A high school girl in Mexico just won a gold medal for putting out fires with a speaker. It plays one deep, low hum and the flames die in about 5 seconds. DARPA tried the same idea back in 2012 with a rig so big and clunky they never took it any further, and hers runs on a 12-volt battery and a speaker she can carry around. Sound is just air being pushed back and forth. A low note at 30 hertz means the air around the flame is flipping direction 30 times every second. Flames only exist in a thin layer where fuel vapor mixes with oxygen. Shake that layer hard enough and it keeps getting stretched thinner, the heat gets pulled away faster than the flame can make more, and eventually the flame rips right off the fuel. Whatever tears off is too cool to keep burning, so it just goes out. DARPA's report put a number on how little it takes. Once the air at the flame is moving about half a meter per second, which is barely a breeze, the flame goes out, and pretty much any pitch can get there if you crank the power high enough. Low notes just get you there way cheaper, because a speaker moves a lot more air on each push when the pitch is low. Two engineering students at George Mason figured this out by trial and error in 2015. They started with super high pitches, 20,000 to 30,000 hertz, and all the flames did was shiver. It only worked once they dropped down to 30 to 60 hertz, which is basically the bass on a hip-hop track. Ángela Karime's build is a 12-volt battery, a frequency generator, an amplifier and a speaker, tuned to 30 hertz. She named it Vortex Tech and talks about it shooting out little rings of air, and that's real physics. When loud, low sound gets pushed through a narrow opening, the air curls up into rings like a smoke ring, and those rings carry a puff of air way farther than the sound would get on its own. In her tests it put out small flames from about 2 meters away. She went through five prototypes and over 100 tests in about a year, and she says it's put out burning wood, flammable liquids, electronics and cooking oil. The cooking oil one is a big deal. Throw water on a grease fire and it flashes into steam and flings burning oil all over your kitchen, which is exactly the kind of fire where you'd want something that doesn't put anything into the pan. Now, sound can't cool anything down, so a big fire sitting on red-hot fuel can come right back the second the speaker shuts off. The George Mason guys worried about the same thing in 2015, and it's why this is still a prototype for small flames and the red extinguisher under your sink isn't going anywhere yet. Space might end up being where this matters first. Up there water floats off in blobs, while sound works the same with or without gravity, and acoustics researchers have already published experiments on fighting fires with sound on spacecraft. She's still in high school, and she's already working on a version that spots a fire and puts it out by itself. The gold in Indonesia also came with a special award that gets her into another international competition.

The craziest part of the FTX story is that the victims got their money back. All of it, plus 18% interest. And the money came from a bet Sam Bankman-Fried made with their cash. Back in 2021, before anyone cared about AI, SBF quietly put $500 million of FTX money into a startup founded by ex-OpenAI researchers. That startup was Anthropic, the company behind Claude. When FTX blew up in November 2022, the stake was just another line item for the bankruptcy lawyers to pick through. Three weeks after the collapse, ChatGPT launched. Suddenly the most radioactive balance sheet in crypto was sitting on one of the hottest private assets on earth. The estate sold the 8% stake for $1.3 billion in 2024. Jane Street, the firm where SBF worked before founding FTX, was among the biggest buyers. That sale, plus the rest of the recovered assets, let the estate pay creditors between 118 and 143 cents on the dollar. Here's the part I can't get over. Anthropic was just valued at $965 billion. The stake FTX dumped for $1.3 billion would be worth roughly $75 billion today. The hole SBF dug was $8 billion. His stolen-money side bet would have filled it nine times over. One catch the victims know well. Claims were paid at November 2022 crypto prices. If you had a bitcoin on FTX, you got 118% of about $16,000, not your bitcoin. The best trade of SBF's life was made with money that wasn't his. His victims collected the first $1.3 billion. Everyone who bought from the estate collected the rest.

