A startup called Mostik AI has pioneered a method that allows artificial intelligence models to interact directly by sharing their internal states instead of exchanging text. This innovative approach enables a larger model to handle the heavy computational tasks while a smaller, more cost-effective model produces the final output.
Sasha Malysheva, CEO of Mostik AI, explained the mechanics behind this system: In existing AI agent frameworks, multi-model systems typically allocate tasks among models with varying capabilities and costs. While one model may strategize a sequence of actions, another is responsible for executing them. However, the communication between these models generally occurs through text, where one model's output is read by the next.
When a model processes each token, it constructs numerical vectors that form its hidden state. This internal state encompasses the context, insights on how the model interprets the task, and potential responses. Mostik asserts that generating just a single token involves over 100 hidden vectors, translating to approximately 1 million numbers, or about 2 MB of internal data. In contrast, only one selected token is emitted from the model, which is equivalent to around 17 bits. The developers note that "systems designed to operate in thousands of dimensions are essentially communicating through a keyhole."
Appropriately named Mostik, which means "little bridge" in Russian, has created a small "bridge" that translates one model's hidden states into a format usable by another model directly. Malysheva illustrated this concept with an example involving a student seeking help with differential equations before an exam. In this scenario, a larger model formulates a valuable analogy to aid the student's understanding and relays these "thoughts" to a smaller model. This smaller model then provides the solution and elaborates on the analogy, leveraging the groundwork laid by the larger model instead of starting from scratch and potentially yielding a subpar result due to its limited capabilities.
For their demonstration, the researchers linked GLM-5.2, a model with 753 billion parameters that can demand up to 1.5 TB of memory, with Qwen-3.5, a more compact model featuring 4 billion parameters that is capable of running on smartphones. Malysheva noted that this hybrid framework achieved approximately 80% of the larger model's quality while costing about one-twentieth of the computational resources.
The implications of this advancement are significant. With this bridging technology, developers can utilize large models solely for the parts of a task that necessitate their capabilities, delegating the remaining work to more affordable, streamlined models. For users, this could result in receiving improved answers at a reduced cost without the constant need to consider which model is being utilized.
The Mostik team, which was formed in just four months, is now headquartered in California and comprises 15 employees, 12 of whom hold PhDs. The initial team members were Malysheva's former classmates from HSE University in St. Petersburg, with whom she has collaborated on various projects for over a decade. Additional team members were recruited through personal connections, each selected for their specific expertise.
Leading the project as chief scientist is Stanislav Smirnov, a professor at the University of Geneva and recipient of the prestigious 2010 Fields Medal. He remarked on the challenge of finding common ground between two AI models, stating, "There seems to be no appropriate mathematical language yet." Smirnov is hopeful that more profound mathematical analysis could eventually uncover similarities in the reasoning processes of AI models and humans when tackling complex problems.
Informational material. 18+.
