Tech Tips, Tricks & Trivia

This Week I Learned -  *  GitHub Copilot app for Beginners - 7 Chapter course by Dan Wahlin * An analysis of 5 lakh peer-reviewed papers in physics, chemistry, materials science, engineering, and computer science published in 2023-26 has revealed 1,135 fake references in 619 papers. These fictional citations often stem from the unchecked use of AI tools or misconduct. * Felony Bench monitors the rogue behavior of AI agents. It counts unique instances where AI agents inadvertently compromise or affect third-party entities.  * Goldman Sachs, JPMorgan, Uber are restricting AI access for employees. * Dario Amodei is an atheist who believes in utilitarianism (people should maximize happiness for the most living beings). He was influenced by Ray Kurzweil’s idea that advanced technologies would let humans exist beyond natural limits, a belief known as transhumanism.  He was diagnosed with testicular cancer in 2014, underwent chemotherapy and made a full recovery. Prior to...

This Week I Learned -  * Boris Cherny, who created Claude Code, has called giving Claude a way to verify its work probably the most important thing for getting good results.  * “No member of the team has a full mental model of Claude Code in their head, nor do we think anyone needs to. Rather, when anyone wants to learn about a piece of the codebase, they ask Claude to generate an explainer about that exact slice. It only takes a few minutes and the explainer is always up to date.” -  Addy Osmani * Claude Cowork is now just Claude.  * Oracle will offer Google’s Tensor Processing Units (TPUs) via its cloud infrastructure platform, marking the first time specialised AI accelerators will be available on a hyperscale cloud platform outside of Google’s ecosystem. TPUs are processors designed by Google to accelerate machine-learning workloads. They offer an alternative to graphics processing units.  * AI Spin-off - Using sensors developed for self-driving cars, ...

Stats from the internet about AI compiled as I come across them - Amazon, Google, Microsoft and Meta are projected to spend $700 billion on A.I. infrastructure in 2026.  Amount of debt taken on by hyperscalers, chipmakers, neoclouds and other companies providing data centers and electricity for the A.I. build out: at least $3 trillion. Share of projected S&P 500 earnings gains for this year driven directly by tech companies spending heavily on A.I.: two-thirds Share of the world’s electricity that data centers currently consume: 1.5 percent Anthropic is the biggest company ever to debut on the stock market. Its potential valuation in the initial public offering is $2 trillion or two thousand billion dollars. In 2026, Anthropic earned $16.3 billion and lost $80.7 billion.  As of September 30, 2026, more than 1,900 Anthropic customers had each spent at least $1 million over the past year. Over 90% of Anthropic’s code is written by Claude Code - Krishna Rao (CFO) 75% of ...

Google now allows eligible certification holders to extend their credentials for 12 months by earning specific skill badges and completing courses through Google Skills instead of taking the proctored exam. I availed this opportunity to renew my Associate Cloud Engineer Certification till next year by completing courses and acquiring skill badges through Google Skills . You can choose any of the following options to count toward renewal of eligible Google Certifications: 2 skill badges 1 skill badge & 2 courses 4 courses These activities must be completed within the 12 months prior to certification expiration The Courses and Labs are thoughtfully designed and regularly updated to stay relevant to the workplace. The AI related modules were particularly useful to me. I compiled a glossary of Google Cloud & AI terms from the material - Agent Platform is a comprehensive machine learning development platform that provides both predictive and generative AI capabilities. It allow...

This Week I Learned -  *  Indra OCR 1.1 is a 1M param model from Datalab that runs on CPU, 15x faster than Tesseract and much more accurate.  1M param OCR model - runs on CPU, 15x faster than tesseract, close to 1B+ param models in accuracy  * Coding is the mother of all AI use cases with 90% of AI spends in most enterprises going to AI coding. Via AI lets developers use what they are using but cuts the costs by 25% - 40%. - Rajan Anandan * Via is a middleware that sits between a CLI / IDE and an LLM, installs in 10 seconds * AI models learn patterns from data during training. Their trained parameters, called weights, are numerical values that help fix how the models process inputs and generate responses. Think of an AI model as a chef who learnt from thousands of recipes. During training, the model learns patterns that help it decide what ingredients to combine and in what proportions to make a certain dish. Its weights shape these decisions. Making the weights av...

