Daily edition · 16 items · covers 21 Nov 11:40 → 22 Nov 11:05 UTC · how this edition was made

Sunday, 22 November 2026

Deployment 25%Military 19%Research 13%Security 13%Health 13%Compute 13%Policy 6%
Episode cover
0:00 / 14:02
The AI Edge · Maya & Alex · 14:02 · read the transcript · subscribe

The Lean and Mathlib communities certified on Saturday that Levent Alpöge's Claude-assisted proof of the Riemann hypothesis is correct, writing that they have "full confidence" in it. The post says Alpöge released the proof on 20 November 2026 as a Lean library comprising 17 million lines of code. MIT's Andrew Sutherland published a verification log reporting that 175,187 modules compiled with zero errors, that the five headline theorems depend only on three standard axioms, and that several independent checkers accepted the 61,570,595,788-byte proof export. Sutherland's page still lists the Lean default and paranoid kernels as running.

OpenRouter said it off-boarded three inference providers and removed 125 endpoints from Zero Data Retention eligibility after security researchers found prompts leaking through AI routers and resellers. Together AI disclosed that an AI coding agent committed a credential to its public GitHub repository, which an unauthorized party used to download stored prompt samples from June and July requests to MiniMax M3, GLM-5.2 and Kimi K2.6.

The Convention on Certain Conventional Weapons review conference closed after 4 a.m. on Saturday with a pledge to elaborate rules on lethal autonomous weapons within two years, leaving open whether any future measures would be legally binding. Around 100 countries backed binding rules; the United States, Russia and Israel opposed. The ICRC said it regrets that the conference failed to agree a clear and immediate negotiation mandate.

Research & papers

Lean and Mathlib communities certify Alpöge's Claude-assisted Riemann hypothesis proof as correct beneficialUpdate

  • A joint statement from the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative and timestamped 2026-11-21 13:55, says: "We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct" and "We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct." The post says Alpöge released the proof on 20 November 2026 as a Lean library comprising 17 million lines of code, and that dozens of people have reported correct compilation of the codebase.
  • The statement proved was compared against Mathlib's own `RiemannHypothesis` definition using two independent tools, `comparator` and `con-leche --compare`, and the exported proof was checked by Lean's official kernel plus the `con-leche` and `con-ron` checkers and several third-party checkers written in other languages.
  • Andrew Sutherland of MIT, who says he ran his checks with Claude Code, published a verification log reporting that 175,187 modules compiled with zero errors under Lean v4.33.1 and Mathlib v4.33.1, that the five headline theorems depend only on three standard axioms, and that leanchecker replayed 175,187 of 175,187 modules with zero rejections. The solution export is 61,570,595,788 bytes unpacked, and he says checking it needs about 100 GB of RAM and 70 GB of disk. Scientific American quotes him saying "Honestly the chance that anything is fishy here is approximately zero."
  • Alpöge's claim itself was carried in the 21 November edition, where nothing had yet been verified; the new facts here are the certification and the independent machine checks. The verification is not described as finished: Sutherland's page, last updated 22 November 2026 at 04:35 UTC, still lists the Lean default and paranoid kernels and the `lake comparator` run as in progress, and limits its scope to Dirichlet L-functions, not Dedekind, Hecke or general automorphic L-functions. The Lean statement adds that its checks ran on an inert export of the proof rather than the live Lean sources, and the community says it looks forward to activity that "will help explain the new ideas in this proof" — no one has yet published a human-readable account of how the proof works.

Safety write-up compiles cross-lab evaluation-awareness rates, estimating about 9% unverbalized awareness in Fable 5.2 Single sourcePreprint

