I ran across this video a few days ago and couldn’t stop watching it.
It’s about something ordinary & boring, a plastic gas lighter. But it changes how one thinks about manufacturing.
That lighter in so many of our homes, holds pressurised gas. It has over 30 microscopic parts, has to pass international safety codes, & travel 10,000 miles by sea, & the total cost of doing all that, materials, labour, freight, every middleman along the way, comes to fifteen U.S cents.
So how does anyone make money on this?
Turns out almost the entire world’s supply comes from one place: a county called Shaodong, in China’s Hunan province.
It wasn’t always there.
But today, Shaodong has 114 lighter-related companies packed into the place & between them they source more than 200 different components from each other, all within a 20-kilometre radius. They supply something like seventy percent of the world’s disposable lighters. And the industry alone employs over 80,000 people locally.
Nobody there is winning on cheap labour anymore. They’re winning by shaving a thousandth of a cent off the thickness of a plastic wall, or redesigning a base so a few thousand more units fit into the same shipping container.
It took my thoughts back to an old professor of mine, Michael Porter.
His 1980 book, Competitive Strategy, is still the 1st book most MBAs read, the one that gave the world the Five Forces and basically invented modern strategic thinking.
But there’s a quieter piece of his work, on industrial clusters, that never got nearly the same attention, and it is the one that explains exactly what is happening in Shaodong.
His argument was that nations and regions rarely win because of cheap inputs. They win when rival firms and specialist suppliers crowd into the same small geography for long enough that they keep pushing each other past what any one of them could manage alone. He found it in the Swiss watchmaking towns of the Jura, in the German printing press industry and in Italy’s ceramic tile and footwear districts (interestingly, it’s the SAME blueprint which built Morbi, in Gujarat, into the world’s second-largest ceramic cluster, now outproducing Italy by volume. I have posted before, about Morbi)
None of these started out as giants. The neighbourhood made them giants.
Which is exactly why it’s so relevant to India’s climb up the global manufacturing table
I’ve also attached a slide with this post that I saw recently and which shows us breaking into the top 5 manufacturing globally. (A quick reference check told me that we may not have overtaken Korea yet, but the trajectory’s clear)
That climb has happened on the back of scale: bigger plants, bigger parks, more FDI.
I should declare an interest here, because the Mahindra Group set up 2 of India’s first integrated, plug-and-play business cities, in Chennai in 2002 & Jaipur in 2006.
Both have been extremely successful. Chennai’s business zone alone today employs 45,000 people..
But I admit that we need to think differently.
A park brings in investors and hands them a ready plot, power, water & roads
A cluster is a completely different animal: hundreds of small, specialised suppliers, each obsessed with doing a tiny thing better than anyone else, feeding off each other’s presence for years until no outsider can compete with the whole.
I think that’s the work ahead of us now.
Not just more factories, and not just more parks.
Policymakers & developers like us need to start consciously pulling as many of the inputs and resources a sector needs, the toolmakers, the component suppliers, the testing labs, the logistics specialists, into the same neighbourhood.
Shaodong and Morbi both got there by accident, one town stumbling onto a way to shave a thousandth of a cent off a lighter wall, the other discovering it had the clay and, later, the gas pipeline for tiles.
We don’t have the luxury of waiting for accidents anymore.
We need to do it on purpose
Big day for all of us at Sarvam.
I want to start by thanking my team for shouldering this mission with immense belief, urgency, and care.
Reflecting on the last few years of the founding journey, my conviction has only deepened:
- AI will be far more consequential than most of us realize even today
- The value loops of this new world cannot be owned by a couple of companies
- Country of India scale cannot rent intelligence. We have to build it ourselves
We are going to push hard across every layer of the company, but the thing that excites me most right now is our shot at building frontier-class AI systems from India. We are assembling the team, the compute, and the deployment engine to make this happen.
I also want to thank our new investors. HCLTech’s partnership opens joint opportunities to bring our research and platform to many of HCLTech’s clients - this is also a unique template to bring together India’s strengths. BVP brings to the team the rare combination of being at the forefront of India's biggest tech shifts for the past two decades while globally having partnered with category defining enterprise AI companies.
Onwards
Ok, so here is my take on the Fable ban, sovereign AI, Sarvam, etc.
The event is interesting as it has implications from many perspectives.
For AI users, it is clear that you should not confuse access with ownership, or adoption itself as advantage. And if the most significant tech differentiator you are leveraging has external control loops, then you have to accept you are vulnerable.
For AI talent, it is now a precedent that you would be *seen* aligning to national interests more than company interests. And even if its just a whim for now, this trend will be hard to reverse as the world gets more automated…
For AI labs, their offerings will be stratified - general purpose AI would be available as utility, but frontier AI would be gated. This is a fantastic business model for labs - *democratized* AI sucks in all the data liquidity of the world which is locked in higher margin frontier offerings.
