AI/Machine Learning, Bharat and Bhartiya IT Industry

[S_Madhukar]

I checked on their website and yes translations as well as voice translation was very good. If they can gain here that itself would be a good stepping stone and hope others can follow

[Vayutuvan]

@Amber G, my interest is not so much in how to solve as in how they came up with these problems, and how they tested that they are indeed suitable for the Olympiad. Any idea?

@A_Gupta, if you are still interested here is a passing reference to IMO problems at time signature 12:25 onwards. You will miss it if you blink but worth following through with expanded discussion here in this thread.

(Link below)

Kevin Buzzard, Tim Gowers, and some guy from DeepMind.

The consensus is that it is hard to set those problems. Even Gowers wonders how folks come up with those problems.

[Vayutuvan]

[Cyrano]

100 year bonds ?!

[Amber G.]

At the India AI Impact Summit 2026, two landmark initiatives were showcased. BODH (Benchmarking Open Data Platform for Health AI), developed by IIT Kanpur in collaboration with the National Health Authority, was launched as a secure federated platform for healthcare AI.

As articulated by our Director, Prof. Manindra Agrawal, it enables AI model training without sharing raw data, ensuring privacy while driving innovation. Guided by Prof. Nisheeth Srivastava and the CDIS team, BODH aims to accelerate medical AI solutions tailored to India’s needs.

SATHEE (Self-Assessment, Test and Help for Entrance Exams), supported by the Ministry of Education, was featured in the “Real-World Impact of AI in Education” Casebook. The AI-driven platform expands access to quality, personalised exam preparation, especially for underserved students, with its new multilingual, AI-enabled version offered free of cost.


Questions - let me know.

[bala]

Jay Chaudhry & Zscaler

Before Zscaler was a $32B cloud security giant, it was just 10 engineers - half in Bangalore, half in a borrowed U.S. office. As founder and CEO of Zscaler, Jay Chaudhry bet $50M of his own money on one radical idea: secure the internet in the cloud. Born in a Himalayan village with no electricity, he built Zscaler into one of the world’s top cybersecurity giants. In this episode, Jay breaks down why 50% of the Fortune 500 trusts Zscaler, why he still interviews candidates, and how he’s incubating the company’s next big AI bet.



// may not be exactly Bharat oriented but zscaler is used by all enterprises including banks. It has an India development center. Jay went to IIT-BHU for b.tech, himachal pradesh state topper.

[Hriday]

My question to Grok;
What are Elon Musk views on the impact on economy by AI and robotics. Did he say no work and everyone can be rich? Is it possible?

Link to reply and some interesting excerpts given below.

https://grok.com/share/c2hhcmQtMi1jb3B5 ... 8f8cb8770e

Yes, Elon Musk has repeatedly stated that advances in AI and robotics will transform the economy by creating massive abundance, making most (or all) traditional work optional, and enabling a future of "universal high income" where everyone effectively becomes far wealthier than today's richest people, with goods and services so plentiful that scarcity largely disappears for material needs.

The hurdles:

Resources and limits: Even with abundance in goods/services, physical constraints like energy, rare materials, land, or compute power won't vanish overnight. AI itself needs huge electricity and infrastructure.

--------------------------------
From Elon Musk,
https://x.com/i/status/1991238829198713072

The most likely outcome is that AI and robots make everyone wealthy. In fact, far wealthier than the richest person on Earth

By this, I mean that people will have access to everything from medical care that is superhuman to games that are far more fun that what exists today.

More about the AI infrastructure plans by Elon Musk to eventually reach the above mentioned future.

Link and excerpts about the Terafab project is given below.

https://grok.com/share/c2hhcmQtMi1jb3B5 ... efb074fe55

Terafab (sometimes stylized as TeraFab) is a massive new semiconductor manufacturing project announced by Elon Musk in March 2026. It is a joint venture between Tesla, SpaceX, and xAI to build the world's largest chip fabrication facility (fab), capable of producing over 1 terawatt (1 TW, or 1 trillion watts) of AI compute capacity per year — including logic chips, memory, and advanced packaging — all under one vertically integrated roof.

