Reading
Newsletter · Latest edition
Sunday 22 November 2026
Tonight
Tonight, we look at the end of COP31 in Antalya, which broke up on Sunday without adopting a final decision, and at the fault lines that sank it. Then South Korea, where AI cheating in exams has cost the education minister his job, markets losing their nerve over AI, and the French prefect suspected of passing information to the far right. We also explain perplexity, the measure OpenAI wants AI labs to share. And three finds, including a question: do the dead have rights?
Tick subjects to build a research prompt to paste into the assistant of your choice.
Today’s focus
01 · CLIMATE · COP31 ends without a deal: what broke, and what is left
After almost 44 hours of overtime, the UN climate conference in Antalya, Turkey, ended on Sunday 22 November without adopting its final decision, the political text that sums up each COP, known as the “cover decision”. “It seems there are parties opposed to this text. I conclude that there is no agreement; these decisions are not adopted,” said COP31’s Turkish president, Murat Kurum, as quoted by franceinfo. The European Union rejected the Turkish and Australian presidencies’ last draft because it no longer mentioned fossil fuels: “We will not endorse a cover decision that does not even mention fossil fuels, the driver of climate change,” said Ireland’s climate minister, Darragh O’Brien, speaking for the EU, according to DW. Panama and the Marshall Islands also objected. According to Al Jazeera, it is the first time since Copenhagen in 2009 that an annual COP has ended without an agreement; according to franceinfo, citing the Earth Negotiations Bulletin, it was also the longest on record, overtaking COP25 in Madrid in 2019.
The idea
A COP decides by consensus: one group of countries formally objecting is enough to sink a text. Antalya pitted three camps against each other, for different reasons. Producer countries, among them Russia and Saudi Arabia, got fossil fuels removed from the text: Russia, Al Jazeera reports, argued that cutting oil and gas supplies would hurt the poorest. The EU, Panama and small island states refused that step back. But many developing countries accused the EU and other rich nations of blocking the money: according to DW, the rejected text offered only a “pathway” to triple adaptation finance by 2035, with no figure, while noting needs of some $300bn a year; the $120bn target in earlier drafts had gone, Al Jazeera notes. “You can’t ask developing countries to transform their economies while they’re drowning in disasters and debt,” said Harjeet Singh of the Fill the Fund campaign. Pakistan and Ghana, for their part, defended the text as a necessary compromise.
Did you know?
According to Al Jazeera, the rejected text would have been the first UN document to acknowledge that warming is on track to exceed 1.5C, the Paris Agreement threshold. It also called for raising electricity’s share of final energy demand from about 20% to 35% by 2035. That goal survives outside the negotiations, as a voluntary pledge: signed by about 75 countries including China, according to DW, or by more than 90, according to the Australian co-president, Chris Bowen, quoted by Al Jazeera. It does not say whether that electricity must be clean.
Further reading
Murat Kurum said no new text would be negotiated and that the presidency would stick to existing decisions. COP32 will be held in 2027 in Addis Ababa, Ethiopia, whose foreign minister, Gedion Timothewos, has promised to “rebuild trust”.
Sources
- franceinfofranceinfo.fr
- DWdw.com
- Al Jazeeraaljazeera.com
The essentials
02 · SOUTH KOREA · AI cheating costs the education minister his job
South Korea’s education minister, Choi Kyo-jin, resigned on Sunday, the BBC reports, citing the Yonhap news agency: he said he took responsibility for the confusion caused by last-minute measures against AI-assisted cheating in the Suneung, the national university entrance exam sat on Thursday by about 500,000 candidates. President Lee Jae-myung accepted his resignation. It began with a confession: Han Won-seok, top scorer in April’s national tax accountancy exam, admitted photographing the questions with a miniature camera, sending them to an AI and reading the answers on his phone. The government then tried something new: to flush out hidden devices, it set off the national disaster alert system three times on exam day, making every switched-on phone in the country ring, even on silent. According to the ministry, 111 candidates were found with unauthorised devices. The scandal has reopened a deeper debate, the Korea Herald notes: what does a written exam measure when an AI can answer its questions? One of the Suneung’s designers, Park Do-soon, argues that a written test assesses only “about one-tenth” of the curriculum’s goals.