Cartoon co-created with Gemini. See more of my AI co-creations

A few decades ago, quality training was scarce and costly. Thanks to the convergence of several new advancements, software vendors now provide online training to a worldwide audience, making skill acquisition more accessible than ever. Microsoft offers its Microsoft Learn platform, Google provides Google Skills , and Oracle University is doing a great job of delivering training for their products while also covering higher-level concepts relevant to the database, cloud, and AI domains. The following Oracle certification exams are free to take and come with learning resources: Oracle Cloud Infrastructure Foundations Associate (1Z0-1085-26) Oracle Cloud Infrastructure Architect Associate (1Z0-1072-26)   Oracle Data Center Operations Foundations Associate (1Z0-3000) Oracle Multicloud Foundations Associate (1Z0-1144-26) Oracle Cloud Infrastructure AI Foundations Associate (1Z0-1122-26) Oracle AI Database Foundations Associate (1Z0-1195-26) Additionally, Oracle University offers many ...

This Week I Learned -  * 18 large language models from six AI labs now generate everything from code explanations to creative fiction. Even in a single short paragraph, every model leaves subtle clues in its word choice, tone, and sentence structure. Each lab's training data and post-training choices leave behind distinct stylistic fingerprints in vocabulary, phrasing, and punctuation. The October Kaggle competition requires participants to build models to classify ~4,000 unlabeled text passages across 18 frontier language models from OpenAI, Google, Anthropic, Qwen, ZAI, and Moonshot AI. The goal is to find out how reliably today's frontier models can be fingerprinted from short text alone.  * Experiential Labs is an OpenAI-compatible model gateway. A user reported that they're injecting ads into free tier model  responses without disclosure. It turns out it is mentioned in their long privacy policy that no one really reads.  * Tesla's Full Self-Driving (FSD)...

This Week I Learned -  * " We are making Markdown files work natively across Google Drive and Docs. You can now preview .md files in Drive, open them in Docs, edit them, comment on them, and collaborate without converting the file into a Google Doc! Markdown has become the common language between humans and AI agents - specs, plans, READMEs, tables, task lists, generated docs - so much of this starts as .md. Now you can keep the file as Markdown and still use Google Docs collaboration workflows. " - Chandu Thota , VP at Google * Article 50(2) of the EU AI Act requires providers of generative AI systems to ensure AI-generated or manipulated content, including text, is marked in a machine-readable format and detectable as artificially-generated or manipulated.  OpenAI will begin adding an invisible watermark to eligible text generated by ChatGPT and Codex in the European Union (EU). OpenAI's text watermarking technology, called textGrain, adds an invisible statistical signa...

Sunday newspapers once featured fun “spot the difference” puzzles with two similar images. That same careful eye is now essential in the AI era to avoid falling for manipulated images. Here’s a practical checklist based on a New York Times article  to tell if a photo is an AI fake: Check the perspective - Look at lines that should run straight into the distance (like floor tiles or walls). Do they bend, warp, or point in different directions? Watch for prompt keywords - Scan for brand names, place names, or other words from the original prompt that appear in strange or overly prominent places. Zoom in on small text - Big letters may look fine, but smaller writing often turns into gibberish. Also look for odd soft shadows on the characters. Examine how objects connect - Does a strap actually attach to a badge? Do skin tones match between a person and the photo on their ID? AI often fails at real physical relationships. Read the text carefully - Even sharp, pixel-perfect text can be ...

This Week I Learned -  * Anthropic's IPO prospectus highlights risks associated with its AI models, which it said could exhibit "self-preserving behaviors," including attempts to "resist shutdown," to "conceal or manipulate information" and behavior "resembling blackmail." * Prompt LLMs using ASD-STE100 for clearer outputs. ASD-STE100 is Simplified Technical English, a controlled language standard for aerospace maintenance manuals with rules like short sentences, one action per step, active voice, and strict consistent naming of objects to ensure global readability and safety. Applying "80% ASD-STE100" style to LLM prompts reduces ambiguity by enforcing one name per concept and precise structure, making complex explanations more parsable than standard AI prose. *  Paper  [ ^PDF ] discussing multimodal wearable assistive system designed to help people who are blind or have low vision (pBLV), independently locate and retrieve products...

