Piyush (hello/hi)

Loves Engineering, likes tech, does Product Management ||

I look after The Indian Music Diaries website as well ||

It's not power that corrupts, it's fear

[Piyush (hello/hi) @piyushdev709]

I always wanted a way to delete the duplicate and near-duplicate photos from my Google Photos account. I never really trusted the third-party tools out there with my private stuff, so I finally built it on my own this weekend.

- Privacy-first: It runs locally, so private images stay on your machine not someone else's server.
- Fuzzy matching: Identifies both exact duplicates and near-duplicates.
- Technicals: Built using Next.js, TypeScript, and Prisma.
Sharing this for some good karma. If you find it useful, a star on the repo would be great. :)

Link: https://t.co/Ed5uhVk7V2

P.S: Would love feedback, feature requests.

[Piyush (hello/hi) @piyushdev709]

@shantanugoel Hahaha. GLM 5.3 is what I’m waiting for now.

[Anand Ranganathan @ARanganathan72]

62% of Indian resident doctors work more than 36 hours at a stretch. 86% report severe sleep deprivation. 97% earn less than an entry level civil servant. 76% are assaulted while on duty.

We can call ourselves civilised only when we learn to treat our doctors like they treat us.

[Piyush (hello/hi) @piyushdev709]

@claudeai I have an account on claude which is giving me a lot of trouble. It's been more than 15 days since I paid for a pro plan. The plan does not show up, it says "free" and keeps asking me to pay. I have raised a complaint ticket, and not response on it

[Piyush (hello/hi) @piyushdev709]

@championswimmer 95

[Hugging Models @HuggingModels]

Meet Ambuj-Tripathi-Indian-Legal-Llama-GGUF: a specialized AI model fine-tuned for Indian law. This isn't just another chatbot. It's a legal assistant trained to understand the nuances of Indian statutes, case law, and legal language. A game-changer for legal tech in India.

[Piyush (hello/hi) @piyushdev709]

Have you come across the term 'Mixture of Experts' (MoE). How are these different from the 'Dense' models?

In a Dense models, every single parameter is activated for every single word (token) you type.
- How they work: If you have a 90B parameter model (~llama 3.2), the model runs all 90B calculations for every prompt, entire weight matrix is multiplied for every single token, whether you ask for a recipe or ask it to write code.
So obviously it's computationally expensive, latency is also higher.

MoE models are different. As the name suggests, these are sparsely architected.
- How they work: They have sub-networks. Example ~Deepseek-V3 has 671B total parameters, but also only 37B active parameters. So only a subset is used when it works on any problem. Efficiency is higher, they are faster, use less compute and memory. But raw performance may be marginally less.

I have been fiddling around with https://t.co/G42ozaNWmR lately, and I am amazed at how tiny models, optimised for local operation, calling functions, running small tasks for you can be. These are tiny experts, at operating your phone, ~270mb models that can do anything on your phone you program them.
AI on edge, running locally, models for specific tasks, specific outputs, that work on much simpler hardware (fully CPU) are getting ignored in the noise around really big models, that break all benchmarks, and need Gigawatt GPUs. But there is another frontier, about to hit your phones.

[Piyush (hello/hi) @piyushdev709]

Made an app to keep score for board games night with my friends: https://t.co/zGhiPtJGtO
I would love to add games if anyone has suggestions. :)

[Piyush (hello/hi) @piyushdev709]

🤭

[Harshil Mathur @harshilmathur]

Can't blame a guy for trying :)

[Deedy @deedydas]

Claude Mythos just obliterated every single benchmark in AI. I can't believe what I'm reading.

[Piyush (hello/hi) @piyushdev709]

@RamayanaAnime Very eager to buy the 4k Bluray, but it's not available anywhere. Please help

[Piyush (hello/hi) @piyushdev709]

One single check via LLMs is important:
The Inverse Cube Law.

The human heart acts as a magnetic dipole. The magnetic field strength (B) of a dipole decreases with the cube of the distance (r) from the source.

B∝1/r^3 (B is proportional to 1 by r cubed)

Let's look at the numbers:

The magnetic field of a human heart at the surface of the chest (distance ~0.05 meters) is roughly 50 picotesla (50×10−12 Tesla).
If an aircraft or drone is flying just 500 meters away, the distance has increased by a factor of 10,000.
Because the field strength drops by the cube of the distance, the signal strength drops by a factor of 10,0003, or 1 trillion (1012).

At 500 meters, the heart's magnetic field would be 50×10−24 Tesla.

The absolute most sensitive quantum magnetometers currently theorized or in existence bottom out around the attotesla range (10−18 Tesla). A signal at 10−24 Tesla is not just undetectable; it is functionally non-existent. It is orders of magnitude weaker than the fundamental quantum noise floor of the universe and the thermal noise of the sensor itself.
Furthermore, you cannot use AI to filter out noise if the signal-to-noise ratio is effectively zero. If the signal is physically weaker than the random quantum fluctuations of the atoms in your sensor, there is no data for the AI to process.

[Piyush (hello/hi) @piyushdev709]

@bhavintu We connected entire teams to sql and clickhouse. It was unbelievable how data blockers just vanished without a trace. :)

[Piyush (hello/hi) @piyushdev709]

@asxna Super! Please keep the /defendyourself for skill invocation:)

[Piyush (hello/hi) @piyushdev709]

@shantanugoel I’m hoping RAM prices also come down properly this time, I’m stuck half way upgrading my pc

[Piyush (hello/hi) @piyushdev709]

Going to happen a lot more!

[Wes Bos @wesbos]

‼️Do not npm install or deploy anything right now

Supply chain attack on axios 1.14.1 - even if you don’t use axios it may be a nested dep.

Pin versions or wait until this is resolved

[Piyush (hello/hi) @piyushdev709]

This guy is relentless. And that’s a good thing!

[biased indian @RakeshK32229480]

X is deleting my posts ?
Day 20.5 of travelling in faridabad (today off) 🥰🥰
Sir @narendramodi pakistan is continuously attacking faridabad with garbage missiles accordingly municipality here is unable to clear the garbage sir kindly take action against Pakistan do they don't attack us with garbage
@MCF_Faridabad location sector 48 main road > today Pakistan might do another garbage attack

[Piyush (hello/hi) @piyushdev709]

We have reached the critical point when small models that fit 1GB ram have become capable enough to do wonders on mobile devices.
September iPhone launch will be massive.

[Piyush (hello/hi) @piyushdev709]

Have you tried making models speak Hinglish?

[Piyush (hello/hi) @piyushdev709]

Somethings never really change. Open source innovates, but larger companies will offer more security, maybe slightly less features, and the world will shift to paying a subscription.
My dream of having a fully open source stack will never really come true. :(

[Claude @claudeai]

You can now enable Claude to use your computer to complete tasks.

It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk.

Research preview in Claude Cowork and Claude Code, macOS only.

[Piyush (hello/hi) @piyushdev709]

@isawsukul By the way, I finally see your ads now. :)