OpenAI is pretty huge, much bigger than all the existing SaaS Companies combined.

Most people think always-on AI agents run 24/7. They don’t. Here is the actual architecture behind Grok Bot @bot (and why it can keep working after you close your laptop). It’s built to stay asleep most of the time and only wake up when there’s real work. The system has three main parts: - Thin clients (your desktop app, phone, Slack, email) - A thick server where the agent loop actually runs - A cloud computer (Firecracker microVM) the agent uses when it needs to browse, run commands, or edit files The key design: everything sleeps by default. When you send a message, a schedule fires, or an event comes in (Slack, email, etc.), it wakes up. It loads its state from a database, restores the VM from a snapshot if needed, and runs the agent loop (LLM + tools). It uses durable workflows (Temporal), so if something crashes mid-task, it resumes from the last completed step instead of starting over. When the work is done, it saves state, snapshots the VM, and goes back to sleep. That’s why you can close your laptop, and it keeps working. The state lives outside any running process. The model is only one piece. The real engineering is in how it sleeps, resumes, and uses its own computer. I think @poteto can add more if she wishes and educate us!

You are using X, you post something, and it sometimes travels to a wide range of audiences, and others are travelling into the 20s, kind of impression. First, that's a wild difference, but have you ever thought about why that happens? It got me curious about what X actually does before deciding who sees your post. So I went through its published algorithm code, and some of the things I found were really surprising. For example: X's algorithm gives a weight of +0.5 and a report −234. So one report cancels 468 likes? Nope. X literally added a warning in the source code saying that's wrong. Those weights multiply predicted probabilities, not actual engagement counts. And reports are over 1,000x less likely than likes. Here's how the pipeline actually works: 1. Discovery: Thunder finds posts from accounts you follow. Phoenix and SimClusters discover posts outside your network. If your post never enters someone's candidate pool, ranking doesn't even matter. 2. Prediction: Phoenix estimates whether that viewer might like, reply, quote, share, follow, click, keep reading, or react negatively. Different viewer, different score. There isn't one universal popularity score attached to your post. 3. Distribution: Out-of-network posts face a discount. Repeated posts from the same author lose some priority. There's even a mechanism to give underexposed authors a chance. Posting more doesn't automatically mean getting more reach. 4. Visibility: This one gets overlooked. Separate filters can remove posts from recommendations, sometimes only for non-followers. The For You pipeline also filters out-of-network replies and reposts. A post can be visible on your profile but still struggle to reach new people. So those viral engagement-point charts? Misleading. The code also doesn't establish a magic posting hour, a guaranteed 30-minute test, or a blanket penalty on external links. Likes aren't useless. Engagement counts are used elsewhere, but one reply doesn't simply equal ten likes. After going through this, I think most creators are optimizing for the wrong thing. X isn't just counting reactions. It's predicting which reactions are likely from each person, after deciding which posts are even eligible to reach them. That explains something important about the system, even if it doesn't tell me exactly why my own posts performed so differently. Your post isn't competing on likes alone. It first has to reach an audience the system thinks will care.

What a use of technology, watch till end 🤣

The honest opinion: If you want to get serious work done without interruption, fast and in one shot, move to $100 pro plan. The 5-hr limit in $20 plan, is barely getting work done. If you fire sone complex task prompt, it will suck up the whole 5-hr limit and you're waiting for your next 5-hr limit to finish it up. Yup, they said to fix the bugs and asked to use skill file for context, but even after that consumption of token is like water, even for Astra light use. The problem here is that you have been allocated the weekly limit and then you are not allowed to use it. The $20 plan is just for the small experiment work and if you're serious and want to build something, go for $100 plan. That will be much beneficial and you will be given more priority. This is not criticism but the experience what we have felt by using $20 plan.

UPI feels free. The payment rails aren’t. MDR is the fee sitting behind digital payments, and it’s back in India’s UPI debate. Here’s what MDR actually is, who pays it, where the money goes, and why it suddenly matters again. 👇

Jio spent the last decade putting towers across India. Its next network may sit 650 km above them. Reliance Jio is reportedly targeting September 2027 for the first test launch of its own satellite network, and this is much bigger than putting one satellite into space. The plan: 🛰️ ~1,600 satellites in Low Earth Orbit 🌍 ~650 km above Earth 📡 broadband + cellular backhaul 📱 direct-to-device connectivity planned ⚡ roughly 100–150 Gbps capacity per satellite A few test satellites would go up first in September 2027 to validate the hardware, software and network. If those tests work: 2028: larger deployment begins 2030: ~400–450 satellites 2035: full constellation At full scale, the network is reportedly designed to provide around 4–4.8 Tbps of capacity over India. But the more interesting part is what this does to Jio itself. Jio already owns much of the terrestrial stack: fibre → towers → 4G/5G → devices/customers Now add: satellites → direct connectivity from space Remote villages, highways, islands or disaster-hit areas no longer necessarily need a nearby tower or fibre connection first. And for India, this could mean something strategically bigger: its first large-scale homegrown LEO broadband constellation, instead of depending entirely on foreign satellite networks. IN-SPACe has already found Jio's proposed architecture technically sound, with ISRO and DoT/WPC involved in the evaluation. Jio is now working through the international orbital and spectrum process with the ITU. One important unknown remains: Who builds and launches 1,600 satellites? Jio hasn't publicly named the manufacturing or launch partners yet. That may end up being one of the most interesting parts of this story. India's telecom race started on the ground. The next layer is being built in orbit.

Notion nearly died before most people had even heard of it. The idea was ambitious. The product kept breaking. Cash was disappearing. Instead of shipping more features, the founders eventually made the much harder call: start again. That reset became the Notion we know today. 🧵

Apple launches tomorrow. You get ONE wish for the event. What are you picking? • 2-day iPhone battery • AI that’s actually useful every day • a camera upgrade you can genuinely notice • a real redesign • something nobody leaked What would actually make you upgrade?

India’s AI story is moving down the stack. TCS subsidiary HyperVault and partners plan to build up to ₹700B (~$7.4B) in a 1GW AI data center campus in Telangana. This isn’t another software launch. It’s compute infrastructure at scale. (1/4)🧵