Show every history made of two or more stories

Mathematician says Lean code of Riemann hypothesis proof has passed nanoda check

The Lean code passed the kernel check with no "sorry" and no additional axioms, but under Clay Mathematics Institute rules the work must go through formal publication and at least two years of verification.

A run of the nanoda kernel in paranoid mode on the Lean formalization of the Riemann hypothesis proof was completed by Andrew Sutherland, mathematician Bartosz Naskręcki reported on 21 November. According to him, the code contains neither a "sorry" — the placeholder for an unproven statement in Lean — nor any additional axioms. Naskręcki rates the probability that the proof is correct as very high. The claim of a proof produced by Anthropic employee Levent Alpöge together with the Claude model was made public on 20 November, US time. The Lean code is posted in 14 archives on Alpöge's personal site, with the posting running to more than 250 thousand pages. There had been no independent confirmation from the mathematical community as of the morning of 21 November. Under Clay Mathematics Institute rules, the work must go through formal publication and at least two years of verification. Skeptical mathematicians pointed to the absence of such confirmation by the community.

The regulator's review now concerns accounting rather than only supplies: the news itself coincided with a 14% drop in the shares and the worst week since 2008.

Reuters: SEC examining Nvidia's accounting; shares fell 14% in a day

The regulator's review now concerns accounting rather than only supplies: the news itself coincided with a 14% drop in the shares and the worst week since 2008.

The enforcement division of the US Securities and Exchange Commission (SEC) is examining possible accounting problems at Nvidia, Reuters reported on 20 November, citing two people familiar with the matter. The Wall Street Journal wrote on 17 November about a review of the company without naming its subject; the focus is now described as accounting practices. The SEC declined to comment, and Nvidia did not respond promptly to a request. The fact of a review does not in itself imply any violation. Reuters was unable to establish when the review began or which specific accounting issues are being examined. The report coincided with Nvidia shares falling 14.01% to $125.53 at Friday's close and the stock's worst week since 2008.

Sources 1

Assembling several specialised agents into a single flow can now be done entirely in chat, and their results are merged into one answer.

AWS Quick: multiple agents merge results into one answer as flows are assembled in chat

Assembling several specialised agents into a single flow can now be done entirely in chat, and their results are merged into one answer.

Amazon Web Services has updated AWS Quick, an SI workspace for work tasks. Flows can now be created, launched and refined right in chat, with a diagram of the steps shown next to the conversation. Flows have gained deterministic steps for fixed calculations and checks, along with support for long-running processes and large data volumes. Orchestration of multiple specialised agents has been added, with their results merged into a single answer. The updated web app has gained a priority tasks screen with a daily briefing, a Customize section with agents, skills and connectors, a library and Mission Control. Tasks keep running after the tab is closed. Connectors have been added to a Meta ads MCP server (with confirmation before any write: creating a campaign, changing a budget) and to Amplitude. AWS says the features are available in all regions where Quick operates. According to the company's blog, one workflow at customer Pomelo saves 80 hours of work a month, and the number of employees working with data has grown almost 4-fold.

There is no binding treaty on autonomous weapons: work through 2028 is limited to "elaborating" a document, and the question of its legal force remains open.

128 countries agree to "elaborate" a document on autonomous weapons by 2028

There is no binding treaty on autonomous weapons: work through 2028 is limited to "elaborating" a document, and the question of its legal force remains open.

128 states parties to the UN Convention on Certain Conventional Weapons (CCW) agreed to "launch the development" of a document on lethal autonomous weapons systems (LAWS). That is the outcome of the seventh review conference in Geneva: five days of talks, a session that ran until the early hours of Saturday, and work planned for two years — until 2028. The wording "negotiate" was replaced with "elaborate" because of objections from some delegations; the question of whether the document would be legally binding was left open. More than 95 states backed a binding treaty; the US, Russia and Israel were opposed. CCW decisions are taken by consensus, so any single country can block an agreement. The Stop Killer Robots coalition called the mandate "vague", and Human Rights Watch called it "disappointing". Conference chair Julien Thöni spoke of "concrete results", and UN representative Izumi Nakamitsu of a "significant achievement" in a "glass half full" situation.

75% of Micron's 2027 capacity is already sold — customers will have to agree new volumes for 2028 and later.

Micron: 75% of 2027 memory capacity sold, $32 billion in contracts

75% of Micron's 2027 capacity is already sold — customers will have to agree new volumes for 2028 and later.

