Saturday, 21 November 2026 / transcript
Transcript — Sat 21 Nov

0:00 / 14:34
Maya and Alex are AI voices. Each part of the conversation below comes from one item in the written edition — linked above it — and is checked automatically before publishing: every number must appear in that item, every caveat the edition raises must be said aloud, the source must be named, and speculative or hyped language is rejected.
Intro
MayaIt's Saturday, November 21st, and this is The AI Edge, presented by Epilogue.
AlexEpilogue works the way this briefing works. Epilogue checks every claim against the primary source, quotes every figure exactly as the source wrote it, and leaves out whatever doesn't reconcile. Visit epiloguelabs.com to learn more.
MayaI'm Maya.
AlexAnd I'm Alex.
MayaHere's what happened at the frontier of AI in the last 24 hours, with every claim traced to a document we opened.
AlexFirst, a mathematician at Anthropic named Levent Alpöge posted a note on Friday saying that he and Claude have discovered a proof of the Riemann hypothesis, and nobody has verified it.
MayaSecond, the AI trade broke. The Associated Press reports the S&P 500 fell 3.7% on Friday into a correction, and the Nasdaq composite went negative for the year.
AlexAnd third, Amazon says a Chinese-nexus attacker's autonomous AI agents found a zero-day in an open-source medical imaging application on their own, and used it to take root on a university server.
Transition
MayaLet's start with the labs.
Frontier models & labs — Thinking Machines releases Inkling 2 Mini, a 26.4B open-weights looped transformer that runs in about 20GB
AlexThinking Machines Lab released a small open-weights model, Inkling 2 Mini, whose architecture reuses its own middle layers.
MayaMeaning what?
Alex26.4 billion total parameters, 3.4 billion active, and two middle blocks of six layers that run three times for every token, giving 48 layer applications from 24 stored layers. It fits in about 20GB, which the company says lets it run on a consumer laptop.
MayaAnd the benchmarks?
AlexIt says it leads every open-weights model under 100 billion parameters on benchmarks such as DeepSWE v1.1 at 48.2% and Vals Index v2.2 at 46.5%. On whether looping hides the reasoning, a research variant's no-chain-of-thought horizon rose from 16 seconds at one pass to 37 seconds at three.
MayaAnd those are its own numbers.
AlexAll of them — company-reported, not independently verified, and it calls the monitorability work preliminary.
Transition
MayaNow the research.
Research & papers — Anthropic mathematician Levent Alpöge says he and Claude have a proof of the Riemann hypothesis, released only in Lean
MayaLevent Alpöge, a mathematician at Anthropic, posted a note on Friday that reads, quote, I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis. He says the method also resolves the Generalized Riemann Hypothesis.
AlexSo where's the paper?
MayaThere isn't one. He released only a Lean formalization, saying it would be inappropriate to post an unreadable document, and noting he hit 250,000+ pages. Nature says the four-page note he did publish includes a sketch of the proof.
AlexAnd the reaction?
MayaNature quotes James Maynard, a Fields medallist at Oxford, saying that if this fully checks out, it is clearly a historic milestone in mathematics. Andrew Granville of the University of Montreal told Nature, quote, what he released is garbled, and tied the timing to the fall in Anthropic's share value on November 19th. Alpöge says that reading is false.
AlexHas anyone checked it?
MayaScientific American reports Andrew Sutherland of MIT compiled the Lean result and called it a historic moment, while adding there is much work to be done to understand the mathematical details. It's not peer reviewed and not independently verified, and Nature says Anthropic did not immediately respond to a request for comment.
Research & papers — Vakil announces an all-human proposed proof of the Generalized Riemann Hypothesis; AI models then help find problems in it
AlexAnd the same problem got a second claim the same day, from the opposite direction. Ravi Vakil, president of the American Mathematical Society, posted on November 20th that a Stanford mathematician, Alessandro Maria Masullo, has a proposed proof of the Generalized Riemann Hypothesis.
MayaBuilt with AI?
AlexNo. Vakil's words are, quote, it is all human, with no AI. Brian Conrad at Stanford had been reading it closely for months and giving extensive feedback.
MayaAnd then there was an update.
AlexAbout eight hours later. Vakil appended a note saying several people had found problems in the proposed proof, and that those problems were identified with the assistance of different AI models.
MayaSo here the AI is the referee rather than the author.
AlexScientific American says Conrad confirmed the post's details and, asked whether the work holds up, said it's undergoing major revision. It's a preprint, and Vakil makes no claim about whether the strategy is correct.
Research & papers — METR says model time horizons are doubling roughly every 3.61 months, reaching up to 27 hours on verifiable tasks
MayaMETR published a capability update. Time horizons are doubling roughly every 3.61 months across its combined suites, with a 95% confidence interval from 2.08 to 4.94, and it estimates current public frontier models reach up to 27 hours on the verifiable subset of its newer tasks.