For the entire history of software, how much one person could use was capped by a single number. Hours at the keyboard. One screen, one task, maybe eight focused hours a day if you're really grinding. Claude Code Projects removes that cap, and the economics underneath are wild. Here's the mechanism. A project is one standing conversation where Claude splits your work into threads and runs each as a full cloud session in parallel. You send a thought, close the laptop, and five sessions keep working. Boris Cherny, who created Claude Code, says he stopped managing sessions entirely and just sends ideas as they come. The work gets divided without him. That reads like a convenience feature. Now look at what Anthropic's own docs warn you about. Projects "use plan limits faster" because every thread is a complete Claude Code session. Ten threads running overnight means one subscriber pulling ten users' worth of compute while asleep. Demand per user used to be bounded by attention. Now it's bounded by ambition, and ambition doesn't sleep. Which explains the waitlist. Nobody capacity-gates a chat feature. You capacity-gate a feature where a single $200/mo account can suddenly consume like a small team. Every flat-rate software subscription ever sold was priced against the same assumption, that a human can only sit at a screen so many hours a day. That assumption lasted fifty years. It didn't survive this one.

A man left home in 1998 and promised his mom he wouldn't come back until he'd walked around the world. He couldn't use any transportation at all. He just got home. The walk took 28 years, and his son saw him 4 times the entire time. Karl Bushby was a British paratrooper when he left Punta Arenas, Chile in November 1998 with about $500 and paper maps. He figured 12 years. The walking mostly kept that schedule. He went up the length of South America, survived the Darién Gap, and made it through Alaska. Then in 2006 he became the first person in history to walk across the Bering Strait into Russia. 14 days on drifting ice, 150 miles of zigzagging to cross a 58-mile gap. His partner lost a finger to frostbite. Russia arrested him the moment he arrived. Fined him $72 and deported him. The ice took two weeks. The paperwork took the better part of a decade. Russia capped his visas at 90 days at a time, so he crossed Siberia in fragments, leaving when each visa ran out and coming back to the exact same spot so the line never broke. When Moscow stopped letting him in at all, he walked 4,800 km from Los Angeles to the Russian embassy in Washington DC to ask in person. They banned him for five years anyway. Last week, at 57, he hit the final border. Channel Tunnel bosses wouldn't let him use the service tunnel, so he swam the English Channel. 20 hours in the water spread across three days. Then he walked up the beach near Dover. 28 years. 29 countries. 31,000 miles. He says he's coming home with nothing and has to figure out the grown-up world again. Every adventurer talks about finishing what they start. Karl Bushby gave it half his life.

Reliability and auditability usually get used as filler words. @JanOberhauser actually defined them. Reliability turned out to be four things. A fallback model, because everyone goes down and Claude is barely at four nines. Self-hosting, so it runs next to your data with no internet in the way. Integration code n8n writes and maintains, so when an API changes it gets fixed once for everyone instead of by every builder separately, each forgetting a different edge case. And retries with defined storage when something breaks. Auditability is the one he says code can't give you. With code you see what went in and what came out. Everything in between is invisible. Why it chose that path, which branch it took, what the values were halfway through. None of it. n8n replays every past run step by step, showing the data going into and out of each node. You can drop a deterministic check in the middle, then look back and see exactly why it went left instead of right. And you can rerun half a workflow from any point rather than starting over. That last bit is the real one. Debugging isn't reading the output. It's finding the step that broke and running it again. Code gives you a result. This gives you a record.

For three years AI video had the same fatal flaw. You could generate anything and edit nothing. Sora, Veo, and Kling work by dreaming up pixels, 30 frames a second of pure prediction. Gorgeous footage, zero structure. If the logo sits two inches too far left, your only move is regenerate and pray, because there is no logo in there. Just pixels that happen to look like one. You can't ask pixels to move a logo. Syren takes a different path. Opus 5.5 writes the video as code, a 3D scene that Synthesia's renderer executes. The camera is a variable. The logo is an object. The pacing is a timeline you can grab. Chat to edit works because of that architecture. The model changes a few lines and re-renders instead of repainting 1,800 frames. Same characters, same lighting, same brand colors on every pass. Anthropic shipped Opus 5.5 a couple weeks ago and people immediately started making full music videos with it, no video model involved. It writes rendering code, a headless browser draws the frames, FFmpeg stitches the file. Synthesia just productized that pipeline with a proper 3D renderer underneath. Agencies routinely charge $1,000 to $10,000 per finished minute of corporate video, and most of that bill is revision rounds. A video you edit by talking to it goes straight after the expensive half of the invoice. The first company to make AI video editable may end up mattering more than the first one that made it photorealistic.