At the Micron TomorrowX summit, CEO Sanjay Mehrotra and president and CTO Scott DeBoer said 75% of the company's 2027 capacity has been sold, and supplies for 2028 are being worked out. 26 strategic customer agreements (SCAs) have been signed, with prepayments and commitments totalling $32 billion. Some contracts have price ranges or fixed prices, others have no price terms; customer requests five years out exceed what Micron can guarantee. The company is carrying out a US investment programme of more than $250 billion to expand DRAM output. According to DeBoer, new fabs will not bring immediate supply. Micron expects 3D-stacked DRAM to reach the market in 2028–2029 and to become widespread in 2030–2031; heat removal when stacking DRAM directly on a GPU was named the main barrier.

Samsung has to negotiate memory supplies years in advance: contracts cover 2027–2029, and some customers are asking to extend them to 2031.

Economy Tribune: Samsung has locked in most of its memory capacity for 2027–2029

Samsung has to negotiate memory supplies years in advance: contracts cover 2027–2029, and some customers are asking to extend them to 2031.

Most of Samsung's memory capacity for 2027–2029 is already locked in under contracts, and some customers are pushing to extend long-term agreements (LTAs) to 2031, Economy Tribune reported on 21 November, citing sources. On 29 October, at its third-quarter earnings conference, Samsung spoke of LTAs with 10 major customers and negotiations with 10 more; the goal is to cover 60–70% of medium-term capacity with long-term contracts by 2030. At leading memory makers, demand satisfaction stands at 60–70%, and wafer capacity for conventional DRAM for 2027 is almost entirely sold out. DRAM bit output growth is limited to 20–25% a year, and coverage of demand for enterprise SSDs is just above 50%. The reason is expanding investment in SI data centers: HBM takes up an ever larger share of wafer capacity, squeezing conventional DRAM output.

US school districts are choosing between access to Google Search and control over how SI answers students: Chicago Public Schools says it will switch search engines otherwise.

Google demands schools allow AI Overviews or lose Search from 2027

US school districts are choosing between access to Google Search and control over how SI answers students: Chicago Public Schools says it will switch search engines otherwise.

Google is demanding that K-12 school districts allow AI Overviews, or students will lose Google Search on their school accounts starting in early 2027. Chris Phillips, Google's vice president of education, puts it this way: search and SI are "deeply integrated". Since December 22, 2025, Chicago Public Schools has been blocking ChatGPT, Claude, Gemini and other third-party SI tools and is looking for a way to suppress AI Overviews separately from search. Otherwise the district says it will switch to other search engines; on November 11, about 30 people protested outside Google's Chicago office. New York City imposed a one-year moratorium on SI for grades 2-K–8 and a ban on companion chatbots. On November 13, City Council Speaker Julie Menin and Shekar Krishnan demanded written answers from Sundar Pichai by November 19. The New York district has about 1.3 million Chromebooks, 500 thousand of which are assigned to students. LAUSD blocked generative SI on its devices from September 2, 2026, for roughly 378 thousand students for the 2026–27 school year. Miami-Dade, by contrast, provides Gemini for Education to about 100 thousand high school students.

Sources 1

Retail investors are buying TSMC in odd lots at a seven-month low, and since 6 November the National Stabilisation Fund has had a mandate for interventions worth NT$500 billion.

TSMC slides to lowest since April as retail investors bought odd lots

Retail investors are buying TSMC in odd lots at a seven-month low, and since 6 November the National Stabilisation Fund has had a mandate for interventions worth NT$500 billion.

Shares of contract chipmaker TSMC fell 4.76% on 20 November to NT$2,000. That is the lowest since 13 April, when the stock traded at NT$1,990, according to Taiwan Stock Exchange data. The Taiex index lost 2,393.60 points (6.23%) to close at 36,033.29 — the third-largest single-day point drop in its history. Against that backdrop, retail investors stepped up buying in odd lots: turnover in TSMC shares in that mode was about NT$24.12 billion, the highest of any stock on the main board. One standard lot of 1,000 shares would have cost NT$2 million. From its peak on 6 October, the stock has lost NT$590 (22.78%), and TSMC's market capitalisation has shrunk by NT$15.3 trillion to NT$51.86 trillion. Taiwan Premier Cho Jung-tai held a meeting of economic and financial agencies on Saturday on stabilising the market. A mandate for National Stabilisation Fund interventions worth NT$500 billion has been in force since 6 November.

Lower cost and time to first token for self-deployed models can now be achieved automatically: the agent itself checks whether a cache is needed and deploys it.

AWS releases distributed KV cache skill, claiming 40% lower cost and 80% lower latency

Lower cost and time to first token for self-deployed models can now be achieved automatically: the agent itself checks whether a cache is needed and deploys it.