Sources
- BBCbbc.co.uk
- The Korea Heraldkoreaherald.com
03 · ECONOMY · AI rattles Wall Street, and Trump wants data centres kept at home
The week ended with a sharp fall on US markets: on Friday 20 November the tech-heavy Nasdaq lost 5.3%, and the S&P 500 and the Dow Jones entered “correction”, meaning at least 10% below a recent high, the BBC reports. The analysts it quotes point to fears of a bubble: a handful of big companies spent more than $1tn on AI in a year, and investors are starting to demand a clearer path to profit. Anthropic’s shares, which began trading on Thursday, took the hit: priced at $83, they closed their first day at $97, then lost 8.5% on Friday. The company, which raised $81bn, is not yet profitable, DW notes. The same day, Donald Trump threatened a “tremendous tax” on US companies that build data centres abroad, “because China will take all of them”. He acknowledged his own team was hearing the idea for the first time; how it would be enforced is unknown. Meta, the BBC notes, has already broken ground on one in Alberta, Canada.
Sources
- BBC · marketsbbc.co.uk
- BBC · data centresbbc.co.uk
- DWdw.com
04 · FRANCE · Suspected of leaking to the far right, a senior prefect asks to step aside
Hugues Moutouh, the prefect of the Centre-Val de Loire region, the state’s top representative there, said on Saturday 21 November that he had asked to be relieved of his duties. Secretary general of the French interior ministry from April 2025 to May 2026, he is suspected of passing confidential information to the far-right Rassemblement national (RN) about investigations into the party. According to franceinfo, which confirms a Le Monde report through a source close to the case, an email from the secretariat he ran was found in Rome during a search, carried out for a European Public Prosecutor’s Office investigation, at the home of Frédéric Chatillon, a former leader of the far-right student group GUD and a long-standing friend of Marine Le Pen. Paris prosecutors opened a judicial investigation at the end of July into breaches of professional and investigative secrecy and handling the information obtained; the RN’s headquarters was searched on 9 November. Moutouh denounces a “destabilisation campaign” and says he acted “in strict compliance with the law and with absolute political neutrality”. Marine Le Pen says she “never received confidential information”.
Sources
- franceinfo · request to step asidefranceinfo.fr
- franceinfo · the email foundfranceinfo.fr
Explainer
05 · A language model’s perplexity
Why it's in the news. On Sunday, OpenAI’s chief scientist, Jakub Pachocki, published the company’s internal perplexity measurements for the first time, and proposed making them a metric shared by every lab, even the basis for “common limits” labs would commit not to cross until our ability to understand and monitor AI behaviour improves.
What it is. A language model predicts the next “token”, a fragment of a word. Perplexity measures how surprised, on average, it is by the actual continuation of a text it has never seen: technically, it is the exponential of the average negative log-probability it assigns to the right tokens, as Hugging Face’s documentation explains. A perplexity of 2 means the model hesitates, on average, as if choosing between two equally likely options; the lower, the better the prediction. OpenAI publishes the logarithm of that value: about 0.99 for GPT-3.5, 0.91 for GPT-4, and 0.76 for the model from which, according to Pachocki, new mathematical results began to appear (a perplexity of about 2.1). His thesis, which remains his own, is that new capabilities “emerge” as this number falls. He acknowledges its limits himself: the figures are measured on OpenAI’s internal code, so they are hard to compare across labs; they depend on how text is split into tokens; and step-by-step reasoning improves capabilities without lowering perplexity.
Sources
- OpenAI · Jakub Pachockiopenai.com
- Hugging Face · Transformers documentationhuggingface.co
Technology & science
06 · HEALTH & ENVIRONMENT · Weight-loss drugs could free up farmland
What would happen to the planet if GLP-1 agonists, the appetite-suppressing drugs such as Ozempic and Wegovy, were widely used? A study published on 19 November in Nature Communications simulated it with a model of the global food system, to 2050, for people living with obesity. If patients simply ate less, the model estimates up to 6.6 million deaths avoided each year by 2050 and 32 million hectares returned to natural land, slightly more than the area of Italy, through lower demand for cropland, pasture and managed forest. Combined with a healthier diet, the one proposed by the EAT–Lancet commission, the estimate rises to 7.7 million avoided deaths and 46 million hectares. But the study also finds a rebound effect: lower demand pushes prices down, which leads people not taking the drugs to eat more, including unhealthy food. These are model scenarios, not observations. The authors, Roslyn Henry and Peter Alexander, stress in The Conversation that the drugs should not become a technological fix that lets the food environment escape scrutiny.
