The episode opens with Leo Laporte teasing the key stories: a new wave of Chinese AI models, an absurd AWS billing glitch, and a robot teacher in upstate New York. The Black Hat USA sponsor segment runs first, highlighting the conference's August 1–6 Las Vegas dates and offering $200 off with code TWIT. Leo then formally introduces Harper Reed of 2389.ai and Alex Wilhelm of Cautious Optimism, with warmth and light banter about the rarity of getting both on together.

The panel digs into a question that seems obvious but turns out to be surprisingly hard to answer: is AI actually speeding up commercial software development, or are we just seeing PR spin? Harper Reed argues the real signal isn't corporate blog posts about AI productivity — it's companies that are so effectively using AI they'd never tell you, because doing so would hurt recruiting and tip off competitors. Alex Wilhelm points to the sheer volume of simultaneous product launches (like Microsoft Build) as evidence. Harper then introduces the 'Move 37' framework — named after AlphaGo's decisive non-human move against Lee Sedol — suggesting the real proof of AI-native teams will be surprising, non-obvious product flourishes, like Easter eggs, that no traditional product manager would have prioritized. Leo adds that vibe-coded consumer software is proliferating explosively, pointing to his own RSS reader project and a wave of poker apps flooding Reddit.

Linus Torvalds drew a public line in the sand on the Linux mailing list this week, stating that Linux is not an anti-AI project and that anyone who objects should fork it or leave. Leo Laporte frames this as a surprise — many in the open source world assumed Torvalds would be hostile to LLM-generated code. Harper Reed pushes back slightly, arguing the bell can't be unrung regardless of Torvalds' position. The discussion touches on the analogy to self-driving cars: just as some people ignore statistical evidence that Waymos are 10–15x safer than humans, some developers deny AI coding value despite the data. Alex Wilhelm adds a sharp observation: many people who say 'AI is bad at coding' tried ChatGPT once 18 months ago and haven't revisited the question.

This chapter is the most technically dense of the episode, as Harper Reed and Leo Laporte compare notes on their actual agentic coding setups. Harper describes using '/loop' and '/goal' commands to keep agents iterating until tests pass, then passing output to a second model for adversarial review — what he calls 'giving it more loops.' Leo describes a more elaborate multi-agent pipeline: Claude Fable plans the work, Opus 4.8 does the coding, Sol 5.6 reviews it, Grok 4.5 does a final review, with agents communicating via Obsidian text files acting as a shared mailbox. Harper notes his company now writes significant amounts of Rust, Swift, and Go — languages they never would have used manually — because the agent doesn't care. Alex Wilhelm makes the memorable observation that children born today will likely never hand-write code in languages like C++, and that manual coding will become a lost art.

The ZipRecruiter sponsor read highlights the platform's new feature that prioritizes the most interested and qualified candidates, with the stat that 4 out of 5 employers who post get a quality candidate within the first day. Leo plugs the free trial at ziprecruiter.com/twit. Alex Wilhelm jokingly suggests a Tinder-style swipe interface for candidates, leading Leo to note that apparently swipe dating doesn't work well for the young either.

For weeks, developers had been bracing for Fable to be discontinued or restricted, with Leo describing himself as 'rushing and rushing' to complete projects while access lasted. But Anthropic announced Claude Fable is being rolled into Max subscription plans at 50% of normal limits, with a one-time $100 credit for lower-tier users. Harper Reed's analysis is the sharpest: the US government's restriction of Fable access 'nuked' developer trust, and OpenAI's guide to using Sol inside Claude Code made it easy for Claude Code devotees to defect. Once users discovered the models were functionally interchangeable, Anthropic's lock-in evaporated. Alex Wilhelm adds that Anthropic is still not compute-abundant — the 50% limits are a real constraint — and Leo notes the company is buying GPU time from xAI at ~$1B per month.

The central topic of the episode arrives in full force: a wave of Chinese frontier AI models that are closing the gap with the best American closed models. Leo describes Kimi K3 as 'almost — not quite, but almost — Fable quality,' built entirely on 100,000 Huawei chips with no NVIDIA hardware. Kimi's subscription sold out immediately — a move Harper Reed compares to the classic T-shirt marketing trick of doing a small run to generate hype. Alex Wilhelm pushes back slightly, noting that Kimi K3 is not actually cheap to run, and that at current pricing it looks similar to a mid-tier American model. Harper uses GLM as his point of comparison: he was using Claude Code in early 2025 when it was still rough, and today's GLM is at least as capable as that early version — effectively making it a 6-month-behind Claude for free. The Toyota Corolla analogy emerges: it's not whether the model is the best, it's whether it's good enough and the price is rational.