AWS has published a distributed KV caching skill in the Agent Toolkit for AWS. The coding agent itself checks whether a workload needs such a cache, sizes the cluster and deploys it on Amazon ElastiCache (Valkey 9.1) with LMCache and RDMA/EFA. AWS claims the distributed KV cache cuts the inference cost of self-deployed large language models by up to 40% and time to first token by up to 80%. According to AWS, the KVTrace tool assesses the benefit on a customer's own traffic for free before deployment, without recording token identifiers or prompt text. The skill is claimed to be compatible with Kiro, Claude Code, OpenAI Codex and other agents that support Agent Skills.

The stock is back almost to its offering price: those who bought at the first-day close of $97.17 are already down about 13%.

Anthropic shares fall about 13% by second day of trading after IPO

The stock is back almost to its offering price: those who bought at the first-day close of $97.17 are already down about 13%.

Anthropic shares were down about 13% by the second day of trading, to $85 as of 12:20 US Eastern time. The first day closed at $97.17, and the offering price was $83. The offering took place on 19 November: 980 million shares at $83. On the first day, retail investors traded more than $750 million worth of Anthropic shares — a record for an IPO's first day (SpaceX's was $670 million). Yet net inflows were lower than at SpaceX's ($118 million) and Coinbase's ($92 million) debuts, according to Vanda Research. Retail investors received less than 20% of the shares in the offering. The decline coincided with a broad selloff in tech stocks over doubts about returns on SI infrastructure spending.

Sources 1

The demo shows that decoding the intention to move a hand is enough to control a complex game, but Neuralink's devices remain in clinical trials without FDA approval.

Neuralink shows video of a patient playing GTA VI without a gamepad

The demo shows that decoding the intention to move a hand is enough to control a complex game, but Neuralink's devices remain in clinical trials without FDA approval.

Neuralink published a roughly two-minute video on X captioned "Playing GTA VI with a Neuralink". A clinical trial participant controls a character without a gamepad: driving a boat, a car, piloting a plane, taking part in fights and shootouts. The demo relies on the Neuralink implant and the Telepathy system: it reads brain activity associated with the intention to move a hand and converts it into controller commands. The screen shows a neural signal decoding panel and a DualSense gamepad map whose buttons light up when commands are registered. The participant issues commands by mentally imitating taps with his ring fingers; later the team added mapping of foot movements. The video carries a disclaimer: the devices are investigational and not approved by the FDA. GTA VI was released on November 19, 2026. According to TechRadar, judging by the footage, the player uses small physical movements, including finger taps, alongside brain activity. Neuralink did not disclose how neural and physical input are separated.

The Stentrode is delivered through blood vessels without opening the skull, and Synchron has brought the number of implantations to 15, planning to read several brain areas at once.

Synchron implants Stentrode in 15 people and prepares to read several brain areas at once

The Stentrode is delivered through blood vessels without opening the skull, and Synchron has brought the number of implantations to 15, planning to read several brain areas at once.

Synchron founder and CEO Thomas Oxley told Fortune on November 21 that the company has implanted the Stentrode in 15 people, adding five over the year. Over the same period the number of participants with a Neuralink implant grew from seven to about 26, according to Fortune. According to him, Synchron is preparing a more powerful device that reads several brain areas at once without opening the skull (whole-brain sensing). It is intended, among others, for patients with age-related cognitive decline, and Oxley described the timeline for its release as "soon". A Merge Labs representative (co-founder Sam Altman; OpenAI invested in early 2026) told Fortune the product would be unveiled in "a couple of months". Oxley also called China, where the first commercial neurointerface was approved in March 2026, a serious source of concern.

Sources 1

Passenger flights of the EH216-S in Japan remain demonstrations for now: commercial launch depends on certification, safety requirements and infrastructure.

EHang shows air taxi flight in Japan with a passenger and no pilot

Passenger flights of the EH216-S in Japan remain demonstrations for now: commercial launch depends on certification, safety requirements and infrastructure.

The two-seat, fully electric EH216-S from China's EHang made a flight on November 20 at the Japan Aviation Academy in the city of Kai (Yamanashi Prefecture). The vertical take-off and landing aircraft flew a pre-programmed route with no pilot on board and one passenger in the cabin. This was the first demonstration with a passenger for the EH216-S in Yamanashi Prefecture. The aircraft first flew 380 meters along the academy's runway, then a second, unmanned flight covered 1,800 meters over the Kamanashi River. According to Euronews, the EH216-S with 16 rotors is designed for speeds up to 120 km/h and about 30 km on a single charge. The demonstration was intended to popularize urban air mobility and test short flights in Japan. Commercial launch still depends on certification, safety requirements and infrastructure.

