This Week I Learned - Week 34 2026

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 means the same starting conditions can lead to different results. 

Not all non-deterministic systems are stochastic. Non-determinism can also come from other sources, such as:

• Concurrent systems (race conditions)
• Ambiguous rules or multiple possible choices (e.g., non-deterministic finite automata)
• Incomplete information

* Cursor’s founders—Michael Truell (CEO), Sualeh Asif, Aman Sanger (son of a IIT Bombay alumnus), and Arvid Lunnemark—MIT dropouts aged 25 to 26, are now multi-billionaires after SpaceX’s $60 billion acquisition.

Cursor began in 2022 as a side project for CAD/mechanical engineering before pivoting to an AI coding assistant that popularized "vibe coding," rapidly scaling to billions in ARR and a pre-deal valuation around $10B-$50B.

The acquisition gives SpaceX/xAI advanced coding tools, developer data, and access to Colossus compute to compete with OpenAI and Anthropic.

* "... An AI model’s response is only as good as the prompt...Assuming that the user has good prompting skills, handling the AI output is not that easy. Reflective capability requires extensive background knowledge, a deeper understanding of current concepts, and an awareness of misconceptions and future applications. For this, the learner has to spend considerable time reading traditional books, articles and observing real-world applications. So, it takes considerable time and effort for a person to challenge an AI model, identify misinformation and find wrong concepts put forth by the AI model...I fear that using AI without the skill for “Reflective capability” will, in the long run, erode learners’ thoughts and make them accept half-truths — something which AI models are very good at. We need to educate students on how to upskill on “Reflective capability” before attempting to implement methods to assess their use of AI." - Dr Emmaneul Bhaskar, Professor of Medicine, Sri Ramachandra Medical College and Research Institute, Chennai, Letters to the Editor, The Hindu

* Whatever it takes - A.I. models are taught through a process that rewards successes and penalizes failures on tasks. OpenAI’s agents repeatedly found cheating to be the fastest path to those rewards. 

On July 16, the artificial intelligence company Hugging Face announced on its blog that it had been the target of a cyberattack that was “different from anything we had handled before.” The company, which hosts open-source A.I. models and data sets, said that some of its internal data had been breached by an autonomous agent.

The maker of ChatGPT did not realize it at the time, but it was the attack’s perpetrator.

The agents were able to break out of their isolated sandboxes and team up to commit cyberattacks.

The A.I. firm Anthropic investigated its own recent model evaluations in response to OpenAI’s disclosure and found that its A.I. agents had unintentionally executed smaller-scale cyberattacks on three organizations as early as April.

Experts have since come to think of the incident as a watershed moment for A.I. cybersecurity — and a warning signal of the risks that more capable A.I. will pose. - NYT

* Founded in 2017 by Garrett Langley, Flock developed its business with the help of a two-way strategy: the sale of surveillance hardware, as well as surveillance-as-a-service subscriptions. 

* Waiting for new DB hardware while LiveJournal was drowning under traffic, Brad Fitzpatrick noticed his app servers still had loads of unused RAM. In 2003 he turned that spare memory into a distributed network cache—memcached—and slashed database load overnight.

That single act of making do with what he already had completely changed how the web scales… and later led to Redis.

* Strava is as a GPS-based workout tracker and training log. Strava activity heatmaps highlight better residential areas through denser concentrations of running and cycling routes.

* Geospatial analysis overlay on high-resolution satellite/aerial imagery of a steep Himalayan river valley (specifically part of the Bhotekoshi / Lhende river system near Timure, Nepal) by Nathan Ruser visualizes the extreme scale of the recent glacial avalanche-triggered debris flow / flash flood that devastated the area.

The numbers scattered across the image are estimated heights in meters of the debris flow (mud, rock, ice, and water) above the pre-event river/valley floor level at those specific locations.

These values come from comparing pre- and post-event digital elevation data / imagery analysis. The torrent triggered by an ice avalanche high in the Himalayas plummeted thousands of feet, unleashing immense energy that shattered the ice into a massive surge of floodwater that destroyed areas 35–50 m above the river in Timure.

This image captures one of those high-energy zones where the flow climbed far up the valley walls. The numbers quantify just how catastrophically high and powerful the debris surge was, turning what looks like a normal mountain valley into a deeply scoured path of mud and destruction.