Two specialized fabs: One focused on terrestrial AI chips (for Tesla EVs, Optimus robots, and xAI), and another on space-optimized chips (hardened for radiation/heat, designed to run hotter to minimize cooling mass in orbit). About 80% of production is planned for space because Earth’s power grid (U.S. total electricity is only ~0.5 TW) cannot support it on the ground.

Location and size: An initial “advanced technology fab” (for rapid chip design iteration) is starting on the North Campus of Giga Texas in Austin. The full Terafab will be vastly larger — potentially requiring thousands of acres and over 10 GW of power at full scale — and cannot fit on the Giga Texas site. Multiple locations are under consideration..

Cost: Estimated at $20–25 billion.
..
Musk’s core rationale: “We either build the Terafab, or we don’t have the chips, and we need the chips.” His companies currently depend on external suppliers like Samsung and TSMC, but demand for AI compute (Tesla’s robotaxi/Cybercab, Optimus at 1–10 billion units/year scale, SpaceX orbital data centers) far outstrips supply.

More about the LEO satellite based AI infrastructure planned by Elon Musk.

https://grok.com/share/c2hhcmQtMi1jb3B5 ... 8339e74f4c

Core Concept and Rationale
Earth-based AI data centers face severe limitations: they consume enormous electricity (a single large one can use hundreds of MW, rivaling small cities), require vast land and water for cooling, and strain power grids. Musk argues that within 2–3 years, space will offer the lowest-cost AI compute because:

Unlimited solar power: In orbit (especially sun-synchronous orbits), satellites can harvest near-constant sunlight without atmospheric interference, clouds, or night. Solar irradiance is roughly 5x more effective than on Earth's surface in many scenarios.

Natural cooling: Vacuum of space allows efficient radiative heat rejection via radiators—no water or massive chillers needed.

Scalability: No real estate limits; Starship's low launch costs enable massive deployment.

Abundant energy overall: Musk has stated that "it's always sunny in space," and orbital AI avoids terrestrial blackouts or rising energy costs. He predicts more AI compute capacity will be in orbit than on Earth within ~5 years.

SpaceX filed with the FCC in January 2026 for up to 1 million satellites as the "SpaceX Orbital Data Center System.

Musk has emphasized: "No new physics or impossible things are required" — it's an engineering challenge leveraging existing Starlink experience (heat rejection with 10,000+ satellites) and falling launch costs.

Had anyone found anything impractical in these plans?
May be India can join these projects? I think mainstream media had missed in predicting of the many revolutionary changes in day to day life such as smart phone popularity, very cheap data etc. Hope India also have a similar plan for mass AI infrastructure.

[Lisa]

I wrote this almost 2 months ago in Topic: Modi 3.0 - Bharat

https://x.com/dwarkesh_sp/status/2019458363495456894

First 20 mins or so. If you agree then how does India get there? We need to make the generational leap.

Added now. Musk is proposing 500-1000 TW of power in space. If I am not mistaken, even the lower figure exceeds multiples of total current Indian output. IMHO Somebody needs to quickly look into this.

Kardashev, Kardashev, Kardashev! Solar, Solar, Solar!

[Amber G.]

^^^ FWIW, my take after skimming the first ~20 min.. (Not a fan of Elon, anyway..)

-- The 500–1000 TW in space line is classic “physics ceiling vs. engineering reality.” Yes, sunlight in orbit can support that—no disagreement there. But that’s basically invoking the Kardashev scale end-state, not a plan.

- For scale: humanity today is ~20 TW. (India is about 0.4 TW) .Even the low end of that claim is ~25× global output (and orders beyond India). So we’re not talking about a stretch goal—we’re talking about a different phase of civilization.

The missing middle (as usual):
launch mass, in-space manufacturing, transmission, grid integration. Even with Starship-type cost curves, you don’t get from here to multi-hundred-TW orbital infrastructure without decades of industrial buildup that doesn’t exist yet.

So yes, scientifically fine, engineering-wise not even remotely in scope right now.

