Lina Bessonova Presented Russian Cognitive Architecture as Alternative to Transformers

Developer Lina Bessonova presented the concept of "Russian cognitive architecture," which she proposes to consider as an alternative to the familiar race for large language models. The main thought of the text is simple: sovereign AI, according to the author, does not begin with a local dataset or purchases of expensive accelerators, but with an attempt to redefine the architecture of the system itself. Instead of predicting the next token, Bessonova proposes to rely on a biomimetic approach, internal states, and accumulation of experience.

The text is constructed as a polemic against what is often called "sovereign AI" in Russia today. The author raises hard questions: is Russian-language data, one's own brand, or access to scarce H100s enough to consider a model truly independent? Her answer is no.

If the foundation remains foreign, and the basic logic is borrowed from American models, then localization, in her opinion, does not transform the product into an independent technological school. In this sense, the article looks not like an announcement of a ready-made mass service, but as a manifesto about what a national AI stack should be in general. Bessonova describes the practical part of the concept through her own stack based on Python, NumPy, and SciPy.

According to her, the system does not require massive training on huge GPU clusters and can work in real time on a local machine, from a home server to a Mac Mini M4 Pro. This is important not only as an engineering solution but also as a political and economic argument: the less dependent on sanctioned and scarce hardware, the higher the chances for autonomous development. The author particularly emphasizes that such architecture is potentially better suited for edge devices, where compactness, power consumption, and the ability to work without the cloud are critical.

The key technical idea of the article is biomimetics instead of statistical prediction. Bessonova contrasts to transformers a system in which the behavior of an agent should be determined not only by word probabilities, but also by the dynamics of internal variables. Among them, she mentions analogues of excitation, inhibition, adaptation, and resource, which are constantly recalculated and influence the model's response.

According to the author's design, this allows us to speak not just of text generation, but of a more integral cognitive scheme, where the response is born from the system's internal state. As scientific support, Bessonova refers to the Russian physiological school and the names of Bekhterev, Chizhevsky, and Pavlov, emphasizing that she seeks the foundation not in the Californian engineering tradition, but in the domestic scientific line. Another important element is learning through "sedimentation of experience."

Unlike models that first absorb giant masses of internet data and then fine-tune, a different path is proposed here: an agent should accumulate personal experience from interaction with the user and the surrounding environment. Such logic brings the system closer to a developing organism than to a conventional language model. At the same time, the author bets on the legal independence of the project: the architecture, according to her, is being prepared for registration in Rospatent as a computer program, and then for patent protection of individual inventive solutions.

Thus, the concept is presented as an attempt to create not only a new tech stack, but also a formally independent Russian intellectual asset. The sharpest part of the text is devoted to criticism of corporate mainstream. Bessonova believes that models of large players, built on the Transformer architecture, remain dependent on foreign "blueprints," even if the interface, language, and brand are local.

RLHF receives particular criticism—learning with reinforcement based on human feedback. The author interprets this approach as a mechanism that makes assistants safe and predictable, but at the same time smooths out cultural acuity, controversy, and intellectual distinctiveness. In her interpretation, the problem is not only in ideology but in infrastructure: if the entire strategy is tied to large data centers and supplies of rare chips, then any talk of technological sovereignty becomes vulnerable.

The conclusion of the article comes down to a change in the very framework of discussion. The question, in essence, is posed as follows: is national AI the one who catches up with leaders faster on their field, or the one who offers a different basic model of machine thinking? While the described architecture looks more like a research program and a worldview statement than a product ready to compete with mass LLMs, the very formulation of the problem is important: the author calls for debate not about the number of GPUs and the size of datasets, but about who sets the initial principles of future intelligent systems.