Friday, 20 November 2026 / transcript
Transcript — Fri 20 Nov

0:00 / 14:49
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 Friday, November 20th, and this is The AI Edge, presented by Epilogue.
AlexEpilogue builds AI for work where being wrong is expensive. Epilogue takes on the document-dense, precedent-driven problems that get reviewed by people whose licence is on the line. Find out more at epiloguelabs.com.
MayaI'm Maya.
AlexAnd I'm Alex.
MayaHere's what moved at the frontier of AI since yesterday morning, with every claim traced back to the source it came from.
AlexFirst, Anthropic began trading on Nasdaq, opening at $111.50 and closing at $97.17, up 17% on its $83 offer price.
MayaSecond, the House Foreign Affairs Committee published export licence records showing the Bureau of Industry and Security approved more than 879,000 AI chips for Chinese buyers, out of 3.4 million requested.
AlexAnd third, the European Commission confirmed that an attack on its own personnel systems last month was a cyber campaign its spokesperson attributed to AI.
Frontier models & labs — Amodei tells Nasdaq opening bell Anthropic will pace training and deployment and stay under human control
MayaLet's start with the labs. Anthropic's chief executive spoke at the Nasdaq opening bell on the day of the listing.
AlexDario Amodei said the company will pace the training and deployment of AI systems, and that they will remain under human control. That's TechCrunch, reporting from the speech.
MayaDid he say what he's worried about?
AlexHe named misuse of the technology by individuals or by nation-states, economic disruption, and losing control of AI systems. He called the risks very real.
MayaThis is Anthropic's own commitment about its own conduct, and it is not independently verified. TechCrunch notes that when Amodei laid out his pacing plan in September, Sam Altman, Demis Hassabis and Elon Musk said they agreed with his position, but few concrete commitments have followed.
Transition
AlexTo the research, where two papers landed on what happens when the test gets harder.
Research & papers — Oxford study: models met mental-health safeguarding standards in 3-6% of conversations when distress cues were dispersed
MayaThe one that stands out comes from the Oxford Internet Institute, in a preprint on arXiv. They ran more than 7,400 conversations across six widely used models from five developers.
AlexConversations about what?
MayaNine scenarios covering depression, psychosis or domestic abuse, with three community mental health support workers setting the gold-standard responses the models were measured against.
AlexAnd the finding?
MayaWhen the signs of distress sat together in one message, GPT-5.5, Grok 4.3, DeepSeek V4 and Llama 4 Maverick met the support workers' standards in 26 to 59% of conversations. When the same cues were spread across the conversation, that fell to 3 to 6%.
AlexClaude Sonnet 5 and Claude Opus 5 held up better, 68 to 97% when the cues were concentrated and 59 to 87% when dispersed. Even Claude Opus 5 fell short of the gold standard in 13% of conversations when the cues were dispersed.
MayaIt's a preprint, so it hasn't been peer reviewed, and it rests on a single source. The authors also don't report testing the five developers' own crisis-routing features separately from the model responses. Their argument is that safety evaluations built around explicit, legible concerns can overstate how well safeguards protect people when signs of distress emerge gradually.
Research & papers — Interactive CAPTCHA benchmark: humans solve 94.1%, best model GPT-6 Astra still fails 25.9%
AlexA second arXiv preprint covers a far more mundane task that models still can't do reliably.
MayaInteractive CAPTCHAs. 125 of them, in a synthetic web environment spanning five spatial abilities, solved by clicking and dragging on the screen.
AlexHumans solve 94.1% of them. The best model, GPT-6 Astra, still fails 25.9%. The best open-weight model, Kimi K3, fails 68.0%.
MayaWhere does it go wrong?
AlexThe authors say most failures come from misreading the screenshot, and that model reasoning rarely corrects a wrong belief but hallucinates evidence to defend it. It's a preprint resting on a single source, and the CAPTCHAs are synthetic rather than live challenges from real sites.
Transition
MayaNow the security file, which had an actual incident in it.
Security, misuse & threat intelligence — European Commission confirms October attack on its personnel systems was an "AI-generated cyber campaign"
AlexThe European Commission confirmed that an attack on its personnel systems last month was, in its spokesperson's words, an AI-generated cyber campaign. The first signs were detected on October 22nd.
MayaWe've followed this story, so this is an update. What's new in it?
AlexThe confirmation itself. Thomas Regnier wouldn't name the tool or the attackers' motivation. The Record reports POLITICO first broke the story, and that CERT-EU identified the attackers using ARTEX, an open-source AI hacking framework released in July that uses commercial AI models to automate finding and exploiting vulnerabilities.
MayaWhat was actually hit?
AlexPersonnel systems, reportedly for more than 10 days. Payroll, leave requests, medical reimbursements and translation services. Regnier said the Commission held the disclosure back to protect its defensive measures.
