Hello Betamax, I don't fully understand Kannada, the language spoken in Karnataka, the Indian state where I live. But I rarely have trouble getting by, as its speakers often incorporate English into conversations, making my interactions with them easy to follow. This kind of code-switching is common across the country. People routinely use two or even three languages in one sentence. One good example is Hinglish, a blend of Hindi and English. For large language models (LLMs) trained primarily on English, parsing user inputs in a variety of languages is a surprisingly difficult problem to solve. That's where Sarvam AI, whose name is from the Sanskrit word for "everything" or "whole," comes in. The Bengaluru-based startup is building models designed to match India's linguistic complexity, with support for all of the country's 22 official languages alongside English. Backed by investors and the government, Sarvam is part of a broader global shift as countries look to build their own AI systems rather than rely solely on Big Tech. For today's first top story, I spoke with AI startups to understand whether they would choose Sarvam over global models. While founders agree that there's a clear demand for localized AI, Sarvam's success will ultimately depend on whether it can drive real-world adoption. As for my Kannada, my son has now stepped into the role of my teacher. While models like Sarvam bet on advances in language, some believe the focus on LLMs is overhyped. In our second featured piece, AMI Labs co-founder Saining Xie argues that the real breakthrough will come from "world models" that learn from real-world sensory data, not text. Samreen Ahmad, Journalist | 
