This Week I Learned - Week 10 2026
This Week I Learned -
* "Kubernetes had become a badge of engineering sophistication.
And we wore it proudly — until it started hurting.
I still use Kubernetes. I respect it deeply.
But I no longer romanticize it."
- Azeem Telli, Hotstar
* When the final New Zealand wicket fell in the 19th over, giving India a 96-run victory, the concurrent viewership was 74.5 crore and reached 82.1 crore (0.82 billion) by the end of the post-match presentation. The number broke the previous record for an ICC tournament final at 5.9 crore peak concurrent viewership, registered during the India vs Australia match in the 50-overs format on November 20, 2023. Peak concurrency is the highest number of viewers during a live stream at a time. Hotstar uses Akamai + AWS CDN for delivery.
* On March 2, strikes hit two AWS data centres in the UAE and one in Bahrain, disrupting local banking apps, airport systems in Dubai and Kuwait, and the UAE stock exchange. AWS and Microsoft Azure may reroute data center workloads from Dubai, Abu Dhabi, and Oman to India and Singapore amid the West Asian conflict.
* Experts estimate that roughly one-third of India’s westward data traffic travels through networks passing through or close to the Strait of Hormuz.
* Why knowledge cutoff dates for models matter - Claude Opus 4.6, possibly the best coding model currently, has a knowledge cutoff date of May 2025. This means in code samples Opus provides involving API calls, it will reference API versions available as of the cutoff time.
* Ammaar Reshi has reverse engineered a DOS game with no source code using Codex 5.4 and shared the source code on GitHub
* Vibe engineering is "the grown up version of vibe coding, where expert programmers use coding agents in a professional and responsible way to produce high quality, reliable results." - Simon Willison
* AI landscape from March 2023 to March 2026 as a chart showing how intelligence became sharper and cheaper over time. The question now is no longer what AI costs. It’s what you would build if intelligence cost a dime.
A rough estimate of the quality of an LLM is the ELO score on the LMSYS Leaderboard. This is like the chess ELO score, but for LLMs, where people compare 2 LLMs on the same task.
The competition is not just between humans and AI models stealing jobs, but also between older models and newer, more efficient ones.
* Moonshot AI's Kimi API's prompt caching feature with a 90% Cache Hit Rate cuts input costs by 75%. Cache hit occurs when parts of your prompt have been processed recently. This isn't magic—it's pattern recognition:
- Multi-turn conversations reuse system prompts and earlier context
- Agent workflows repeat tool schemas and instruction sets
- RAG applications overlap on retrieved document chunks
- Code agents persist file trees and project structure across calls
In coding workflows, cache hit rates often reach 85-95% because prompts contain repetitive context like system instructions, file contents, and tool definitions.
Cache hits don't just reduce your bill—they also reduce latency. When prompt prefixes are already processed, responses arrive faster. Same tokens, less time, lower cost.
* Free, open source course - GitHub Copilot CLI for Beginners
* From The Batch:
- Context Hub (chub for short) is a new tool to give to your coding agents the API documentation they need to write correct code. If you’re building AI systems using modern technologies, your coding agent will often use outdated APIs, hallucinate parameters, or not even know about the tool it should be using. This happens because AI tools are rapidly evolving, and coding agents were trained on old data that does not reflect the latest tools. Context Hub, which is designed for your coding agent to use (not for you to use!) provides the context it needs. chub has been populated with an initial set of documentation for some of the most popular tools, like common LLM providers, databases, payment processors, identity management solutions, messaging platforms, and so on.
- Moltbook is a social network for AI agents where AI agents share, discuss, and upvote. Humans are welcome to observe.
- Nano Banana 2 (formally designated Gemini 3.1 Flash Image) is an image-generation system that takes advantage of Gemini 3 Flash’s speed and strengths in language and reasoning. It uses a Mixture-of-experts transformer architecture based on Gemini 3 Flash.
- Nano Banana 2 currently competes with OpenAI's GPT Image 1.5 and is cheaper.
- OpenAI signed a contract with the U.S. military to provide AI systems that securely process classified information, displacing Anthropic’s Claude.
- Palantir performs data analytics for military and law-enforcement agencies.
- OpenAI's Frontier is a platform designed to help orchestrate corporate cadres of agents, including building them, sharing information and business context among them, evaluating their performance, and managing their interactions with employees and each other. Frontier provides a unified user interface for managing agents regardless of the frameworks and models involved.
- Microsoft's Agent 365 is a similar platform that integrates with Microsoft applications like Word and Excel. Agent 365 focuses more tightly on security and governance, while Frontier offers more features for building, evaluating, and improving agents.
