In this talk, the speaker introduces IndicGPT, a large‑scale multilingual foundation model built specifically for India’s linguistic diversity. The presentation begins by diagnosing the problem of Anglocentric bias and the resulting "Babel Barrier" that marginalises non‑English speakers, especially in a country with 22 official languages and over 1,600 mother tongues. To overcome this, the IndicGPT team constructed four core pillars: (1) the Bharatvani Corpus – a 15‑trillion‑token multilingual dataset harvested from digital libraries, news, government archives, social media, and film scripts; (2) the Brahminet tokenizer – a phonetics‑based tokenizer that leverages the shared Brahmi script ancestry to avoid inefficient sub‑word fragmentation; (3) a mixture‑of‑experts architecture that assigns specialist expert groups to language families (Indo‑Aryan, Dravidian, etc.) rather than a monolithic model; (4) a Sanskrit Knowledge Graph that provides cultural and mythological grounding for entities like "Arjuna". The model is further aligned with a Dharmic Alignment Framework rooted in Indic philosophy (Satya, Ahimsa, Asteya, Daya, Anekantavada) to ensure value alignment beyond Western AI safety norms. The speaker reports benchmark results on a newly created Daksh Eval suite, where IndicGPT scores 82 versus 45–58 for baseline and specialized models, demonstrating superior cultural and linguistic understanding. Real‑world implications are highlighted across education, agriculture, healthcare, and language preservation, culminating in the broader vision of "sovereign AI"—nationally owned foundational models that resist algorithmic colonisation. The talk ends by urging the community to replicate this culturally‑grounded approach for other cultures.
Chapters
Hey everyone, today we're diving into something truly fascinating— IndicGPT. This is a huge project all about building an AI that doesn't just, you know, process words; it actually understands culture. We're talking about giving a truly intelligent voice to a billion people. So, let's just jump right in! But let's start with a question for you. Think about it: when you chat with an AI, does it really get you? You know, the little inside jokes, the cultural references, the real context behind what you're saying? Or is it just doing a super clever translation job? This, right here, is the core problem we're gonna unpack. Well, there's actually a term for this: Anglocentric bias. You see, most of the big, powerful AI models that we use every single day, they're built and trained on just mountains of English-language data. And what happens is, you end up with a digital world where the English-speaking point of view becomes the default, and pretty much everyone else can get left behind. And this creates what the researchers behind IndicGPT call the Babel Barrier. Honestly, it's the perfect way to describe this massive language divide that's being built by modern AI. Okay, so here's the bottom line: when the very foundation of AI is English, it just makes everything harder for billions of people. We're talking about access to information, economic opportunities, even just being a part of the digital world. It's this invisible barrier that's quietly making global inequality even worse. And if you want to see just how big of a problem this is, there's no better place to look than India. I mean, we're talking about a nation with 22 official languages and over 1,600 mother tongues. A one-size-fits-all, English-first model? It just completely falls apart when it comes up against that kind of diversity. So, let's break down exactly why these models fail. First up, you've got tokenization inefficiency. That's a fancy way of saying the AI breaks Indian words into tiny, useless pieces. Then there's cultural acontextuality. The AI might know the word Diwali, but it has no clue what it actually means to people. Of course, there's a huge scarcity of good data for most of these languages. And then you get Anglocentric evaluation. This is a big one. It's basically judging these models by English standards. It's like judging a fish on its ability to climb a tree, right? And finally, value misalignment, which tries to force Western ethics onto situations where they just don't fit. So, how in the world do you fix a mess like that? Well, you don't just try to put a band-aid on the old system; you have to build a whole new one, right from scratch. And that is exactly what the creators of IndicGPT did, building their model on 4, frankly, revolutionary pillars, designed specifically for India. And where did they start? With the data. And folks, the sheer amount of it is just mind-blowing— 15 trillion tokens! Let that sink in— this isn't just a big dataset; it's one of the largest multilingual datasets ever put together. It is the absolute bedrock of this entire project. So, this is their first pillar— the Bharatvani Corpus. And to really capture India's true voice, they looked everywhere. We're talking digital libraries, local news