Why AI is thinking small

howindialives.com

howindialives.com is a database and search engine for public data

When the U.S. ordered artificial intelligence (AI) giant Anthropic in June to suspend access to its Mythos and Fable models to foreign nationals, it triggered a debate in India over the need for sovereign AI models (bit.ly/Shaastra-sovereign). Some argued that India should prioritise building small language models rather than depending on frontier systems from a handful of companies abroad.

The debate has been triggered at a moment when the economics of large language models (LLMs) is becoming harder to ignore. For some years now, AI companies, especially in the U.S., have competed to build ever larger models, spending billions on training and infrastructure. Training a frontier model can cost more than $1 billion, and running it is expensive too. Token prices have fallen, but the cost of running AI at scale has not followed.

Many businesses don't need that scale. A chatbot built for customer relationship management or sales does not need a model trained on the breadth of human knowledge. This gap between what frontier models offer and what most tasks require has pushed companies toward small language models, generally defined as models with fewer than 40 billion parameters. They stand in contrast to LLMs, and the difference between the two is now shaping how companies and countries plan their AI investments.