To conserve data centreenergy usage by generative AI, Manish Gupta, Senior Director at Google DeepMind, said the company is working on a Matryoshka doll model wherein smaller AI models are embedded within a larger model for efficiency.

Noting that data centre energy consumption has spiked with generative AI providers signing with nuclear-powered plants, Gupta, at the Mumbai Tech Week event, said, “Our teams have been developing methods to make these large models efficient, one of which is called the Matryoshka models. The name was inspired by Matryoshka dolls, which have dolls nested one inside the other. We built the models in this nested manner where you have the largest model with progressively smaller models embedded inside it.”

With this approach, a complex task can be assigned to the biggest AI model, a moderate query can go to an intermediate-sized model, and a simple query can be carried out by a small model. This allows for tremendous savings in energy and computational cost, said Gupta.

Need to bridge language gap

When discussing the inclusivity of AI, Gupta said that the lack of Indic language data online poses a challenge in building multi-language AI.

“Roughly 10 per cent of the world’s population speaks Hindi and the percentage of worldwide content in Hindi is 0.1 per cent. So many of these Indian languages spoken by millions of people are simply not present on the web,” said Gupta, adding that even the company’s model that covered 29 Indian languages showed significant gaps compared to capabilities in English.

“We are quite determined to ultimately close that gap. My dream is to make our model do better on Hindi than even English but for now, it needs to perform at the same level as English,” said Gupta.