AlexHow does that compare against people?
MayaOn 15 programming contests, the median placement of the strongest models METR tested was the 96th to 98th percentile of human participants. The models had 48 hours on a 28-core machine; the human participants had about three months.
AlexAnything cutting the other way?
MayaTwo things. On one speedrun the best agent run reached a 15.7% speedup over METR's starting record in 48 hours, where a human record-holder reported about 60 hours of work for a 41% speedup over the previous record. And METR reports that AI graders are overgenerous when scoring tasks written by models stronger than themselves.
AlexAnd this is METR's own work.
MayaIt is, and it's not peer reviewed.
Transition
AlexNow to security.
Security, misuse & threat intelligence — Amazon says a Chinese-nexus actor's autonomous AI agents found a zero-day on their own and took root on a university server
AlexOn September 10th, 2026, Amazon Threat Intelligence found a server running a publicly hosted, unauthenticated instance of an open-source autonomous penetration-testing framework, exposing the operator's prompts and the agents' output to anyone on the internet.
MayaWhat had the agents been told to do?
AlexVariations of, quote, obtain shell access, against IP addresses of universities in East Asia outside mainland China, over roughly 48 hours beginning September 8th. The prompts were in Simplified Chinese and the script timestamps were set to China Standard Time.
MayaAnd then one of the attempts worked.
AlexAgainst an open-source medical imaging application, the agents mapped the API, found an upload endpoint with a filename parameter and tested it for path traversal. Amazon says that in about 11 minutes they went from mapping the API to arbitrarily writing files on the server, later writing a reverse shell to get root.
MayaAmazon calls it a zero-day?
AlexWith high confidence, based on the absence of any corresponding fix, issue, advisory or prior mention in vulnerability databases. Moderate confidence on the Chinese-nexus attribution. The Register names the software as MONAI Label and says the maintainers released a fix; Amazon's own post doesn't name it.
MayaThe failed attempts are the ones I'd look at.
AlexThose are the real measurement. Roughly 16 hours and 148 sequential actions against an unauthenticated Redis instance, including a failed Log4Shell injection, all unsuccessful. And more than 6 hours against a university GPU server, where the agents built about 80 real usernames, tested more than 30,000 credential combinations, and never found a working login. This is Amazon's own report, and these are its own numbers.
Security, misuse & threat intelligence — Bell Canada says an "agentic threat actor" breached its systems on 17 November and took enterprise customer data
MayaAnd a Canadian carrier's breach statement used the word agentic.
AlexIn what context?
MayaBell told BleepingComputer it is investigating an incident on November 17th in which, quote, an agentic threat actor accessed some of our systems, including a system used for ordering and provisioning circuits, and exfiltrated some enterprise customer data.
AlexWhat might have been taken?
MayaPer a notification a customer shared, an unauthorized party reached a pre-production environment of Bell's Maximo system, and the possibly stolen information includes business contact details, information in ticket titles, and some IP addresses tied to network elements. Bell says no credentials, no highly sensitive personal information, and no impact to customer services.
AlexDoes it say who, or what it means by agentic?
MayaNeither. Its words are, we currently do not know the attribution, and it hasn't named any tooling. BleepingComputer reads the word as suggesting AI-powered tools were used. The company says all of this itself, and only one outlet has it.
Transition
AlexNext, defense and geopolitics.
Military, defense & geopolitics — House chairmen urge removing PRC open-weight AI models from federal systems and federal contract work
AlexTwo House committees published a joint investigation into American companies running Chinese open-weight models, and it names three of them.
MayaWhich?
AlexAirbnb, which it says uses Alibaba models in customer support. Anysphere, which it says built the flagship proprietary model in its Cursor coding platform on a Moonshot AI base model. And DoorDash, which it says reported that multiple Chinese model families had been authorized and deployed.
MayaAnd what do the chairmen want done?
Alex12 recommendations. Bar Chinese-developed models from federal systems, federal contract work and high-risk critical infrastructure. Build an American open-weight initiative, create a model provenance standard, and hold what the report calls illicit distillers accountable.
MayaThe foreword is blunt.
AlexIt says federal investigations found that major Chinese developers used proxy networks and fraudulent accounts to harvest outputs from American frontier models, and that Beijing is using stolen intelligence to close the gap. The Record puts the report at 50 pages. These are recommendations. Nothing has been adopted.
Transition
MayaHealth and science next.
Health, science & medicine — Seven large drugmakers tell Reuters AI cut early-stage development time by as much as half this year
MayaSeven large drugmakers tell Reuters that AI, according to their top scientists, has sped up early-stage drug development work this year, cutting by as much as half the time it takes to produce new molecules and move them into the clinic.
AlexAny specifics?
MayaSanofi says it saved at least a year on an asthma treatment by skipping a dose-selection study. Novo Nordisk says it cut design-make-test cycles from around eight to three, and the time per cycle from eight weeks to four.
