" philosophy, where frontier labs like OpenAI, Google, and Anthropic race to build monolithic models with trillions of parameters. However, a new breakthrough from a startup called Mostik—the Russian word for "bridge"—is challenging this scaling orthodoxy. By developing a method described as "machine telepathy," Mostik has demonstrated that different AI models can interact and share intelligence directly through their internal mathematical weights, rather than through slow and computationally expensive text-based communication. This approach allows a smaller, more efficient model to inherit the capabilities of a much larger one, potentially shifting the industry toward a more modular and cost-effective future.
The Mechanism of Machine Telepathy
In traditional machine learning workflows, if one model needs to utilize the capabilities of another, the process is sequential: Model A generates a text output, which is then fed as a text input into Model B. This "chatter" between models is inefficient, as it requires the full computational overhead of generating and parsing natural language. Mostik’s approach bypasses this linguistic layer entirely.
The startup’s technology focuses on the "weights" of a model—the numerical values that determine how a neural network processes a prompt into an output. By creating a mathematical bridge between the weight structures of two different models, Mostik allows them to "communicate" in their native tongue of high-dimensional vectors. This enables a large, highly capable "teacher" model to inform the operations of a smaller "student" model in real-time.
In a recent demonstration, Mostik applied this technique to two prominent Chinese open-weight models: the GLM-5.2, a massive architecture boasting 753 billion parameters, and a mobile-friendly 4-billion-parameter version of Qwen-3.5. By bridging these two disparate systems, the team created a hybrid model. The results were startling: the hybrid system achieved a performance level exactly halfway between the two original models, yet it operated at only one-twentieth of the cost of the full GLM model. This suggests a new pathway for achieving "frontier-class" intelligence on consumer-grade hardware.
Validation through the ARC-AGI Challenge
The efficacy of Mostik’s bridge approach is not merely theoretical; it has been validated on one of the most rigorous testing grounds in the AI world: the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI). Unlike standard benchmarks that test a model’s ability to recall facts or predict the next word in a sentence, ARC-AGI—created by Google researcher François Chollet—is designed to measure a model’s ability to learn new concepts and solve novel problems it has never seen before.
ARC-AGI is widely considered a "hard" benchmark because it resists the brute-force memorization that often inflates the scores of large language models. Mostik utilized their weight-bridging technique to build a specialized model for the competition, which quickly ascended to the top of the leaderboard. While the team remains tight-lipped about the specific architecture of their ARC-AGI entry to maintain their competitive edge, the success of the model underscores the potential of weight-based ensembles to handle complex reasoning tasks that stump even the largest monolithic models.
The Mathematical Foundation: Ensembles and the Wisdom of Crowds
The philosophical core of Mostik’s work is rooted in a well-established concept in both statistics and machine learning: the "Wisdom of Crowds." Sasha Malysheva, the CEO of Mostik and the architect of the bridge approach, often cites a classic mathematical anecdote involving the estimation of a pig’s weight. In math circles, it is a known phenomenon that while an individual expert might struggle to guess a pig’s weight accurately, the average of guesses from a diverse group of random people is often remarkably precise.
In the context of AI, this translates to "ensembling." It is a known fact in machine learning that a collection of models (an ensemble) typically outperforms any single model within that group. However, the cost of running multiple large models simultaneously has historically been prohibitive for most applications. Mostik’s breakthrough lies in making these ensembles "talk" to each other at the weight level, drastically reducing the latency and compute requirements that previously made high-level ensembling impractical for real-world use.
Stanislav Smirnov, Mostik’s chief scientist and a 2010 Fields Medalist—the highest honor in mathematics—emphasizes that the project is filling a significant void in mathematical theory. According to Smirnov, there is currently no formal mathematical language that allows two different neural networks to understand each other’s internal representations. Mostik is essentially building that language from the ground up, creating a "bridge" where none existed.
Timeline of Development and Industry Reaction
The rapid development of Mostik’s technology has surprised many in the AI community. While many established labs have spent years attempting to optimize model communication, the Mostik team has moved from concept to a functional, high-performing system in just a matter of months.
- Initial Concept (Late 2023): Sasha Malysheva begins exploring mathematical methods to link neural network weights without re-training models from scratch.