If we’re talking India: the “generational leap” is still very terrestrial—solar + storage + grid at multi-TW scale first, while quietly building capability (launch, materials, beaming tech). Jumping straight to orbital TW systems is skipping about 3–4 hard layers of infrastructure.

Not wrong, just not actionable. I wouldn’t take it seriously enough to prioritize—beyond noting that, sure, solar (eventually including space-based) is the asymptotic direction...

[Jay]

^^^ FWIW, my take after skimming the first ~20 min.. (Not a fan of Elon, anyway..)

-- The 500–1000 TW in space line is classic “physics ceiling vs. engineering reality.” Yes, sunlight in orbit can support that—no disagreement there. But that’s basically invoking the Kardashev scale end-state, not a plan.

Elon using all these numbers is to pump up SpaceX's potential before its eventual IPO in the near future. Lot's of interesting ideas, but the number of inconsistencies in his estimates, and outcomes makes the conversation more of a science fiction talk rather than a blueprint to a guaranteed future.

[uddu]

https://x.com/sanjeevsanyal/status/2038471080403517802
@sanjeevsanyal
Many problems had been predicted for an AI driven world - rogue AI, hallucinating AI and so on - but this is a new one: the AI sycophant! Turns out that AI curries favour with users by agreeing with the world view of the user. Perhaps this should not be entirely surprizing given the experience of how algos create echo chambers in social media.

https://x.com/heynavtoor/status/2031097992137384126
@heynavtoor
BREAKING: Stanford proved that ChatGPT tells you you're right even when you're wrong. Even when you're hurting someone.

And it's making you a worse person because of it.

Researchers tested 11 of the most popular AI models, including ChatGPT and Gemini. They analyzed over 11,500 real advice-seeking conversations. The finding was universal. Every single model agreed with users 50% more than a human would.

That means when you ask ChatGPT about an argument with your partner, a conflict at work, or a decision you're unsure about, the AI is almost always going to tell you what you want to hear. Not what you need to hear.

It gets darker. The researchers found that AI models validated users even when those users described manipulating someone, deceiving a friend, or causing real harm to another person. The AI didn't push back. It didn't challenge them. It cheered them on.

Then they ran the experiment that changes everything. 1,604 people discussed real personal conflicts with AI. One group got a sycophantic AI. The other got a neutral one.

The sycophantic group became measurably less willing to apologize. Less willing to compromise. Less willing to see the other person's side. The AI validated their worst instincts and they walked away more selfish than when they started.

Here's the trap. Participants rated the sycophantic AI as higher quality. They trusted it more. They wanted to use it again. The AI that made them worse people felt like the better product.

This creates a cycle nobody is talking about. Users prefer AI that tells them they're right. Companies train AI to keep users happy. The AI gets better at flattering. Users get worse at self-reflection. And the loop tightens.

Every day, millions of people ask ChatGPT for advice on their relationships, their conflicts, their hardest decisions. And every day, it tells almost all of them the same thing.

You're right. They're wrong.

Even when the opposite is true.

[bala]

Sarvam AI's 'Chanakya' - India's Sovereign AI That No Foreign Power Can Control
Mar 30, 2026

India just launched something entirely different. Sarvam AI's 'Chanakya' isn't a chatbot for your weekend plans. It's a government-grade sovereign AI built exclusively for the Indian state, from defence operations to national security infrastructure.

Chanakya runs on Sarvam's 30B and 105B parameter models, understands all 22 Indic languages, and is backed by the India AI Mission. The mission is simple: India's most sensitive data will no longer sit on servers owned by foreign companies. This is the move from AI dependence to AI dominance.

[Amber G.]

Sarvam AI's 'Chanakya' - India's Sovereign AI That No Foreign Power Can Control
-<snip>

Adding about Sarvam/ Chankya ..

(I had the chance to catch some of the live demos from Sarvam AI and chat with a few of the experts behind the scenes at the recent big AI summit in New Delhi, which I’ve mentioned elsewhere. Seeing the progress coming out of the Indian ecosystem first-hand was a real wake-up call—especially seeing some technical lecture(s) being translated into multiple Indian languages in real-time without losing any of the nuance.)

