Amit Bhat

guy with an computer 🖥

[Amit Bhat @iamAmitBhat]

@TheKricKid Yes. Today WAL writes are synchronous with sync_data, so slow disk I/O can block the event loop and Raft heartbeats. That’s a known learning-version tradeoff.

[Amit Bhat @iamAmitBhat]

I built a distributed task orchestrator from first principles.Rust core engine, Bun gateway, custom binary protocol, append-only WAL, SWIM membership, and Raft consensus. Checkout here : https://t.co/nqZesSHSUH

[Amit Bhat @iamAmitBhat]

interesting part: no Raft crate, no async runtime on the core path, no database-backed log.
Just the primitives, implemented directly.

[Tehxi @yajnshri]

Neuroscientists have found that a thought held for just 17 seconds begins to attract others like it activating a frequency of resonance.

[Amit Bhat @iamAmitBhat]

Damn bruv! 🥶

[David Ondrej @DavidOndrej1]

if you are not burning 500M+ tokens per day...

you are falling behind

[Amit Bhat @iamAmitBhat]

Here you go another win for rust
https://t.co/aor6jNnNxX

No bs,this is a good one

[shadcn @shadcn]

Starting today, we are making @base_ui the default component library in shadcn/ui.

First, a bit of history. When shadcn/ui launched in 2023, it was built on Radix. At the time, nothing else came close. Headless. Accessible. Composable.

Fast forward a few years and the team who built Radix are building something new: Base UI.

[Qian Li @qianl_cs]

Postgres is a superpower.

[Amit Bhat @iamAmitBhat]

Enjoying zennotes and Lumary is also cool. Oh man! how r u building all such cool stuff!?
@adibhanna
#zennotes #lumary #screenanotater

[New Mentalities @NewMentalities]

It's going to suck for a while.
But not forever.

[Abhishek🌱 @Abhishekcur]

One thing I've started to notice the longer I study, it's superb.

whether it's physics, math, systems engineering, ML, or AI inference, the same patterns keep showing up. different names, different notation but underneath, it's usually the same few ideas wearing new clothes.

you end up asking the same questions every time:

- Where's the bottleneck?
- What's the trade-off?
- What are the real constraints?
- Can this run in parallel?
- Where is time or memory being wasted?
- What's the simplest version that still works?

The domain changes. the vocabulary changes. but the way you think stays surprisingly the same.

That's why the start of anything feels so overwhelming. everything looks new, and it's easy to assume everyone else is just smarter than you.

Then one day it clicks. you realize you were never learning thousands of unrelated things, you were just seeing the same patterns show up in different places.

So the hard part was never "learn more." It's staying with the craft long enough, through the boring and confusing parts, until those patterns stop being things you look up and start being things you just see.

Here's some example:
The memory-vs-compute trade-off.

Say you're computing Fibonacci numbers. The naive recursive version recalculates the same values over and over, fib(50) takes billions of calls and crawls.

So you store each result the first time you compute it, and just look it up afterward. Suddenly it's instant. You spent a little memory to save a huge amount of time.

That exact trade-off shows up everywhere once you notice it:

- Caching a slow database query so you don't hit the database twice.
- Dynamic programming filling a table so each subproblem is solved once.
- A GPU keeping activations in fast memory instead of recomputing them during training.

Different fields, different vocabulary: memoization, caching, DP, activation storage, but it's all the same idea: spend memory to save time.

That's the pattern. Once you've seen it in one place, you start recognizing it in all the others.

If you're early in something and it feels like too much, that's not a sign you're behind. keep going. give it good time.

[Ben Dicken @BenjDicken]

Know how long every operation takes on a computer.

Once you do, identifying good vs bad vs acceptable performance becomes way easier.

It all boils down to knowing your napkin math.

[Amit Bhat @iamAmitBhat]

Reliability and Verification at scale with decentralized model . #decentralizedsystems
#dhiway #verification #digitaltrust

[dhiway @dhiwaynetworks]

Blockchain is moving beyond buzzwords — toward verifiable trust.

Glad to see @dhiwaynetworks featured by Moneycontrol on NSRCEL-recognised startups driving real blockchain adoption in India.

https://t.co/Dta42q07cF @smohan

[Malay Krishna @Malay4Product]

Not many are talking about it, but this is one of the most underrated things India is shipping right now and every Indian must know what this is all about.

Let me explain;

The system is called DIGIPIN and the username layer sitting on top is called DHRUVA. Built by the Department of Posts in partnership with IIT Hyderabad and ISRO's National Remote Sensing Centre.

Officially launched on May 27, 2025.

Here's how it works.

