Sunday, 20 September 2026 / transcript
Transcript — Sun 20 Sep
0:00 / 16:12
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 Sunday, September 20th. This is The AI Edge, presented by Epilogue.
AlexThe last day in frontier AI. What shipped, what was measured, and how the technology is being used, for good and for harm. Every claim here points back to a source.
MayaI'm Maya.
AlexAnd I'm Alex. Both of our voices are AI-generated, and nothing we say goes beyond what's in the written edition.
MayaThree things lead today. President Trump says he's forming an AI Force, modelled on the Space Force, and that he'll name an AI czar. NBC News reports he called warnings that AI could destroy humanity a hoax.
AlexSecond, Nvidia's Jensen Huang told CBS News that we should go as fast as we can irrespective of anybody else. CNBC reports Nvidia's revenue reached $215 billion in the latest fiscal year, up from $17 billion in 2021.
MayaAnd third, a New York Times investigation found DraftKings scored gamblers by how much they were expected to lose, then aimed about $400 million of automated promotions at the highest scorers.
AlexLet's get into it.
Frontier models & labs — Nvidia's Jensen Huang tells CBS News AI should go "as fast as we can", rejecting slowdown calls
AlexStart with Nvidia. Jensen Huang was asked by CBS News whether AI development needs to slow down.
MayaHis answer was that we should go as fast as we can irrespective of anybody else. He added, we would never ever, and never should, ship products before they're ready, deliver products that are unsafe.
AlexThat is the sharpest public rejection yet of Dario Amodei's call to slow capability gains.
MayaCNBC puts the context around it. Nvidia's revenue surged to $215 billion in the latest fiscal year, up from $17 billion in 2021. And at a House hearing on Monday, Treasury Secretary Scott Bessent said the president is completely aligned with Jensen Huang.
AlexAny counterweight in the reporting?
MayaHuang said in Scotland on Thursday that the answer to safety concerns is good old-fashioned engineering, and that there were incidents, and those incidents, thankfully, did no harm. Nvidia declined to comment to CNBC, and the White House didn't respond.
Frontier models & labs — Alibaba publishes Qwen-Image-2.1 with a 7B generation component under a research licence
AlexAlibaba published a new image model, Qwen-Image-2.1.
MayaThe Hugging Face model card lists 7 billion parameters in the visual generation component, across 32 single-stream layers. It generates transparent images, and it edits using up to 10 reference images, with local edits marked by circles or painted annotations.
AlexWhat do the benchmarks say?
MayaThere aren't any. The model card publishes no numbers and no comparisons, so every capability described is the company's own account, not independently verified. It's also a single source. One model card, no outside evaluation yet.
AlexAnd the licence named on the model card is the Qwen Research License Agreement, so check that before you build on it.
Transition
MayaFrom what shipped, to what Washington did about it.
Policy, regulation & law — Trump says he will form an "AI Force" modelled on Space Force and name an AI czar
AlexPresident Trump posted on Truth Social on Saturday that he is forming what he calls the AI Force, much like he did Space Force, and that he'll be announcing an AI czar in the near future. CNN has the post.
MayaWhat's the actual policy content?
AlexHe wrote that the administration will not in any way hinder or stifle the growth of the industry, and that misconduct can be handled with the already existing criminal and civil justice system. NBC News reports he characterised warnings that AI could destroy humanity as a hoax.
MayaIs the AI Force a military branch, or a figure of speech?
AlexUnresolved. CNN says it asked the White House whether it would be a branch like the Space Force, which Trump created in 2019 as the first new US military service since 1947, and that question is unanswered. No appointee was named.
MayaAnd there is no paperwork behind it. We checked the White House presidential actions listing: nothing dated September 19th or 20th, and the most recent entries are two September 18th items on visas. This is a social media post, not an executive order.
Policy, regulation & law — Newsom signs two California election-deepfake bills, extending the ban's sunset from 2027 to 2031
AlexContrast that with California, where something actually became law.
MayaAB 686, from Berman, was approved and chaptered on September 19th as Chapter 334. The Office of Governor Gavin Newsom says it extends the sunset date, from January 1st, 2027, to January 1st, 2031, on the prohibition on distributing materially deceptive audio or visual media.
AlexDeepfakes, in plain terms.
MayaRight. And AB 502, from Pellerin and Berman, was chaptered the same day as Chapter 333, with an urgency clause, so it takes effect immediately. The Governor's office says it modifies the law barring anyone, with malice, from knowingly distributing a campaign advertisement or election communication containing materially deceptive content, unless certain conditions are met.
AlexBe precise about the scope, though. Neither bill binds model developers. They govern the distribution of synthetic election media, not the systems that produce it.