- Proof of Concept (Early 2024): The team successfully bridges small-scale open-weight models, proving that performance gains are possible through weight-sharing.
- Scaling and Benchmarking (Mid-2024): Mostik applies the technique to the GLM and Qwen models and enters the ARC-AGI competition, achieving top-tier results.
- Industry Recognition (Present): Experts from leading AI labs begin to take notice of the startup’s unconventional approach.
Vladimir Arustamian, the tech lead at the AI software company Lovable and an observer of the Mostik team, noted that the speed of their progress is anomalous. "This team has been at it for a matter of months and already has something running that I would have guessed was years out," Arustamian stated.
Similarly, Karl Tuyls, a former computer scientist at Google DeepMind, highlighted the practical utility of the bridge. He noted that the ability to approach large-model quality without requiring the large model to handle the entire computational loop is a "no-brainer" for companies looking to optimize their AI infrastructure.
Challenging the Scaling Hypothesis
For the past several years, the "Scaling Hypothesis"—the idea that simply adding more data and more compute to larger models will eventually lead to Artificial General Intelligence (AGI)—has been the guiding light for Silicon Valley’s AI giants. This has led to a massive surge in demand for Nvidia’s H100 GPUs and a global race for energy resources to power massive data centers.
Malysheva and the Mostik team represent a growing counter-movement. They argue that the future of AI will not be defined by a single, all-knowing monolithic model, but by a "network of specialists."
"I personally do not think we will have a monolithic model [in the future] or that the capabilities of models will come from scaling," Malysheva says. Instead, she envisions a future where a general-purpose frontier model is bridged with highly specialized models—one for biology, one for fluid dynamics, one for legal reasoning, and so on. Mostik’s technology provides the connective tissue that allows these specialists to work together as a single, cohesive unit.
Broader Impact and Implications for the AI Ecosystem
The implications of Mostik’s "machine telepathy" extend far beyond mere cost savings. If the technology becomes widely adopted, it could significantly alter the power dynamics of the AI industry.
1. Empowering Open-Weight Models
Currently, closed-source models like GPT-4 and Claude 3 maintain a lead over open-weight models because their creators can afford the massive compute required to train and run them. However, if Mostik’s bridge allows open-weight models (like Meta’s Llama or Mistral) to be combined into high-performing ensembles at a fraction of the cost, the gap between proprietary and open AI could close rapidly. This would democratize access to high-level intelligence, allowing smaller companies and academic researchers to compete with tech giants.
2. Efficiency and Sustainability
The environmental and economic costs of running massive AI models are becoming a point of contention. By enabling "halfway" performance at 5% of the cost, Mostik provides a path toward sustainable AI. This is particularly relevant for edge computing—running AI on smartphones, medical devices, or autonomous vehicles—where battery life and processing power are limited.
3. Insights into Human Cognition
From a scientific perspective, Mostik’s work may offer clues into the nature of intelligence itself. Chief Scientist Stanislav Smirnov believes that a deeper mathematical analysis of how models "bridge" their reasoning could reveal commonalities between artificial neural networks and the human brain. If a mathematical language can be found to link two different AI models, it might eventually provide a framework for understanding how different regions of the human brain coordinate to solve complex problems.
Overcoming Skepticism and the Road Ahead
Despite their early success, the Mostik team faces significant hurdles. Critics in the field argue that as models become even more complex, the mathematical "bridge" may become too difficult to maintain, or that the performance gains may plateau. Furthermore, the startup is operating in a highly competitive and geopolitically sensitive environment, where talent and capital are concentrated in a few global hubs.
For Malysheva, however, the skepticism is a familiar motivator. Her journey into mathematics began as a response to a challenge from her brother, who doubted her ability to solve Math Olympiad problems. Later, when she proposed the bridge approach, peers warned her that the task was too difficult—some even suggesting it was "too hard for a young girl."
By proving these skeptics wrong and placing at the top of the ARC-AGI leaderboard, Malysheva and her team have demonstrated that the next great leap in AI might not come from a bigger data center, but from a more elegant equation. As the industry continues to grapple with the limits of scaling, the "bridge" built by a small team of mathematicians may well become the standard architecture for the next generation of intelligent systems.
About the Author
