The head of the Yandex artificial intelligence laboratory,

[tanjimajuha20]

According to him, a business can take a general model from the market and slightly modify it so that it meets the company's needs. "If a company is not ready to invest serious funds, including equipment, over several years to get a product that will probably be used only once, then it is better to hand over the task to specialized companies," concluded Valentin Malykh.

Alexander Kraynov, explained that large, fundamental models - Chat GPT (from Open AI), LLaMA (from Meta* AI), Mistral AL (from French developers), Giga spain whatsapp resource Chat (from Sber), Yandex GPT - are expensive in terms of creation, they require an investment of $150 thousand and personnel, of which there are almost none on the market, so this is a solution for companies with a large budget and serious calculations.

According to Alexander Kraynov, companies should try various models already on the market, including open source, compare them and further train them as an alternative. In his opinion, a team of even three people - developers already on staff - can handle the further training. It is important to create competence within the company so that there is an opportunity to apply ideas for the main business, as well as to be able to cope with the problem when a third-party boxed solution is not suitable.

"You don't need hardware to further train a large model. This can be done in the cloud. However, you shouldn't spend money on clouds for computing if this is not part of your core business. It's better to turn to external cloud services. The only difficulty you might encounter is passing the information security service, which will prevent the transfer of data outside the company. But in general, using external clouds is convenient and effective," noted Alexander Kraynov.

The economic feasibility of using generative AI may not always be obvious. But as Alexander Kraynov said, it is a question of whether a business will gain or lose a competitive advantage in the future.

"In search, using a generated answer loses out to search results in terms of monetization. In search results, there are, say, 50 links where the necessary information is located, and a person spends time on a search resource to find it - he interacts with the service and sees advertising. And there is a "Neuro" search, which immediately displays a compressed answer, and a person spends a minimum of time on the resource, which is much worse monetized. On the one hand, this leads to a drop in profits, the business only bears costs, but does not earn. But on the other hand, all other companies will add this function, and only those who are more useful to the user will survive," concluded Alexander Kraynov.

* Meta is recognized as an extremist organization in the Russian Federation and is prohibited.

Read also