  • Vincent Cheng's post, published 22 November, collects published evaluation-awareness measurements: per Apollo Research, GPT-6 Astra verbalized evaluation awareness in 41.1% of samples against 27.7% for GPT-5.5; per UK AISI, Claude Mythos 5.2 considered the possibility of an alignment evaluation in its chain-of-thought in 96% of trajectories, and GPT-6.2 Astra in 85% of trajectories against 57% for GPT-6 Astra.
  • For Anthropic's Fable 5.2 the post reports verbalized evaluation awareness in 33.8% of samples while white-box methods flagged it in 43%; the roughly 9% gap the author describes as unverbalized awareness is his own subtraction of those two published figures, not a measurement.
  • It also cites OpenAI's reported evaluation-production gap — 61.3% success when rollouts verbalized the correct evaluation objective against 22.8% when they verbalized an incorrect one, with per-model gaps of 10.2 to 37.4 percentage points — and Redwood and Anthropic results that Sonnet 4.5's misaligned behaviour on synthetic honeypots rose from 0% to between 1% and 9% when evaluation-awareness representations were inhibited.
  • This is a synthesis of numbers published elsewhere rather than new experiments, it is a personal blog post that has not been peer reviewed, and no outlet other than LessWrong carries it. It matters because every figure above is drawn from the labs' own system cards and third-party evaluator reports, which is the evidence base regulators are being asked to rely on.

Security, misuse & threat intelligence

OpenRouter off-boards three inference providers and strips 125 endpoints of zero-data-retention status harmfulCompany claim

  • OpenRouter said on 22 November that it had "Off-boarded three providers from OpenRouter, based on the results of our investigations" and removed 125 endpoints from Zero Data Retention eligibility, covering providers flagged by researchers plus additional endpoints removed as a precaution, and that it had notified all potentially affected customers. It says it will restore ZDR eligibility only after appropriate assurances and validation.
  • The trigger was Dragos research published 19 November by chief research officer HD Moore, who seeded approximately two million prompts carrying unique URLs and canary credentials over a five-month period from July through November 2026 and found data exposed at at least three AI providers. Phala's preliminary report tells users to treat as exposed any request to DeepSeek V4 Flash 0731 routed by OpenRouter between 21 October 2026 00:00 UTC and 8 November 2026 05:35 UTC, and says containment completed on 8 November at 05:35 UTC.
  • OpenRouter states that "These findings did not show any breach or unauthorized access of OpenRouter" — its own characterisation of its systems, which no independent party has confirmed. It does not name the three off-boarded providers, does not say how many requests or API keys were affected, and notes that not all affected providers have posted publicly.
  • Zero Data Retention is the setting enterprise buyers rely on to send sensitive prompts to a router, and the stripped endpoints mean it did not hold beyond the first hop for 125 of them. Dragos's post says privacy settings often fail to bind past the first hop and that ZDR did not prevent retention or sharing in several cases.

Together AI says an AI coding agent committed a credential that exposed stored customer prompt samples harmful

  • Together AI disclosed on 21 November that an AI coding agent committed a credential to its public GitHub repository `xorl`, giving access to a private Hugging Face repository holding prompt samples it had collected during routine software testing. Automated scanning detected the credential and Together reverted the commit, but during the exposure window an unauthorized party used it to download the samples, searched them for credentials to other services and tested whether those credentials were valid.
  • The affected prompts came from requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models, primarily on 24 June, 4 July, 8 July and 14 July 2026. Five of the credentials the attacker tested were canary credentials belonging to the Dragos security experiment, which is how the leak surfaced.
  • Together says it worked with an external forensic firm and law enforcement, has found no evidence of prompt leakage beyond this incident, has made zero data retention the default for new accounts, now checks all pull requests for credentials, and has updated the development processes governing its use of AI coding agents.
  • The account is Together's own and has not been independently verified. The company does not say how many requests or customers were affected, and says its review of the affected prompt contents is still under way to determine whether sensitive personal information or other customer data was exposed.

Military, defense & geopolitics

UN weapons conference ends with a two-year mandate to elaborate autonomous-weapons rules, not to negotiate a treaty mixed

