Russian AI models may lose competitive advantage due to lack of data.
AI Institute of ITMO presented a study on the market of the Large Russian Language Model (LLM) with the participation of 14 experts from Moscow University, Higher School of Economics, RANEPA, AIRI, Yandex, Sever, T-Bank, VK and other organizations.
Russian models are not inferior to foreigners in working with the Russian language, and in some cases they are superior to them in understanding the cultural code and local reality. But for further development they lack quality data.
According to experts, the open Internet in the Russian language is 10-20 times smaller than the global Internet. At the same time, international sources are also required for training universal models: the optimal share of the Russian language in the training corpus is 5-30%.
Where to look for new data?
In libraries, archives, universities, scientific and professional organizations. Such materials can strengthen Russian models in medicine, law, construction and other industries. Expert and synthetic data are also promising: small but well-prepared sets can be more useful than huge sets of plain texts.
However, due to the lack of data, new rules for their use are being discussed.
On November 12, the Presidential Council for Codification approved draft amendments that establish an agreement with the copyright holder, an open license or the use of works with payment of compensation. It is also proposed to give copyright holders the option to prohibit the use of their works for the purpose of training artificial intelligence.
For Russian developers, this may mean additional costs and restrictions. It turns out to be a paradox: to compete with world leaders, local models need more quality data, but new regulatory requirements can make access to it more difficult and expensive.