
Building something new. Co-Founder @TurboML @puch_ai. Previously @Google, @awscloud, @NUSingapore
United States
Joined February 2019
- Tweets1.4K
- Following718
- Followers20.7K
- Likes2.5K
Siddharth Bhatia retweeted
Research News! Announcing the 1st release of fundamental AI research from the Open-Endedness Team @LilaSciences: a vision-focused blog + 4 preprints. The blog details a broad research agenda for achieving open-endedness while each paper unveils its own novel research direction.
The ideas covered in the papers include:
• a novel approach to training for creativity designed to sidestep a newly-identified problem called the “creative paradox,” which means that directly reinforcing creative outputs can sometimes limit reaching full creative potential;
• an investigation of the advantages of seeing the world through fresh eyes by training an LLM that is intentionally naïve in a certain area (in effect presenting less data as a potential advantage);
• a new kind of training regularization technique for SGD called unified factored organizers (UFOs), which is designed to achieve unified factored representation (UFR) as seen in Picbreeder CPPNs;
• and finally a unique study of catastrophic forgetting that shows that networks with UFR (like those from Picbreeder) are intrinsically resistant to catastrophic forgetting. The inset video on this post is a visualization from that catastrophic forgetting work.
The blog post also includes an interactive “wheel of science” demo.
The implications of these ideas are broad and diverse, but the fun part is that each one opens up its own new research playground - there is a lot of room for further research here!
Moreover, we hope this release also highlights one other important point about AI today: while it may seem that the road ahead is already paved and all that’s left is to drive mindlessly forward, there actually remains a lot of room for new directions and insights in the pursuit of open-endedness and creativity.
This research was possible because @LilaSciences appreciates the critical role of open-ended innovation in the pursuit of scientific superintelligence. The advances you see here are part of this much larger company-wide ambition, which includes AI Science Factories and many other unprecedented components. When this engine is fueled by open-endedness, the result will be much more than a step change.
More advances in open-endedness are coming!
Blog link below 👇
We’re sharing how GLM-5.3 helped build and optimize the inference infrastructure serving GLM-5.3-Flash.
The system went from its first successful run to production readiness in less than two weeks, with end-to-end throughput tripling relative to the initial baseline.
The key was dense feedback: local correctness tests, execution traces, microbenchmarks, and end-to-end measurements that enabled targeted hypothesis testing rather than reliance on aggregate performance metrics alone.
z.ai/blog/glm-built-its-infe…
Thoughts About Scaling Law
Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions.
The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed.
Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter.
Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it.
This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count.
Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.
Announcing Discovery Loop!
I am very excited to announce that, along with my longtime friends and collaborators @Sanjay_Ghemawat, @OriolVinyalsML and @quocleix, we are founding Discovery Loop (@DiscoLoopAI), a Public Benefit Corporation whose mission is to automate machine learning, science, and engineering to accelerate discoveries and progress. The four of us have worked together for 14 to 30 years, and have helped build some of the world’s most used products, infrastructure and AI models, and we’re excited to turn our attention to this ambitious endeavor.
♾
Learn more at: discoveryloop.com
Clarifying a few things, starting with the MoU.
Are you taking taxpayer money?
No. The MoU signed with the UP Government is structured as a public-private partnership. It does not involve any cost to the taxpayers of Uttar Pradesh. On the contrary, it brings investment into the state. The project will be executed in phases, with support from external investment partners. We have not taken any money, any GPUs, or any other form of support from the Government. The only part of this MoU that involves citizens is that they'll receive free access to AI, in their own language and through familiar interfaces, via AI Commons.
Is your revenue ₹42.9 lakh?
No. This is a bug in Google’s AI, which has confused Puch AI with another company called Pucho AI. The ₹42.9 lakh revenue figure belongs to Pucho AI, not Puch AI. Puch AI’s revenue is not public.
What’s your valuation/annual revenue? Are you bootstrapped?
We’ll share those details when the time is right. For now, as a private company, we’re not required to make our valuation or revenue public. But, if you’re curious, we're not bootstrapped. We’re a well-funded startup.
Do you have a foundational model of your own?