Doctors in 77 regions already work with SI services through MosMedAI, and the Federation Council is ready to take part in connecting the remaining regions.

Federation Council: 77 regions connected to MosMedAI platform

Doctors in 77 regions already work with SI services through MosMedAI, and the Federation Council is ready to take part in connecting the remaining regions.

Since 2024, 77 federal subjects have been connected to the federal MosMedAI platform, said Alexei Sinitsyn, head of the Federation Council Committee on Social Policy. The committee meeting was held on November 21 at the Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies of the Moscow Department of Health. Sinitsyn said the upper chamber is ready to take part in connecting the remaining regions. Federation Council Deputy Speaker Inna Svyatenko spoke about access to vetted SI services through the federal platform. Yuri Vasilyev, chief physician of the center, said about 850 radiologists from more than 25 regions work at the center, describing over 180 thousand studies weekly. More than 20 million studies have been analyzed, 45 clinical areas are in operation, and more than 200 companies took part in the experiment.

Sources 1

The $1 trillion purchase of SK hynix will not happen: SoftBank stopped the discussion at an early concept, and any foreign acquisition in Korea requires prior government approval.

The Information: SoftBank discussed buying SK hynix for $1 trillion and dropped the idea

The $1 trillion purchase of SK hynix will not happen: SoftBank stopped the discussion at an early concept, and any foreign acquisition in Korea requires prior government approval.

SoftBank Group head Masayoshi Son approached Wall Street banks to discuss financing a possible purchase of SK hynix for about $1 trillion, The Information reported. A SoftBank representative told the publication that the deal is "no longer under active consideration". SK hynix said it was unaware of such plans and recalled that under Korean law any foreign acquisition requires prior government approval. SK hynix's market capitalization is about 777 trillion won (roughly $577.7 billion). According to The Information, the idea never advanced beyond an early concept and was discarded as unworkable. SoftBank's SI portfolio has no memory maker: it includes Arm, Ampere Computing and DigitalBridge, as well as stakes in Intel and OpenAI.

The planned standards establish the token as a single unit for billing and quality assessment of SI services; the agency did not name a timeline for publishing them.

China announces plans to standardize measurement, payment and quality of SI tokens

The planned standards establish the token as a single unit for billing and quality assessment of SI services; the agency did not name a timeline for publishing them.

China's National Data Administration published on November 21 a statement by director Liu Liehong on plans to improve standards for measuring, paying for and assessing the quality of SI tokens. The agency intends to expand use cases for token services and to counter low-quality duplication of developments. The remarks themselves were made on November 20 at the opening of the World Computing Conference in Changsha; media reports appeared on November 21. According to Liu, an industry of token production, circulation and consumption is taking shape. "Token factories", he said, turn compute rental, model supply and data into a token-production service. In March 2026 Liu spoke of more than 140 trillion tokens per day (growth of more than 1,000 times since the start of 2024), and by June, according to the agency, the daily volume exceeded 500 trillion.

The SI-cheating scandal has moved into politics: the opposition demands the education minister's resignation, while universities hold exams with screening and disposable aprons.

Korea: universities search applicants over SI as opposition demands minister's resignation

The SI-cheating scandal has moved into politics: the opposition demands the education minister's resignation, while universities hold exams with screening and disposable aprons.

The first written exams after the national CSAT were held at 14 South Korean universities, including Sogang, Kyung Hee, Konkuk, Sungkyunkwan and Sookmyung. They were conducted under heightened scrutiny amid a SI-cheating scandal. At Kyung Hee, applicants were given disposable aprons to conceal cameras on clothing and buttons and were allowed to write only with pens issued to them. Glasses and cuffs were checked, earplugs were banned and applicants were required to keep their hair back so their ears stayed visible; some halls were walked with detectors for recording equipment. The day before, the Ministry of Education and KACEA checked measures at 46 universities holding their own exams and set up an emergency response team. At Kyung Hee suspicious situations involving electronics were recorded, but no violations were confirmed. The same day, applicant associations and the opposition People Power Party demanded the resignation of Education Minister Choi Kyo-jin. They claim the anti-SI measures at the CSAT were formal: button stickers were not handed out everywhere, and glasses checks were done for show. According to Ministry of Education data as of November 20, 111 violations involving electronic devices were identified at the CSAT on November 19, against 98 in the first count. Five of them were actual use of smart glasses, mini cameras and wireless earbuds. The minister said he would step down if the body that appointed him demanded it.

