Tech Tips, Tricks & Trivia

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

This Week I Learned -  * Thomson Reuters launched its proprietary LLM "Thomson" built on Alibaba's open-source Qwen model, investing $40 million over two years with a final training run costing $450,000 using its legal data from Westlaw and Practical Law. This approach highlights how enterprises with strong proprietary datasets can efficiently create customized AI, pressuring premium API revenue models of companies like Anthropic by demonstrating viable lower-cost alternatives.  * Large language models are called “stochastic parrots” as they do not truly understand language or the world. LLMs are extremely good at statistically mimicking patterns from the massive amounts of text they were trained on. They can produce fluent, coherent-sounding text that looks intelligent but are essentially “parroting” (repeating and recombining) patterns without genuine comprehension, reasoning, or grounded knowledge. * All stochastic processes are non-deterministic because randomness me...

This Week I Learned -  * Microsoft's Phi-4 Mini is widely regarded as one of the most capable models under 4 billion parameters. It punches significantly above its weight class, outperforming older 7B and even 70B models in specific reasoning tasks. It runs at a smooth 13–18 tokens per second on flagship devices with 8 GB+ RAM (like the iPhone 16 Pro/17 Pro or Galaxy S25/S26 series). * From The Batch : - PrismML, a Caltech spinout backed by Khosla Ventures, publicly released compressed versions of Alibaba’s Qwen model on Tuesday, shrinking it from 54 GB to under 4 GB so all 27 billion parameters run on an iPhone (15 or newer). PrismML achieves the compression by reducing the models to ternary or binary quantization, reducing each value from 16 bits to just two or three possible values.   - Two people asking the same model for business feedback, one in Hindi and one in Russian, may receive substantively different advice not because they’re using different models, but because th...

This Week I Learned -  *  arXiv offers HTML papers alongside the existing PDF to improve accessibility . * " Students need to learn how to work with AI, recognize weak output, think through choices, test results and decide what is worth doing. I described my own approach as “delegate maximally”. Instead of keeping a fixed list of AI-proof skills, keep giving AI everything it might be able to do. Whatever remains is the human role for now. " - S Anand * " The reason I dislike this emphasis on term “LLM Visualization” or “Claude Viz” or even “Agentic Viz” is the focus on the tool. No one goes out and says “Excel Viz” // “PowerPoint Viz” // “Figma Viz”. The tool is there to augment the human and has never been the important part.  The craft and skill of the  human using the tool is what we should acknowledge and celebrate. " - Amit Kapoor * Draw with Data by Schubert De Abreu   * Bharat Atlas is a lightweight open-source visualiser for India administrative boundar...

This Week I Learned -  *  Brain Moore  prompted nine different AI models to programmatically generate World Clocks to show how some of them struggle with the rendering. Every minute, a new clock is displayed. Each model is allowed 2000 tokens to generate its clock.  *  2025 State of AI   * India’s first sovereign LLM Sarvam AI's 30B-parameter model is pre-trained on 16 trillion tokens and supports a 32,000-token context length, enabling long conversations and agentic workflows while keeping inference costs low due to fewer activated parameters. It is a mixture-of-experts (MoE) model and has just 1 billion activated parameters, meaning that in generating every output token, it only activates 1B parameters. * Alibaba Group's Qwen3.5-397B-A17B is a new open weight multimodal model built to be faster, cheaper, and more agent capable than its predecessors. It combines text, image, and video processing in a single architecture and uses a mixture of experts design...

  See more of my AI co-creations

This Week I Learned -  *  Oracle and Google Cloud have announced a Multicloud partnership to simplify cloud migration, multicloud deployment, and management * A nice comparison of AI tools by PashooTech *  How LLMs Work, Explained Without Math * The average hospital generates 137 terabytes of data daily.   *  AI Toolkit for Visual Studio Code simplifies generative AI app development by bringing together cutting-edge AI development tools and models from Azure AI Studio Catalog and other catalogs like Hugging Face.  * In an innovative approach, MIT Media Lab has developed a website where users can interact with a personalized digital representation of their future selves . This artificial intelligence-driven experience is meant to inspire more long-term thinking and behavior by allowing users to imagine and plan for their future. The future self, however, is not a crystal ball; it's simply a model generated from user-provided survey data and existing ...

