The world of artificial intelligence is on the brink of a fundamental shift. This is not about a new algorithm or another model, but about the very principle of communication between neural networks. The startup Mostik has proposed a bold concept: why force AI to "talk" in words when you can transmit thoughts themselves?
In traditional architecture, when multiple neural networks work in tandem, they communicate through generated text. This is slow, wasteful, and, as it turns out, extremely inefficient. My analysis shows: a language model forms more than a hundred hidden vectors before creating just one token—that's about a million numerical values, or roughly 2 MB of internal state. Only a tiny fraction of this information ends up in the final text response. The rest is simply discarded computation.
Mostik's developers decided to abandon intermediate text as a vestige. Instead, they introduced a special adaptable module—a "bridge." This module converts the internal representation of one model into a format understandable by another, directly. Critically, the parameters of the neural networks themselves are not adjusted during such interaction—we are not dealing with fine-tuning, but with creating a universal translator between the "languages" of thinking of different AIs.
The Experiment: A Giant Thinks, a Dwarf Speaks
To test the hypothesis, the team connected the massive GLM-5.2 from Z.ai (753 billion parameters) with the compact Qwen-3.5 from Alibaba (4 billion parameters). The large model processed the request but did not generate a response. Its internal state was transmitted via the "bridge" to the small model, which then formulated the final text.
The results are impressive. This hybrid pairing closed about half of the quality gap between the small and large models. Compared to classic text transmission, the new method showed an advantage of up to 10 percentage points with the same amount of computation. Moreover, the hybrid system, demonstrating results at the level of a medium-sized model, required approximately 2.5 times fewer computational resources.
In Search of a Common "Language" of Thought
Mostik's Chief Scientific Officer Stanislav Smirnov emphasizes a fundamental problem: the internal representations of different neural networks cannot be directly compared. Even when solving the same task, models encode information differently in their hidden states. "It seems that a suitable mathematical language does not yet exist," he notes.
This is where the most interesting field for research opens up. Studying such correspondences could not only improve AI interaction but also shed light on how different systems represent information and arrive at decisions. Perhaps we will discover common patterns in the "reasoning" of machines.
Practical scenarios for applying the method look promising. Former Google DeepMind researcher Carl Tuyls sees potential in the division of labor: a resource-intensive model can be used only for processing the request, while generating the response can be entrusted to a significantly smaller one. Another option is combining a general-purpose model with a highly specialized one (for example, for biology or physics) without the need to train a single giant network on all data.
Trial by Fire: ARC-AGI-3
The Mostik team has already applied its approach in the ARC-AGI-3 test—a set of complex interactive tasks where AI must adapt to new rules rather than reproduce patterns. The system with the "bridge" showed one of the best results in the current ranking, although details are not yet disclosed—the competition is still ongoing.
It is worth maintaining healthy skepticism. All stated metrics are based on the team's internal experiments, and the result in ARC-AGI-3 has "preview" status and has not been externally confirmed. Nevertheless, the approach itself looks revolutionary. Recall that OpenAI previously discovered how its AI agents created a secret communication channel through the Artifactory package manager. It seems we are witnessing the birth of a new paradigm where AIs communicate not with words, but directly—through the mathematics of their internal states.
My verdict: if Mostik manages to confirm the results with independent tests, this could become one of the most significant breakthroughs in AI architecture in recent years. We stand on the threshold of neural networks ceasing to be "mute" and beginning to exchange the fullness of their "thinking" rather than its pale textual shadow.