MTS is testing a machine learning platform on third-party devices

Kommersant discovered that the MTS structure is testing Roy 9, a product that involves the use of third-party devices for the development and training of neural network models. MTS assured that this is about testing the company’s “hypotheses” and there are no plans for a real launch yet. Meanwhile, experts consider the approach promising, as it will save computing power.

Kommersant discovered in the SPARK-Interfax database that the MTS structure that deals with cybersecurity, Serenity Cyber ​​Security LLC (MTS RED), registered the domain roy9.ru on November 1. Judging by the description on the page (“Kommersant” saved copies of it), the company is testing the product “Roy 9”: we are talking about users renting out the computing power of their devices, including mobile ones, for distributed training of machine learning models (machine learning, ML).

From the information on the website it follows that four neural network models are currently being trained on the platform: “New AI models will help people make new discoveries, you can become part of this movement. Plus, by helping to train models, you make the world a more interesting place.” The Security Code company, whose representative, at the request of Kommersant, got acquainted with the site, suggested that the essence of the project is to use the power of third-party devices to train neural networks.

MTS assured Kommersant that the site was created only for testing hypotheses in the field of artificial intelligence and machine learning: “The resource is intended for internal needs, there are no plans for a product on this basis.” A company representative clarified that the testing is carried out by the MTS Future Crew innovation center. After Kommersant’s request, MTS closed access to the site.

“Security Code” expert Maria Fesenko notes that a similar concept for training neural networks is already practiced abroad: “Only devices or open platforms are usually used so that language models can be trained on user data. In the case of MTS, the neural network models will use only the power of the devices of the project participants.”

Kommersant’s interlocutor in the IT market explains that the required amount of computing power for such projects depends on the level of maturity of language models: “For example, training GPT-4 required 25 thousand video cards, which in total worked about 57 million hours, and the project MosaicML MPT-7B – 440 similar cards that have worked for more than 100 thousand hours.”

Large Russian IT companies have already created platforms for the development of ML systems. Last year, VK launched an open service for developing solutions based on machine learning, VK Cloud Solutions. It is necessary for conducting experiments with data, testing hypotheses, etc. The service is suitable for both training and working with neural networks, VK explained to Kommersant. Yandex has also been working in this direction for several years: its platform for creating and testing machine learning algorithms DataSphere appeared in 2020.

MTS is also developing services based on machine learning: in April 2022, the company’s vice president, Evgeniy Chereshnev, spoke about a self-learning automated pentest system (testing the company’s infrastructure for stability.— “Kommersant”) CICADA 8: “When connected, CICADA begins to break the infrastructure, just like hackers do. She’s still working worse than hackers, but you can’t imagine how quickly she learns.”

“The shortage of available computing power for training AI models is a reality, and the idea of ​​using some kind of crowdfunding would solve the problem,” says Nikita Tsaplin, CEO of the hosting provider RUVDS. He admits that if the product is successfully tested, the operator will provide a discount to subscribers who allow them to use the power of their smartphones, for example, at night – “in the regions this may be quite in demand.” Smartphones, the expert clarifies, are comparable in power to computers; for the company this would be a “low-cost and simple way” to cope with the lack of resources.

Tatiana Isakova, Nikita Korolev