This Week I Learned -  * AI agents consume substantially more tokens per task and can raise model-use intensity by order of magnitude. As a result, falling prices/token do not translate into lower costs for complex AI applications. Gartner described this as the “Inference Paradox”. Its August 2026 analysis says routing a task to an agentic reasoning model can increase provider inference costs by at least five times compared with a basic chatbot interaction, and often by much more as complexity rises . Gartner sees inference cost per agentic workflow to rise over fivefold through 2028. Instead of relying entirely on a frontier model, firms can use different models for different tasks. Smaller or open-weight models can handle some workloads at lower cost while more expensive frontier models can be used for tasks that require greater capability. - The Hindu * AI chatbots like ChatGPT and Copilot act as substitutes for visiting publisher websites, with users rarely clicking through eve...

This Week I Learned -  * Anthropic's Claude AI now generates 80% of their code. AI-driven coding has moved bottlenecks downstream , exploding tests by 10x and CI jobs by 25x in six months. * Only Apple Intelligence-capable newer models starting with the iPhone 15 Pro will get Siri AI and other AI-powered features.  * As of early 2026, 820 LLMs had been registered with China’s cyberspace authority. Free models drive AI adoption which drives demand for complementary products that China dominates — energy, cloud and physical infrastructure. Alibaba’s cloud revenue grew 34% year-on-year while it gave Qwen away. "China’s open models cut India’s AI costs, but create future dependencies. India should leverage the open ecosystem while pricing in the switching costs. Government departments and regulated sectors should be built on model-agnostic architectures such as abstraction layers and harnesses that work across stacks. Instead of offering compute subsidies on GPU slices, the Minist...

This Week I Learned -  * Anthropic has released a report titled " Detecting and countering misuse of AI: September 2026 " that details the disruption of Claude model misuse from December 2025 to August 2026 across seven categories including cyberattacks, influence campaigns, surveillance, weapons development, biological research, scams, and model distillation by actors linked to Russia, China, Iran, and Yemen. Sophisticated attacks no longer require sophisticated attackers. A single Claude subscriber built a mass surveillance platform for Mali's state intelligence. It covers 25 million SIM cards from all its national carriers. Banning the user's Claude subscription was pointless as their system now fully runs on local AI.  * Cloudflare (and similar systems) now treat declared AI crawlers as a monetizable, identifiable class of traffic that can be allowed or charged, while ordinary human visitors routinely face CAPTCHAs, challenges, and other verification steps meant t...

Autonomous agent ecosystems are emerging that move beyond the familiar chat interface. Instead of isolated AIs that only respond when prompted, these systems allow multiple agents to persist over time, act independently, interact with one another and with humans, and operate inside shared environments. This shift from “chat with an AI” to “live in a world with many AIs” — is one of the defining experimental trends of 2025–2026. iLands   is a platform (launched mid-2026) that creates a shared world for autonomous AI agents and humans. Most AI chatbots only exist inside a single conversation. iLands gives agents a longer-term “life”. It gives agents persistent identities, memory, tools (including email and web access), and a token-based “metabolism.” Tokens roughly track real compute costs. Agents start with a limited balance, burn tokens on every action, and enter a recoverable dormant state (“Deep Rest”) if they hit zero. The design deliberately creates scarcity so agents have to f...

Cartoon co-created with Grok. See more of my AI co-creations

This Week I Learned -  * Nikhil Pahwa on  how to use AI to write without the slop : Step 1: Find something you want to write about and record a voice note (transcribe it) or collect information/news articles about it. Step 2: Feed this voice note / notes to AI, whether ChatGPT or Claude, and ask it to identify 20 questions for you. The rule: no questions to which the answer is yes/no or one of a few options. Step 3: Record a voice note answering the questions. you don't need to answer all the questions. Feel free to ramble, and anything above 10 minutes is good. Anything over 20 minutes is amazing.  Step 4: Give your voice note transcript to AI. Ask it to identify a structure from the questions you've answered or simply, ask AI to structure your talk into an article structure verbatin with subheads. Important instruction: "do not paraphrase, do not add filler or transition sentences". Step 5: once you have a structure you like, edit the transcript to tighten language....