India needs to up the ante on AI research: Google’s Manish Gupta

While I have a lot of admiration and respect for work being done globally, I wouldn’t think of languages as not high tech. Languages is a very important area, because we want AI to work for everyone, not just the privileged few. There’s barely 10 to 11 percent of people in India who are comfortable with English. We want to enable, for example a farmer’s daughter the same level of access to information as you. I see this as a huge deficiency of these large language models today. So that is actually a cutting edge problem. There are a lot of very interesting challenges that come up in the context of languages. By the way we are also doing a lot of work and in public health.  

 That’s a great point. In fact the reverse is happening. For instance, I asked our team to focus on making these AI models work on low-end, and mid-range mobile phones because these are the only digital device used by hundreds of millions of people in India. And interestingly, the work that we did on making these AI models is now being used at the high end.  

 I wouldn’t say so. Large language models have already established capabilities to solve a very broad range of tasks. The problem of energy consumption and high cost of serving is a real problem. So we build these models in a nested manner. This approach is called the Matryoshka Representation Learning, inspired by these Matryoshka dolls which are nested one inside the other. Now we are saying don’t just use a small language model, use the appropriate size model depending on the complexity of the task at hand. That way you get the best of both words where you get the higher quality of a larger model as well as the lower cost of a smaller model. Most of the time, your queries are going to be solved with fairly small size models.

 There is no doubt that India needs to up its game. When you think about the scientific output on AI research, India still needs to up its game to get into the same league as some of the leaders like US and China. However, I do see very, very positive signs. I see tremendous activity because we have so many problems where AI Technologies can make a huge difference. I see tremendous creativity from so many start-ups that we interact with, whom we try to support through access to our technology. I see a lot of very interesting work from Indian companies. The level of AI research and the quality continues to grow. While the gap (with other leading countries) is very real and we need to up our game as a country, but at the same time I see tremendous promise and progress.

 It is already happening. Some of the work that we have done on understanding Indian languages has been used to make these models understand many global languages. Google Translate added 110 new languages and this was based on the work our team had done here. The work on the Matryoshka Representation Learning approach is being deployed globally. And even our solutions for the agriculture sector, which we are initially launching in India, we are already seeing a lot of requests from outside India for some of those capabilities.

 You have a very good open question but at least part of our AI principles is to pledge that we are not going to let our technology be used to cause harm, for building weapons etc. That is part of our AI principles.