This Week I Learned - Week 8 2026

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 where only 17 billion of its 397 billion parameters activate per query, helping reduce compute costs while keeping performance high.

* Claude Sonnet 4.6 is now available across all Claude plans, including Claude Cowork, Claude Code, the API, and major cloud platforms. Sonnet is designed for everyday coding and enterprise tasks, while Opus handles more complex reasoning and analytical work.

* Yann LeCun criticizes LLMs for lacking true reasoning, planning, persistent memory, and physical world grounding—they're good at language but not intelligent like humans or animals.

His bets: Joint Embedding Predictive Architecture (JEPA) for self-supervised learning of hierarchical abstractions; world models that simulate real-world dynamics for better prediction and action.

Progress till now (2026): LeCun founded AMI Labs to advance world models; JEPA prototypes show promise in video prediction. DeepMind and others are racing on similar tech, with early integrations into robotics.

Future trajectory: Hybrid systems combining LLMs with world models for AGI, scaling via more multimodal data.

Limitations: High compute needs, incomplete physical simulations, early-stage—far from human-level understanding.

Progress on JEPA/world models (2026): AMI Labs (LeCun's venture) raised €500M for hierarchical JEPA; prototypes excel in video prediction and action planning (e.g., arXiv 2601.00844). DeepMind's Genie 3 enables real-time 3D sims; integrations in robotics show promise.

Incomplete physical simulations: Models approximate Newtonian physics but falter on quantum effects, chaotic systems (e.g., turbulence), or multi-scale interactions due to compute limits and simplified abstractions—leading to errors in complex scenarios like fluid dynamics or material behaviors.

* According to Google DeepMind chief executive Demis Hassabis, AI could reach artificial general intelligence within five to eight years. At the India AI Impact Summit 2026 in New Delhi, he highlighted limitations of current AI systems like inconsistency, lack of creativity, and true general intelligence—echoing LeCun's critiques of LLMs.

* Google's AI Food tool, Food Mood is a playful fusion recipe generator that creates unique recipes inspired by multi country cuisine. Select the type of food (starter, soup, main course, or dessert), number of people to serve, and two countries for cuisine inspiration. 

* ElevenLabs manages around 60,000 support calls for one client in English, Hindi and other Indian languages, and makes nearly 50,000 outbound calls a month for IDFC Bank.

* Sarvam Kaze is an AI-powered wearable spectacle can respond to inputs, interpret surroundings and record what users see. Custom experiences can be built for the device through the Sarvam platform. 

* The AI-powered app, Shwaasa, uses an algorithm-based platform to screen patients for chronic obstructive pulmonary disease (COPD), one of India's major causes of illness. Last year, AllMS-Delhi conducted a validation study on 460 people at its Ballabhgarh unit. When compared with spirometry— the gold standard for lung function testing — the tool showed moderate overall correlation and strong agreement in severe cases. 

Key hurdles in making spirometry widely available in India include high equipment and test costs, unaffordability for many patients, limited access in primary/rural care, shortage of trained technicians, time-consuming procedures, and challenges with patient cooperation and interpretation. 

False positive rates in AI health apps vary widely (5-30%) based on the app, condition, and study. For Shwaasa, a Nature study shows 75% specificity, implying ~25% false positives—meaning 1 in 4 healthy users might get a false alert. Apps like this are for screening, not diagnosis; professionals should confirm results. Data collection often aids AI improvement but can benefit pharma companies—review privacy terms.

* A McKinsey & Company analysis estimates that by 2030, around 70 per cent of all new data centre capacity will cater to AI-heavy workloads. 

Researchers at the University of California found that after generating a simple 100-word email, an AI chatbot’s cooling systems consume between half and three bottles of water (200–1,500 millilitres), depending on local conditions. Training these models requires exponentially more: Meta reportedly used 22 million litres of water to train its LLaMA-3 model. 

