The Daily Brief
Friday 20 November 2026  ·  covering Thursday  ·  a 13-minute read
5 Things to know
1  Coal India is selling a tenth of its biggest mine company

The Mahanadi Coalfields IPO opened on November 19th and closes on the 23rd: up to ₹8,981 crore at ₹129–136 a share, which values the company at about ₹90,000 crore at the top of the range. It is entirely an offer for sale, so the money goes to Coal India, not to MCL. The book was 38% subscribed on day one. MCL mines Odisha’s Talcher and Ib Valley fields and produced 218.3m tonnes in 2025–26, about 21% of India’s coal, down from 225.2m the year before.

Context  MCL does not set most of its own prices. In 2025–26, 82% of its sales went through long-term fuel-supply agreements at prices Coal India notifies; the rest goes to e-auctions, where prices float. That is why the prospectus shows dispatches rising from 199m tonnes in 2023–24 to 211m in 2025–26 while profit fell from ₹11,842 crore to ₹10,698 crore. Finshots puts the shrinking auction premium at the centre of that drop.

2  Postgres 19 shipped two months late, because AI found too many old bugs

PostgreSQL, the open-source database behind a large share of the web, normally releases in September. Version 19 came out on November 19th, delayed by what Simon Willison calls “the ongoing AI-reported vulnerability apocalypse”. Robert Haas, a long-time committer now at Databricks, counted the cost. Between April 8th and September 25th, commits back-patched to fix bugs in already-released versions rose by more than 60% on the year before, while commits to the development branch fell slightly. Three large features were pulled between August 27th and September 15th.

Context  The bottleneck is human review: AI tools can find a bug and even draft its fix in minutes, but a committer still has to check the fix. The release’s headline feature, WAIT FOR LSN, lets a read-only copy of the database wait until it has caught up with a given write. That gives “read-your-writes” consistency, which Wikipedia’s software has built by hand on MySQL for years.

3  Mira Murati’s lab has released a laptop-sized model that thinks in loops

Thinking Machines Lab, founded in 2025 by OpenAI’s former chief technology officer, released Inkling 2 Mini with open weights under an Apache 2.0 licence. It has about 26bn parameters in total but uses only about 3bn for each token, and it accepts a million tokens of context. Its 24 stored layers act like 48 because the middle 12 run three times in a row, reusing the same weights. Artificial Analysis says it is small enough to run on a MacBook. Hugging Face’s early tests put it roughly level with Alibaba’s Qwen3.8-27B.

Context  Looping buys depth of reasoning without more memory. On the same day, on the a16z Podcast, Nick Bostrom named looped transformers as the kind of capability gain that is costly for safety. A looped model does more thinking between the outputs humans can read, so less of its reasoning is visible. Bostrom suggested labs might agree to compete on other dimensions instead.

4  Haematologists have doubled the threshold for iron deficiency

The American Society of Hematology’s guideline on diagnosing iron deficiency, published in September, suggests that adults count as deficient when blood ferritin is at or below 30 nanograms per millilitre, not 15. That includes people who menstruate or are pregnant. For children aged nine months to four years the threshold rises from 12 to 20. NPR’s Life Kit covered it on Thursday. The doctors it spoke to said ferritin is not checked at routine visits, and that home finger-prick tests are not validated.

Context  Ferritin is the protein that stores iron, and its blood level stands in for the body’s reserves. It also rises with inflammation, which can hide a deficiency. For adults with inflammation the panel therefore suggests ferritin below 100 or transferrin saturation below 20%. The panel rates the evidence behind the new thresholds as low-certainty, but raising them means more people will be diagnosed.

5  Fairfax is buying IDBI Bank

Fairfax Financial, the Toronto insurance and investment group run by Prem Watsa, has been chosen to buy 60.72% of IDBI Bank for about ₹52,878 crore ($5.6bn), or ₹81 a share against a previous close of ₹90.39. It plans to fold its Go Digit insurance and IIFL Capital holdings into the bank, which needs further approvals. To help pay, it is selling its whole stake in IIFL Finance.

