Nandan Nilekani says small language models trained on specific data are ‘quite effective’