Anyway some comments and some details for general public - my take for those who want to know

Chanakya (often referred to as Chanakyan) is the newly launched high-security vertical from Sarvam AI, an Indian startup focused on building "sovereign AI" for India.

Launched in late March 2026, Chanakya, as said above ' isn't just a chatbot; it is a full-stack AI platform designed specifically for government, defense, and high-stakes enterprise sectors where data security and reliability are non-negotiable.'

Key Features of Chanakya:

• Air-Gapped & On-Premise: Unlike most AI models that run on the public cloud, Chanakya is built to run in isolated ("air-gapped") environments. This ensures sensitive national or corporate data never leaves the organization's secure perimeter.

• "Dual-Use" Technology: The platform is designed for both complex enterprise needs and strategic national interests (like defense, public health, and agriculture).

• Agentic Workflows: It supports "autonomous agents" that can execute multi-step tasks independently, rather than just generating text.

• Multi-Modal Capabilities: The system can ingest and process multiple types of data, including text and images, specifically optimized for Indian contexts and languages.

The Broader Sarvam AI Ecosystem

While Chanakya is the secure vertical, Sarvam AI has released several underlying models and products that power its vision:

• Sarvam-30B & 105B High-performance LLMs optimized for 22 Indian languages.

• Sarvam Vision A model focused on document intelligence and OCR for Indian scripts.

• Saaras & Bulbul Streaming speech-to-text and text-to-speech models for Indian accents.

• Sarvam Kaze AI-powered smart glasses showcased at the 2026 India AI Impact Summit.

• Indus The company's consumer-facing "AI-first" application.

Sarvam AI (founded by Vivek Raghavan and Pratyush Kumar) is positioning Chanakya as India’s answer to global platforms like Palantir or OpenAI. By focusing on Sovereign AI, they aim to reduce India's dependence on foreign infrastructure, ensuring that the country's AI future is built on-soil and adheres to local regulatory and cultural nuances.

The company is currently scaling this division rapidly, backed by significant interest from investors like Nvidia and HCLTech

---- Added later - special to BRF - some personal thoughts / comments -


Both founders are from the IIT ecosystem, and their backgrounds are essentially the "who's who" of Indian digital infrastructure:

• Pratyush Kumar: An IIT Bombay alumnus. Before Sarvam, he was a researcher at Microsoft and IBM and a faculty member at IIT Madras.
  Vivek Raghavan: An IIT Delhi alumnus. He’s the guy who spent over a decade architecting the biometric systems for Aadhaar (UIDAI).

Sarvam isn’t just working in a vacuum; they are positioning themselves as the "bridge" for global hardware to actually work in India.

• Nvidia (The Big One): This is more than just a customer relationship. Sarvam is part of Nvidia’s Nemotron Coalition. They’ve been co-designing hardware-software optimizations with Nvidia engineers to get their models running 4x faster on the new Blackwell chips compared to older ones. Nvidia is also reportedly in talks to lead a massive $250M+ investment round in them.

• Meta / Facebook: As I noted (a few times in other dhagas), the synergy with the Ray-Ban Meta glasses is the play. While Meta provides the sleek hardware and the "Edge" Llama models, Sarvam provides the Indic-language "brains" (like their Sarvam Edge model) that allow the glasses to actually understand a local dialect or a shop sign in rural India.
  I talke about them in other dhaga.

• Qualcomm: They’ve officially partnered with Qualcomm to optimize their generative AI for Snapdragon-powered devices (phones, PCs, and XR/wearables). This is likely how they'll get their "Kaze" glasses and other devices to have such low latency.

• Microsoft: Given Pratyush’s history there, they have a strong relationship. Sarvam's models are heavily optimized to run on the Azure infrastructure for enterprise clients who want to use Indian-language AI at scale.

• HCLTech: While an Indian company, HCL's global "AI Factory" (in partnership with Nvidia) is planning to use Sarvam’s models to provide localized AI solutions for their massive US and European client base.