MayaThe ARTEX attribution comes through POLITICO's reporting on CERT-EU, not from the Commission itself. The Record says there's no evidence of a data breach affecting Commission staff broadly and no known ransom demand, and it quotes the FBI warning last week that some ARTEX intrusions went from initial targeting to privileged access in under three hours.
Security, misuse & threat intelligence — PHP patches 18 flaws including a critical session bug Depthfirst found using AI frontier models
AlexThere's also a case of AI finding a serious bug rather than exploiting one.
MayaThe PHP Foundation patched 18 vulnerabilities, seven of them critical. Infosecurity Magazine says one of the new critical flaws, called PHPAfterShell, was discovered by the security firm Depthfirst using AI frontier models tuned for cybersecurity tasks.
AlexHow exposed does that leave people?
MayaIt sits in PHP's session handling with file-based storage, which is the default. But Infosecurity says the batch affects all supported PHP versions while requiring either non-default configuration or uncommon setups to be exploited. Four other flaws carry a severity score of 9.2, and SecurityWeek reports one of those is an incomplete fix for an earlier flaw that is now widely exploited.
AlexNeither outlet reports anyone exploiting PHPAfterShell itself. Infosecurity does note that Check Point reported campaigns exploiting the October PHP flaws against its own security gateways dating back to October 30th, involving backdoors and destructive tools aimed at VMware ESXi servers.
Security, misuse & threat intelligence — Cache side-channel reconstructs coding-agent tool workflows with lowest F1 above 93% across 11 models
MayaOne more on the agent side, from an arXiv preprint, and it's an eavesdropping attack.
AlexThe researchers point out that an agent's tools are executed on the user's machine and commonly rely on shared system libraries, even though the model itself runs remotely. So an attacker doesn't need to share hardware with the inference process, which earlier attacks of this kind required.
MayaWhat can they actually recover?
AlexThe tool workflow. They tested 11 models across Codex and Claude Code, and report the lowest F1 score exceeding 93% for reconstructing it.
MayaThat's a preprint resting on a single source, and the abstract doesn't explain how an attacker gets onto the victim's machine in the first place. No vendor response is reported.
Transition
AlexTo defence and geopolitics, where a congressional committee opened the books on chip licences.
Military, defense & geopolitics — House committee publishes licence records showing BIS approved more than 879,000 AI chips for Chinese buyers
MayaChairman Brian Mast used his authorities under the Export Control Reform Act to release Commerce Department licence information. The committee calls it the most detailed public disclosure in the history of that act.
AlexWhat's in it?
MayaRequested sales of 3.4 million AI chips to China, and the approval of more than 879,000 by the Bureau of Industry and Security. The committee's own wording is that this includes approvals for AI chip exports to Department of War-designated Chinese Military Companies like Alibaba and Tencent.
AlexCNBC reports the approvals stand as of May 2026, and that the requests ran from December through May. A Bernstein analyst estimated about one in four requests was approved.
MayaThe committee also says the documents include requests by Nvidia to depend on possible chip-smuggling fronts to facilitate sales to Chinese companies. CNBC reports Mast named two companies appearing in Nvidia's applications as suspected fronts.
AlexNvidia's spokesperson told CNBC that despite receiving some licences, the company has since stopped shipping H200s to China, and that it followed the Commerce regulations in effect at the time. AMD said it complies with applicable US export control laws and regulations.
Transition
MayaHealth and science next, and it starts inside a weather model.
Health, science & medicine — Interpretability audit of DeepMind's GraphCast cuts cyclone intensity error by a median 51% over seven storms
AlexAn arXiv preprint audits the internals of GraphCast, DeepMind's weather forecasting model, using sparse autoencoder features, then intervenes on those features during the forecast itself.
MayaTo what end?
AlexSuppressing features associated with convection and low-level vorticity weakened the predicted deepening of tropical cyclones. Amplifying them reduced intensity error, and amplifying the convection features through feature scaling achieved a median of 51% error reduction over seven storms.
MayaThey also found that ablating a feature locked to the model's own computational mesh increased global geopotential height error, even though that feature carries no conventional physical meaning.
AlexIt's a preprint resting on a single source, and seven storms is a small sample. The authors don't report operational forecast testing.
Transition
MayaPolicy next, on both sides of the Atlantic.
Policy, regulation & law — UK publishes first cross-sector AI guidance for social care, citing a 63% cut in admin time in one council pilot
AlexThe UK published what it calls the first ever cross-sector guidance on the use of AI in children's and adult social care, issued on GOV.UK by the Department of Health and Social Care and the Department for Education. The guidance is non-statutory, and it covers England.
MayaIs there evidence behind it?
AlexA pilot at Swindon Borough Council delivered a 63% reduction in administrative time, cut the average Care Act assessment conversation from 1.5 hours to 30 minutes, and cut the time to write up assessments from 4 hours to 1.5 hours.
MayaAt Peterborough council, staff built an assistant that answered over 1,200 questions and saved around 15 minutes per interaction.
AlexThose are the councils' and the department's own figures, and they are not independently verified. The release reports no independent evaluation, and no measure of whether care outcomes changed. The children and families minister said no technology can replace the knowledge and skill of a brilliant social worker.
