When a “reasoning” neural network from the “Yandex” appears
Yandex (MoEx: YDEX) works on a “reasoning” neural network, the output of which is expected in 2025. The new artificial intelligence (AI) can be used in the « Search with the Neuro » and « Alice ». Analysts believe that total development costs can amount to more than 10 billion rubles. Other companies intend to use the algorithm in new generations of their AI.
Yandex is developing a “reasoning” neural network, and also tests the possibility of “reasoning” in various products, they told Kommersant in the press service. They added that Yandex continues to develop its own generative models and teaches them to reason, analyzing the information deeper. The Kommersant source in Yandex says that you can expect the product output in 2025. Another source familiar with the company's plans adds that the most likely products where a new “reasoning” neural network can be used is “Search with Neuro” and “Alice”.
According to the definition of Openai (the American developer ChatGPT), the “reasoning” AI build a long inner chain of reasoning, check themselves and recognize their mistakes. Also, thanks to this, they can consider several approaches to solving a particular problem within the framework of one answer.
Now several similar AI are represented on the international market. Among them are O3 from Chatgpt, Deepseek R1, QWQ-32B-PreView from Alibaba, Gemini 2.0 Flash Thinking from Google and the latest Claude Sonnet 3.7 from Anthropic (belongs to Amazon), which can work both in reasoning and in the reasoning mode and in the mode normal mode.
Nevertheless, such models have not yet been presented on the Russian market. “We expect that in 2025,“ reasoning ”models will appear in most companies. MTS AI also has mature achievements in this direction, and we will use them in new generations of corporate models of the Cotype family, which will be released this year, ”said Andrei Kalinin, CEO of MTS AI. The “SBER” “Kommersant” said that they are developing their own Gigachat language model, which “ultimately allows you to develop different opportunities, including reasoning”. The company was not called specific dates for the launch.
According to the results of last year, more than half of the largest Russian companies introduced artificial intelligence at least one business function. At the same time, the cost of developing AI increased by 20-30%, and competition for domestic language models is foreign but learned in Russian in Russian (see “Kommersant” of January 21)
“Given international examples, annual total costs of developing such a functionality can be 10 billion rubles. And more. Such models can be monetized not only by the classic subscriptions of B2C users, but also significantly better solving the problems of B2V segment, ”says Maxim Bolotsky, partner, Yakov and Partners.
The initial investment in the new Yandex product, the WMT Group, Natalia Lisitsina, estimates from 500 million to 1 billion rubles. She also adds that it is worth counting on the release of the first version for one or two years. However, most likely, it will be an alpha version of the product that is not ready for mass use. It may require a lot of improvements and long -term training so that II really “reason”.
Nevertheless, today you can create such a model quickly, notes the co -founder of Just AI Kirill Petrov. “In early February, an article by researchers from Stanford and the University of Washington, where they described the method of creating a reasoning model from the usual one, applying which they received such a model in just 26 minutes of training on 16 video cards. The model called S1-32B surpassed the version from Openai O1 Preview for several benchmarks, ”adds Mr. Petrov.