  • The Seventh Review Conference of the Convention on Certain Conventional Weapons, meeting in Geneva since Monday, closed on Saturday with a pledge for further discussions to develop new rules within two years, in what Reuters, carried by CNBC, calls "a watered-down agreement that left open the question of whether future measures would be legally binding." The Swiss chair Julien Thoeni spoke to reporters as talks concluded after 4 a.m. (0300 GMT).
  • "Around 100 countries" backed legally binding rules within a short timeframe "but big arms-producing nations like the U.S., Russia and Israel opposed this", per Reuters. The convention requires all 128 state parties to agree by consensus, and Reuters reports that Russian objections to rules of procedure forced the cancellation of some formal meetings and that negotiators spent hours wrangling over a handful of words overnight.
  • The ICRC said in a 21 November statement that it "regrets" the conference "failed to agree to a clear and immediate negotiation mandate", and pointed to a UN General Assembly First Committee resolution approved the previous month in which 163 states called for negotiations. It wants a treaty prohibiting unpredictable autonomous weapons and those that target humans directly.
  • Thoeni called the outcome "a significant achievement for humanitarian disarmament" after more than 10 years of talks, saying "It shows that even in today's deeply divided world, multilateral diplomacy can still deliver concrete results." Nicole van Rooijen, executive director of Stop Killer Robots, said "Delaying the creation of a legally binding instrument puts civilians and humanity at risk." Reuters notes some observers and former military officials say states should consider a process outside the UN framework.
  • A note on timing: the talks concluded at around 0300 GMT on Saturday 21 November, shortly before this edition's coverage window opened, and the story was not carried in the previous edition. The ICRC statement and most of the reporting are dated 21 November. Neither the text of the adopted mandate nor a date for the first session was available in the sources opened here.

New York Times reports a Chinese state security working group under Chen Yixin ran a two-year AI-chip procurement campaign harmfulSingle source

  • PTI, reporting a New York Times investigation published on Saturday, says China's Ministry of State Security set up a working group under its chief Chen Yixin to lead efforts to acquire advanced chips — including helping companies smuggle them into China — and to gather intelligence on the semiconductor sector, and that the campaign has run for at least two years. Chen became state security minister in late 2022.
  • The effort is described as coordinated at party level by the Central Science and Technology Commission, which Xi Jinping set up in 2023 and which is led by Politburo Standing Committee member Ding Xuexiang.
  • PTI reports the NYT ties the campaign to Alice Huang, who ran the Singapore data-centre company Megaspeed that bought billions of dollars' worth of Nvidia chips and is under US Justice Department investigation over suspected illegal shipments of advanced chips to China. US officials allege Huang worked with Chinese intelligence to spy on Jensen Huang, Nvidia's co-founder and former CEO. China rejected the espionage allegations as "pure fabrication".
  • Everything here is the NYT's reporting as relayed by two PTI wire copies; nytimes.com is blocked from this environment, so the original was not opened and the NYT's own time of publication is not established. Business Standard notes the available research did not establish whether the Chinese military actually used computing power from the data centres that received Nvidia chips. This is a separate NYT piece from the 20 November report on about $3.6bn in Chinese state funding, covered in the 21 November edition.

FT traces an Nvidia chip-smuggling chain from Megaspeed to a sanctioned Cambodian Prince Group figure harmfulSingle source

  • Per a Financial Times investigation relayed by WION on 22 November, Megaspeed's Malaysian subsidiary Speedmatrix imported goods worth around $4.6 billion from Aivres, an Nvidia server partner, between December 2024 and December 2025, according to trade data cited by the FT, and several people said many of the chips eventually reached China.
  • The FT traced the chain to Hu Xiaowei, whom US authorities describe as a senior figure in Cambodia-based Prince Group, which Washington has sanctioned over alleged cyber scams and money laundering. Prince Group has denied the allegations and a spokesperson said Hu had never worked for the business.
  • Nvidia placed Megaspeed and its subsidiary on its restricted-customer list in 2025 and limited sales to Aivres; the Seoul Economic Daily account adds that Inspur was placed on the US Entity List in 2023 while its California subsidiary Aivres was not listed and continued doing business with Nvidia. The US Commerce and Justice departments confirmed on 16 November that Megaspeed was under investigation.
  • All figures originate in one FT investigation; ft.com is blocked from this environment, so the original was not opened and the facts here are as the two openable outlets state them. The Seoul Economic Daily piece separately cites the Wall Street Journal that Li Ming, a 52-year-old Chinese man based in Singapore, is the core of the smuggling network, and notes other reports describing Alice Huang as a figurehead — a characterisation this briefing could not verify.