Two models list the same price: $0.10 in, $0.50 out per 1M tokens. I ran 20 podcast excerpts through both. GPT-6 Luna cost 40% less than Haiku 5.5 on this job, cheaper on 20 of 20 requests. On code generation their gap was 12x, on my summarization job 40%. The job determines the size. For me it came from token counts for the same text, plus default thinking charged as output. For the same text, Haiku counts 119,723 input tokens and Luna 87,055. Haiku charged for 2.1x Luna's output tokens, mostly hidden thinking, while visible words were only 1.28x. The price list shows none of it.

One episode after Zapier's CEO sat in the same chair, I asked @JanOberhauser the obvious question. He didn't take the bait, which was more interesting than a swipe would have been. Right tool for the right use case, and Wade's a great guy building a great product. Then he drew the line where he actually sees it, which is power and flexibility. Specifically, never feeling locked in. Look at what that means in practice. Code nodes for when the visual layer runs out. Agents where you define your own memory. Output parsers that force a guaranteed format. A different model per step. Fallback models. Human approval gates. Every one of those is a door for the moment the abstraction stops fitting your problem. That's the difference that never shows up in a comparison table. Simple workflows look the same everywhere. You find out which tool you picked the day you need something slightly unusual and there's either a door or there isn't. He's clear that you should start small and build up. The question is whether the thing lets you keep building once small stops being enough. Plus self-hosting, which decides it for a lot of companies before anyone compares a single feature.

Sabi is building a baseball cap that turns thoughts into prompts for AI agents. An early wired helmet version hit up to 77% accuracy in the lab, the company says. That error rate would sink a keyboard. An AI agent might live with it. Brain-computer interfaces were built for a harder job. Neuralink, valued at $42 billion, implants chips so people with paralysis can control a computer by thinking. Precision like that has mostly meant hardware inside the skull. Khosla, who led the round, put the scale problem bluntly to Forbes. “You’re not going to do a billion brain surgeries.” A cap trades surgery for a weaker signal. A University of Chicago neurology professor told Forbes that signals filtered through skull and scalp come out lower quality. An agent asks for much less. Sabi’s pitch is that you think of booking a flight and the model fills in the details you left implied. The company also says accuracy climbs as the model learns each wearer. That 77% came from people with more than two weeks of use. I suspect the harder problem starts after the thought is read. A typo stays on screen until you catch it. An agent acting on a misread thought may have already booked the wrong flight. The pitch even has agents moving before you finish a thought, which skips the moment you’d normally notice. If I were designing it, the cap’s planned voice assistant would read the plan back first. A rough signal can start the work, and the cap that asks before it buys is the one people will keep wearing.

T-Mobile spent five years trying to get rid of the exact spectrum Elon just called "the last critical piece" of complete phone coverage in America. This 800 MHz block is Sprint's old Nextel spectrum, the stuff that powered push-to-talk walkie-talkie phones in 2004. The DOJ forced T-Mobile to divest it as a condition of the $26 billion Sprint merger. Dish had first rights at $3.59 billion and walked away because it couldn't raise the money. T-Mobile finally offloaded it to an investment firm called Grain Management last year for $2.9 billion plus a spectrum swap. Grain's original plan was selling it to utilities for private networks. Fourteen months later, SpaceX buys the whole portfolio, and shares of the carriers who passed on it fall as much as 8% in a single afternoon. Here's why it matters so much. Starlink's satellite signal at 2 GHz can reach your phone outdoors, but it dies at the first wall. 800 MHz penetrates buildings. Most phones already have the band built in. So satellites handle the open sky, this low-band layer handles indoors, and suddenly Starlink Mobile covers everywhere a normal carrier does without the tower network. Stack the moves. $17 billion for EchoStar's 50 MHz of 2 GHz spectrum. FCC approval for 15,000 Gen2 satellites this week. Now the low-band coverage layer. Verizon, AT&T, and T-Mobile paid $81 billion in a single auction for C-band spectrum in 2021, then spent years and tens of billions more deploying it on towers. SpaceX assembled a nationwide carrier out of a distressed-asset sale, a forced divestiture castoff, and rockets it already owns. The walkie-talkie spectrum from 2004 ended up being the piece that lets satellites replace cell towers. Nobody who touched it for twenty years saw that coming.

