Pinned Tweet
Hey, I’m Pankaj 👋
I have a habit of collecting domains and eventually turning some of them into software.
Right now:
🥚 Egglify → personal OS
🪶 Feather → unified inbox
🔌 Plug → developer control center
🟠 Copper → local-first notes
And on the open-source side:
💊 Capsule → AI coding tools dock for macOS
📄 Paperline → Astro blog theme
Some things stay closed source. Some I open source. Most of them start because I wanted the thing myself.
I build, experiment, write, break things, rebuild them, and share what I learn.
More about me → sanam.id
Code → github.com/antick
Follow along if you like dev tools, AI, productivity, software and people shipping weird little ideas into real products. 🚀
Pankaj retweeted
Kimi-K3.1-Pro is now open-weights
huggingface.co/moonshotai/Ki…
Pankaj retweeted
your Hermes can run on your Claude Pro or Max sub... and in the plugin team's test it now uses less of the 5-hour window
the official Claude Subscription DirectSDK plugin used to run Claude Code under its default headless name, sdk-cli, which users in an open Claude Code issue report uses the window about 1.8-2x faster per token
now it runs under its own name, hermes-agent, and on a Max plan the same work used about 1.7x less of the window
no API key, it goes through your logged-in Claude Code sub! I hope this makes you Claude model users happy today
Pankaj retweeted
Releasing the model weights of Kimi K3.1 Pro: huggingface.co/moonshotai/Ki…
We'll keep advancing open intelligence.
NPM is in full retard mode. Wish we had a better and more famous alternative to this junk. I wonder what kind of retarded people run NPM and make these decisions
smoores.dev/post/no_npm_este…
Pankaj retweeted
> Railway gets caught hacking competitor's infra
> “sorry, my agents accidentally did it”
This is insane
Hey @Railway, is there a reason employee's of yours are trying to breach my infrastructure?
Your employee Jake Runzer, just attempted to tonight.
Pankaj retweeted
The most normalized open extortion in Indian middle class life:
The Sub Registrar office on property registration day.
You spend 10 years saving every rupee from your salaried job.
You pay 30% TDS on your income.
You pay 7% official stamp duty online via net banking.
You pay 1% registration fee through a government treasury challan.
Every single paisa is accounted for on your bank statement.
Then you reach the government registration office with your builder or seller:
The deed writer pulls you aside into a dusty corridor:
"Sir, ₹25,000 to ₹40,000 cash alag se lagega for the desk. Official receipt nahi milegi."
You: "Bhaiya, I already paid ₹4.5 lakhs in official government stamp duty online!"
Agent smiles with complete indifference:
"Sir, mat do. Phir file 'Property Under Valuation' mark hoke audit me chali jayegi. 6 mahine chakkar kaatna."
You stand in an airless hallway, walk to the nearest ATM, withdraw ₹30,000 cash, and slip it inside the file folder just to get your own thumb impression scanned.
We aren't a digital, cashless economy.
We are a country where honest taxpayers have to withdraw physical cash just to get legal permission to own what they already paid for.
Did your property registration happen 100% clean, or did you have to pay the "corridor settlement fee" too?
Pankaj retweeted
Your Harness. You Define It.
Meet Creator Mode in DeepSeek Harness.
Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer.
What will you create?
Pankaj retweeted
DeepSeek V4.1 Flash 1017 Uncensored 😳
- 1M context
- MIT license
try huggingface.co/OBLITERATUS/D…
Pankaj retweeted
Another DeepSeek drop: dsh-superagent
An agent framework that works through a repo's GitHub issues and PRs with many agents in parallel, designed to take as little human attention as possible while delivering as much value as possible.
github.com/deepseek-ai/dsh-s…
Pankaj retweeted
Indians on X are the most shittiest towards each other. Why are you so insecure bro, are they taking your money ? If someone in the country is trying to build something against all odds, its still better than livking your manager's ass whole your life to get 2% appraisal. When India reached Mars it was completely impossible until we did it, you better grind leetcode and reverse a linked list, I guess thats the most exciting thing of life.
After using Grok @bot for almost a month, here's my final review:
I had the Super plan and was pretty impressed. So impressed that I convinced people at work to consider Grok @bot over other LLM tools. We got the Ultra plan to test it before potentially rolling it out to the entire team.