When moving to GPT-6.2, developers will have to rewrite old system prompts and account for temperature and top_p no longer being supported.

OpenAI releases system prompt templates for GPT-6.2 Astra and Sol

When moving to GPT-6.2, developers will have to rewrite old system prompts and account for temperature and top_p no longer being supported.

OpenAI published ready-made system prompt templates for GPT-6.2 Astra and Sol across three blocks — writing style, initiative and code. The company recommended that developers review old instructions when migrating. According to the company's description, Astra responds more briefly and directly by default, and when a request is unclear the models tend to suggest the next step instead of acting. The templates require making reversible assumptions and stopping only when the task's scope changes, there is a real blocker, or the action is irreversible. The code template says to reuse existing logic, not add abstractions and write new tests only on request or for a specific regression. On the API side with GPT-6.2, the Responses API discards previous reasoning content when the conversation history is edited, the model is changed, or the end user and organization change. The temperature and top_p parameters are no longer supported; for new integrations the "medium" reasoning level is recommended.

SI data centers in Korea are part of three state megaprojects, and the government promises to help construction sites with labor shortages.

South Korea's president visits SK's SI data center construction site in Ulsan

SI data centers in Korea are part of three state megaprojects, and the government promises to help construction sites with labor shortages.

South Korean President Lee Jae-myung visited the construction site of an SI data center operated by SK Inc. in Ulsan on November 21. It came on the day after his four-day trip to the Philippines and China. He called SI data centers a major source of demand for semiconductors and key infrastructure for physical SI (robots and autonomous vehicles). The president expressed hope that the Ulsan project would attract investment from global SI companies. The project is part of three state megaprojects — turning the Chungcheong, Honam and Yeongnam regions into hubs for SI data centers, physical SI and advanced semiconductors. The government promises to help with the shortage of construction workers; the president was accompanied by SK Telecom CEO Jung Jai-hun, science minister Bae Kyung-hoon and officials from relevant agencies.

Sources 1

According to the Chief Justice of India, regulating SI in legal education should rest on disclosure rather than prohibition, and a student should not be found guilty on an SI detector's output alone.

India's chief justice: banning SI in legal education would be unreasonable

According to the Chief Justice of India, regulating SI in legal education should rest on disclosure rather than prohibition, and a student should not be found guilty on an SI detector's output alone.

A blanket ban on generative SI in legal education would be "impractical and unreasonable", Chief Justice of India Surya Kant said. He spoke on November 21 at the first annual international conference of the Commonwealth Legal Education Association (CLEA) in Hyderabad. Instead of a ban, he called for clear rules on which uses are permitted, which require disclosure and which violate academic integrity. As a model he cited the Supreme Court of Singapore: court users there may use generative SI but are fully responsible for the materials they submit. Kant warned against SI detectors as sole evidence. According to him, citing a 2023 study, seven common detectors wrongly flagged more than three out of every five essays by non-native English speakers as machine-written. "No student should be found guilty on the word of a machine alone," he said, and proposed that CLEA develop model rules for disclosing SI use in legal research. The position of supporters of a full ban on generative SI in legal education was not found in the verified sources.

SI adoption in China's HR services is being consolidated onto a single full-cycle site — from an industry model and datasets to the handover of ready-made solutions.

China opens state pilot SI base for HR services in Chongqing

SI adoption in China's HR services is being consolidated onto a single full-cycle site — from an industry model and datasets to the handover of ready-made solutions.

A state pilot base for applying superintelligence to HR management services has officially opened in Chongqing. The announcement was made at the opening of the International Conference on Talent Exchange on November 21. The announced plan includes one computing platform, one industry large model, 27 high-quality datasets and one service operating system. The base operates on a full-cycle model of "technology development — scenario validation — transfer of results." Its task is to close the "last mile" of SI adoption in the HR industry and accelerate its digital transformation.

The statement was not accompanied by any decisions or directives on SI regulation — it introduces no new rules.

Russian presidential aide Medinsky calls SI a threat that dictates behavior

The statement was not accompanied by any decisions or directives on SI regulation — it introduces no new rules.

Russian presidential aide Vladimir Medinsky called "neutrally depersonalized superintelligence that is beginning to dictate and determine human behavior" a threat. In his words, that danger is no less than the one posed by "moral perverts." He named the church and the family as a support for people: "nothing so far but the church and the family can replace this." The statement was made on the program Slovo pastyrya (The Word of the Shepherd), marking Patriarch Kirill's 80th birthday; the publication with the quote came out on November 21.

Sources 1