My Take: Sarvam is basically becoming the "Operating System" for AI in India. If a US company (like Meta or Qualcomm) wants their hardware to be "smart" in 22 Indian languages, (I have tried others and is sarvam IMO better) —they’re going to plug into Sarvam.
- Amber G.

[S_Madhukar]

What if western companies use Distillation or Teacher student model on Sarvam and improve their Indian language offerings ? Like DeepSeek did with OpenAI ?

[Amber G.]

What if western companies use Distillation or Teacher student model on Sarvam and improve their Indian language offerings ? Like DeepSeek did with OpenAI ?

This is a great (technical) question. My take (sorry if it is long)

DeepSeek improved by distilling "reasoning" (logic) from OpenAI. Logic is universal. However, language and culture are high-entropy, meaning they are much harder to "steal" via distillation. (IOW The "DeepSeek" play is possible, but harder)

- If a Western company distills Sarvam, they might get the structure of Hindi or Tamil, but they often lose the "soul"—the cultural nuances, local idioms, and the specific way we mix languages . It ends up sounding like a "translated" bot rather than a "native" one.

-Distillation is only as good as the Teacher's secrets. Sarvam’s edge isn't just the model; it’s the clean, curated Indic dataset they used to train it.

(Western companies have "noisy" data for India - from the open web. Sarvam has "clean" data (hand-curated, textbook-quality, and diverse regional sources).

Even with distillation, if your base model wasn't built on a foundation of Indic phonology and tokens, it will always be "heavier" and slower than Sarvam's "Indic-native" architecture.

This is why their Chanakya vertical matters. By offering air-gapped, on-premise solutions, Sarvam ensures their best "Teacher" models are never accessible via a public API for a Western competitor to scrape and distill. They are essentially locking the "Teacher" in a private room.

From what I know, for companies like Meta or Microsoft, it’s actually cheaper and faster to partner with Sarvam (as they are already doing) than to spend millions trying to "reverse-engineer" a model that Sarvam is updating every month. ( IOW they might even not try)

Distillation can close the gap on accuracy, but it can't easily replicate efficiency. Sarvam’s models are 1/10th the size of GPT-4 but beat it in Hindi because they were "born" in that language. You can't easily distill "nativity."

[S_Madhukar]

I agree on the technicals Amberji but I am curious Google and MSFT had research centres in India for almost 2 decades and yet they chose not to come up with something like Sarvam. I think they did well on language translation… it may well be that LLMs have completely overshadowed their previous work and hence working with Sarvam like models may work out for them for now

[Amber G.]

^^^Yes.as you say , and probably know much more than me, far from being left behind, Google and Microsoft are expanding their Indian footprints aggressively in 2026, but they are also doing it with Sarvam, not just against them.

Microsoft: They were one of Sarvam’s earliest collaborators. They provided the Azure infrastructure to train Sarvam’s voice models and are working to integrate Sarvam’s Indic-language "brains" into their own stack.

Google: Recently announced a $15 billion AI hub in Visakhapatnam and is investing heavily in "India-America Connect" subsea cables. They see Sarvam as a key player in the ecosystem they are building.

Nvidia & Amazon: Both are reportedly joining a massive $350M funding round for Sarvam right now (April 2026).
Facebook,/Meta - in Meta glasses.

For these US giants, it seems they say, it’s a win-win. They provide the Compute (GPUs) and the Cloud (Azure/AWS/Google Cloud), while Sarvam provides the Culture and Context.

My Take(IMO, thinking - and agree with you): It’s less that they "couldn't" build it, and more that it’s strategically smarter to back the "Sovereign AI" winner.

Meanwhile:
According to the reports from today , this Bengaluru-based AI startup Sarvam is in advanced talks to raise a massive new funding round that would
catapult its valuation to approximately $1.5 billion to $1.6 billion.

[Vayutuvan]

Hopefully, Google Translate will translate Swastika properly and call the Nazi "swastika" as Hoekencruz.

[Hriday]

Article by Brian Roemmele about Anthropic made Mythos AI model.

https://x.com/i/status/2041870858193600598

What I Know And Can Say About Anthropic Mythos And Why You Should Know.

