India's DeepSeek Moment & Your AI Clone Is Already LIVE!

At the start of 2025, DeepSeek reset everyone's assumption about who gets to build Frontier AI, a Chinese lab working under chip restrictions trained a model that went toe-to-toe with the best in the world and open-sourced it. It wasn't just a technical story, it was a signal the US doesn't have a monopoly on this anymore. Now over a year later, India is making its own version of that argument. This week, New Delhi hosted the India AI impact Summit, a government-backed event that drew delegations from over 100 countries, more than 20 heads of state and close to 300,000 attendees. French President Mrron was there. The UN Secretary General was there. Sam Alman and Dario, CEOs of OpenAI Anthropic, were there as well, awkwardly trying to hold hands on stage at Modi's request or not holding hands depending on which camera angle you watch. The point is the world showed up for India's AI moment and India had something to show them. Three models all built in India announced on Indian soil. First Serb 30 billion a 30 billion parameter large language model built from scratch by Bangalore's based startup Serb AI. It uses a mixture of experts architecture the same efficiency first design philosophy that made DeepSeek turn heads in January. The 30 billion model activates only 1 billion parameters at a time during inference which means it's powerful but lean. It was pre-trained on 16 trillion tokens of text optimized for Indian languages and this is the part that matters for adoption efficient enough to run on low-end devices feature phones not just data centers second 105 billion the flagship 105 billion parameter model again mixture of experts activating around 9 billion per token and this one is built for serious reasoning math code research analysis long document processing it supports 128,000 token context window. And at the summit, Serum CEO claimed it outperforms Gemini 2. 5 Flash on Indian languages and beats GPT 120 billion on MMLU Pro, which is one of the harder general intelligence evaluations in use right now. Both models are trained to work across 22 Indian languages, not just Hindi, but regional languages most global models have largely ignored. Third, Bajen Pam 2. This one comes from the government side. Bharat genen is a stateback initiative and pram 2 is their 17 billion parameter model developed in collaboration with Nvidia. It's multimodal supports 22 Indian languages and is positioned specifically for public sector use governance healthcare agriculture education. It's less a commercial product and more India saying we want sovereign infrastructure for how the government runs. When Deep Seek dropped in 2025, the story wasn't just China made a good model. The story was a team that wasn't supposed to compete trained a model that could at a fraction of the cost and open-sourced it. It reframed what building frontier AI actually requires. India's announcement carries a different but parallel energy. Sorum AI was founded in 2023 by two researchers, Vim Ragavan and Pratushkumar. They raised 41 million in series A. They were selected by India's government under the India AI mission, a 1. 2 billion program and given access to 4,000 H100 GPUs to actually build this. The models were trained from scratch on Indian soil on Indian compute with Indian language data as first class priority, not fine-tuned from something else. The Deep Sea Peril isn't about parameter counts or benchmark comparison. It's about intent. China proved you don't need to be open AI to build something that matters. India is now making the same argument but with a different angle. Not just we can be efficient but we can build for 1. 4 billion people that global AI has largely been trained to ignore. Think about what it means to build a voice first AI that works in TAML or Odia or Bojri for a population where a huge percentage of people interact with technology through speech, not keyboards. That's not a niche edge case. That's the actual use case for most of India. And serum isn't just doing software. At the summit, they showed off case AI powered smart glasses. Modi even tried them on. It's early hardware, but the signal is clear. This is a company thinking about the full stack. To be fair, the 30 billion and 105 models are still listed as coming soon on Serb's website. The benchmarks were shared by Servom, not independently verified yet, and outperforms Gemini 2. 5 Flash on Indian languages is a meaningful but narrow claim. These models haven't been stress tested at scale across the broad ML community yet. The gap between announcement and deployment is still real. India has made big AI announcements before. The question is always does this end up in production at scale or does this stay at a summit talking point? What's different this time is there's actually a funding structure, a government mission with teeth, compute infrastructure tied to it, and a company in serv that has been quietly shipping real products, speech models, translation APIs, enterprise tools for 2 years already. This isn't starting from zero. The race to build sovereign AI is no longer a two-player game between the US and China. India just put three models on the table in front of the world at their own summit on their own terms. Whether the 105 billion holds up under real world scrutiny, whether the government followthrough materializes, whether any of this reaches the farmer in Bihar or the student in Chennai the way the pitch promises, that part is still ahead. But the starting gun has been fired. And this time, India pulled the trigger

themselves. You probably know Pika as the AI video tool. Text in, video out. One of the better ones in a crowded field. But they just announced something completely different and depending on how you feel about the direction AI is heading is either fascinating or quietly unsettling. PA launched something called AI selves. The pitch straight from their site create a living AI version of you or whoever you want it to be. Your AI self talks, posts, remembers, and grows so you can quote live your best life without all those silly human limits. Here's how it works. You upload a selfie. You record your voice. You answer some personality question, shy or bold, how you write, how you make decisions. From that, Pika generates a digital version of you that can be deployed across platforms. It handles your Slack. It posts your social content. It answers emails at 11 p. m. It sends voice memos to people while you're busy elsewhere. They call it a living AI, not a chatbot, not an assistant. A version of you that acts on your behalf, learns over time, and adapts to different contexts. professional on LinkedIn, casual on Instagram, while supposedly staying true to how you actually come across. This is a big swing for Pika. Their original product, the video generator at pika. art, has been one of the more popular tools in that space. They launched in 2023 and went viral on Discord, raised serious money, and kept shipping. Pika 2. 5 is genuinely good. Physicsaw aware video generation, solid consistency, fast turnaround. Creators use it constantly for social content, but AI video is an arms race they probably can't win long term given Seed Dance 2, Sora, Runaway, and a handful of wellunded competitors who are all fighting for that same ground. So, Pika is going somewhere different. Instead of making videos, they're now making you or a version of you that can operate in a world while you step back. It's a completely different product category, digital identity rather than content creation. And the market they're going after is enormous. Anyone who feels like there's never enough time and would offload parts of their professional social presence to something that sounds and acts like them. This is where it gets interesting to think about. Pika is careful to say on their site that the testimonials are quote absolutely fake, but the use cases are real, which is an odd thing to lead with, but it signals they know that this product is going to invite scrutiny, and it should. A few things worth sitting with. When your AI self sends a message, does the person receiving it know? If your AI is managing investor emails or team syncs, are those relationships built on real interaction or simulated ones? There's an authenticity question here that's hard to wave away. There's also the data angle. Training your AI self means feeding it your text, your voice, your personality. That's a different kind of data exposure than using a productivity tool. You're not just giving a company your usage data, you're giving them you. And then there's the deeper thing. What does it mean when the friction of human interaction becomes something you can outsource? The pitch is efficiency. The risk is that you start hollowing out an actual texture of relationships in exchange for scale. None of that means that this product fails. In fact, it might do extremely well. A lot of people will look at this and see exactly what they've been waiting for. A way to be in more places, do more things, stop missing their daughter's soccer game because they're stuck on a late call. Pika is betting that identity can be a platform that people will want a persistent deployable version of themselves running in the background that the concept of showing up will shift. Whether that's liberation or something we'll look back on differently, that's the question this product is quietly asking. Make sure to

subscribe to our channel. We do real tests, not just headlines. Make sure you're also subscribed to the world of AI. And don't forget to check out our newsletter for deeper breakdowns you won't see on YouTube. and I'm growing my Twitter following, so make sure you follow me on Twitter as well.