This Week I Learned -  * Microsoft 365 Education is a productivity cloud suite. Microsoft 365 builds on top of Office 365 apps such as Word, Excel, PowerPoint, and Outlook and provides enhanced management, security, and compliance tools. Microsoft 365 Education has three tiers of academic plans: A1, A3, and A5. A5 has the most comprehensive plan features [ PDF ]. Microsoft partnered with experts to develop Learning Accelerators tools Reading Progress and Reading Coach. *  DBRX is Databricks' Foundation Model * In Forrester's 21-criterion evaluation of AI Foundation Models for Language providers, Google, Databricks, NVIDIA emerged as Leaders in Q2 2024. * The first car was driven by a neural network in 1988. Twenty-five years later, Andrej Karpathy took his first demo ride in a Waymo. A decade after that, the company received its driverless permit. That’s thirty-five years of rigorous  engineering, testing, refinement, and regulatory navigation to go from p...

The DataIsBeautiful subreddit is a thriving hub for data enthusiasts, boasting over 21 million members passionate about visualizations. What truly distinguishes this community is the captivating blend of wisdom and humor found in the comments section. The feedback from the community is a delightful blend of insightful observations and witty, snarky remarks. The creator's on how to read this scatter plot chart,  Protein vs. Calorie Density: A Visual Guide , shown below. On the axis, as you move towards the right, the food becomes more calorie-dense. Similarly, moving upwards indicates increasing protein density. Consequently: Top left indicates foods with high protein content per calorie and low calorie count per 100g. Bottom right represents foods with low protein content per calorie but high overall calorie count per 100g. Top right denotes foods with high protein content per calorie and high calorie count per 100g. Bottom left signifies foods with low protein content per ca...

As a self-taught nutrition enthusiast , I have been contributing to the crowdsourced India database of Open Food Facts - "the Wikipedia of food", by adding and editing products. The decentralized open source project has data on more than 3 million products (as of April 2024) from countries around the world. Open Food Facts (OFF) data is released as Open Data: it can be reused freely by anyone, under the Open Database License ( ODBL ).  I've been cleaning up & compiling the data by food categories to extract insights . There are multiple ways to get the data : CSV or Excel format using the advanced search feature MongoDB daily export - more than 30GB uncompressed! JSONL daily export CSV daily export API As I was interested in select product features and macronutrient details, I decided to use the OFF API. This is how the process for compiling data for each food category currently goes: 1. I construct the following URL with the fields I require to fetch the nutrition ...

This Week I Learned -  *  Streamlit is an open-source app framework for Machine Learning and Data Science teams *  FoodData Central is an integrated data system that provides expanded nutrient profile data and links to related agricultural and experimental research. It provides nutrition data of ~500K branded food products. The FoodData Central API provides REST access to FoodData Central (FDC). USDA FoodData Central data are in the public domain and they are not copyrighted. They are published under CC0 1.0 Universal (CC0 1.0) * Gemini succeeds PaLM 2, the current foundation model from Google. * Internal tests showed that Q, Amazon's AI-powered assistant that enables employees to query documents and corporate systems, could leak confidential information from Amazon such as internal discount programs, unreleased features, and locations of AWS data centers. Amazon spokespeople called such scenarios hypothetical and denied that Q had leaked such information. Amazon ...

This Week I Learned -  * HTTP/3 builds on the foundations laid by HTTP/2 but introduces significant changes, primarily by shifting from TCP (Transmission Control Protocol) to QUIC (Quick UDP Internet Connections) as the underlying transport protocol. - The Valley of Code * From the Chrome Developer Tools Network panel, you can override HTTP response headers and web content , including XHR and fetch requests, to mock remote resources even if you don't have access to them or the web server. * A VSIX package is a .vsix file that contains one or more Visual Studio extensions, together with the metadata Visual Studio uses to classify and install the extensions.  *  Replicate lets you run machine learning models with a few lines of code, without needing to understand how machine learning works. *  Chatbot Arena lets you chat with any two models among these side-by-side: GPT-3.5: GPT-3.5 by OpenAI GPT-3.5-Turbo-1106: GPT-3.5-Turbo-1106 by OpenAI GPT-4-Turbo: GPT-4-T...