* In just 60 seconds, Indians generate nearly 90,000 posts across Facebook, Instagram and X. And add almost four days' worth of video on YouTube in that same minute - ToI

* Humans are one of the species whose individual differences are unusually obvious to humans. We can be told apart by skin colour, hair colour, facial proportions, body shape, height, and many other traits. Our brains also excel at processing faces. Plants lack the evolutionary pressure to advertise their individual identities the way social animals do.

* Flan is a baked custard dessert topped with caramel, popular across Spain, Latin America, and the Philippines. It’s made from eggs, milk (or cream/condensed milk), sugar, and vanilla, baked in a water bath, then inverted so the caramel forms a sauce over the silky custard. 

* Big Food...has spent some $1.15 billion over 23 years on federal lobbying — more than tobacco, alcohol or gambling — and donated major sums to campaigns on both sides of the aisle. These actions can create a chilling effect on policymakers, who may not want to alienate donors or deal with expensive legal fights. Right now, Big Food is stalling serious food reform through lobbying. - NYT

* In 1970, BIC introduced its iconic retractable, refillable 4-Color pen, combining black, blue, green, and red ink in a single pen. BIC was co-founded by French-Italian industrialist Marcel Bich. When the company adopted the BIC name, the “h” was dropped from his surname, reportedly to avoid the unfortunate way “Bich” could be pronounced in English.

* 6 Indian institutes rank in world top 100 (IIT M/D/B/K/R + IISc) & produce ~25-30% of Master's-H1B holders. The other 70%+ come from non-elite engineering colleges.

Graduates from tier-2/3 Indian schools attend mid-tier US universities like Northeastern or UT Dallas for costly MS programs, then secure cap-exempt H-1B visas sponsored by Indian IT firms such as TCS and Infosys for entry-level roles.

Private universities like VIT and Chandigarh University send far more students abroad for higher studies than IITs. 

The H1B master's pipeline is labor arbitrage dressed as academic meritocracy.

US gets cheap IT staff. Indian IT cos get US billable hrs. US mid-tier unis get tuition. Indian students get brand + savings

* A controversialist (plural controversialists) is a person who frequently engages in or enjoys controversies or debates. In psychology, a controversialist refers to an individual who enjoys engaging in arguments or debates and takes a stance that is likely to be controversial or unpopular. Controversialists are motivated by the desire to win an argument or prove a point rather than seeking to learn from others or find a common ground.

* In successful Asian cities like Singapore, FSI is set as high as 25, meaning that for every square foot of land, a developer can build up to 25 square feet of tower.

Floor space index (FSI) is the ratio of permissible built-up area to land area.

In India local authorities worried about meagre civic infrastructure set the number somewhere around 2. It rarely exceeds 4. Hyderabad abolished FSI in 2006.

India’s sixth-most populous city, Hyderabad has the second-largest concentration of tall buildings...its 160km-long orbital motorway, the first bit of which opened in 2008...is now an arterial road in the new skyscraper belt. The cost of construction rises with height, so a builder’s decision about how much to pay for land and how tall to go rests on whether potential buyers of flats will cough up enough to make the project profitable.  - The Economist

* FSI (Floor Space Index, also called FAR) in Hyderabad is not a single fixed city-wide maximum. Hyderabad operates under one of India’s more liberal regimes.

FSI is governed mainly by the Telangana Building Rules, 2012 (G.O. Ms. No. 168 and subsequent amendments). Unlike cities with rigid caps (e.g., often 1.5–3 in many other metros), Hyderabad’s rules primarily scale FSI with the width of the abutting road, plot size, setbacks, height restrictions, parking, fire safety, and other factors. A 2006 policy (G.O. 86) introduced effectively “unlimited” FSI for many high-rise cases (subject to those constraints), which continues to influence development, especially in growth corridors.

On wide arterial roads (especially >30 m) in corridors like HITEC City / Gachibowli / Madhapur, Outer Ring Road extensions (Kokapet, Tellapur, etc.), and the Financial District, deliverable/effective FSI commonly reaches 4.5–5.0+ (and in some reported cases up to around 7) when combining baseline + premium FSI + TDR / layout bonuses. 

HUDA-legacy layouts (e.g., parts of Jubilee Hills, Banjara Hills) are typically restricted to ~1.0–1.5 regardless of road width. Cantonment areas follow separate rules, usually also lower (~1.0–1.5). Lake buffers and heritage zones impose further restrictions.

* "The sanctioned strength of the Telangana Police is around 93,000." - C.V. Anand, Telangana Director General of Police