Leo Laporte walks through Dean Ball's influential tweet thread, now that Ball has become OpenAI's head of Strategic Futures. Ball's argument has four points: Kimi K3 is genuinely impressive and cannot be explained away by distillation; China allowing its release is strategically surprising; open-weight models deter AI capex by making private investment in frontier models irrational; and an open-weight dominant world is essentially AI communism — state-provided digital infrastructure rather than a market product. Harper Reed takes sharp issue with Ball's framing, calling the 'China is just too stupid to know what it's doing' angle a 'fatal flaw' in most US analyses of China. Leo reads Ball's final warning: at some point these models will be capable enough that a non-living, self-replicating agent escaping from a Chinese lab will be a real concern. Harper responds that US labs haven't been blameless about training data copyright either, and that the real question is how America competes in a game it's not currently playing.

Harper Reed's most geopolitical argument of the episode: China's open-weight AI releases aren't just a commercial play, they're soft power — the digital equivalent of building roads and ports across the developing world. Countries in the Middle East building massive data centers will choose models they can customize to their own values, and American closed models won't even be in the running. He points to the book 'Breakneck' by Dan Wang for the structural explanation: China's Politburo is staffed by scientists and engineers; America's leadership are lawyers. When US tools are regulation and laws, and China's tools are science and technology, the long-game advantage is structural. Alex Wilhelm adds a crucial counterpoint: Zhipu AI and Minimax are both public companies that he has personally reviewed, and both are deeply unprofitable — so the sustainability question for Chinese labs is real, absent state backing.

Leo describes Project Golden Eagle — the Trump administration program that would give the government a 30-day window to review and approve new frontier AI model releases before public availability. The frontier labs are reportedly going along with it. Harper Reed is skeptical: having worked for the Obama administration, he jokingly suggests he'd never get access to any approved models. More seriously, he argues that US government actions around AI are driven by shareholder value and political interest, not safety — drawing an analogy to COVID response failures and school shooting non-action. Alex Wilhelm expresses disappointment that what seemed like a 'free-range AI' administration has ended up in the same place as a more regulatory one — government control over model access.

The Ethos sponsor segment focuses on the experience of becoming a new parent and suddenly feeling responsible for someone else's future. Leo recounts getting his first life insurance policy at 35 when his own children were born. Ethos offers quotes in seconds, applications in minutes, same-day coverage with no medical exam, up to $3M coverage from $30/month, at ethos.com/twit.

Satya Nadella posted a 'reverse information paradox' warning on X: when you pay for AI tokens, you're also paying with something more valuable — the business knowledge, correction patterns, and workflow signals you feed into the model. Alex Wilhelm digs into the nuance: even if labs honor their 'no training on your data' policies, the metadata around usage is rich enough to learn a great deal about any enterprise customer. Harper Reed notes the sharp irony that Microsoft — which built and owns GitHub Copilot, Azure AI, and co-invested in OpenAI — is the one raising the alarm. Alex coins 'SaaSpocalypse' to describe what happens when AI labs, growing horizontally and vertically, eventually eat the entire enterprise software stack. Leo notes the Pentagon runs Claude on private AWS infrastructure precisely to avoid this problem.

Alex Wilhelm poses the question that's been nagging at tech observers: Microsoft had GitHub, had GitHub Copilot at $10/month, had a massive developer user base — and still lost the agentic coding wave to Anthropic and smaller startups. Harper Reed's answer is blunt: Copilot sucked, and it launched the old way — big release, iterate slowly — right at the exact moment Claude Code, Cursor, and a dozen competitors compressed the entire product cycle to weeks. Alex adds that Microsoft effectively outsourced its intelligence to OpenAI, then saw Anthropic leapfrog. But Harper also credits Microsoft for surprising reinvention under Satya Nadella, noting that younger developers who've only known the post-Ballmer Microsoft have no idea how monopolistic and hostile the company once was.

The Arctic Wolf segment highlights the company's Aurora AI platform combining agentic AI, generative AI, and machine learning for threat detection, and promotes the State of Cybersecurity 2026 Trends Report, available at arcticwolf.com/trends. The report covers AI adoption, threat detection, and what security professionals are prioritizing in an AI-driven threat landscape.

Ursula von der Leyen's EU push for social media limits for teens triggers a broader conversation about digital surveillance. Harper Reed cuts straight to it: any age verification system requires a database with the real name and birthday of every internet user — and that database will be breached or weaponized. He points to the EU's failure to stop Chat Control (warrantless mass scanning of private messages) this week, with a full debate on Chat Control 2.0 coming in September. Alex Wilhelm connects the dots: from EU regulations to US state laws (more than half of states now have teen social media restrictions) to China exporting Great Firewall technology, everything is pointing toward centralized government control of digital life. Signal's Meredith Whittaker is cited as threatening to pull out of the EU if encryption backdoors are mandated.