Context  IDBI began in 1964 as the state’s industrial-development lender and became a bank in 2004. After heavy bad loans, LIC took control in 2019. The cabinet approved selling it in 2021, and the stake on sale belongs to the government and LIC. For scale, Tata bought Air India, India’s last big privatisation, for ₹18,000 crore in 2021.

4 Lines worth keeping
“One could equally maintain that if nobody builds it, everyone dies. In fact, most people are already dead.”

It answers Eliezer Yudkowsky and Nate Soares, whose 2025 book If Anyone Builds It, Everyone Dies calls for a global ban on advanced AI. Bostrom’s frame is “not like playing Russian roulette; it is more like undergoing risky surgery for a condition that will otherwise prove fatal.” The same argument featured on Thursday’s a16z Podcast. The paper’s numbers: 170,000 people die every day, and if mortality stayed at the rate of a healthy 20-year-old, life expectancy would be about 1,400 years.

“Every feature that we write from now until the end of our careers will be scrutinized by AI.”

The sentence continues: things that “would have escaped forever unnoticed in an earlier era will now be found, perhaps in a matter of hours.” Earlier the post gives the cause bluntly, “The short answer is that LLMs happened”, and says the bugs were always there: “What’s different now is that it’s become vastly easier to find those bugs.” Haas calls the delay a sign of care for quality, not a worry. The open question is how the project pays down decades of old bugs and still builds new things.

“The world is a sacred vessel. It cannot be worked on. Whoever tries to work on it spoils it. Whoever grips it loses it.”

The essay offers it after a trip to Shanghai, where young workers describe “996”, nine to nine, six days a week, “as if it were the weather”. The essay’s reading: much modern fatigue comes from treating life “as a project to optimise”. Against it the essay sets the old people writing calligraphy on the pavement in water, which evaporates within minutes. “The practice was never meant to leave a trace.”

“Sovereignty doesn’t mean isolation, it means optionality.”

Apellund, whose Finnish firm processes card payments, was quoting a line heard at a conference in Amsterdam, in reply to Tom Borgers of Walapay, who had warned against swapping many non-European payment systems for one European one. Optionality here means several interchangeable rails accepted everywhere. Later Apellund adds the condition that makes it hard: “if they are not accepted anywhere, it makes no sense.”

3 Arguments, and the case against
Was going direct Ola’s mistake?

For  Ola Electric led India’s electric-scooter market in 2024 by selling like a phone maker, through its own stores. It has now signed 255 dealers running 319 stores and is closing most of its own. Zerodha’s Daily Brief lists what a dealer does better. Dealers pay for stock when they order, so unsold scooters stop tying up Ola’s cash. A scooter carries 5% GST but many of its parts carry more, and the refund of the difference is easier to claim once the sale is booked when the dealer buys. And a dealer with money in the workshop loses income while scooters wait. Ather, which expanded through dealers, now registers more than twice as many scooters. It made its first quarterly profit, ₹1.26 crore, in July–September, the same quarter in which Ola lost ₹362 crore on revenue down 39% to ₹421 crore.

Against  Tesla has always sold its cars directly, and kept doing so as it grew. What hurt Ola was execution: the brief itself grants that product quality explains much of the gap. The stores also brought legal trouble. In April SEBI issued a show-cause notice alleging that Ola’s December 2024 announcement of 4,000 locations with service centres was misleading. Ola has applied to settle without admitting anything.

Test  Ola’s monthly registrations against Ather’s from January, once the dealer network is fully running.

Does Europe need its own payment rails?