DIGIPIN divides all of India into 4 metre by 4 metre squares. Every single square gets a unique 10-character code like 829-4G7-PMJ8. That's down to the level of your front door, your shop counter, your hospital entrance, your village home, even a fishing boat in territorial waters. The entire country is now a digital grid.

But remembering a 10-character alphanumeric code is hard. So DHRUVA sits on top of it. You convert your DIGIPIN into a simple readable handle like rajesh@dhruva. The handle stays with you for life. If you move houses, only the underlying DIGIPIN updates. Your handle doesn't change.

Exactly like UPI replaced 16-digit bank account numbers with simple handles. malay@ybl instead of remembering an account number.

But why is our government building this?

Today roughly 20-25% of Indian addresses are unstructured. Slums, tribal areas, unplanned colonies, rural homes without proper street names.

An average Indian spends 8-12 extra minutes on an average in finding an address in India versus 2-3 in the West.

Ambulances reach late because nobody can describe the lane. Banks reject mortgages because they can't verify the property location. Insurance claims get delayed because addresses don't match across documents. Quick commerce loses crores in failed deliveries every day.

DIGIPIN solves all of this with one open-source standard.

The full source code and documentation are on GitHub. Any government department, private company, or startup can integrate it for free.

This is exactly the India Stack playbook. Aadhaar (identity), UPI (payments), ULPIN (land), DigiLocker (documents), and now DIGIPIN (address) are all open public infrastructure that private companies build on top of.

Of course developed countries already use a version of this. But India is building the best of the lot.

> UK uses postcodes plus house numbers. Works because they have structured street planning from the 1800s. We don't.

> Dubai built Makani numbers. 10-digit codes tied to building entrances. Government-only, not open.

> Japan uses block-based addressing that relies on physical signage and local familiarity.

India just built the best version of all of these.

Open-source, geo-coded, privacy-first, with a human-readable layer that even a non-tech grandparent can use. And it's free to integrate.

Once this gets rolled out, the government expects that;

> Ambulance response times improve by 40-60% in unplanned areas.

> KYC verification becomes instant. No more manual address proof.

> Rural credit unlocks. Banks can verify property and ownership in seconds for loans.

> Disaster response improves. Floods, fires, earthquakes. Rescue teams know exact homes to reach.

> Insurance pricing becomes location-precise. Same building, ground floor versus third floor, different flood risk, different premium.

> E-commerce delivery accuracy goes from approximate to exact. Failed deliveries drop sharply.

> Privacy too gets better. You share your DHRUVA handle, not your physical address. The delivery agent gets the GPS coordinates without seeing your full address. Less data exposed, less misuse.

Boring infrastructure rarely gets any hype. Everyone laughed at UPI for the first two years. Now it processes 16 billion transactions a month and seven countries have adopted it.

DIGIPIN will be the same story. In 5 years we'll wonder how we ever functioned without it. In 10 years it'll be quietly running underneath every delivery, every emergency call, every loan approval in India.

[Amit Bhat @iamAmitBhat]

Came across mindsdb tool and got some hands on with it. Here is a blog post I wrote as an part of my learning
https://t.co/T02VYT2hFn

@dhiwaynetworks

[dhiway @dhiwaynetworks]

From lamps that light homes to trust that powers networks - Team Dhiway wishes you a bright and joyous Diwali!

[dhiway @dhiwaynetworks]

Got ideas? We’ve got opportunities. Become a @dhiwaynetworks #Intern.

Send your resume to [email protected]

[Brodha V @BrodhaV]

My new song - 'OM' is out now! This one's very personal to me. I hope this resonates with all of you! 🙏🏼
#OM #newmusic

[Andrej Karpathy @karpathy]

+1 for "context engineering" over "prompt engineering".

People associate prompts with short task descriptions you'd give an LLM in your day-to-day use. When in every industrial-strength LLM app, context engineering is the delicate art and science of filling the context window with just the right information for the next step. Science because doing this right involves task descriptions and explanations, few shot examples, RAG, related (possibly multimodal) data, tools, state and history, compacting... Too little or of the wrong form and the LLM doesn't have the right context for optimal performance. Too much or too irrelevant and the LLM costs might go up and performance might come down. Doing this well is highly non-trivial. And art because of the guiding intuition around LLM psychology of people spirits.

On top of context engineering itself, an LLM app has to:
- break up problems just right into control flows
- pack the context windows just right
- dispatch calls to LLMs of the right kind and capability
- handle generation-verification UIUX flows
- a lot more - guardrails, security, evals, parallelism, prefetching, ...

So context engineering is just one small piece of an emerging thick layer of non-trivial software that coordinates individual LLM calls (and a lot more) into full LLM apps. The term "ChatGPT wrapper" is tired and really, really wrong.