All news

China's DeepSeek neural network could revolutionize AI market — NTI experts

Anton Averyanov, CEO of the ST IT group of companies and an expert at the Technet NTI market, told TASS that the Chinese spent much less to develop DeepSeek than OpenAI did on ChatGPT

MOSCOW, January 28. /TASS/. China’s DeepSeek neural network could revolutionize the AI market, potentially shifting the balance of power in the sphere, experts from the National Technology Initiative (NTI) told TASS.

As they noted, the Chinese neural network can train AI models much more efficiently and is unique in the way it processes requests, allowing analyzed data to be refined.

Earlier it was reported that Nvidia shares on the Frankfurt Stock Exchange lost almost 7% amid the success of the new version of the Chinese chatbot DeepSeek, which, by all accounts, turned out to be more effective than its American competitor ChatGPT of OpenAI (USA).

Anton Averyanov, CEO of the ST IT group of companies and an expert at the Technet NTI market, told TASS that the Chinese spent much less to develop DeepSeek than OpenAI did on ChatGPT.

"DeepSeek, according to [preliminary estimates based on the developers' statements], <…> works even better than GPT-4o. This shows that one can create neural networks with far less resources than it was previously believed.

Moreover, Chinese developers used Huawei chips for this model. We need to wait for independent benchmarks. <…> In general, if the situation is really as described in the project, then the chatbot can be called truly revolutionary in the field of neural networks, AI. Its emergence will upend the top of the global [AI] race," Averyanov believes.

The expert also pointed out that the developers have provided a free model of DeepSeek for users.

"They put the model in the public domain (open source) - a distribution model that makes it possible to freely host the neural network on their own capacities, which are much lower than those of OpenAI and other large language models. As a consequence, the shares of Nvidia and many American companies collapsed," he noted.

More thrifty

The amount DeepSeek spent on training one of its models ($5.5 mln) is several dozen times lower than competing American companies, including Microsoft, Amazon, and Google, said Timofey Voronin, deputy head for technology transfer at the NTI Competence Center for Big Data Storage and Analysis Technologies based at the Lomonosov Moscow State University (MSU).

"In the context of restrictions on the export of advanced chip models to China, which are necessary for creating large language models, Chinese companies are presenting a solution that shows that there is no need to invest such a significant amount of funds in the development of AI models. At the same time, it is more adjusted for the Chinese market, which allows for a significant increase in the accuracy of processing requests, the response to which must take into account the specifics of the region," he told TASS. He added that the model allows downloading text files up to 100 MB for free, which is impossible when using other services.

Voronin also pointed out that when solving logical and mathematical problems, DeepSeek "makes thoughtful decisions, and doesn't just give ready-made answers, as ChatGPT often does."

Innovative new AI

As for the features of the new AI model, key differences are embedded in the arrangement of the neural network training process, says Alexander Bukhanovsky, head of the School of Translational Information Technologies at the ITMO University.

"Firstly, this is an initially balanced arrangement of training data (Internet, books, codes, etc.) in such a way as to fully cover a number of selected tests. Secondly, this is the use of various heuristic rules and models that provide preliminary cleaning of data from various "garbage," as well as the removal of numerous duplicates. A kind of "refining" of data for training occurs. Thirdly, this is the effective use of the memory of the computers on video cards. As a result, due to technical improvements <…> the actual volume of data becomes smaller, and the training speed is higher," he noted.

MSU representative Voronin believes that for now DeepSeek can be the best choice for simple tasks, including source analysis and information search. With more complex tasks that require a high level of elaboration, "ChatGPT is currently better at handling," he added.

Russian chatbots

The expert reminded that Russia also has competitive large language models, for example, GigaChat from Sber, YandexGPT from Yandex and JustGPT from the developer Just AI.

"If we compare these models, we can note a certain balance, since if JustGPT surpasses its competitors in copywriting, then it lags behind in speed, YandexGPT is inferior to GigaChat in speed, but surpasses it in creativity and originality. To compete with foreign solutions, we need to build up infrastructure, including the construction of data centers, and increase the volume of private investment in the development of AI," Voronin believes.

In turn, Bukhanovsky believes that it is more effective not to develop Russian counterparts of foreign neural networks, but to create tools for their adaptation and additional training for specialized tasks.

"It is also possible to develop more complex systems based on them. For example, multi-agent systems, where each agent is responsible for a separate task and contains a model specially trained for this purpose," he concluded.