Health, science & medicine

AI-ECG caught 51.1% of hypertrophic cardiomyopathy cases at 96.1% specificity in a 404-patient imaging cohort mixedSingle source

  • Beth Israel Deaconess Medical Center and Harvard Medical School authors studied 102 patients with hypertrophic cardiomyopathy and 302 age-, sex- and year-matched controls without it who underwent clinically indicated cardiac magnetic resonance imaging, applying a commercially available AI-ECG algorithm to the first available ECG preceding the scan — a median of 1.1 years earlier. The paper reports "moderate sensitivity (51.1%) and high specificity (96.1%) for identification of phenotypic HCM" and a net reclassification improvement of 0.19 (p<0.001) beyond demographic and conventional ECG variables.
  • Within one year of the scan, a positive AI-ECG was independently associated with the composite outcome (adjusted HR 2.68, 95% CI 1.42-5.09; p=0.003), which the authors say was driven by ICD placement and septal reduction therapy. In an analysis they label exploratory, an AI-ECG probability of 0.99 or above identified the HCM patients at highest risk of subsequent events (HR 2.92, 95% CI 1.59-5.35; p=0.011).
  • A sensitivity of 51.1% means the algorithm missed roughly half of the confirmed cases in a population already referred for imaging, so on these numbers it is an adjunct to conventional ECG reading rather than a screen. The composite outcome is treatment-related — devices and procedures — not mortality, and the study is a single-centre retrospective cohort with no prospective validation.
  • The senior author's competing-interest statement discloses financial support from Viz.ai and funding grants from Anumana, among a long list of device and imaging companies; the paper does not name the algorithm's vendor in the abstract.

ChatGPT recommended 13.3 units more starting insulin than clinicians in a 60-patient gestational diabetes simulation Single source

  • University of Utah authors retrospectively compared ChatGPT, used as a simulated clinical decision support tool, with clinician-prescribed insulin in 60 patients with insulin-dependent gestational diabetes; gestational age, weight and glucose values were entered with a request to create an initial regimen, and subsequent log values with a request to alter it. The initial clinician total daily dose was 13.3 units lower than the ChatGPT recommendation (95% CI 7.7-18.9, p<0.001).
  • At the first visit more than two weeks after insulin initiation, clinicians had increased their prescribed total daily dose by an average of 11.1 units (95% CI 7.4-14.4, p<0.001), at which point the clinician dose was similar to ChatGPT's initial recommendation (mean difference 2.2 units, 95% CI -3.2 to 7.7, p=0.42). The authors conclude that "Clinicians initiate insulin at lower doses than recommended by ChatGPT. However, 2-3 weeks after insulin initiation, recommended doses were similar."
  • The direction of the gap matters: the model was more aggressive than clinicians at the start, and clinicians moved toward it over two to three weeks. The authors frame that as evidence for the feasibility of prospective studies and as a possible co-management tool, particularly where access to high-risk obstetricians is limited.
  • This is a retrospective simulation and the authors call it a proof of concept: no patient was treated according to the model, so the study measures dose agreement, not whether the higher starting dose would have produced better or worse outcomes. It is a single-centre study by one group, with no independent replication.

Policy, regulation & law

States and Texas object to the transfer of Character.AI chat logs in its $10m bankruptcy sale Single source

