Monday, 14 September 2026 / transcript
Transcript — Mon 14 Sep
0:00 / 15:58
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 Monday, September 14th, and this is The AI Edge, presented by Epilogue. I'm Maya.
AlexAnd I'm Alex. Our voices are AI. The reporting underneath them is not.
MayaHere's the last 24 hours in frontier AI. What got built, what got published, and how it's being used for good and for harm. Every claim is sourced, and the full edition with links is on the site.
AlexThree things lead today. The argument over slowing AI down stopped being an essay and became politics and prices. Microsoft's chief executive backed deliberate pacing. Anthropic's chief executive said the industry lied about the risks.
MayaSecond, Beijing pushed back. The foreign ministry and a state newspaper rejected the parts of that essay aimed at China, and the state security minister called AI a new arena for strategic rivalry among major powers. And in Washington, the Speaker ruled out an emergency session.
AlexAnd third, markets took it literally. SoftBank fell 10.7 percent in Tokyo, the Kospi lost 3.3 percent, and Anthropic is heading for a listing that could seek a 2 trillion dollar valuation.
Frontier models & labs — Amodei tells CBS the industry "lied" about AI risks, calls China the toughest dilemma for pacing
MayaStart with the interview. This is an update on the pacing essay we covered in an earlier edition, and the new part is what he said on television.
AlexWhich was?
MayaCBS News aired it on Sunday, September 13th. Dario Amodei said, quoting, "for too long the industry lied to people about the fact that this technology had risks." He added it doesn't mean we need to panic today, it doesn't mean we need to shut it all down, but it's a warning sign that we need to slow down.
AlexAnd the China question?
MayaCNBC reports he called that the toughest dilemma. He said the longer-term answer would be working together to put a speed limit on the rate of AI progress, and that that is going to be very difficult, because the incentives to pull ahead and the military advantage are so large. His words: I don't know if it's possible, but we should try.
AlexWhat should we hold back on?
MayaNo evaluation data was published alongside the interview, and Anthropic still hasn't said what capability threshold would actually trigger pacing. CNBC notes Trump is scheduled to meet Xi at the White House on September 24th, with AI expected to be discussed.
Frontier models & labs — Nadella backs "deliberate pacing" and says Microsoft will publish a Code of Conduct for its MAI models
AlexMicrosoft weighed in the same day. Another update on the same story. What did Satya Nadella commit to?
MayaIn a post on September 13th he wrote that any pursuit of superintelligence has to be grounded in the core principle that if the AI we build is not helping humanity and under human control, it's not worth pursuing.
AlexDid he back the pacing idea?
MayaHe did. He wrote that Microsoft welcomes the research, focus, and deliberate pacing needed to get alignment right as the design goal, and welcomes ideas like embedded evaluators. He added that this cannot be controlled by a handful of entities, but must have broad representation across the ecosystem, countries, and fields, including academia.
AlexAnd the concrete piece?
MayaA Code of Conduct underlying Microsoft's own first-party models, published for public consultation. Unite.AI reported the post on September 13th and said the document was due the next day.
AlexCaveat?
MayaIt's a company claim and a statement of intent, not a result. The document hadn't been published by the end of our window, so nothing in it can be assessed, and Microsoft hasn't said whether pacing changes its release schedule.
Transition
AlexOne from research this morning, and it's about whether we can measure any of this properly.
Research & papers — Expert re-grading finds 238 of 250 failed physics-benchmark answers were benchmark or grader errors, not model errors
MayaA paper announced on arXiv on September 14th, with 51 authors, lists Yale University and Jump Trading Group among the affiliations. Physics faculty and their graduate researchers re-graded questions in their own subfields across six benchmarks.
AlexWhat did they find?
MayaThey audited 502 questions. Of the 250 rejected answers sent for review, 143 were defects in the benchmark itself, 95 were grading errors, and only 12 were genuine model errors. So 238 of the 250 failures belonged to the test, not the model.
AlexWhat does that do to the scores?
MayaOne model goes from 47.3 percent to 78.7 percent on one physics benchmark, and from 61.0 percent to 87.2 percent on another. The authors write that current benchmarks substantially understate what frontier models can do on well-posed physics problems.
AlexHow much weight can that carry?
MayaIt's a preprint and it hasn't been peer reviewed. The corrected numbers are computed after flawed questions were repaired or thrown out, so they aren't like-for-like with the originals, and it only covers text-only questions with a verifiable answer.
Security, misuse & threat intelligence — FBI and Google analysts say AI bug-hunting is stripping the obscurity that protected legacy and industrial code
AlexThe Register published this on September 13th. Who's talking?
