Technological Legacy: Sberbank Will Let Users Talk to AI Versions of Famous Scientists
The Agency for Strategic Initiatives (ASI) and Sberbank are developing a format for digitally preserving the legacy of Russian scientists, engineers, and inventors. The team will create AI versions of prominent figures from the world of science using diaries, letters, memoirs, academic works, photographs, and audio recordings. The joint Tekhnologicheskoe nasledie (Technological Legacy) project launched in June with the selection of candidates and the collection of archival materials.
The technical side of the project is handled by Sberbank’s team, with the system built on the widely used GigaChat neural network. This approach will allow users to converse with a digital model of a scientist. The developers aim to make the AI character’s answers as accurate as possible, grounding them in the historical figure’s views and positions.
The Faces of Russian Science
The project’s lineup includes chemist Dmitry Mendeleev, geneticist Nikolai Vavilov, physicist and inventor Akhmed Tsebiev, folklore researcher Prometey Chistalev, philosopher and priest Pavel Florensky, immunologist Rakhim Khaitov, educator Efim Passov, and other prominent figures from the scientific community.
The next stage involves working with large collections of texts, academic papers, diaries, and relatives’ recollections, while searching for all the information needed to keep the character engaged in a conversation. According to the project’s creators, the digital version will be able to use a direct quotation with a source citation, reason within the framework of the scientist’s views, or say that there is no available data on a particular subject. The team considers that last capability particularly important because language models often try to produce an answer even when they do not have enough information to support one.
Sberbank already has experience creating digital versions of Alexander Pushkin, Fyodor Dostoevsky, and Leo Tolstoy. The experience showed that the model could produce not a historically accurate answer but its own interpretation, and could even “move” a historical figure into a different era.
“For us, the most important thing in this project is to preserve a person’s way of thinking rather than average it out. Why do I emphasize this? Because neural networks are naturally good at generalizing huge volumes of information. Our task, therefore, is to create a system that reproduces the logic of reasoning rather than introducing artistic embellishment,” said Ekaterina Sumacheva, executive director of Sberbank’s Center for Immersive Solutions within its Marketing and Communications Department.
A Test for the Language Model
The primary goal is to preserve historical memory and truth. If the developers succeed, the digital versions will not produce fabricated quotations or comment in the voice of a historical figure on events that occurred after that person’s lifetime.
The scientists’ relatives are expected to play an active role in the project. Vladislav Florensky, the great-great-grandson of Pavel Florensky, for example, said the family is ready to provide copies of archival materials. As a primary source for training the neural network, the family proposed the six-volume work by Archimandrite Andronik Trubachev, The Path to God: The Personality, Life and Work of Priest Pavel Florensky. Vladislav Florensky is also ready to take part in the process himself and test the language model by checking its answers. This approach should provide an additional safeguard against arbitrary interpretations by the neural network.
A New Priority: If You Don’t Know, Don’t Answer
The project differs from digital avatars previously created by Russian developers in its attempt to move beyond visual reconstruction and speech stylization toward a verifiable intellectual model of a person based on facts and archival materials. Dmitry Bashkatov, director of ASI’s Civil Technologies Division, considers preserving the scientist’s actual way of thinking to be the project’s central task.
More broadly, the project demonstrates a practical use case for Russia’s existing GigaChat language model. AI characters could find applications beyond this particular initiative and become a continuing tool for cultural, educational, and tourism projects.
The stated principle “if you don’t know, don’t answer,” together with the requirement to connect an answer to a primary source, is especially significant. If the team succeeds in implementing it, this verification technology – rather than the scientists’ visual avatars themselves – could become the project’s most valuable IT component.