As India races to power its AI revolution, massive data centres are being built in areas already struggling for water, raising fears that the country’s digital boom could deepen local water crises - Down To Earth 

* "AI is a big growth booster for IT services, not a cost-cutter. Coforge now integrates AI in all new projects, shifting from trials to real-world delivery. Despite lower renewal rates (30-50% discounts), tech spending is rising, creating opportunities." - Sudhir Singh, Coforge

* Elon Musk - Flight code for the SpaceX rockets and Starlink satellites is written in C and C++. 

Python is used where runtime performance is less important than rapid iteration and ease of use.

* A list of physical visualizations and related artifacts 

* Hyderabad Urban Observatory from Hyderabad Urban Lab is envisaged as a set of projects around:

• data repositories, archives and libraries;
• research and advocacy;
• physical and social interventions around communities and places.

* The Eternal Group includes Zomato, Blinkit Quick-Commerce, Blinkit Ambulances, District, Nugget, and Feeding India.

* Galgotia started as a book store on Ansari Road, Daryaganj, in Delhi in 1952. It branched into publishing of engineering textbooks and other competitive exams eg., business management.  It was allotted 52 acres of land in Greater Noida in 2011 to set up a University by the Uttar Pradesh Government when Mayawati was the Chief Minister. 

* Pharmacognosy is the oldest branch of pharmacy that studies natural medicines from plants, microbes, and animals. 

* Colonel Sir Ram Nath Chopra (1882 – 1973) – aka the 'Father of Indian Pharmacology' kickstarted the experimental validation of Ayurvedic and traditional remedies. He didn't dismiss traditional Ayurveda or Unani medicine, but he didn't blindly follow it either. He was a bridge-builder.

One of his most famous works was on Rauwolfia serpentina (Sarpagandha). While it had been used for centuries for "insanity" and snake bites, Chopra scientifically proved its effectiveness in treating high blood pressure.

He spent his life proving that nature is a pharmacy, but only when handled with chemistry. 

His significant work includes the book "Indigenous Drugs of India: Their Medical and Economic Aspects" published in 1933, which discusses the medical and economic aspects of indigenous drugs in India. 

* Affordable smartphones, low-cost data and widespread 4G connectivity have dissolved the tyranny of distance. Small manufacturers, traders, artisans and service providers in Tier-3 and Tier-4 towns can now access national marketplaces through platforms such as e-commerce portals, social commerce and WhatsApp-led selling.  The authors of The Power of Tier 3 and 4 Cities - Gateway to $10 Trillion Economy estimate that most of the 7,935 towns in the country are Tier-3 or Tier-4 towns.

* Defence spending now accounts for 14.67% of the Union government’s total expenditure — the highest among all Ministries. 

* The Union Budget has increased allocations for the AYUSH Ministry by 20 per cent. Critics led by the Indian Medical Association (IMA) have argued that traditional systems like Ayurveda often lack the rigorous, empirical evidence required to meet the thresholds of modern medicine. Allopathic drugs must clear randomised controlled trials that test their safety and efficacy. Many AYUSH treatments however are based on ancient texts and observational history, and either haven’t been tested or whose specifics remain unverifiable. Perhaps the most controversial aspect of AYUSH has been “mixopathy”, the state-sanctioned blurring of lines between traditional and modern medical practice. In 2020, the Central Council of Indian Medicine authorised postgraduate Ayurveda students to be trained in 58 surgical procedures, including general surgery. The IMA called this “legalised quackery”, arguing surgery is a complex discipline requiring deep knowledge of anatomy, anesthesia, and perioperative care that AYUSH curricula don’t sufficiently cover.  - The Hindu

* India’s new free trade agreement (FTA) with the European Union has opened the door for Indian doctors and products to enter the European market more easily. The FTA recognises India’s Traditional Knowledge Digital Library, a database of formulations, thus preventing companies from wrongly claiming ownership of traditional Indian remedies.

* "Success in life is not how much money you make, it is how little money you need" - Dr. Arokiaswamy Velumani