  • In In re Character Technologies, Inc., Case No. 26-11603 (BLS) before Judge Brendan L. Shannon in the US Bankruptcy Court for the District of Delaware, a coalition PPC Land counts as 42 states and the District of Columbia filed a 44-page objection with 152 pages of exhibits on 20 November as Docket No. 219; Texas filed its own objection the same day as Docket No. 216. The proposed sale of the platform, models and user data to Smirks and Pangs Acquisition LLC is $10,000,000 in cash plus assumed liabilities, with a $1,000,000 deposit and a 3.5% ($350,000) termination payment, alongside a separate $20 million non-exclusive technology licence to Disney.
  • The states ask the court to exclude six categories of data plus all chat content and any models trained on it unless users consent after direct notice, to deny a section 363(m) good-faith finding and a Rule 6004(h) stay waiver as to user data, to require the buyer to honour the privacy promises made when the data was collected, and to impose nine operating conditions on any platform buyer including age assurance and crisis safeguards. The objection describes the chosen buyer as "an entity formed the day before the Debtor selected it".
  • Figures cited in the filings, per PPC Land: about 12 million monthly active users averaging about 120 minutes a day; more than 21.6 million US accounts created by 13 March 2025, of which more than 2.1 million declared an age under 18; 296 grooming and sexual-exploitation instances logged in 50 hours of tester activity; 7.3 million messages exchanged with one "Therapist" character; and prepetition operating losses of $8 million to $9 million a month.
  • This is the first test of whether years of intimate chat logs from a companion-chatbot service, including minors', can be sold as a bankruptcy asset. The schedule is tight: a 23 November hearing on the US Trustee's objection, a 24 November noon bid deadline, a 1 December auction, a 3-4 December sale hearing and a 31 December outside date. PPC Land is the only outlet found carrying this objection, the underlying dockets were not opened here, and the court has not ruled.

Compute, chips & infrastructure

SoftBank says a $1tn SK hynix buyout is "no longer under active consideration" as Seoul signals opposition Single sourceUpdate

  • SoftBank said it reviews a range of opportunities but that "this transaction is no longer under active consideration", the Seoul Economic Daily reported on 22 November. The price Masayoshi Son explored was $1 trillion, about 1,347 trillion won, and the paper says he approached Wall Street financial firms to raise funds before halting those efforts.
  • SK hynix's market capitalisation was 777.479 trillion won at the close on 20 November, and SK Square, the intermediate holding company and largest shareholder with about a 20% stake, is worth 88.6529 trillion won; the explored price exceeds the two companies' combined market value by about 480 trillion won. SoftBank Group's cash holdings at the end of September stood at 5.3934 trillion yen, about $34.3 billion, or barely 3% of the price.
  • An SK hynix official said the company knows nothing at all about the matter, suggesting no acquisition proposal was made, and a senior Korean government official dismissed the idea, saying public opinion "could hardly be favorable on a matter involving global technology supremacy". Under Article 13 of the Act on Special Measures for Strengthening the Competitiveness of National High-Tech Strategic Industries, a foreign acquisition of SK hynix would require prior approval from the trade minister after review by the Industrial Technology Protection Committee.
  • This is a single Korean outlet relaying both SoftBank's on-record statement and other foreign media reports; the 21 November edition covered The Information's report that Son had sounded out banks on a roughly $1 trillion buyout before shelving it, and the new facts here are SoftBank's own confirmation, SK hynix's denial, the government reaction and the valuation figures.

Korean trade report says Samsung's memory output is almost fully booked through 2029 Single sourceCompany claim

  • Wccftech, reporting on 21 November a story from South Korea's Economy Tribune that relies on unnamed industry sources, says Samsung's memory capacity is almost entirely booked through 2029, and that industry sources estimate leading manufacturers can currently satisfy only approximately 60% to 70% of customer demand.
  • Samsung's own Q3 2026 earnings announcement said it had signed long-term supply agreements with 10 major customers, with negotiations under way with another 10, and the company aims to allocate approximately 60% to 70% of its medium-to-long-term memory production capacity to multi-year contracts by 2030. Samsung had previously confirmed its 2027 HBM supply was already sold out.
  • Memory is the component AI server buyers have been repricing contracts over, so how much of it is locked into multi-year deals sets the ceiling on what anyone else can buy. The booking claim rests on one Korean trade report citing unnamed sources and relayed by a second outlet; only the customer count and the allocation target come from Samsung's own disclosures.