No, and we do not believe it is necessary for our mission to build one at this stage.
What is your mission?
We want to bring AI to everyone in India. Today, many people already use tools like ChatGPT and Gemini, but millions still cannot. Your parents, shopkeepers, drivers, and many others are being left out because AI, in its current form, is not built for them.
What exactly is Puch AI doing?
Puch AI is building the domestic consumer AI offering for the masses. We're bringing AI to people through simple, familiar interfaces like WhatsApp and voice calls, so that it is easy to use even for those who cannot type and prefer to interact through voice.
Why don’t you have your own app?
We understand the frustration people feel about this, and at times it is frustrating for us as well (currently fighting Meta on the WhatsApp policy). But so far, this has been a conscious choice. A new app becomes one more thing to learn and one more barrier for the millions of people who are currently not using AI. Try teaching your parents a new app!
So are you asking me to not use ChatGPT or other AIs in the name of nationalism?
No, not at all. People should use whatever AI works best for them. Our effort is simply to expand access to AI by bringing it to millions who, today, are unable to use the existing tools.
Is Puch AI just a wrapper?
Wrappers are generally defined as making API calls to AI providers like OpenAI and Anthropic. So no, Puch AI is not a wrapper. We do not use any APIs. We have built our entire infrastructure in-house on top of open-source models, and engineered it for the Indian context. However, if you define a wrapper as any company that does not have a foundational model pretrained from scratch, then yes, Puch AI is a wrapper.
What is the point of Puch AI if you haven’t even built a model?
The problem is data. Today, ChatGPT and Gemini are the default go-to AI apps for almost everyone in India who uses AI. As more people use them, these companies keep getting more and more data to improve their models. That data will never help Sarvam, BharatGen, or any other domestic AI model builder. More importantly, the bulk of India will end up accessing information through OpenAI or Google: their systems and their narratives that may ultimately be shaped by foreign actors. With Puch AI, we are trying to build sovereign distribution, so that India has at least one alternative, no matter how hopeless it may seem to compete with OpenAI or Google’s marketing.
Puch AI seems relatively simple. I can build it in a weekend. Are you even doing anything?
The choice of a simple interface is deliberate, but the underlying infrastructure is far from simple. Happy to give ₹50 lakh to anyone who can build it in a weekend, match our performance in Indian languages, and offer it for free at our consumer scale.
How was your experience building for India?
It is a very interesting situation, where the Government understands the importance and wants to do good, but it becomes difficult to do any good when so many bad actors are rooting for you to fail. Thousands of paid accounts are ready to try to shape narratives to fit their own political agenda. Everything from Community Notes to actual news articles and journalism gets weaponized against you, because your failure gets more views.
Will you keep fighting?
Not sure. Quite exhausted. Maybe this is the time to sign off.
Talking to @howardlutnick will teach you more about business than any business school ever could.
The energy at @wef in Davos is electric. Sovereignty and AI are all anyone’s talking about.
Great spending time with the Birla family: @NeerjaBirla, @ananya_birla and @AdvaiteshaFc. So much warmth and grace.
As a BITSian, it felt extra special 🤗
Siddharth Bhatia retweeted
Watching Meesho quietly win in Bharat 🇮🇳 while Amazon and Flipkart throw everything they have at India, is a good reminder.
Distribution & local beats generic brute force. That got me thinking about AI.
Tried a few local GPTs including Puch AI.
Very colloquial. Very Indian. Use cases most global models don’t think about.
India doesn’t just consume platforms. It reshapes them 🇮🇳🤖
Curious to see who gets this right.
Siddharth Bhatia retweeted
This is a pretty bold gamble, releasing a front-page ad with zero context and only the brand name in bold. Not even a QR code or a website URL! It forced me to Google and find out what the company does and it was an interesting surprise :)
I also figured that it is not 'Puch' as in 'Hero Puch', but the Hindi word for ask - Pooch.
#advertising #marketing #creativity
Siddharth Bhatia retweeted
Watch out for @puch_ai. They've cracked Tier 2/3 distribution through enabling WhatsApp voice calls at scale. @siddharthb_
Caught up with Prof. Ajay Kumar Sood, @PrinSciAdvGoI. Super energizing to see the planning and momentum around AI.