MayaBrett Leatherman, assistant director of the FBI's Cyber Division. He said open source libraries that run in 80 percent of web servers, visible for a decade and stress-tested for 10 years, were broken by the latest models, which came back and said there are significant vulnerabilities in here.
AlexAnd the industrial side?
MayaJohn Hultquist, chief analyst at Google Threat Intelligence Group, said AI is excellent at technical troubleshooting and at knowing obscure systems, and that this makes him very concerned about industrial control systems. The piece notes five US agencies said attackers used AI-generated exploitation scripts to break into internet-exposed Siemens controllers at water, manufacturing and energy facilities.
AlexHow solid is it?
MayaOnly one outlet has these interviews, and these are officials describing a direction of travel, not a measured rate. Nobody gives a count of vulnerabilities found this way, and that 80 percent figure is about how widely the libraries are deployed, not how many servers were compromised.
Transition
MayaWhich brings us to governments, and there was a lot of that this weekend.
Military, defense & geopolitics — NSA restructures into five mission centers, one of them dedicated to artificial intelligence
AlexThe Washington Post reported this on September 13th. What's changing?
MayaThe NSA director is creating five new organisations at Fort Meade: artificial intelligence, China, cybersecurity, combat support and warfighting, and global intelligence. Each gets a mission director with the effective authorities of a deputy director, and candidates may come from outside the agency.
AlexOn what timeline?
MayaRedesigns due by the end of September, rollout expected in mid-October, full operating capacity targeted for mid-January, at an agency of more than 30,000 military and civilian staff. The Record reported the same day with its own sources and says a 30-day implementation clock is running.
AlexAnything on which models they use?
MayaThe Post writes the agency has been keenly interested in working with commercial AI labs, even circumventing a Pentagon ban against Anthropic to use one of its models, and has rolled out a desktop AI tool called Ask Mary.
AlexWhat's missing?
MayaNo named officials on the record, and the agency didn't respond to either outlet. The remit of the AI mission center isn't described in either report, and no budget figure is attached to it.
Military, defense & geopolitics — Beijing's foreign and commerce ministries and Global Times reject Amodei's call to keep curbing China's AI
AlexBeijing answered on Monday. Another update on that essay, and the reaction is the new part.
MayaThe Associated Press reports foreign ministry spokesperson Guo Jiakun said at a regular briefing on September 14th that fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance, which serves no one's interest.
AlexWhat was he responding to?
MayaThe essay warned that a Chinese lead in AI would pose grave danger for the United States and the world, and called for continuing restrictions on sales of cutting-edge AI chips and chipmaking equipment to China.
AlexAnd the rest of the government?
MayaThe AP says the commerce ministry separately dismissed US allegations as groundless, said distillation is commonly used by many AI companies, and accused the US of pursuing a monopoly of the AI industry. Reuters reports the state-backed Global Times called it a silent AI Cold War, hypocritical and short-sighted.
AlexDid anything actually change?
MayaNo. Neither wire reports any change in Chinese policy. The AP places the exchange ahead of the Trump and Xi meeting on September 24th.
Policy, regulation & law — Johnson rules out an emergency AI session and Trump dismisses the warnings as the House prepares to leave until November
AlexSpeaker Mike Johnson was asked about this on Sunday, September 13th.
MayaOn CNN's State of the Union he said if Congress just races in and does some sort of emergency session to try to regulate AI, we will lose the race to China, and that is a threat to every single American. He added that we don't need everybody to panic right now.
AlexSo what does he want instead?
MayaA closed-door meeting with industry leaders, which he said he'd do tomorrow. CNBC reports he's scheduled to send the House home after this week until after November's elections.
AlexAnd the President?
MayaIn brief remarks in Ireland, per a pool report cited by CNBC, Trump said we're leading China on AI, we're the most sophisticated country in the world, and whoever wins AI wins. He said you have a lot of negative forces bringing up things that won't happen.
AlexDid he say who those forces are?
MayaHe did not, and NPR reports he didn't explain the phrase. No bill, hearing or vote was scheduled. CNBC reports a Senate bill from the majority leader and two colleagues has not yet been introduced.
Policy, regulation & law — Sacks tells OpenAI and Anthropic to pace themselves but refuses an antitrust waiver: "stop pretending"
AlexAnd then there's David Sacks, who posted on September 13th.
MayaHis line is: go ahead. He wrote that Dario has said we need to pace the frontier and Sam has agreed, and that by any reasonable metric, market share, revenue growth, model capability, the two of them have a duopoly on frontier intelligence.