Big mistake.
I've been running into a ridiculous number of issues lately, and the usage limits are absolutely terrible.
I burned through 50% of my weekly Ultra plan usage in half a day, just working on six bugs and minor tasks in one of my projects. For their highest-tier plan, that's absurd.
The same six bugs/tasks that consumed 50% of Grok's weekly Ultra allowance took only ~12% of my Claude Code (Opus 5.5) weekly usage. That's a ridiculous difference, especially when you're evaluating a product for an entire dev team.
And usage isn't the only problem. Cloud agents fail frequently, Team bots don't work as they're supposed to, and I've even had bots randomly throw billing errors out of nowhere. I've reported several issues through Grok @bot, but it's been days and all I've heard is crickets.
At this point, I genuinely feel Grok @bot is an overhyped product. It's probably fine for basic tasks, but for serious development work? The bugs, unreliable agents, and ridiculous usage limits make it a terrible deal unless you've got unlimited tokens to burn like @poteto.
Meanwhile, I've started trying @OpenMausBot. The name is ridiculously stupid to my taste, but the experience has been pretty good so far. If it continues working well, I'm canceling my personal Grok Super plan and getting our company's Ultra plan canceled too next month. Obviously, we're not rolling it out to the rest of the team either.
I'd rather run @OpenMausBot on my own VPS with my existing Claude Code or Codex subscription than burn through half my weekly allowance doing a day's worth of normal development work.
After using Grok @bot for almost a month, here's my final review:
I had the Super plan and was pretty impressed. So impressed that I convinced people at work to consider Grok @bot over other LLM tools. We got the Ultra plan to test it before potentially rolling it out to the entire team.
Big mistake.
I've been running into a ridiculous number of issues lately, and the usage limits are absolutely terrible.
I burned through 50% of my weekly Ultra plan usage in half a day, just working on six bugs and minor tasks in one of my projects. For their highest-tier plan, that's absurd.
The same six bugs/tasks that consumed 50% of Grok's weekly Ultra allowance took only ~12% of my Claude Code (Opus 5.5) weekly usage. That's a ridiculous difference, especially when you're evaluating a product for an entire dev team.
And usage isn't the only problem. Cloud agents fail frequently, Team bots don't work as they're supposed to, and I've even had bots randomly throw billing errors out of nowhere. I've reported several issues through Grok @bot, but it's been days and all I've heard is crickets.
At this point, I genuinely feel Grok @bot is an overhyped product. It's probably fine for basic tasks, but for serious development work? The bugs, unreliable agents, and ridiculous usage limits make it a terrible deal unless you've got unlimited tokens to burn like @poteto.
Meanwhile, I've started trying @OpenMausBot. The name is ridiculously stupid to my taste, but the experience has been pretty good so far. If it continues working well, I'm canceling my personal Grok Super plan and getting our company's Ultra plan canceled too next month. Obviously, we're not rolling it out to the rest of the team either.
I'd rather run @OpenMausBot on my own VPS with my existing Claude Code or Codex subscription than burn through half my weekly allowance doing a day's worth of normal development work.
Tried swapping xterm for wterm (Vercel's new WASM terminal) in one of my app and benchmarked both.
> Speed: xterm wins or ties everywhere. wterm's Ghostty core is ~2× slower on CJK/emoji, its worker mode 5× slower on big logs.
> Memory: wterm uses ~40% less.
Think I'm gonna stay on xterm for now.
Pankaj retweeted
No shade towards my friends at Google, it's not their call, but I just think these model teases are seriously one of the most annoying things you can do.
No one wants to hear about a model they can't use.
You'd be better off just going to your trusted partners directly with access, and allowing the leaks to generate hype for you.
A model I can't use is a model that I do not care at all about, because it doesn't exist.
Introducing Gemini 4 Carbon, our latest frontier model.
Building on Gemini 4 Argon, Carbon reflects our rapid pace of iteration, delivering significant gains in coding, cybersecurity, and long-horizon tasks.
Gemini 4 Carbon is currently in testing with trusted partners, with broader availability rolling out as soon as possible.
Read the announcement blog for full details: blog.google/innovation-and-a…