For  On Fintech Insider, Victor Mithouard of Mambu, a cloud core-banking supplier, said Europe has been “digitally colonized” by American infrastructure. The question is not only who controls a system but whether you can be cut off from it. Mithouard pointed to Brazil, whose central bank’s Pix (2020) took everyday payments away from the card networks, and the host added India’s UPI. Europe’s answer is linking its national wallets. In February EPI, the bank-owned company behind the Wero wallet, agreed with Bizum, Bancomat, MB Way and Vipps MobilePay to connect them through a central hub.

Against  Borgers, on the same panel: “if we’re just replacing bunch of non-European payments with one European system, well, then all of a sudden we have another single point of failure.” Mithouard conceded that Europe’s sovereign clouds lack the depth and funding of the American hyperscalers. Today, choosing sovereignty can mean accepting less resilience. Brazil and India each built one national system, which is the model Borgers warns about.

Test  Whether a Wero user can pay a Bizum merchant in Spain by the end of 2027, and what share of European card spending moves.

Should we pause before superintelligence?

For a short pause only  Bostrom models the question for people alive today, weighing safety progress, discounting and quality of life. For many settings, the best course is to move quickly to AGI capability and then pause briefly before full deployment: “swift to harbor, slow to berth.” On the podcast Bostrom gave the logic. If safety improves fast during a pause, a few months is enough. If it improves slowly, waiting costs more lives than it saves. Only in between does a longer wait pay. The bigger fear is that a pause built on stigma hardens into never building it at all, “itself… a kind of an existential catastrophe”.

Against  The model counts only people alive now. A LessWrong critic calls that “person-affecting” stance under-justified, since it gives no weight to everyone who might come later. Bostrom’s own paper says the analysis “does not on its own imply support for any particular policy prescriptions.” Yudkowsky and Soares reject the premise that safety research can keep pace at all.

2 Models to think with
The appropriability problem

An invention is information: once it exists it is cheap to copy, so the inventor captures only part of its value and markets spend too little on invention. Patents and public funding are the standard fixes. Patents work for drugs, because a company can own a molecule and charge for it. They fail for commodity sellers, and Zerodha’s line gives the reason: “A power company can’t charge more for a better electron.” A barrel or a kilowatt-hour sells on price, many utilities earn returns set by regulators, and a better grid helps every rival.

Today. Rich-country governments spent about 0.1% of GDP on energy research after the 1970s oil shocks. The IEA says they spend about half that now, and estimates that global public spending fell 2% in 2025, to about $55bn. The research that lasts now comes from equipment makers that can own what they invent. BYD spent ¥54.2bn on R&D in 2024, more than its net profit.

Where it breaks. The model says who underinvests, not who invests instead. In energy that turns out to be whoever owns the factories, and that is mostly China. Watch whether India’s subsidised cell and module plants start spending on research, or only on capacity.

Normal accidents

Perrow, a Yale sociologist, wrote it after Three Mile Island. In a system that is both complex, with interactions no one can see in advance, and tightly coupled, with no slack between steps, accidents are not anomalies but normal. Adding safety devices can make things worse by adding complexity. The prediction held on 19 July 2024, when one faulty CrowdStrike security update crashed about 8.5m Windows machines and stopped airlines, hospitals and card terminals. The dependency they shared was one most of them had never mapped. It fails for “high-reliability organisations” such as aircraft carriers and air-traffic control, which run complex, tightly coupled systems with very few accidents.

Today. On Fintech Insider, Borgers described European money movement running through “a very constrained set of banks”, naming ClearBank and Banking Circle. E-money firms sit on top of them, sometimes several layers deep, so one bank’s outage has a wide “blast radius” that is “very hard to get visibility on until it’s really too late”. Apellund: the worst risks are the shared dependencies “that we don’t know”. The costly remedy is a second, independent system. Monzo’s Stand-in runs on Google Cloud while its main bank runs on AWS, with 18 services to the main platform’s 30,000-plus.

Where it breaks. Perrow’s remedy for the worst systems was to abandon them, and nobody can abandon payments. The workable version is loosening the coupling: a separate backup, offline fallbacks, cash. Watch the EU’s Digital Operational Resilience Act, in force since January 2025, which lets supervisors oversee the technology firms that banks share.

