It was fun presenting at NODES 2025 ! Thanks @neo4j @JulesHitz

Why is everyone asking the government about the AI models ? Don’t we have flourishing Investors and already billionaires in India who can actually spend and build the Frontier Lab ? Not one SARVAM - but in fact more. Why is the MD and GPs of top investment firm tagging the Indian government ? Open AI / Anthropic didn’t have US government backing when they were founded. What am I missing ? GPUs - Billionaires have money - they will buy - that’s where the investment is. Infrastructure - it is already evidently available. For building the models - why putting it on the government ? Curious to understand this.

I said this then. I shall say this now. Your intelligence is your commodity. Your data is your sovereignty. You need to chat with one of the authors of PyTorch and understand why local compute is the key. Perhaps that will instill the confidence of the point. Meet / Talk with @sasank51 You would need to open your intelligence.

Expect a hit on cloudflare today. Everyone will be on the experimental path of self hosted models. @CloudflareDev @Cloudflare -you might be seeing a high traffic in next 48 hours. Might.

Everybody challenging Nandan Nilekani - before you challenge the dynamics of Sovereign AI - if you are a startup using agents - and you are using any of the closed source models towards your revenue generation, perhaps first see to make your agents / business work on opensource model , hosted compute. You never know when the ground slips and your business would not have access of let’s say the proprietary closed source model access. Have you thought that through ? If you aren’t in local compute, if you aren’t building on private and opensource models , self hosted / cloud hosted non closed sourced model / sovereign models ( like Sarvam AI ) completely - you ought to take a back seat and fix your structure first. Say In worst times , say there is a war - typically , US has all the right to stop complete access on all models in defence of its national security. And there goes your agents poof !

Unless you are building upon local private models / local compute and private ecosystem - every single code that you are writing using openAI / Anthropic / Gemini is as good as being controlled. For many startup founders being emotional and sudden epiphany of sovereign AI and call for it - your startup is thriving on the very model that is at disposal by Anthropic/ OAI / Google. Take that away what are you ? Definitely not much. So, before you call for sovereign AI , build on local AI and open-source models. Talk is easy when you use intelligence as a parity. Bring the A-Game on your builds and then shout for sovereign AI. I asked this a while back - Has anyone started using SARVAM AI for production grade agentic tasks ? If there is a strong call for sovereign AI , why not include the available models and prove that you can make things work even on local compute and open source models or Sovereign Models ? Talk is Cheap, Show me the code - Linus Torvalds.

The cacophony of pseudo-intellectualism and talking about sovereign AI and how one country should be building in these waking times - An absolute disarray among those folks. Control is an aspect that is needed but also a regulated approach. And building something as Anthropic is not just CAPITAL , but a TALENT issue as well.

As Ramin said it right, VC bros being VC bros. They will invest on building rockets and invest on space datacenter but would turn a blind eye on local compute / edge AI / on device startups. They forget the roots or the basic.

Trillionaire. Damn. Imagine.

Here for the quoted posts. Just do a favour. Read the quotes on this post. Damn.

Mad respect for @zatlyn @eastdakota and Lee Holloway.

Watch “The Witness” on @netflix . Will ya ! Please !

Energy will be the next routing element for model efficiency. From “Is your model efficient for the task” to “is your task intelligent enough to burn that joule” Excerpt from my presentation at NIST Code Gen AI discussion.

I knew rPPG via camera was coming 😊. x.com/officialkrishd/status/…

“Don’t be a Homelander.” This is how your agents feel and act when your tokens are exhausted. So, go local. Build upon the compute. Make the most out of local models. In courtesy of @TheBoysTV. Issued in the interest of private compute and local models @huggingface

6 GB VRAM - 16 Cores - 16 GB RAM one machine that routes models and also gives GPU precendence and CPU precedence - MedGemma1.54B Q8, Gemma4E4B, Gemma4E2B, Presidio Analyzer and Anonymizer, Gemma3 270M-it_q8 finetuned, and Qwen3.5_0.8B-Q8. All these models takes precendence on tasks. The Video Processing is handled by Gemma4E4B on a separate 16GB VRAM machine when needed. 0nly for video processing. Everything else runs via the 6GB VRAM machine. TLDR : It is upto you how you juice out your compute. It is just there, Max it out, on inference.

The scavengers of LinkedIn are at it again. If you are in cybersecurity domain, you might have heard about the potential and disastrous leakage of answer sheets by one of the most critical examination bodies of India. [ CBSE ] LinkedIn Scavengers - Talking about the leak without giving credit to the actual folks who found it. That’s why we have limited genuine raw talent. Since many are just trying to copy / paste known big facts without giving the due credit. Credit Hogging is a skill that should be added on LinkedIn.

If your agent can do 1/2 of the jobs that it is doing now, but without internet and low resource compute - you built yourself a moat. Else it will be still a glorified API call only this time instead of a wrapper - they call your agent a tad an empty suit. TLDR : works on local models with low compute ? You built it right.

What @FactoryAI built is truly amazing and is equally a true embodiment of people believing in them. Flashback : year 2025 - prior to presenting at Linux Foundation in August 2025- I showed my work to a “certain well known VC” from India around June 2025. The feedback was - “who will believe this or invest in this ? You are doing a side project. model routing through intelligence and self hosted compute is a big no. It will not make any difference as enterprise grade projects works on closed source models”. I was completely disheartened and honestly felt demotivated. I felt it never would make sense. I doubted my intention with this project scope and idea. Nevertheless , I presented my work at @linuxfoundation AI_Dev Summit 2025. My lesson learnt : Arrogance and Ignorance has a face too. Today : I now want to reopen the project and create the intended scale of impact it can have. Give this for free. OSS. As always intended. The 2025 demo and framework not just focused on model route intelligence , but also looked into compute intelligence. Local models do require compute and efficient computing is not a choice but a necessity when you use a private local model router.