sites, government archives, but they also dove into social media and even film scripts to understand how people actually talk. Including all the cool ways they mix languages, like Hinglish. Okay, pillar number 2 tackles this tokenization problem head-on. On the left, you can see the old, clunky way. It chops up a beautiful word like Dharma into a bunch of meaningless letters. But the new Brahminet tokenizer? It's way smarter. It understands that most Indian scripts actually share a common phonetic ancestor, the ancient Brahmi script. So, it uses that shared root to unify them, which keeps the meaning intact and makes the whole AI way, way more efficient. The 3rd pillar is all about the architecture. Now, instead of having one single giant AI brain that's trying to be an expert in everything and ends up being a master of none, IndicGPT uses what's called a mixture-of-experts model. Just think of it like this: it's got a whole team of specialist translators, each an expert in a specific language family, like Indo-Aryan or Dravidian. It's a team of pros, not one totally overworked generalist. And the 4th pillar? This one might be the coolest of all. It's called the Sanskrit Knowledge Graph. You can think of this as the model's cultural brain. It's this huge, interconnected encyclopedia of Indian history, mythology, art, philosophy— all of it. So, when the AI sees a name like Arjuna, it doesn't just see a word; it can instantly look it up in this knowledge graph and understand all the deep, rich context behind it. But look, building a truly powerful AI isn't just about the tech, right? It's also about its values. And this is where IndicGPT takes a really radical turn. They decided to move beyond the standard ideas of AI safety and create a new framework for ethics that's actually rooted in Indic philosophy. They call it the Dharmic Alignment Framework, and it basically acts as the AI's constitution. It's built on 5 core principles: Satya, which is all about truthfulness; Ahimsa, or non-harm; Asteya, which means non-stealing and respecting intellectual property; daya, which is compassion, and then finally this really critical principle called anekantavada. And that last one, anekantavada, is just so important. It's a concept from Jain philosophy that basically requires the model to understand that truth can be really complex and have many sides. It teaches the AI to acknowledge and represent multiple viewpoints, especially on tricky subjects, instead of just spitting out one single dogmatic answer. I mean, this is a profound change in how we even think about building an unbiased AI. Alright, so that's the theory— a massive dataset, smart tokenization, a team of experts, and a cultural brain with a moral compass. But does it actually work? Well, the team put IndicGPT to the test. And the results? They suggest this isn't just another model; it's more like a blueprint for a whole new kind of AI. So, they built a whole new benchmark test called Daksh Eval, and it was designed specifically to test for this kind of deep cultural and linguistic understanding. And the results? They're just stunning. On this test, the baseline models scored around a 45. Even a pretty specialized model like IndicBERT only got to 58. IndicGPT? It scored an 82. That is not a small improvement, people. That is a monumental leap forward. But what does a score of 82 actually mean in the real world? It means you have an AI that can do things that were basically impossible before. It can explain the meaning of a complex proverb in Bengali and then find you the equivalent one in Tamil. It can figure out which Arjuna from mythology you're talking about just from the context of your sentence. It can even, and this is amazing, it can pick up on the subtle sarcasm in a Hinglish tweet. That's true understanding. And this is why all of this matters so much. The real-world impact is just enormous. Think about education. Students can suddenly access all of the world's knowledge in their own mother tongue. Or agriculture. A farmer can get life-saving advice in their specific local dialect. In healthcare, this could be the bridge that closes communication gaps between doctors and patients. And maybe most importantly, it becomes an incredibly powerful tool for preserving and even reviving languages that are in danger of disappearing. All of this leads to a really powerful new concept: sovereign AI. And this is really the end goal here. It's the idea that a nation, or a culture, can and should build its own foundational AI models. Models that reflect its own unique wisdom, its own values, its own language. It's the direct answer to what some are starting to call algorithmic colonization. And that really leaves us with one final big idea to chew on. The work done on IndicGPT is proof. It proves that AI can be built to reflect the deep, rich, complex nuances of a culture. So the question for all of us is, if it's possible for one culture, what's stopping us from building an AI that can truly reflect all of ours?
