MTS AI has developed a language model for the corporate sector

MTS AI has developed a large language model designed for use in the corporate sector – in particular, for creating extracts from documents and telephone conversations. The model supports only text queries, but, as the company clarified to Kommersant, products for photo, video and code generation are in development. Market participants admit that the language model can also be used in the consumer sector of MTS products (MOEX: MTSS). They suggest that when developing it, the company relied on data collected by other divisions of the group.

The MTS subsidiary in the field of artificial intelligence (MTS AI) has developed a large language model (LLM) MTS AI Chat for use in the corporate sector, the company told Kommersant. Among the scenarios in which the development can be used are drawing up job descriptions, extracting information from documents, compiling summaries of telephone conversations, etc. The project is offered according to the deployment model on the customer’s equipment (on-premises). “MTS AI will inform you additionally about the release of the technology into public access.”

In July 2023, MTS was looking for employees on hh.ru for the position of data scientist to “train the MTS-ruGPT and MTS-ChatGPT models” for use in business. The latest model, according to the description of the now deleted vacancy, had “support for pictures and videos.” MTS clarified that MTS AI Chat supports text queries, but the company is working on “a separate product that can recognize both videos and images.”

A code generation and completion service is also being developed. Sber launched a similar GigaCode in early access mode in 2023; Yandex and Rostelecom are developing in this direction (see Kommersant, September 27, 2023).

According to a joint study by Yakov and Partners and Yandex, published in December, 20% of large companies are already using generative AI based on OpenAI technologies (the company does not officially provide access to them in the Russian Federation), YandexGPT (for text generation) and Kandinsky from Sberbank (for images). A significant economic effect from the introduction of LLM is expected “in agriculture, healthcare, industry, transport and construction,” the AI ​​Alliance (Yandex, VK, MTS, Gazprom, etc.) told Kommersant.

When creating the language model, MTS could use anonymized data collected by other divisions of the group, says Larisa Malkova, managing director of the Data and Applied AI practice at Axenix. According to her, it is impossible to assess the competitiveness of the company’s development without access to it. As a general rule, Ms. Malkova clarifies, the success of AI products for business depends on a combination of factors: “Firstly, the developer must be able to conduct fundamental research himself, secondly, to anticipate the specific needs of customers, and thirdly, to help them overcome barriers to implementation.” None of the developers offers this fully, she believes.

For use in a corporate environment, it is also important whether data leaks from the customer’s circuit, as well as hardware requirements, says Just AI Business Development Director Svetlana Zakharova: “The customer needs to understand how much hardware is required to service the model.”

The AI ​​developments of Yandex and Sber are presented both in the companies’ consumer products and in the corporate sector (see Kommersant, December 7, 2023). In the future, Ms. Zakharova believes, MTS can integrate the basic technology into its consumer projects (b2c), where there are “existing clients and an audience on which to test the technology.” But the first b2c applications of MTS AI Chat, in her opinion, will be limited: “I think that MTS will start with MVP (minimum viable products.— “Kommersant”), will look at the reaction of the target audience and begin to further train the model.”

Yuri Litvinenko, Nikita Korolev