September 20, 2026 — The Strike the Algorithm Didn't Stop

Host0:42

Somewhere in a military database, there's a note from twenty nineteen. It flags a building in the Iranian town of Minab as no longer what it used to be. That note sat in a system that never talked to the system planners actually used. Fast forward to February. The United States and Israel are launching strikes across Iran, moving through thousands of potential targets in a compressed window. A Tomahawk hits Shajareh Tayyebeh Elementary School in Minab. More than 150 people die. Reporting from Bloomberg puts the child death toll at around 123. An internal Pentagon review, described to Bloomberg and summarized by Gizmodo, calls it a cascade of preventable failures. Outdated imagery. A targeting timeline that left almost no room to double check anything. And an overreliance on an AI tool called the Maven Smart System, supplied by Palantir, which had the Minab site listed as an old Revolutionary Guard facility and recommended it. Here's the part that should bother you more than the software failure. The Pentagon's civilian harm mitigation staff, the people whose entire job is catching exactly this kind of mistake, had shrunk by roughly ninety percent department-wide. Central Command's team went from ten people down to one. No one on that team looked at Minab before the missiles launched. Think of Maven like a GPS app that's still routing you down a road that got demolished last year. The app isn't lying to you, it's working off a map nobody updated, and the human who's supposed to notice the detour sign got laid off. Palantir told Bloomberg it isn't responsible for the underlying data or for finding intelligence gaps, and that there's no evidence its software was actually at fault. The company says it's since added features that flag inconsistencies in a target before a strike. Separately, a United Nations fact-finding mission went further than the Pentagon has publicly gone, saying there are reasonable grounds to believe the United States committed a war crime by failing to verify the building wasn't a military target. The casualty numbers themselves aren't settled. Different outlets cite a range, from about 120 children up to 123 and beyond. What is settled is that this review exists and it hasn't been made public. Watch whether the Pentagon shares the full findings with Congress, or whether this stays another line in an unreleased report.

3:20

If you bought a Wi-Fi router in the last decade, there's a good chance it was built somewhere you never thought about. The FCC just decided that has to change. The agency updated what's called its Covered List to include consumer-grade routers and mobile hotspots made outside the United States. New models built overseas can no longer get the approval they need to be sold here, unless the manufacturer applies for what the FCC calls Conditional Approval, which requires disclosing your ownership, your supply chain, and a plan to bring manufacturing onshore. The router already sitting in your house is fine. Nobody's coming to unplug it. And the transition window for existing authorized models has already been pushed out to at least January of twenty twenty-nine. But for new products, the clock's running, and companies like Netgear and Eero have already lined up conditional approvals while most of the rest of the industry scrambles. The FCC's justification is national security. It points to attacks it calls Volt, Flax, and Salt Typhoon, all tied to foreign-made networking gear that got exploited to spy on or disrupt American networks. That's a real problem. The question is whether banning new foreign-built routers wholesale is the right tool for it. The Electronic Frontier Foundation's William Budington put it bluntly: this is an extremely blunt instrument, and it's going to catch a lot of harmless products while trying to stop a real one. Almost every consumer router relies on overseas manufacturing somewhere in its chain, so the practical question becomes how strictly the FCC defines foreign-made. Design versus assembly versus where the chips themselves came from. That definition is still murky, and it's going to decide who actually gets hurt by this rule.

5:17

Four hundred twenty-four thousand, nine hundred eleven CT scans. That's the training set behind a new AI model out of Alibaba's DAMO Academy, and this week the company just gave the whole thing away for free. It's called RADAR, and it's built to read contrast-enhanced abdominal CT scans and flag findings across eighteen different organs and roughly 150 possible conditions. The research behind it was peer reviewed in the journal Science, and the code, the trained model, everything, is now sitting on GitHub under an open license anyone can use. What makes this different from the usual medical AI story is the scale and the openness. Most diagnostic models get built for one disease at a time. This one was trained as a generalist, learning from clinical reports instead of manual labeling, and Alibaba is handing the whole thing to hospitals and researchers worldwide instead of locking it behind a licensing deal. That's the flattering version. Here's the catch that applies to basically all of this technology, not just Alibaba's. Reporting out of the Financial Times this week has clinicians warning that medical AI is being pushed well past diagnostics, into actual clinical decision-making, without the kind of trial evidence that shows it improves what actually matters, which is patient outcomes. Clinicians say their routine AI use roughly tripled over the past year, but a lot of them are still double-checking the AI's answers against the original source material by hand. That's the gap worth watching. A model can be excellent at reading a scan in a research paper and still be an open question in an actual hospital hallway. RADAR hasn't been tested yet in the kind of messy, varied, real-world conditions that would prove it generalizes. Being open-source doesn't settle that. It just means more people get to try to find out.

