Google DeepMind India Chiefs on Making AI Cheaper, Safer, and Better at Code

In an interview, Manish Gupta (head of research) and Seshu Ajjarapu (head of applied AI) of Google DeepMind India discuss the company's top priority: closing the coding gap in AI. They emphasize that improving coding ability enhances overall model performance. The Bengaluru team developed a Matryoshka-inspired transformer technique, first used on Pixel phones, which nests smaller models inside larger ones to save compute and battery. This efficiency focus, rooted in India's price-sensitive market, is being extended to server-side workloads. On enterprise AI, they discuss shifting from token-based pricing to task-based pricing and the importance of distinguishing between public and private data for IP protection. The article also touches on finding the optimal amount of 'thinking' for models to balance performance and cost.