I think for the world to be a better place, all three of the above are bad vectors. We need to have more countries and companies owning their own destinies. And in the post AI world, that means being able to use and improve AI systems within their own perimeters - what one may call Sovereign AI.
At Sarvam, Sovereign AI in India was the founding thesis a couple of years back, and continues to remain the core operating principle. From our vantage point, it is super clear that India will build, leverage, and create massive business value and societal impact with sovereign AI. The following is precisely how we at Sarvam are contributing to make that happen.
RoPE is one of those things that everyone uses in LLMs but most explanations jump straight into equations.
I wanted to understand it from first principles.
So I wrote a blog explaining RoPE from scratch.
🎉 We just shipped a major redesign of https://t.co/WqEUat9rdc.
"How do I run model X on hardware Y for task Z?" now has a clickable answer.
What's new:
- URLs mirror HuggingFace: just swap https://t.co/7AFxaoRYOi → https://t.co/WqEUat9rdc in any model URL to jump straight to its recipe (e.g. https://t.co/O5r5hHjQOR)
- Interactive command builder: pick hardware, variant, strategy (tensor, tensor+expert, or data+expert; single or multi-node; or a prefill/decode disaggregated cluster), toggle features → get the exact `vllm serve` command
- Pluggable hardware: NVIDIA + AMD already integrated. One-click switch between Hopper/Blackwell and MI300X/MI355X, and the right flags and env are applied automatically
- JSON API for agents: every recipe is also published at /<org>/<repo>.json (e.g. https://t.co/ck2MqvYYlb), so tools and agents can consume recipes without scraping
- Contribute a new recipe end-to-end with the agent skill shipped in the repo: https://t.co/dS7f84rLQc
🔗 https://t.co/WqEUat9rdc
Enjoy! ✨
Mistakes happen. As a team, the important thing is to recognize it’s never an individuals’s fault — it’s the process, the culture, or the infra.
In this case, there was a manual deploy step that should have been better automated. Our team has made a few improvements to the automation for next time, a couple more on the way.
Claude Code leaked their source map, effectively giving you a look into the codebase.
I immediately went for the one thing that mattered: spinner verbs
There are 187
Software horror: litellm PyPI supply chain attack.
Simple `pip install litellm` was enough to exfiltrate SSH keys, AWS/GCP/Azure creds, Kubernetes configs, git credentials, env vars (all your API keys), shell history, crypto wallets, SSL private keys, CI/CD secrets, database passwords.
LiteLLM itself has 97 million downloads per month which is already terrible, but much worse, the contagion spreads to any project that depends on litellm. For example, if you did `pip install dspy` (which depended on litellm>=1.64.0), you'd also be pwnd. Same for any other large project that depended on litellm.
Afaict the poisoned version was up for only less than ~1 hour. The attack had a bug which led to its discovery - Callum McMahon was using an MCP plugin inside Cursor that pulled in litellm as a transitive dependency. When litellm 1.82.8 installed, their machine ran out of RAM and crashed. So if the attacker didn't vibe code this attack it could have been undetected for many days or weeks.
Supply chain attacks like this are basically the scariest thing imaginable in modern software. Every time you install any depedency you could be pulling in a poisoned package anywhere deep inside its entire depedency tree. This is especially risky with large projects that might have lots and lots of dependencies. The credentials that do get stolen in each attack can then be used to take over more accounts and compromise more packages.
Classical software engineering would have you believe that dependencies are good (we're building pyramids from bricks), but imo this has to be re-evaluated, and it's why I've been so growingly averse to them, preferring to use LLMs to "yoink" functionality when it's simple enough and possible.
PyTorch 2.11 is now available, featuring 2,723 commits from 432 contributors since PyTorch 2.10. This release prioritizes performance scaling for distributed training and next-generation hardware architectures.
Highlights include a FlashAttention-4 backend for FlexAttention on Hopper and Blackwell GPUs, Differentiable Collectives for distributed training, and performance optimizations for Intel GPUs via XPU Graph. This release also delivers comprehensive operator expansion for Apple Silicon (MPS) and RNN/LSTM GPU export support.
🖇️ Read the PyTorch 2.11 release blog and release notes: https://t.co/JZ4xkjEiNQ
#PyTorch #OpenSource #AIInfrastructure
We released 🤗 Kernels 0.12.3 to support Flash-Attention 4! This means we now support kernels written in `cutlass.cute`.
```
from kernels import get_kernel
fa4 = kernels.get_kernel("kernels-community/flash-attn4" version=0)
```
and you're ready to go!
Diffusers and Transformers have already been PR'd to support FA4 exclusively through 🤗 Kernels.
Find the release notes:
https://t.co/ic681vn7Xs
@Amank1412@meszmatew Llama was once leading the open-source LLM space.
They have strong dense models, but they seem to have fallen behind in the MoE race, if I’m not wrong.
I think the initial releases of Qwen were based on Llama’s architecture.
I’m hoping for a comeback from Meta.