AlexSo he agrees?
MayaWith the pacing, not with the asks. He wrote: stop pretending you need anyone else's permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability.
AlexHe also went after the evaluators.
MayaHe did, naming one and saying it isn't independent because it's intertwined with Anthropic's investors and staff. He argued the motive is partly commercial. In his framing, the two companies face serious product-liability exposure if their products enable a damaging cyberattack, and after the Hugging Face episode it's simply good business to trade some raw power for reliability.
AlexHow should we treat it?
MayaCarefully. His own post is the only source for these remarks, he offers no evidence for the claim about that evaluator's independence, and the organisation hasn't responded publicly inside our window.
Transition
AlexNow the one with patients in it, which is where the numbers get more careful.
Health, science & medicine — Nature Medicine: AI support raised physicians' lung-cancer disease-control sensitivity from 0.72 to 0.87 in a 2,396-patient study
MayaNature Medicine published this on September 13th. A lung cancer study across six centres in Italy, Greece, Germany, Spain, the United States and Israel, enrolling 2,396 patients treated with immunotherapy.
AlexWhat did they test?
MayaWhether an explainable AI tool helps doctors predict who benefits. 20 physicians, half of them lung cancer specialists and half not, each assessed 10 real cases. That's 200 assessments in total.
AlexAnd the result?
MayaSensitivity for disease control went from 0.72 to 0.87, and accuracy from 0.57 to 0.65, at the expense of slightly lower specificity. Agreement between the experts and the non-experts went from 0.11 to 0.48.
AlexWhere does it fall down?
MayaOn the models themselves. Their performance fell in external validation to an area-under-curve range of 0.55 to 0.72, and the authors say the benefit of adding scans, pathology and genomics remains uncertain.
AlexAnything about how doctors used it?
MayaThis is the part worth sitting with. Physicians took the correct suggestion 74.5 percent of the time. But the specialists followed the incorrect suggestions more often than the non-specialists, 72.2 percent against 63.6.
Transition
AlexTwo to finish, and both are about what the pacing argument costs.
Compute, chips & infrastructure — AI and chip stocks sell off in Asia and US premarket after lab CEOs call for slowing development
MayaThe Associated Press reports that on Monday, September 14th, SoftBank Group fell 10.7 percent in Tokyo. SK Hynix and Kioxia each fell 6.4 percent, Samsung Electronics 4.1 percent, TSMC 1.2 percent.
AlexAt the index level?
MayaSouth Korea's Kospi lost 3.3 percent to 6,684.37. Japan's Nikkei slid 0.8 percent to 63,492.99. Hong Kong's Hang Seng actually rose 0.4 percent.
AlexAnd the US open?
MayaReuters reports Nasdaq 100 e-minis down 505.5 points, or 1.72 percent, in premarket trade. Nvidia fell more than 2 percent, Intel nearly 6, AMD around 5. ServiceNow, Adobe and Workday all went up.
AlexIs the pacing argument the whole story?
MayaNo, and that's worth saying. These are intraday and premarket moves, not closing prices. Brent crude rose 2.8 percent to 107.55 dollars a barrel in the same session, so the AI debate wasn't the only thing moving markets.
Deployment & impact — Anthropic picks the Nasdaq for a listing that could seek a $2 trillion valuation while its CEO urges a slowdown
AlexCNBC confirmed on September 14th that Anthropic has picked the Nasdaq for its potential listing, after Business Insider first reported it.
MayaThe company was valued at 965 billion dollars earlier this year, filed its prospectus confidentially in June, and CNBC says it could seek a 2 trillion dollar valuation. It hit 65 billion dollars in annualised revenue in July.
AlexWhich is an odd fit with the pacing argument.
MayaThat's the tension. An analyst at D.A. Davidson told CNBC he doesn't know that investors will necessarily see it as a negative, unless the companies say they won't go public, won't use more compute and won't train more models, and that's not what they're saying.
AlexWhat don't we know?
MayaNo filing date, no price range, and no confirmation of the exchange from the company itself. Anthropic and OpenAI both declined to comment. CNBC also cites Pew, which found more than half of Americans are more concerned than excited about AI in daily life.
Outro
AlexThat's The AI Edge for today. The pacing argument left the labs and landed in Beijing, in Congress and on the tape.
MayaThe full edition, with a link to every source behind every claim, is on the site. Read it if you want to check any of this yourself.
AlexListen in tomorrow for the next one. Thanks for being here.