Policy, regulation & law — Khanna calls for a federal AI agency "like the FDA" and a Trump-Xi ban on recursive self-improvement
MayaAnd on Face the Nation with Margaret Brennan on September 20th, Congressman Ro Khanna, the top Democrat on the Select Committee on the Chinese Communist Party, said we need to set up an AI agency like the FDA.
AlexModelled on what argument?
MayaGeoffrey Hinton's. That AI can be regulated the way we regulate nuclear energy, electricity, aviation, with independent technical experts making sure it's safe. On the tech executives who say Washington can't follow it, his line was, give me a break, we have regulated very complex things.
AlexAnd on China?
MayaHe wants two principles in a US-China agreement. Banning any self-improving AI, recursive self-improving AI. And safeguards so AI can't be misused by people for biological nuclear weapons. He said it needs to be the first agenda item on the Trump-Xi Jinping summit.
AlexHave the Chinese firms he wrote to replied?
MayaThey have not. And he conceded the distillation point. They're right that they are engaging in distillation, he said, basically they are copying American models, which he argues is itself a reason to regulate American models. CBS News has the full transcript.
AlexOne thing to keep straight. Brennan's citation of FBI, NSA and CISA warnings is the host's characterisation. We did not open those warnings, and the White House has published no position on Khanna's proposals.
Security, misuse & threat intelligence — BragJack browser-assistant hijack earned more than $20,000 in bounties and produced two CVEs
MayaAn update now on a story we covered on September 17th. The BragJack browser-assistant hijack.
AlexWhat is new since then?
MayaThe accounting. BleepingComputer reports the research earned more than $20,000 in bug bounties from the five vendors, ranging from $600 to $7,000, and produced two CVEs.
AlexAnd have the vendors fixed it?
MayaGoogle and Microsoft have both since resolved the flaws they were assigned, according to BleepingComputer. Chrome paid the $7,000 at the top of that range.
AlexThe researcher, Gal Weizman, calls the method Prompt Forcing. Instead of smuggling instructions into content the agent reads, the attacker writes the whole prompt and the follow-ups, and the agent carries them out with the browser privileges it already has.
MayaOne caveat on detail. Forever Security's own table doesn't match BleepingComputer's attribution of the second CVE, so we're only reporting the Chrome assignment.
Transition
AlexTo defence, and to autonomy with weapons attached.
Military, defense & geopolitics — US Air Force plans 500 highly-autonomous collaborative combat aircraft by 2032, Defense One reports
MayaDefense One reports the US Air Force plans to field 500 highly-autonomous collaborative combat aircraft by 2032.
AlexAir Force Secretary Troy Meink is quoted calling autonomy one of, if not the key technology enabling cost-effective combat power.
MayaThere is one figure from the vendor side. Shield AI's president and co-founder Ryan Tseng told Defense One his company worked with Ukrainians to improve the strike capability of one of its drone weapons by 70 percent.
AlexSo what's the constraint on getting there?
MayaThe article's argument is that it is doctrine, training and defined permissions for autonomous decisions, rather than airframes. Michael Gregg of the Air Force Research Laboratory frames it as a trust and autonomy question.
AlexOnly one outlet carries this, and the Air Force isn't quoted giving a programme cost or a contract award, so treat the plan as reported rather than budgeted.
Military, defense & geopolitics — Chinese state-TV-affiliated account accuses Anthropic of 13 privacy-policy revisions easing US intelligence access
MayaAlso under geopolitics. A Chinese state-television-affiliated account has gone after Anthropic.
AlexGeopolitechs reports that Yuyuan Tantian, an account affiliated with China Central Television, went through 13 revisions to Anthropic's privacy policies since 2023, and characterised them as progressively enabling transfers of user data to US intelligence agencies.
MayaWhat does the post point to?
AlexA February 2026 Anthropic report on Chinese distillation that said it was sharing technical indicators with relevant intelligence agencies. A June 2026 statement to the US Senate describing analysis of 28.8 million user conversations and 25,000 user accounts. And a September 10th report analysing approximately 200 million interactions.
MayaThe caveat matters more than usual here. These are a state-affiliated account's characterisations of Anthropic, not verified findings about what Anthropic does. Anthropic isn't quoted in the post, and only one outlet carries the account of it.
Transition
AlexNow the machines, and the money behind them.
Compute, chips & infrastructure — CXMT says its fifth-generation DRAM platform is in mass production at an 11.95 nm half-pitch
MayaChina's CXMT said on Sunday that its fifth-generation DRAM platform has entered mass production.
AlexThe number that matters is the half-pitch. Global Times puts it at 11.95 nanometres in the active area, and Seoul Economic Daily reports it was reached using quadruple patterning.
MayaWhat does that buy them?