Sources
- Nature Communicationsnature.com
- The Conversationtheconversation.com
07 · OPEN SOURCE · LLVM closes an eight-year relicensing effort
LLVM, the sprawling open-source compiler infrastructure, has declared its relicensing effort, begun in 2018, finished, Kristof Beyls announced on the project’s forum. Since 19 January 2019, new contributions have come in under Apache-2.0 with an “LLVM exception”. For older code, released under the previous licences (UIUC/NCSA, plus MIT for some runtime libraries), each author’s permission was needed. Today more than 99.5% of all lines ever contributed are either under the new licence, too small for copyright to apply, or no longer in the repository. Rather than “chase the last tiny fraction forever”, the project is stopping; pre-2019 contributors who never signed can still ask for their code to carry the appropriate licence identifier. It is a rare concrete measure of what changing the licence of a large collective project costs in time and coordination.
Sources
- LLVM Discourse · Kristof Beylsdiscourse.llvm.org
- LWN.netlwn.net
In brief
- The FBI is investigating an apparent breach of the NCIC, the criminal database it shares with local police for warrants and background checks, according to US media cited by the BBC.
- New York’s mayor, Zohran Mamdani, has unveiled a plan to close the Rikers Island jail complex by 2032, build four new jails and cut the detainee population from about 6,500 to 4,100, the BBC reports.
- Meta is “re-evaluating” its 2030 net-zero goal after its emissions rose from 15.6 to more than 21.7 million tonnes of CO₂ equivalent between 2024 and 2025, according to its sustainability report as read by Ars Technica.
- Codex, OpenAI’s coding agent, reaches version 0.171.0, with GPT-6.2 Sol as the default model and a command to run non-interactive sessions on an existing app server, according to its release notes.
Worth a click
08 · Do the dead have rights?
A Glasgow museum has taken an Egyptian priest’s mummy off display on the grounds that he never consented to being exhibited. In Philosophie Magazine, Octave Larmagnac-Matheron uses that contested decision to ask a real question: can you wrong someone who no longer exists? He shows that the law mostly protects the dead through obligations placed on the living (“the respect due to the human body does not end with death”, says the French civil code) rather than through rights they would hold, then surveys the philosophical positions. A short read for a November evening, in French.
Sources
- Philosophie Magazinephilomag.com
09 · Before you criticise, restate
How do you criticise without caricaturing? The Marginalian revisits the four rules the philosopher Daniel Dennett borrowed from the game theorist Anatol Rapoport in “Intuition Pumps and Other Tools for Thinking”. Restate your opponent’s position so clearly that they thank you for it; list the points you agree on; say what you have learnt from them; only then rebut. The point is not politeness for its own sake: it is that the other person can then actually hear the criticism. Worth trying before your next comment.
Sources
- The Marginalianthemarginalian.org
10 · Vim, without the AI
Vim Classic is a long-term support fork of Vim, maintained explicitly without help from generative AI tools. Its second release, 8.3.1, takes in Vim’s fixes up to patch 8.2.3237 and drops Python 2 support. Like Vim, it remains “charityware”, inviting donations to the Ugandan children’s charity championed by Vim’s creator, Bram Moolenaar. Appealing if you want a stable editor with no surprises; less so if you are waiting for Vim 9’s new features.
Sources
- Vim Classicvim-classic.org
- LWN.netlwn.net
Watch next
11 · Spain’s general election, Sunday 29 November
Spain votes next Sunday. Polls published on 20 November, compiled by ARA, give the conservative People’s Party (PP) a lead of 0.7 to 3.6 points over Pedro Sánchez’s Socialists, and an absolute majority to the PP and the far-right Vox combined. A Cluster 17 poll for Le Grand Continent puts Vox at 18.4%, up from 12.4% in 2023.
Sources
- ARAen.ara.cat
- Le Grand Continentlegrandcontinent.eu
To close
Enjoy the rest of your Sunday, and see you tomorrow.