n8n was the hottest AI product of 2025. Then Claude Code ate its lunch. Or did it? I had on CEO Jan Oberhauser. It turns out, @n8n_io is actually doing better than ever. They grew over 10x the last year. You see, this isn't the first time n8n has been declared dead. They were declared dead when ChatGPT came out. Then they become the tool to build AI on. They were declared dead when OpenAI launched AgentKit. Then they had their biggest week ever. They were declared dead with Claude Code. And they've been growing relentlessly anyways. It's an insane story with great product lessons. I got to the bottom of it all in today's episode:

In Tavus's own study, 120 people watched a 90-second video call and 45% decided the person on screen was real. It was an AI. A separate 120 watched two real humans talk, and 28% decided one of them was the machine. The company gave me early access to the smaller model it tested, Griffin Lite. The 28% is the number I keep coming back to. More than one in four viewers doubted two real people. My read is that looking human on video has become a timing problem. People leave gaps of roughly 200 milliseconds between turns, the figure usually cited from a 2009 study of ten languages. Turn-based voice AI waits for silence before it answers, and that pause is often what gives it away. The company describes this model as full duplex, meaning it watches and listens while it talks. It says the model can be interrupted, cut in on your rambling when asked, and notice someone walking into frame behind you. NVIDIA researchers built a benchmark for this from 237 clips of real two-person video calls. Real humans score 3.92 out of 5 on its generation track. The company reports 3.83 for its model, against 2.80 for the best system in the published paper. It also reports responses in about 1.9 seconds, where that system took 2.8. Humans answer in 0.9 seconds on the same test. Nearly a full second still separates the model from a person, and 45% is still under half. Language models learned what to say. Face-to-face AI is learning when to say it.

Harvey was valued at $15.5B in September, selling AI to lawyers. Many of them still get paid by the hour. Every hour the software saves them is an hour they can no longer bill. Thomson Reuters ran the math. If AI cuts a one-hour task to 12 minutes, a firm billing hourly loses 80% of the revenue on that task. So an hourly firm has little reason to pass the speed on to the client. My guess is that's a big part of why legal still feels slow to the people paying for it. The gain goes to the firm that changes the pricing. A firm that charges per document keeps every hour AI saves, so it has every reason to get faster. The client gets a known price and a quicker turnaround, and the firm keeps the difference. Arceus launches today as that kind of firm. It says AI prepares every output and a licensed attorney approves each one before it goes back. Clients send work over Slack and approve a flat price up front. The firm says a contract review typically comes back in 4 to 6 hours. Big firms could move to flat fees tomorrow. Alternative fee arrangements have sat around 20% of law firm revenue since roughly 2014, per Thomson Reuters. When partner pay rides on billable hours, I think a flat fee looks like a pay cut until the firm gets fast. Legal AI keeps selling time savings to firms whose product is time.

I showed Zapier's CEO three versions of the same PM work and had him grade each one live against Zapier's AI fluency rubric. A PRD written with ChatGPT: Capable. He named what was missing. No prototype, which he now expects because AI makes one cheap. And no evidence. Where did customers ask for this? Gong calls, Zendesk tickets, Reddit threads, posts on X. His line: "If there are things I know that you don't know, that's probably not a great signal." A working prototype plus a PRD that cited the evidence: Adoptive. An operating model for a whole product team: Transformative. Skills hold the recurring work. Subagents review in parallel, including a skeptic that argues nothing should ship. The agent reads Zendesk, Gong, Amplitude and Linear directly, and a human gate sits before anything merges. "This is not how product and software was built three years ago. Not even close." The catch he added: about 1 PM on a team of 40 should hold that rating. The floor moved. What was transformative in 2024 is adoptive now. The rubric, the three artifacts and his notes on each:

When SeatGeek launched ChatGPT company-wide, IT got 300 requests on day one, per Serval's case study. By day three, 60% of the company had filed a ticket. The tool was the easy part, and the tickets were the job. Every new AI tool arrives with licenses, access requests, approvals and a slightly different workflow for each team. The same case study counts 406 Claude licenses provisioned with zero human touches. My guess is adoption creates operational work faster than most teams can hire for it. Automation Engineer sounds like one more hot AI job title. Prompt engineer paid six figures at the peak of the hype, and by May 2025 Fortune was calling it obsolete. I think the difference is what the knowledge is about. Prompt engineering was knowledge about a model, and the models absorbed it. This role runs on knowledge about a company. Who approves access, which step legal refuses to skip, where the spreadsheet everyone depends on actually lives. A better model doesn't learn that from a release note. Building keeps getting cheaper. Serval describes a new automation as a sentence and an afternoon of work. The expensive part is knowing which process to automate and getting a team with real politics to use it. For PMs, it looks a lot like discovery with coworkers as the users. Every automation starts as a few lines of code and ends as a negotiation over who gets to approve it.

