The neural network jammed gears – Kommersant

The emergence of new technologies, the principles of which have not yet entered the body of general knowledge, historically awakens different reactions in people: in some cases this is a desire to understand the principle of operation, in others – rejection based on prejudices. Recently, recommender algorithms on the Internet have been such a popular, but few really understandable topic, now the turn of generative neural networks has come.

The other day, the chairman of the Just Russia – For Truth (SRZP) party, Sergei Mironov, was indignant at the results of the new version of the Kandinsky neural network from Sberbank. He did not like that some “representatives of the expert community” could not generate with its help the flag of Russia in the correct colors or the image of the “Z-patriot” they expected. It is unclear whether the SRZP leader tried to verify this information himself, but he asked the Prosecutor General Igor Krasnov about this, fearing that the development of Sberbank “forms a negative image of Russia.”

Active users of generative neural networks have long been aware of their limitations and the specifics of working with them: in order to make the same Midjourney or Stable Diffusion depict something specific, sometimes it is necessary to set very precise parameters – up to specifying the image style (for example, “real photo”). In addition, many neural networks are trained on data sets that do not include the latest realities. For example, for GPT-4, released in March, September 2021 serves as such a cut-off date.

Deputy testing of computer technologies has a long tradition. Charles Babbage (1791-1871), inventor of the difference and analytical engines (the latter is commonly considered a protocomputer), wrote in his memoirs that members of the British Parliament repeatedly asked him: “If you enter the wrong values ​​into the machine, will it give the right answers?” – “I am not able to correctly understand the confusion of concepts that could cause such a question,” the inventor was amazed.

Sometimes parliamentarians do not limit themselves to questions to the authors of new technologies, but immediately move on to answers. So, during a public hearing in the US Congress in 2018, Congressman Lamar Smith asked Google CEO Sundar Pichai if something prevents company employees from manipulating search results. After that, without even allowing the interlocutor to answer, he said: “My time is up, so I’ll just answer that I don’t agree with you.”

The mere fact that a particular technology is progressive cannot automatically save it from criticism. In the end, the technology of the future is made by people of the present with their own set of priorities and ideas about the world. But perhaps, if you put the right questions in front of the machine, it will work.