A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Alright kiddos, Tony DeLuca here on Barely Possible, and today's menu is about one question that keeps showing up wherever you look: who's got their hand on the off switch. Pull up a chair.
Let me start with a number, because it stuck in my head like a popcorn kernel in your teeth. Twenty-seven dollars. Per student. Per year. That's what one school paid, according to a 2020 piece by the writer and editor Maria Bustillos, for continued access to digital copies of The Diary of Anne Frank. Not to buy the book. To keep renting it. Every year. For every kid.
Think about that for a second. When I was in school, the book was in the library. It had a stamp card in the back. Somebody spilled grape juice on page forty in 1981 and it was still there for you to wonder about. That book was owned. The library bought it once and it sat on the shelf until it fell apart. Now the deal is different. You don't buy, you license. And if the license lapses, or the publisher changes the terms, or the platform decides something, the book just isn't there anymore.
That's the frame for today's show. We've got a library e-book project trying to put ownership back on the table. We've got a holiday gift guide that, if you read it sideways, is really a map of who controls what in your pocket. We've got a Chinese self-driving company talking openly about its Plan B for when the Nvidia chips stop coming. We've got a deep dive into a small startup that says it doubled revenue while leaning hard on OpenAI's coding tools, and what that actually tells a founder, and what it doesn't. And yes, we've got Saturday Night Live putting the whole tech elite in one sketch for a lesson on spycraft. So a real mix. Let's go.
Start with the e-book story, because that's where that twenty-seven-dollar number comes from.
TechCrunch's Anthony Ha wrote up a Kickstarter campaign from Flaming Hydra, which is an online culture and commentary magazine. The campaign is called The Phoenix Project, and the tagline is a big swing: the "return of the human internet." When somebody pitches me the return of anything, my first instinct is to check my wallet. But let's actually look at what they're building, because underneath the slogan there's something concrete.
If you back The Phoenix Project, you're supporting three things. One, a one-off print magazine meant to celebrate, quote, "the values and fun of the early internet." Fine, that's the merch, that's the nostalgia piece. Two, a new e-book service for libraries. And three, a newsletter subscription and archiving program, also built for libraries.
Bustillos, who's the lead editor at Flaming Hydra, says these are, in her words, "basically all one thing." The thread tying them together is what she calls "protecting freedom of information" and fulfilling the "promise of the early internet, where you're making information available to as many people as possible, while still protecting the right of artists and authors to make a living."
Now here's where it gets sharper. She brought up a case that Flaming Hydra itself covered, where a writer named Tal Levin found what the piece describes as countless AI-generated clones of his book on Amazon. You write a book, and then a pile of machine-made knockoffs shows up next to it on the shelf, competing for the same buyer. That's not a hypothetical. That's somebody's livelihood getting diluted by a content mill.
Bustillos's answer is not "ban AI." Her answer is build a fenced yard. She talks about "creating an ecosystem that is based in people working together, who are offering their stuff in a context that is controlled and doesn't allow" others to create and sell those AI clones. And then the line I want you to sit with: "AI is not this unstoppable force. If people just set up publishing pipelines and networks and organizations that are dedicated to the human internet, people who want to read stuff can go to them, and it's actually completely isolated from mechanistic interference."
You can agree or disagree with that, but notice what kind of argument it is. It's not a moral argument. It's a distribution argument. She's saying the problem isn't that AI exists, it's that the big storefronts don't care what's on the shelf. So build a different storefront, with a different set of rules about what gets in.
And then the ownership piece. Bustillos puts it flat out: "E-books are books. They're not some other thing. They are books, and people should be able to own them, and libraries should be able to own them, and media should be able to be owned. All digital media should be able to be owned."
And her test for ownership is the cleanest one I've heard in a while: "If you are at the mercy of somebody who can shut the book off, you do not own it."
That's it. That's the whole test. Not who paid. Not what the receipt says. Who can shut it off.
The piece notes there have been cases of e-books being altered or removed from readers' devices long after purchase. And Bustillos concedes that for any single book, maybe that doesn't feel like a big deal. "But if it's all the textbooks for everybody, if it's all the books," then you've got a real problem. That's the scale argument. One rental is an annoyance. A whole school system on rentals is a dependency.
So what's the actual product? Flaming Hydra wants libraries to truly purchase and own e-books, and then lend them to patrons through an open source "library in a box" system called Lenny for Libraries. They're starting small. Five new e-books from Flaming Hydra writers, including collections of writing from The Awl, and essays by Tom Scocca and by Bustillos herself.