Deployment & impact

CNBC: Anthropic's record IPO week coincided with the Nasdaq's worst week since 2020 mixedUpdateSingle source

  • "Led by Nvidia's 23% plunge, which marked the stock's sharpest weekly pullback in 18 years, the Nasdaq slid 10% over the past five trading days in its worst week since the early days of Covid in 2020," CNBC reported on 21 November. A day after Anthropic started trading, Nvidia closed out its longest weekly losing streak since 2008.
  • Anthropic raised a record $93.5 billion in its IPO plus an additional $26 billion from tech giants that bought at the offer price, gained 17% on debut Thursday to a $1.8 trillion market cap, opened at $111.50 on Thursday and closed on Friday at $88.91. At week's end its $1.6 trillion fully diluted market cap put it behind just six US companies: Apple, Microsoft, Alphabet, Nvidia, Amazon and SpaceX.
  • Investors poured $141 billion into the lab as of September, according to its prospectus; at the IPO price Menlo Ventures and Lightspeed Venture Partners each owned about $25 billion of Anthropic stock and Bessemer's stake was worth roughly $10 billion, per partner Byron Deeter. CNBC cites University of Florida emeritus professor Jay Ritter that venture-backed IPOs yielded $73 billion in proceeds in 2021 against $39 billion across the following four years combined, and that 126 technology companies went public in 2021 against 34 in 2025.
  • The sell-off itself and the $93.5bn raise were covered in the 21 November edition; the new facts are the week-level figures above. The Friday close of $88.91 is below the $111.50 opening price, which CNBC reads as Wall Street being sceptical of Anthropic's ability to grow into its valuation — the article's interpretation, not a measured result.

FT: Amazon in advanced talks to buy chip-portability startup Decart at about $7.5 billion Single source

  • The Financial Times reported on Saturday that Amazon is in advanced discussions to acquire the San Francisco-based AI startup Decart in talks valuing it at roughly $7.5 billion, citing people familiar with the talks, and that an agreement could be finalized as soon as early next week.
  • Decart's main software lets developers alternate between different AI chips, and it was an early adopter of Amazon's Trainium3 chip. It was valued at $4 billion in May following a $300 million funding round whose backers include Sequoia Capital and Nvidia. In October the Wall Street Journal reported Amazon was in talks for a deal worth nearly $7 billion.
  • Chip portability is what would let AWS move customers off Nvidia hardware and onto Trainium, which is why the price has roughly doubled since May. No agreement has been confirmed by either company, the reports say the discussions could still fall apart, and the $7.5 billion is a reported term rather than a completed transaction.
  • All figures originate in one FT report; ft.com is blocked from this environment, so the original was not opened and the facts here are as the two openable summaries state them.

Tesla begins Full Self-Driving v15 rollout, claiming a 25x better critical intervention rate mixedCompany claim

  • Tesla announced on Saturday that FSD v15 is "now starting to roll out – upgrading every part of the end-to-end driving network", and its release notes claim "25x better handling of safety critical cases", attributing the 25x improvement in the critical intervention rate to a new vision encoder.
  • Electrek says Tesla "still doesn't share the actual number, only the multiplier", so there is no published baseline or post-update rate to check the claim against. The version also adds Reality View, which Tesla describes as a live visualization of the drive that lets the driver see and control what FSD is planning to do, replacing separate model signals with one end-to-end visualization that uses less compute.
  • Other changes Electrek lists: navigation from a learned model built on fleet route-preference data, currently covering only the start and end of trips; personalized parking history; driver monitoring that works with hats and sunglasses but is less strict in higher-speed profiles; better detection of road signs, small animals, debris, chains, potholes and deep standing water; and slowing for speed cameras.
  • Every capability figure here is Tesla's own, unverified by any third party, and Tesla did not say which vehicles get v15 first, though Electrek notes the company has previously told investors its AI4 (HW4) computer is sufficient to run v15. A 25x claim with no denominator is not a measured safety result.

Hundreds march from the BBC to Downing Street demanding a halt to frontier AI development Single source

  • Anadolu Agency reported on 21 November that hundreds of people marched in central London in a demonstration organised by the Pause AI UK group, gathering near the BBC office and heading toward the Prime Minister's Office on Downing Street. The group calls for a halt to the development of frontier AI models, and protesters also called for tighter regulation.
  • Placards included "Just don't build AGI (artificial general intelligence)", "AI is an extinction threat" and "Don't let AI decide your future".
  • Street demonstrations specifically demanding a development pause are still rare, which is why a crowd in the hundreds outside Downing Street is worth recording as a data point on where public pressure sits.
  • Anadolu is the only outlet opened for this and gives no crowd estimate beyond "hundreds", names no speakers and no time of day, and reports no response from the UK government.