- Aletheia is an agentic workflow for math research that uses the latest update of Gemini 3 Deep Think, a specialized reasoning mode of the Gemini 3 Pro model for subscribers. Google bills Deep Think as its most advanced reasoning mode, geared toward multi-step tasks in math, science, and engineering.
- Aletheia is an agentic workflow with three parts — generator, verifier, and reviser — all powered by Gemini 3 Deep Think.
* Microsoft has open-sourced inference engine bitnet.cpp (based on but forked from llama.cpp) for CPU-Powered AI on everyday devices. It outperforms llama.cpp in speed on x86 and ARM chips. It enables 100B models on CPU & offline AI on phones and edge devices. Frontier-scale models and heavy training will stay cloud-bound for now. Hybrid wins: local for speed/privacy, cloud for peak power when needed.
* Groq's free tier gives you far more audio per day than you'd ever use for dictation. No credit card needed.
* Claude makes you sound like an expert so that you don't get ignored. A citizen got a civic service implemented promptly by getting Claude to frame a request email that impressed the approving authority.
* Claude Community Ambassador, Vikram Pawar created a offline voice dictation tool for his Linux machine to dictate his thoughts to Claude Code. Everything runs locally on CPU. No cloud services, no subscriptions, no internet required.
* "I'm choosing outcome over process, and for my particular relationship with code—nostalgic but not existential—that feels right." - Samarth Bansal, Head of Content, The Whole Truth
* A 11 year old girl had an idea for a website, and instead of asking Claude AI to make it for her, she asked it to make a tutorial that would teach her how to make it. She then wrote the code for her site Rhyme Finder
She followed that up with another website StorySparks that helps you find your next great story idea. Some of the premises suggested in the Comedy section are hilarious -
- An alien ambassador arrives on Earth and becomes obsessed with reality TV dating shows.
- You accidentally become the leader of a cult, and you're desperately trying to turn it into a book club.
- A villain's evil plan keeps getting thwarted by their own bad luck. They've started going to therapy.
- A genie grants wishes but interprets everything literally. You wished to be 'outstanding in your field' and now you're stuck in a cornfield.
- You're a ghost trying to haunt your house, but the new family thinks you're just a weird roommate.
- A superhero's power is incredibly inconvenient: they can fly, but only backward.
- The chosen one of prophecy just wants to be a mediocre accountant and live in peace.
* 8 year old Hardware + AI Systems Builder Laksh's stack: ESP32, Arduino, Raspberry Pi, Vicharak, BBC Micro:bit, Python, C++, TensorFlow, PyTorch, TinyML, OpenCV, MediaPipe, Fusion 360, Git, and more. According to his father, Captain Venkat, he "uses OpenAI, Claude, Gemini, Cursor, Replit. He didn't learn any of this in a classroom. He learned it because he wanted to build something and had to figure out what stood between him and that thing."
* Hungarian “freelance” mathematician, Erdős proposed nearly 1,200 problems between the early 1930s and his death in 1996. Fewer than 500 have been solved, but AI models have helped to solve around 100 of them in the past six months alone!
* A prompt library is a centralized repository designed to store, organize, and manage predefined prompts for use in AI applications, particularly with large language models (LLMs). These libraries streamline workflows by providing ready-to-use, optimized prompts that can be reused across projects, ensuring consistency and reducing redundant work. Using libraries from reputable sources and regularly updating them can mitigate risks. Monitoring AI outputs for anomalies is also a best practice to ensure ethical and effective use.
* Data from LinkedIn's 2025 reports shows 68% of US tech roles prioritize experience over alma mater. Practical credentials like GPAs and accelerators increasingly trump institutional hype.
* In March 2024, Andres Freund's discovery of a backdoor in XZ Utils versions 5.6.0 and 5.6.1 was uncovered while investigating a 200ms SSH connection delay during PostgreSQL tests on Debian unstable.
The embedded malicious code enabled remote code execution on vulnerable SSH servers, representing a sophisticated supply chain attack likely orchestrated over years by a single contributor using a pseudonym.
Freund's alert to upstream maintainers halted distribution in major Linux repos, averting potential global disruptions and underscoring how routine performance anomalies can expose existential software threats.
* Amche Atlas is a digital public library of maps maintained by the open data community in Goa. The development of amche.in was largely done with Cursor and Claude. The website has a simple static site structure with minimal Javascript libraries and is intentionally developed to be as simple as possible to setup, contribute and test with minimum dependencies.