Back in 1993, the Senate hauled video game executives into hearings over Mortal Kombat and threatened a federal ratings law. Within a year the industry created the ESRB, its own ratings board. Congress shelved the bill. Thirty years later there's still no federal video game law. Hollywood ran the same play in 1930 with the Hays Code. Comic books did it in 1954 with the Comics Code. The pattern is nearly a century old. When Washington starts drafting rules for your industry, you hand them a rulebook you wrote yourself. That's the play that just ran at the White House. Read the four commitments in the accord closely. Internal controls, an internal oversight team, an external evaluator, a board committee. The frontier labs already do most of this. Anthropic publishes a Responsible Scaling Policy. OpenAI has a board safety committee. Google publishes a Frontier Safety Framework. The page largely commits these companies to what they were doing in 2024. The sentence that matters is the last one. "Over time, it may make sense to codify these steps into laws or regulations." So if Congress ever writes an AI law, this page is the first draft. And the companies it would regulate wrote it. The ESRB kept Congress out of video games for three decades. This single page is built to do the same job for AI.

"Hard choices, easy life" came from a Polish firefighter who joined Solidarity, watched martial law crush the movement, and fled the communist regime in 1986 with his wife and almost nothing else. Jerzy Gregorek landed in America as a political refugee at 31. No English, no money, no credentials that transferred. What followed was two decades of the hardest available choices. He became a 4x World Weightlifting Champion and set a world record. He went back to school in his 40s and earned an MFA in poetry. In 2000, he and his wife Aniela founded UCLA's weightlifting team. The line only reached your feed because Tim Ferriss had Jerzy on episode 228, with Naval in the room, and both of them have been spreading it ever since. People quote it about 5am alarms and cold showers. The man who coined it used it to decide whether to leave his country forever.

Passengers just pulled a hijacked 737 out of a dive at close to the speed of sound. The plane fell 17,000 feet in under a minute. Everyone on board is alive. A C-17 pilot ran the math on the flight data and got a peak around Mach 0.98. For context, the 737 MAX is certified to cruise at Mach 0.82, and in flight testing Boeing dives it to roughly 0.89 and stops there. This airframe blew past all of it with a stabbed captain bleeding in the cabin and a passenger hauling back on the yoke. So why didn't it break apart? Certification limits sit nowhere near the actual limits. Regulators require every structure on a commercial jet to survive 1.5 times the worst load it could ever see in service. When Boeing tested the 777 wing, they bent it 24 feet upward before it snapped, at 154% of maximum design load. The engineers cheered when it broke. Something close to this happened once before. Back in 1985, a China Airlines 747 lost control over the Pacific and fell 30,000 feet in about two minutes, pulling nearly 5g and flying well past its maximum operating speed. The wings bent. Pieces of the tail tore away. It landed in San Francisco and almost everyone walked off. Every extra knot this airframe absorbed today was paid for decades ago by engineers stress-testing for a day nobody could predict. A copilot with a knife found that day. The margin was waiting.

YouTube has paid creators over $100 billion since 2021. Every challenger that tried to match that check is dead. Quibi raised $1.75 billion before launch, one of the biggest war chests in startup history, and folded in six months. YouTube pays out roughly that much to creators every single month. Vine is the cleanest autopsy. Back in 2016, about 20 of its biggest stars offered to keep posting in exchange for $1.2 million each. Vine said no. They took their audiences to YouTube and Instagram, and Vine shut down within the year. Here's why nobody breaks through. Video is the most expensive content on the internet to serve, so a challenger bleeds money on infrastructure for years before ads can work. YouTube itself didn't clearly make money for close to a decade after Google bought it for $1.65 billion in 2006, a deal people called the dumbest in tech. In 2025 it did over $60 billion. Only Google could afford that wait, because search profits covered the losses. And once YouTube could fund a 55% revenue share, creators became the defense budget. Beating YouTube was never a product problem. You'd have to outbid $22 billion a year in creator checks with no ad machine to pay for it. Twenty years in, nobody has even tried twice.