The head of the Russian Academy of Sciences said that it will become more difficult for a person to control AI and that it needs to be entrusted to another AI

The head of the Russian Academy of Sciences, Gennady Krasnikov, said that the development of artificial intelligence will accelerate, and it will become more difficult for humans to control it.

To ensure safety, it is necessary to create models in which one AI controls another.

The basis of artificial intelligence is language models, and in this direction, solutions from domestic companies are preferable, while for applied aspects more accurate models trained on large databases are needed.

MOSCOW, September 27 - RIA Novosti. Artificial intelligence technologies will continue to develop, and it will be more difficult for a person to control it - this process must be entrusted to another AI model, the head of the Russian Academy of Sciences, a scientist in the field of micro- and nanoelectronics, Gennady Krasnikov, told RIA Novosti.

“It is impossible to stop the development of artificial intelligence, and at a certain stage it will be more difficult for a person to control this development. But this must be done by another artificial intelligence,” the academician said.

“We must create models where one AI controls another AI, maybe a third one, in order to ensure greater security for ourselves in this matter,” Krasnikov added.

It is necessary to very carefully analyze the vulnerabilities of the frameworks that make up certain language models, the head of the Russian Academy of Sciences pointed out. Therefore, rules are needed to regulate the creation of various machine learning methods and applied models for artificial intelligence.

The basis of artificial intelligence is language models; they are needed to create sovereign AI, and solutions from domestic companies are preferable in this direction, Krasnikov explained.

“And when it comes to applied aspects - for example, the creation of drugs for many serious diseases, then there should already be more accurate models trained on large databases. They should be an order of magnitude more accurate than language models for these tasks,” the academician said. And here you need to have the best models - either create your own or take them abroad. The only requirement is that they all must be trusted, that is, they must be protected and produce correct results, the head of the RAS concluded.

Earlier in the public space there was a discussion about the risks associated with AI. A couple of employees of the artificial intelligence company Anthropic shared their opinion that in 10 years, AI has a 10% chance of destroying humanity.