7:21

More than seventy five lawsuits. That's how many cases are now working through courts against AI companies over harm linked to chatbots, a lot of them involving kids. And that flood of lawsuits is exactly why states started writing their own safety laws, since Congress hasn't passed anything close to a federal standard. NPR did the digging on where those state bills actually came from, and the answer is uncomfortable. In states including Hawaii, Iowa, Oregon and Washington, some of the bills modeled on California's twenty twenty-five law were drafted with input from the tech industry itself. And that input shows up as language that could exempt exactly the chatbots most people actually use. Policy analysts NPR spoke with called some of these provisions get-out-of-jail-free cards, built into laws that are supposed to look protective on their face. The pattern is a state legislature moves fast because parents are furious and a lawsuit just made headlines, and the resulting bill has definitions narrow enough that the biggest players slide right underneath them. It's the same underlying problem as the medical AI story, just wearing a different outfit. The technology moves at one speed, oversight moves at another, and the people writing the rules often have more access and more lawyers than the people the rules are supposed to protect. With medical AI, the gap is missing evidence. With chatbot laws, the gap is language nobody outside the room fully understands until it's already signed. The companies that helped shape this language will argue the carve-outs exist so the laws don't accidentally regulate backend systems that have nothing to do with consumer chatbots. That's a fair technical point. Whether it's also a convenient one is the fight to watch as these laws start actually getting enforced, or not enforced, in the months ahead.

9:24

Five stories, five sentences each, let's move. CXMT's new memory chips. The Chinese memory maker says its fifth-generation DRAM platform has entered mass production, launching twenty-four gigabit mobile memory chips it claims are already in domestic flagship phones. Company press releases aren't independent yield data, so file this under promising until someone outside China checks the math. China's price crackdown. Beijing says it's stepping up enforcement against what it calls malicious price competition among online sellers and platforms. Translation: the government thinks some of that brutal e-commerce discounting has gone from competitive to destructive, and it's about to start saying so with penalties attached. Meta versus the UK regulator. Meta is formally challenging Ofcom's decision to classify WhatsApp and Instagram as the highest-risk category under Britain's Online Safety Act. That classification comes with the steepest compliance duties on the books, so this is a fight over exactly how much oversight the UK's biggest platforms have to accept. Polymarket's laundering problem. The Wall Street Journal reports that sources say more than ten million dollars moved through Polymarket wagers in an attempted money-laundering scheme back in February, and that the company's CEO dismissed early warning signs. The company is reportedly also facing a CFTC investigation, which is an awkward time to be eyeing an IPO. Spain blocks an archive. Spanish authorities ordered blocks on archive.today and its mirror sites, using a fast-track administrative process instead of a full court case. Archivists and journalists rely on that service to preserve pages that get deleted, so blocking it without a judge weighing in is the kind of shortcut that tends to outlive its original excuse. And that's Fast Boot. Five stories, zero patience.

11:32

So that's today: a Pentagon review linking a deadly Iran strike to an AI targeting tool, the FCC's new ban on new foreign-made routers, Alibaba open-sourcing a CT-scan AI while doctors question the evidence behind medical AI broadly, and state chatbot safety laws that may have loopholes built in by the industry they're meant to regulate. For the sources behind every claim in this episode, check the show notes on our website. If this show is useful to you, follow The Daily Reboot wherever you're listening, and send it to one person who needs to hear it. That's the show. Talk to you tomorrow.