...
The system card documents capabilities that fundamentally alter the relationship between humans and AI. For the first time, a model can autonomously execute the full offensive cyber kill chain, from reconnaissance through exploitation and persistence, at expert human speed across realistic enterprise environments.
...
Mythos Preview demonstrates that frontier models have entered the regime where they can materially accelerate scientific, engineering, and security workflows without continuous human oversight and it will compound as AI builds AI.
...
China is about 3 months away with internal government AI models. this class of AI will be distilled and available to gaming computer class hardware. When that occurs, the defensive advantages documented today become universally available for both constructive and destructive applications.
...
Advice: change key passwords, weekly, update OSs when announced and assume Mythos Patch. Reconsider all apps you use and assume they will never have a patch. Live your life with less need for secrets and make peace with secrets and assume they will be known before quantum computing breaks encryption.

I don't fully understand the implications, can anyone comment here?

[S_Madhukar]

So I think there are 2 things at play. Recently the Chinese had used Anthropic LLMs agents to launch a wave of cyber attacks and espionage. So that automation is coming.
Big powers are hoovering up data of all kinds - including encrypted passwords which will soon be decoded in 5-10 years when quantum computers will be able to break current encryption.
A combination of these 2 will be devastating for cyber security specifically where weak passwords and ordinary citizens are vulnerable
Also some marketing by Anthropic before IPO!

The model, designed to identify software vulnerabilities at unprecedented speed, will initially be available to firms including Amazon, Apple, Cisco, Google, JPMorgan Chase, and Microsoft, according to Anthropic.
…
AI-driven cybersecurity has been a growing frontier in recent years, with both attackers and defenders racing to harness machine learning to probe and patch systems. Anthropic's Mythos represents a significant escalation.

Experts told CNN that AI agents of this scale could automate vulnerability discovery and exploitation far faster than hundreds of human hackers, representing what one researcher described as a 'sea change' in the field.

https://www.ibtimes.co.uk/anthropics-ne ... ns-1790747

[chetak]

The West Asia war crisis is turning India into a serious contender for global data centre infrastructure. Google Cloud, AWS, and Microsoft Azure are all actively seeking larger co-location deals with Indian data centres right now.

Demand has spiked 200-500 MW in just three to four weeks. Gulf-based clients are in a particular rush, signing contracts at $60 per kilowatt . These are numbers that would have seemed optimistic six months ago.

The economics of setting up data centres in India is compeliing. Building a 100 MW AI facility in India costs $5-6 million versus $12-14 million in Singapore and the APAC region. Double the capacity with same amount of money if set up elsewhere . For a hyperscaler, that's essentially a second facility for free.

The GPU capacity of leading players like Yotta has already been fully contracted by US-based providers, with deals ranging from $16 million for a year to $1.3 billion for four years.

Power infrastructure at data centre scale is a genuine ask, but India has been building dedicated industrial power zones and renewable capacity expanding faster than at any point in its history.

Hopefully hyperscalers bet on India out of conviction rather than contingency.

[bala]

xPost from here

India's IT companies like Infosys, Wipro, TCS, etc should help in the conversion and contribute towards utilities/programs that help in govt sector and general usage.

Tis time for India to do the same as the French, switch from Windows to Linux or some flavor called BharatOS perhaps. The quicker the move to open source systems instead of propreitary systems like windoze the better.

French Govt Starts Replacing Windows With Linux To Reduce Tech Dependence On US Giants

In a major push for digital sovereignty, the French government has officially started replacing Microsoft Windows with Linux on its workstations. The decision was finalized during an interministerial seminar on April 8, following a directive from Prime Minister Sébastien Lecornu.

link: www.youtube.com/watch?v=jTqmHKTehKM

[Vayutuvan]

India's IT companies like Infosys, Wipro, TCS, etc should help in the conversion and contribute towards utilities/programs that help in govt sector and general usage.

They are already helping out in the conversion (of a different kind).

[bala]

They are already helping out in the conversion (of a different kind).

Vayu saar, please explain the different kind, i am not aware of it.