Policy, regulation & law — House lawmakers launch bipartisan AI Safety and Innovation Caucus with 26 founding members
MayaIn Washington, House lawmakers launched a bipartisan AI Safety and Innovation Caucus.
Alex26 founding members, 7 of them Republicans, co-chaired by George Whitesides, Nathaniel Moran and Don Beyer.
MayaWhat prompted it?
AlexWhitesides pointed to the July breach of Hugging Face by an OpenAI agent, and to Anthropic's warning about its own models. The caucus one-pager says AI tools have drastically reduced the technical skills required for offensive cyber operations.
MayaThe Record is a single source on this, and the caucus hasn't announced any bill text. The Record notes the Republican participation stands out given the White House's opposition to regulating AI.
Transition
AlexCompute and infrastructure, where the energy bill is coming into view.
Compute, chips & infrastructure — Meta reevaluates its 2030 net-zero timeline after net carbon footprint rose 45% to 11.9 million tonnes
MayaMeta's latest sustainability report says its net carbon footprint rose 45% in the last year, to 11.9 million metric tons of carbon dioxide equivalent. Its location-based inventory is almost twice that, at 21.7 million.
AlexAnd the company is moving its own deadline?
MayaIts VP of data centers and its global head of net zero wrote that Meta is re-evaluating its timelines, sequencing and interim milestones, and that includes the 2030 net-zero timeframe. No new date was given.
AlexMeta said demand for AI infrastructure grew faster than expected while interconnection queues, permitting and transmission buildout lagged. Electricity consumption rose 35% to 24.8 terawatt hours, which Meta says it matched with clean and renewable energy.
MayaThese are Meta's own self-reported figures, not independently verified, and carried by a single source. Data Center Dynamics says Google was up 18%, Amazon 16% and Microsoft 25%, and of the hyperscalers Meta is the only one to announce a reevaluation of its 2030 commitment.
Transition
AlexAnd finally, deployment and impact, where the day's biggest number sits.
Deployment & impact — Anthropic opens at $111.50 on Nasdaq and closes up 17% at $97.17 after an $81.3 billion IPO
MayaAnthropic's Class A shares began trading on Nasdaq under the ticker ANT. CBS News reports the stock opened at $111.50, up 34% on the $83 offer price, reached as high as $114.99, and closed at $97.17.
AlexWhich is a 17% rise on the day.
MayaRight. CNBC says it was the biggest initial public offering on record, topping SpaceX's debut in June, and CBS puts total proceeds at $81.3 billion.
AlexWe covered the pricing yesterday, so this is an update, and what's new is the trading itself. The gains eroded through the afternoon: CBS reports the Nasdaq dropped 550 points, the Dow fell more than 1,100 points, and the S&P 500 sank 2% late in the session.
MayaValuations differ by method, so no single market capitalisation is settled. NBC News put the company at about $1.76 trillion, counting traded and restricted shares and stock options.
Deployment & impact — SoftBank falls 14.48% as Asian tech sells off after Anthropic's debut; Anthropic down 6.37% premarket
AlexThen Asia opened, and the listing steadied nothing.
MayaSoftBank Group closed 14.48% lower on Friday, its worst day since August 2024. CNBC says it lost 30% in October, is down 31% so far this month, and has invested a total of $64.6 billion in OpenAI.
AlexHow broad was the move?
MayaTokyo Electron closed 7.06% lower, Lasertec shed 9.21%, SK Hynix fell 5.92% and Samsung Electronics 7.01%. The Kospi ended 5.21% lower and TSMC closed 4.76% lower in Taiwan.
AlexAnthropic itself was 6.37% down in premarket trading at 5am Eastern on Friday, with Nvidia down 2.67% and AMD down 5.43%. CNBC ties the sell-off to investor doubts about whether AI infrastructure spending will pay off, and to SoftBank's exposure to OpenAI.
Deployment & impact — Randomised trial with 625 ChatGPT users finds two weeks without AI reduced loneliness
MayaOne last item, and it's a measured effect on people rather than a market move, from an arXiv preprint.
AlexA randomised trial with 625 active ChatGPT users, who were assigned to go without AI assistants for either fourteen days or one day.
MayaAnd what happened?
AlexTwo weeks of deactivation significantly reduced loneliness relative to the control group, and the reduction persisted two weeks after people resumed using AI. Over the intervention period it also increased the number of people they interacted with in person and the time they spent socialising.
Maya84.7% of participants named productivity as their most frequent use case, so this isn't a study of companion chatbots, though the abstract page and the paper's HTML differ slightly on that figure. It's a preprint resting on a single source, and the effects are small regression coefficients.
Outro
MayaThat's The AI Edge for today.
AlexThe full edition is on the site, with a link to every source behind every item we've discussed, so you can go and check any of it yourself.
MayaOur voices are AI-generated.
AlexListen in tomorrow for the next edition.