1 To read in full
Kashish & Krishna, The Daily Brief by Zerodha, 19 November · 14 minutes

Fifty years of government energy-research budgets, company accounts and Indian budget papers, tested against this year’s Hormuz shock. The finding: governments spend on energy research when they are frightened and stop when the fright passes. The spending that lasts comes from firms selling batteries and solar panels, which are mostly Chinese. It is the Arrow model above with the data filled in. If you have ten minutes, read “Why energy companies barely do research” and “Today’s research follows factories”. The India section is the weakest, by the authors’ own account: there is no official figure for energy research, so they build one from a single study and budget documents.

Read the piece →

Also in the reading

The rest of the Taoism essay draws two more keys from the Tao Te Ching. Chapter 63 describes the sage as one who “tastes what has no taste”, which the essay reads as retraining attention toward plain things. Chapter 80 describes villagers who can hear the next village’s roosters and never feel the need to go there. The conclusion is modest: none of this asks you to leave the city, only to keep one hour of the day for yourself.

Tuning a model to its harness. Hugging Face trained small LoRA adapters with reinforcement learning inside three coding-agent harnesses: the scaffolding of tools and prompts a model works within. Under the pi harness, Qwen3.8-27B’s score on Terminal-Bench 2.1 rose from 39% to 55% while average tokens per task fell from 2,900 to 1,500, mostly because it reasoned less and called fewer tools. The three adapters also merge into one that works under all three.

Sources
The Daily Brief by Zerodha, “Who pays to invent the energy future?” and “Ola finally lets the middleman in”
Finshots, “The Mahanadi Coalfields IPO Explained”
11:FS Fintech Insider #1117, “Sovereignty vs. resilience”
a16z Podcast, “How Fast Should We Build Superintelligence?” with Nick Bostrom
Words of Taoism, “Is Taoism Too Slow for Modern Life?”
NPR Life Kit, on iron
Simon Willison, “PostgreSQL 19 Released”
Hugging Face, “Welcome Inkling 2 Mini”
Hugging Face, “Introducing harness adapters”
Chittorgarh, Mahanadi Coalfields IPO details
Chittorgarh, where MCL stands among coal producers
SEBI, MCL prospectus filing
ETV Bharat, MCL subscription status
Robert Haas, “Postgres 19 is late, and it’s AI’s fault”
Artificial Analysis, on Inkling 2 Mini
American Society of Hematology, new iron-deficiency standards
ASH 2026 guidelines for diagnosis of iron deficiency, Blood Advances
Ascendants, the IDBI Bank sale explained
Whalesbook, Fairfax to acquire 60.72% of IDBI Bank
Nick Bostrom, “Optimal Timing for Superintelligence” (working paper, 2026)
Yudkowsky & Soares, If Anyone Builds It, Everyone Dies
LessWrong, a critique of the person-affecting stance
Inc42, Ola Electric’s second quarter
Business Today, Ola’s ₹362 crore loss
CarToQ, Ather against Ola
Acko Drive, Ola’s SEBI settlement application
Pix, Brazil’s instant-payment system
Bancomat, EuroPA and EPI agree a pan-European hub
Kenneth Arrow, “Economic Welfare and the Allocation of Resources for Invention”, NBER, 1962
IEA, energy innovation
China Daily, BYD’s 2024 results
Charles Perrow, Normal Accidents
Microsoft, on the CrowdStrike outage
High-reliability organisations
Monzo, “Tolerating full cloud outages with Monzo Stand-in”
EIOPA, the Digital Operational Resilience Act
Read in full, Thursday 19 November: nine pieces. The 11:FS, a16z and Life Kit episodes were machine-transcribed. Skipped: The Ken and a Stratechery interview (paywalled), and an a16z fellowship advertisement.