AlexGlobal Times reports the number of dies produced per wafer rising by at least 50 percent against the prior generation. Two LPDDR5X products of 24 gigabits each are in mass production, 50 percent more capacity than the previous equivalents.
MayaWhere does that leave them against the incumbents?
AlexSeoul Economic Daily reports CXMT held 9.5 percent of global DRAM revenue in the second quarter of 2026, fourth behind Samsung, SK hynix and Micron. The company declined to disclose yields or monthly output, so this is a company claim.
MayaAnd Seoul Economic Daily notes that figure is hard to compare directly with the 12-nanometer-class process names the incumbents use, so the parity it implies is not independently established.
Compute, chips & infrastructure — FT: technology companies provided as much as $300 billion in guarantees financing AI data centres and chips
AlexStaying with the money. Seeking Alpha, summarising the Financial Times, reports technology companies have provided as much as $300 billion in guarantees to finance AI data centers and chips during the past year.
MayaAnd most of that isn't recorded immediately as debt?
AlexThat's the point of the structure. The guarantee promises a minimum asset value through a special-purpose arrangement, so capacity gets added without the exposure landing immediately as debt on the guarantor's books.
MayaWhat's the risk?
AlexSeeking Alpha writes the guarantees can obscure potential liabilities if demand fails or equipment devalues, and that losses could rise if AI hardware becomes obsolete quickly or the industry overbuilds.
MayaWe could not open the Financial Times piece itself, so only the figures printed on the Seeking Alpha page are in our edition, and that single source is what this rests on.
Transition
AlexLast stretch. Deployment, and what it is doing to people.
Deployment & impact — New York Times: DraftKings scored bettors by expected losses and aimed about $400 million of promotions at them
MayaThe New York Times published an investigation on September 19th into DraftKings. We're reading it through Yahoo News, Startup Fortune and Tech Times, because the Times itself wouldn't open for us.
AlexWhat did they find?
MayaA machine-learning model built in 2023 that scored each online casino player by how much they were expected to lose after getting a free bet or a bonus. It weighed play frequency, daily account balances, and the ratio of losses to wagers. Staff called the output an elasticity score.
AlexAnd how much money moved on the back of it?
MayaRoughly $400 million in automated promotions aimed at the highest-scoring accounts in 2025. Executives credit data science and personalised promotions with lifting the sportsbook margin by 13 percent that year.
AlexThere's a second model in this story.
MayaThere is. Former employees said a parallel model meant to flag customers sliding toward gambling harm, started by a data scientist in 2024, was shelved. The chief responsible gaming officer said leadership decided against predictive risk tools because the approach was not evidence-based.
AlexDraftKings says it rejects any implication that its marketing is unfair, and that promotions go to customers showing sustained, engaged use of the platform, not customers picked out by their losses. All of it traces to a single source, one Times investigation.
Deployment & impact — Flock Safety offers voluntary buyouts across its 1,500-person workforce after surveillance backlash
MayaFlock Safety is offering voluntary buyouts across its 1,500-person workforce.
AlexTechCrunch, reporting on a Wired story, says the packages were announced on Friday and described internally as the most generous the company has ever offered. Flock expects a significant portion of staff to express interest, and said that without them it would almost certainly need to lay people off.
MayaWhy is the company under this much pressure?
AlexTechCrunch cites 46 documented cases of police officers misusing Flock's licence plate recognition technology, identified by The Washington Post in August 2026, Florida and Texas halting use of the system, and 90 cities terminating Flock contracts in August alone, a fourfold increase on the previous month.
MayaChief executive Garrett Langley said the biggest damage from the backlash has been to internal morale. The story originates with Wired, it's a single source, and no headcount reduction target has been published.
Deployment & impact — AI evaluation startup Vals raises a $40 million Series A led by Andreessen Horowitz
AlexOne funding round worth your time. TechCrunch reports Vals, an AI evaluation startup founded in 2024, raised a $40 million Series A led by Andreessen Horowitz.
MayaWhat do they sell?
AlexModel developers pay Vals to test their models. They evaluate on law, finance and coding tasks, and they're extending into recursive self-improvement, mental health, cybersecurity, biosecurity and the law of armed conflict. They recently launched a programme providing evaluations to federal agencies.
MayaHeadcount went from 8 at the start of the year to 25, with 10 to 15 more planned, and revenue is 8 times last year's. Those are the company's own numbers, a company claim, not independently verified.
AlexAnd Vals doesn't publish its test materials, to stop models being trained against them, which also means the scores can't be reproduced by anyone outside. One outlet has this.
Outro
MayaThat's The AI Edge for today.
AlexThe full edition is on the site, with a link to every source behind every claim we've made.
MayaNothing we said here goes beyond what is written there.
AlexListen in tomorrow for the next edition.