‘AI is a tremendous tool to augment you’: Manish Gupta of Google DeepMind
Manish Gupta, Senior Director at Google DeepMind, believes AI can supercharge human potential, especially in India.
“There is a big difference between what is potential availability of content (in local languages) and what is digitally available… because there are so many languages spoken by millions of people where there is no digital content,” Gupta explained, adding that Google is trying to unlock these sources and also encouraging the government to open up its archives to train the models.
“A lot of information sitting in Tamil is not available in English. So you have to translate. Now, instead of having to pre-translate all of this, the benchmark suite handles cross-lingual question answering, where you answer a question, let’s say in Hindi, and the information is sitting in English, and then it comes back to you in the right language,” Gupta added.
He explains that his team at DeepMind is also working on plugging that language gap where they know there is a difference in the answer if the same question is asked in English versus a local language. “If you learned math in Tamil medium, you can answer even if a problem is asked in a different language. Humans do that very easily. But AI models often struggle because it’s stuck in their capabilities across languages,” he said, adding that his team has made “significant progress in making sure Gemini really does well on a whole bunch of languages”.
Google has also announced an updated Gemini 2.5 version as well as a text-to-video Veo 2 product, Deep Research which scans the entire web to answer questions, and Astra which comes with the ability to engage with the world through visuals like humans. Gupta reminded us that Gemini has set a “new bar on all leaderboards like ChatBot Arena” with its new versions. “It’s like a pretty significant lead, in particular this reasoning ability. So when you want to plan things and so on, reasoning is going to be the new frontier. And that’s an area where we have really made serious inroads into.”
Asked if all the new content generation capabilities will create a vicious cycle of synthetic content infiltrating the training data, Gupta highlighted how Google DeepMind’s SynthID is now partnering with Nvidia to embed a signature in all AI-generated content from text to video. “We now mark all content that we create with our models so that at least a model can very quickly identify whether the content is human- or AI-generated. And then, when you are training future versions, you can appropriately use it or not use it.”
Khosla, meanwhile, explained how new use cases were emerging out of India, like that of a retired employee who scans all documents to help him file his income tax returns or another person who used Gemini to learn how to master the functions of his new DSLR. “Gemini is making a real-time meaningful impact in the daily lives of people to do fundamental daily tasks better, get more productive, get more creative, and get more confidence in life. And I think that’s what excites us.”