AlexIs it showing up in the pipeline?
MayaBarely. Reuters says researchers counted 129 AI-enabled drug candidates across 69 companies in March, up from around 117 eight months earlier, and most are still pre-clinical.
AlexAnd what did the people counting them make of that?
MayaReuters quotes Waqas Haque, an oncology fellow at the University of Chicago who worked on the count. He said he doesn't know whether those drugs will be more successful than non-AI drugs, and doesn't think they'll be approved any quicker, because of bottlenecks with human trials and regulators. Every timeline figure is company-reported and not independently verified, and only one outlet has the story.
Transition
AlexOn to policy.
Policy, regulation & law — GAO sets out 16 policy options on AI in cybersecurity, including requiring human authorization for consequential actions
AlexThe Government Accountability Office laid a menu in front of Congress. 16 policy options on AI in cybersecurity, across four goals.
MayaWhich ones matter most?
AlexTwo would bite directly on autonomous security tooling. One is to require human authorization for consequential AI-initiated cybersecurity actions. The other is to restrict autonomous and nondeterministic AI in consequential functions pending certification.
MayaAnd the underlying problem?
AlexIt found there are no widely accepted evaluation criteria for assessing offensive or defensive capabilities of generative AI. It points at two incidents, including a June case where a threat actor used AI to build a testing framework to refine malware for evading detection.
MayaAnd none of this binds anyone.
AlexNone of it. These are options, not recommendations, and a single source has published it.
Transition
MayaNow compute and chips, where the week ended badly.
Compute, chips & infrastructure — Nvidia falls 14% on report that SEC enforcement staff are examining its accounting, capping a seven-week slide
MayaNvidia fell 14% on Friday, after a Reuters report that SEC enforcement staff are scrutinizing potential accounting issues at the company, according to two people briefed on the matter.
AlexHow much does Reuters actually know?
MayaNot much. It says it could not determine when the inquiry opened or which specific issues are under review. The SEC declined to comment and Nvidia didn't immediately respond.
AlexAnd the scale of the slide?
MayaCNBC says that was a seventh straight down day and a seventh consecutive losing week, the longest such stretch since 2008. Down 46% over the period, and about $2.6 trillion of market value gone.
AlexIs there a specific question in the reporting?
MayaReuters notes that last year Nvidia began reporting geographic revenue by customer headquarters rather than billing addresses, a change the company said better reflects its revenue geography, and which removed Singapore from its geographic breakdown. Reuters does not say whether that change is part of the inquiry, and it says plainly that investigations are not evidence of wrongdoing.
Transition
AlexAnd finally, deployment and impact.
Deployment & impact — Anthropic closes its IPO at about $93.5bn gross after underwriters exercise the full 147 million-share option
AlexAnthropic closed its offering, and the underwriters took the whole overallotment.
MayaWhat are the numbers?
AlexAn aggregate of 1,127,000,000 Class A shares at $83.00 each, including the full exercise of the option to buy an additional 147,000,000 shares. Gross proceeds before discounts and expenses of approximately $93.5 billion, of which approximately $91.6 billion was received by Anthropic.
MayaAnd there was a private placement alongside it.
AlexApproximately $26.0 billion of Class N shares at the same price, to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners. Combined, approximately $119.5 billion.
MayaThose buyers are its own suppliers.
AlexIts chip, cloud and memory suppliers, yes. And that extra block sold while the shares were falling. These are company-reported figures, and the release states no valuation.
Deployment & impact — S&P 500 falls 3.7% into a correction and the Nasdaq composite goes negative for the year as AI stocks sell off
MayaBack to the market, because Friday was the worst of it. The Associated Press says a sell-off in AI stocks pulled the S&P 500 down 3.7%, its worst weekly loss in more than 18 months, and knocked the Nasdaq composite into the red for the year.
AlexHow far down is the Nasdaq?
MayaThe Nasdaq composite slumped 5.3%. The Dow, with less technology exposure, fell 1.9%. And the S&P is now in a correction, more than 10% below its October 6th high, its first since March 2025.
AlexWhat about Anthropic itself, two days into trading?
MayaDown with everything else. Yahoo Finance reports the shares fell nearly 8.5% on Friday to close just below $89, still above the $83 IPO price, for a market capitalisation of $1.17 trillion.
AlexAnd AP doesn't pin it on any one disclosure.
MayaIt attributes the move to doubts about AI valuations and spending, and it does not quantify how much of the fall is attributable to AI names. This is market pricing, not a measurement of AI capability or revenue.
Outro
AlexThat's The AI Edge for today. A day when a claimed proof of the Riemann hypothesis arrived as formal code, with no conventional write-up.
MayaThe full edition is on the site, with a link to every source behind every claim we made.
AlexOur voices are AI-generated.
MayaListen in tomorrow for the next edition.