The ThreatLocker segment highlights the company's expansion from endpoint protection to cloud and network zero-trust enforcement, ensuring that even successfully phished employees can't access cloud services without a trusted, physically verified device. Leo cites Heathrow Airport's Rob Thackeray as a satisfied customer and mentions an upcoming live broadcast from ThreatLocker's Black Hat booth. Free 30-day trial at threatlocker.com/twit.

Leo's live demo of 'Quicksilver' — his home AI agent — turns into one of the most delightful segments of the episode. The agent talks through Sonos speakers throughout the house, maps Wi-Fi access points to the nearest speaker to follow Leo room by room, listens via Apple Watch on port 8839 through Tailscale, transcribes through Whisper, and routes to a local Hermes instance. Harper Reed reveals he built a multi-agent setup where AI agents live in a 'palace,' gossip with each other, and build things overnight while he sleeps. Alex Wilhelm delivers the line that crystallizes the moment: 'Technology is fun again.' Harper adds that the people who have never felt the joy of building software are now discovering it for the first time through AI agents — reclaiming autonomy they were previously denied.

The Wall Street Journal's piece on Meta flooding the market with Ray-Ban smart glasses and the privacy concerns around them leads naturally into Harper Reed's most provocative topic of the day: AI psychosis. Harper knows people using Claude as their primary therapist because they can't afford real therapy, and people whose social batteries drain after talking to AI agents just as they do after talking to humans. He and Leo debate whether this is alarming or just another technological transition — like learning to ride a horse instead of walking. Harper's conclusion is nuanced: the bell can't be unrung, but we should be paying attention to the effect on human connection rather than assuming it's fine.

The Shopify segment features Leo's son Henry — known as Salt Hank, with a large TikTok following — as the success story of someone who used Shopify to quickly launch merch and salt sales without any coding or design background. Leo highlights the templates, Sidekick AI assistant, Shop Pay one-click checkout, and award-winning support. Free trial at shopify.com/twit.

The episode's news rapid-fire covers three standout stories. First: an AWS global billing glitch caused by a unit pricing error in the billing computation subsystem sent a UK nonprofit a $7.8 billion invoice (normal monthly bill: 43 cents) and another user a $1.5 trillion bill. Harper Reed notes it's always units. Second: MLB issued a memo banning teams from using custom AI apps on dugout iPads to make in-game strategic decisions like pitch calling and substitutions — though the panel wonders why this should be forbidden at all. Third: a White House teleprompter operator won $100,000 on prediction market Kalshi by betting on Trump speech content he had advance access to — and the panel gleefully notes that $100K barely covers a year of private school.

Kashmir Hill's 'I Got Slopped' piece in the NYT describes finding a 90-page AI-generated biography of herself on Amazon, written by an account that published 10 journalist biographies in a single week. Harper Reed connects this to the small fiction writing community where AI slop is flooding niche markets, and his brother's point that the best filter is production speed — if someone is publishing thousands of lines a day, it's probably not human. Alex Wilhelm notes the same problem hit music: Spotify removed 75 million AI-generated tracks. His larger worry is niche creative communities — heavy metal bands that already operate at thin margins could be wiped out by AI flooding even small markets.

The episode's final news story is its most surreal: Salamanca City Central School District in New York is deploying an AI robot named Sally as a classroom tutor, manufactured by Realbotics — a company that rebranded from Tokens.com after acquiring hyper-realistic sex doll maker RealDoll in 2024. Leo shows the picture and the panel's stunned reactions provide the comedic finale. Harper quips that Sally is what he'll look like 'once I get on all the peptides,' launching a brief tangent about Chinese research peptides in Silicon Valley. Alex Wilhelm, who just had his third child, closes on a more serious note: he's changing his long-term financial strategy to backstop his children against an AI-disrupted labor market. Harper quotes a John Quincy Adams line about warriors, merchants, and poets — and the show closes with gratitude.

The extended outro becomes an impromptu 2389.ai product demo as Leo pastes skill URLs into Hermes live on air and Harper walks through the most useful skills: Review Squad (adversarial code review panel including 'Well Actually'-ers, Anna Karenina, and a dying Star Trek ensign), Thrifty (using Fable for planning + Haiku for execution to reduce cost), and Fresh Eyes Review. Harper teases upcoming business-focused skills and explains the company's emerging product thesis: the product is the process, with enterprise teams paying to learn how to uplift entire teams rather than individuals. Alex Wilhelm closes by plugging cautiousoptimism.news and noting he covers 'technology, business, and power with a modestly upbeat lens.' Leo signs off with the traditional 'Another TWiT is in the can.'