Yandex is reportedly developing an AI model coordinator system that will ensure intelligent distribution of tasks between neural networks to maximize the quality of AI work at minimal cost Forbes in the company. Experts say such solutions already exist abroad and are popular.
Image source: yandex.ru/company
Yandex engineers will prepare a unified system that will combine the company’s own models and the orchestrator of open neural networks, which will form the basis of Alice AI, said Valery Stromov, head of Yandex’s Alice and smart devices division. The first version of the coordinator already exists and is used in Yandex; the public version should be released in September. “Our strategic tasks now are to move from a paradigm of one model to a paradigm of many models, some of which will be open source, create a system that intelligently distributes tasks in order to achieve the highest quality at the lowest cost, and create a tool (software shell) that provides the model with a “hand” in the physical world – so that “Alice” can use different services, create user or enterprise memories, and run this code.”, Mr. Stromov explained.
When the number of AI models gets larger, the orchestrator starts to struggle and the quality of the work degrades. “In the first version, we will use three to five expert models. The lion’s share of the overall task flow (I think it will always be around 80-90%) will be run on our models, and the highly specialized stuff will be sent to the open source models in our framework. This will actually allow us to surpass players like Google, who currently only have one model, meaning they need to ensure the resources of it everywhere.”he added. Yandex developed this system not out of a desire to follow trends, but out of engineering needs: inference, i.e. launching an already trained model, is not cheap, so “I want to bring the greatest benefits to users and businesses with the least resources.”. Valery Stromov admitted that last year Yandex invested 55 billion rubles in Alice AI, and this number will remain the same this year, but it is a very small amount compared to the world’s giants.
Experts interviewed by Forbes said that the consumer quality of artificial intelligence often does not lie in the billions of parameters of the model, but in its actual application. The accessibility of interfaces and routes helps to bypass limitations and ensure the high-speed operation of the model. With the coordinator, access to the most powerful neural network is not always needed, and models become interchangeable resources. Finally, the problem of storing context is solved – storing it in the model is sometimes too wasteful.
If you find an error, select it with your mouse and press CTRL+ENTER.