Couldn’t agree more with what @amitabhk87 said, India must own its AI distribution.
Congratulations @adityagrover_ and the team at Inception. Looking forward to more breakthroughs from you!
Excited to share that we’re leading Inception’s $50M seed round.
Inception is taking a fundamentally different approach to AI — applying diffusion models (the technology behind image generation systems like Stable Diffusion) to code and text generation.
While most AI companies are iterating on autoregressive models that generate text token by token, Inception founders Stefano Ermon (Stanford professor and diffusion model pioneer), Aditya Grover and Volodymyr Kuleshov are building models that refine outputs through iterative refinement. This architectural difference unlocks real advantages: Their Mercury model achieves over 1,000 tokens per second — significantly faster than traditional approaches. In a world where compute costs and latency are critical bottlenecks, this parallel processing capability is a game-changer.
As codebases grow more complex and AI infrastructure costs become clearer, diffusion-based approaches offer meaningful advantages in processing large volumes of code and managing data constraints.
Thrilled to back Stefano and the Inception team alongside Mayfield, Innovation Endeavors, M12, Snowflake Ventures, Databricks Ventures, NVentures, and angels @AndrewYNg and @karpathy
1
23
9,376
Indian startups used to copy global companies. Refreshing to see the trend flip.
xitter.gq/siddharthb_/status/193…
Too bad @elonmusk couldn’t make it to @allinsummit this year, would’ve loved his take on the billboard outside 👀
Why did we buy India’s most expensive number?
Because India deserves its own AI.
Sovereign AI is a bigger priority than even nuclear energy in ways we can’t yet imagine. China is already releasing powerful open-source models, but every single one of them calls Arunachal Pradesh “South Tibet.” That’s how subtle propaganda seeps in. Relying on the US or any other foreign country is no safer.
That’s why we’re building Puch AI: India’s first AI, the true gateway to AI for a billion+ people. Over the last decade, India has witnessed three defining shifts: Jio made the internet accessible, UPI revolutionized payments, and Quick Commerce redefined shopping. The next leap is mass AI adoption. Why begin with WhatsApp? Because for most Indians, apps are friction. After WhatsApp, we launched AI through voice calls. Will it always be free? Yes, even though many warned us that if it’s free, people would dismiss it as “second-class.” For us, accessibility and empowerment come first.
A few days ago, I was in an Uber and asked the driver to try Puch. He thought for a while and then asked: “What’s my wife doing at home right now?” That’s India, curious about AI, but still new and unsure where to begin.
Today, nearly a million people use Puch every month, but here’s the part I love most: more than 60% are experiencing AI for the very first time. In small towns, people use Puch in their own languages, often through voice notes, because that’s how they naturally communicate. For first-time users, AI has to feel exciting and magical, not lifeless. That’s why we introduced stickers just like ChatGPT did with Ghibli, brought old photos to life through video, and now done something similar for voice calling: just pick up the phone, call Puch on WhatsApp or SIM, and find out.
The obvious crab mentality question: Can Puch AI compete with OpenAI, Meta, and Google? Despite minimal resources, and even with WhatsApp’s restrictions, Puch already offers better video generation than Grok (sorry, Elon) and stronger multilingual support than Meta AI, all of it fully self-hosted without using any external APIs. If we wanted, we’d love to beat them at their game. But the bigger game is having them come to us for distribution. That’s the only game I’m playing, and the one I intend to win. There’s plenty of news about OpenAI, Google, and Meta trying to expand into India. But if we can reject an acquisition worth millions, if one of our own employees can turn down a personal offer from Sam Altman, the only way they beat us here is by bribing government officials. Nothing else would be enough.
Real behavioral change, and yes, unlimited iPhone giveaways, requires capital. If you only move when others clap, watch this and cheer from the sidelines: puch.ai/ashish. But if you want to win the fight for the world's biggest AI market, let’s talk.
PS: If you tag an investor whose money we accept, you win an iPhone.



