* Two brothers, Swaroop and Sharan have created a website that analyzes AR Rahman's discography — 282 albums (across multiple languages) and 2,130 tracks spanning 1992 to 2026. Their data visualization reveals sonic fingerprints, collaborator networks, and patterns across three decades of one of cinema's most inventive composers.
* 23-year-old Anjali Sardana, founder of Bengaluru-based instant home-help startup Pronto, has quietly emerged as one of the youngest entrepreneurs in India's on-demand services space. The company is now valued at around $100 million. Market leaders Urban Company, Snabbit and Pronto together recorded over 2 million monthly orders in February, up sharply from 1.3 million in December, helped by expansion into new cities and micro-markets.
* Chinese internet major, Tencent holds about 5-6% of the Walmart-owned ecommerce giant Flipkart. Walmart acquired Flipkart for $16 billion in 2018.
* The United States remains India's largest trading partner, accounting for nearly 20 % of India's goods exports valued at roughly USD 86.5 billion in FY2025, or about 2.2 % of GDP and generating a trade surplus of around USD 41 billion.
* "Friendships don’t fade because people don’t care. They fade because the city quietly drains the energy required to sustain them.
The affection is still there. The intent is still there.
But the civic friction of the city sits between people.
In cities like Bengaluru, maintaining friendships has quietly become an act of effort. And sometimes, effort is the first casualty of urban and civic exhaustion." - Rohan Babu
* The illusion of global brands - McDonald's in India doesn't serve beef. Coca-Cola in Mexico uses cane sugar; in the US, corn syrup. Nutella—that icon of globalization, sold in 160 countries—is actually dozens of distinct products, each shaped by local laws, local tastes, and local supply chains.
* Asian Indians characteristically develop abdominal obesity despite relatively normal Body Mass Index (BMI), a phenotype associated with higher insulin resistance, ectopic fat deposition, and earlier onset of metabolic diseases. Data from the National Family Health Survey-5 revealed that abdominal obesity affects approximately 40% of women and 12% of men. BMI is an inadequate measure of obesity. Prioritising waist circumference and related indices, including waist circumference-to-height ratio helps in more accurate cardiometabolic risk stratification.
* According to study published in Science magazine, some 8,300 million tonnes of new plastics have been produced, in turn generating over 6,300 million tonnes of plastic waste till date. Of this, only 9% has been recycled, 12% has been incinerated, and some 79% has lingered in landfills or the environment. The highest rates of recycling in 2014 were in Europe (30%) and China (25%). In the U.S., plastic recycling has stayed at 9% since 2012.
* Grey seals must maintain light sleep in water so they can regularly surface and breathe, preventing oxygen deprivation.
* Cancer treatment costs are three times higher than the hospitalisation expenditure for any average ailment in the country. Its medicines itself form the single largest component of expenditure in public hospitals — especially in rural areas. - The Hindu
* India has a 63 million hearing-impaired population as per WHO data.
* India’s prepaid recharge customers account for nearly 90% of the country’s 125 crore mobile users.
* Apple now makes about 25% of iPhones in India after China pivot
* Indian nationals accounted for about 71% of approved H1-B visas in FY24, making them the programme’s biggest beneficiaries.
* Rickrolling is an internet slang term for a prank where a user clicks a link disguised as something appealing but is instead directed to the music video for Rick Astley’s 1987 hit, “Never Gonna Give You Up”. A QR code printed on the Central Board of Secondary Education (CBSE) Class 12 mathematics question paper (2026) led to the music video of “Never Gonna Give You Up” by British singer Rick Astley.
* Hotel star ratings aren't universal; they are a patchwork of national playbooks - A 5-star hotel in Paris (where criteria include a minimum of 220 guest-facing metrics) can vary significantly from a 5-star in a country with self-assigned ratings. While the 21-country Hotelstars Union uses a strict 239-point checklist to harmonize European standards, most regions still vary wildly. Ultimately, modern travelers prioritize "trust symbols"—like real-time guest reviews—over official certifications to judge a stay.
* Ayurvedic doctors who have completed PG courses in surgical studies can perform 39 Shalya Tantra (general surgery) and 19 Shalakya Tantra (diseases of ear, nose, throat, eye, head, oro-dentistry) procedures in Andhra Pradesh according to order issued by AP Minister for Health, Medical Education and Family Welfare Satya Kumar Yadav.
* "For me...success is defined by meaning. Meaning comes from the fact that you do something in your day-to-day life which has an impact, which you believe is important." - Alexander Stubb, President of Finland
Comments
Post a Comment