We want to democratise AI access, says Google DeepMind’s Manish Gupta at HTLS
Manish Gupta talked about making AI more proficient with Indian languages, as an example of making it relevant for more users
NEW DELHI: Google’s DeepMind AI (artificial intelligence) research labs is working on an universal AI agent called Project Astra with the aim to scale relevance and application to new areas including agriculture and medicine, a top company official said.
The tech giant, which has already built the Gemini chatbot for smartphones and computing devices, as well as cutting edge video generative tool Veo ,and the image generator Imagen, is building the next evolution. Two key elements to this, are speed and wider scope.
Manish Gupta, senior director, research, Google DeepMind, talked about the importance of building and deploying AI for the greater good in a session at the Hindustan Times Leadership Summit 2024 on Saturday, and stressed that the first step towards that was what he called a “really powerful universal agent”. In recent weeks, Google DeepMind has made the AI protein prediction tool AlphaFold 3 open source. Alongside, they’ve developed an Habermas Machine AI tool, which is designed to resolve real-world conflicts. All this, while the Gemini team has now become an integral part of DeepMind too — and the Gemini app for the iPhone released soon after.
In the works is Project Astra, its key element being situational awareness capabilities; it will be available on a smartphone. “Take the instance of a learning scenario where somebody just points an AI assistant at a question in your textbook and asks for help in answering the question or leaning more about the concept. Or let it scan everything in your refrigerator before you go shopping and it’ll tell you if you’re out of milk,” explained Gupta. Quite when Project Astra makes it to the real world, as a standalone tool or as part of the Gemini chatbot, will become clear in the coming months.
“I think of it along two dimensions. One is, how do we make sure that that AI itself is being developed in a manner that is not just working for a privileged few but pretty much for every human on the planet. This includes doing all of the foundational work in terms of languages, making sure the models can scale to billions of users at a small cost, as well as applying them right in areas like agriculture,” Gupta said.
In the AI space, Google’s AI initiatives are competing with companies like OpenAI, Microsoft, Anthropic and Perplexity, across a variety of tools and applications.
Gupta talked about making AI more proficient with Indian languages, as an example of making it relevant for more users, than just those who know English. “People say, just Google it and you have that information. How do we provide that same level of information to the farmers in Assam, for instance, in their native language?” he asked.
Google is tackling the challenge to build AI proficiency in more than 125 Indian languages and dialects. Challenges include making AI proficient those languages, which are considered ‘zero conscious languages’ or ones about which there is no digital data; another challenge are languages that are spoken in different ways.
Gupta talked about the DeepMind initiative to make AI more relevant for agriculture. “A farmer may be unable to get a loan from a bank, and therefore ends up paying very high rates of interest to a loan shop. They are unable to get reasonably priced insurance, and if there is crop failure, they caught in that debt trap. Our team took this approach of analysing satellite imagery, which enabled us to build the first model that for all of India through satellite imagery, is able to now identify all of the field boundaries, the crops which are being grown in,” explained Gupta.
“This provides the foundational layer on which we can build a unique identity to each person, and then on top of which, you build an Aadhaar stack, unified payment interface and more. We can unlock all kinds of use cases for the farmers, like loans, recently, prices, stock insurance or supplement programs for the government,” he added
HT asked Gupta about how he plans to address ethical considerations raised by the rapid evolution of generative AI, given how realistic it already is.
He referenced Google DeepMind’s policy to label all generated images and videos, which would define them as separate from a real image. “There’s a very good indication whether a content is natural or a generated image. We extended that same technology to text, videos and so on. We are making technologies like these available so that when there is misinformation, these tools will help catch that,” he said
There have been many instances of artificial intelligence getting it wrong with information, context and facts. To the question about whether it is an innate human inability to correct technology when it gets things wrong, Gupta said, he believes that it is more because of a lack of education.
“I think part of it is also educating people about these technologies, and it is not just about the potential, but also implementations. In fact, some of my colleagues had done the study earlier, and the title was ‘Because AI is always correct’. That was because some of the people who were using these technologies, believe that AI is an infallible tool. There needs to be that application, first and foremost, that this assumption is not true. In all Google products, we do provide those mechanisms, whether it is a thumbs up or comes down feedback option, and there’s also a capability for you to correct the AI model output,” he said.