Five books. I'm not going to pretend that's a revolution. But Bustillos says she's talking to library consortiums and author groups, and the stated goal is to "change the conversation" around e-books, digital archiving and lending. In other words, five books is a proof of concept. It's a working demo you can point at when you walk into a room with a library consortium and say, look, this model can exist.
The second piece is the one I find more interesting from a builder's seat. It's called The Phoenix Foundry Archive, and it's aimed at paywalled newsletters. Bustillos makes the point that "so much of the best journalism, criticism, essays" now lives behind newsletter paywalls, which means libraries can't get at it, and neither can the Internet Archive. So a lot of the best writing of this era is effectively invisible to the institutions whose job is to keep things around.
The Phoenix setup lets newsletter writers offer subscriptions to libraries, the way libraries already subscribe to newspapers and magazines. It's starting with writers on Ghost, with other platforms promised later. Once it's set up, it automatically converts whichever posts the writer chooses into EPUB files and sends them to libraries for digital lending.
Now, why should a founder listening to this care about a thirty-thousand-dollar Kickstarter for library e-books? A couple reasons.
First, look at what they actually built. A converter. A pipe from the newsletter platform into a format libraries already know how to handle. That's a classic picks-and-shovels move. They're not trying to replace Ghost or replace the library system. They're building the adapter between two worlds that didn't talk to each other. If you're hunting for a business, adapters between institutions that already have budgets and platforms that already have content are an underrated place to look. Libraries have line items for subscriptions. Newsletter writers want more paying readers and, frankly, some permanence. Nobody was connecting the two.
Second, the money. The funding came from a mix. Bustillos thanked backers including Anand Raganathan, Brewster Kahle, and Paul Ford for financial support before the Kickstarter. But she went to crowdfunding on purpose. Her words: "We don't have to be at the mercy of anybody, and it also protects editorial independence. Those things go hand in hand."
Notice she used the same phrase twice in one interview. At the mercy of somebody who can shut the book off. At the mercy of anybody. That's not an accident. The whole project is one idea applied twice: don't build on a foundation somebody else can pull out from under you. Not for the reader, not for the library, and not for the magazine itself.
Now the skeptic's corner, because you know I've got one. Thirty thousand dollars is not a lot of runway. The campaign ends November 27th. Library procurement moves at the speed of a city council budget hearing. And the big publishers who prefer the license model are not going to roll over because five essay collections went up on an open source lending system. The honest read is that this is a small group trying to establish a precedent, not a company about to eat anybody's lunch. But precedents matter. Somebody has to build the thing that proves the alternative works before anybody bigger copies it.
And that phrase, who can shut it off, turns out to be a pretty good lens for the next story too, which on the surface is just a holiday shopping list.
TechCrunch put out its holiday gift guide, written by Amanda Silberling, rounding up the gadgets their reporters actually liked enough to buy themselves. And I want to be clear, I'm not turning this show into a shopping channel. But if you read this list with the Bustillos test in your head, it's a remarkable little document about dependency.
Start with the readers, because it connects straight back. Silberling says 2026 was the year the team fell in love with the pocket-sized e-reader. The top budget pick is the Xteink X3, fifty-five dollars during their Black Friday sale. It's small enough to fit in your wallet or magnetize to the back of your phone. Lovely idea. Read a book in the coffee line instead of scrolling Instagram.
But here's the catch, in her words: it "takes some fiddling with to set up, and it has no integration with Kindle, Libby, or any other apps." Only buy it for someone willing to do some Reddit browsing and tinkering. The bigger X4 Pro, ninety-nine dollars, adds a backlight, which she calls a game changer for reading in bed.
Then she says the quiet part plainly. She's, quote, a hardcore Libby power user, and it frustrates her she can't get library books onto these devices. And then: "The same goes for apps like Kobo and Kindle, where any e-book you already own cannot be transferred out of the Amazon ecosystem unless the publisher agrees, which most do not."
There it is. In a gift guide. The book you bought can't leave the store you bought it from. If you want a phone-sized e-reader that does talk to outside apps, the guide points to the Boox Palma Pro 2 at four hundred dollars, from her colleague Aisha Malik's review. For a full book-sized screen, her favorite is the Kobo Libra Colour, two fifty-nine. And she ends with a little jab: the Paperwhite line is great, but, "c'mon, Amazon. Branch out!"
So look at the shape of that market. On one end, a cheap little independent device that's free of the big ecosystems but also cut off from them, so you've got to hand-load your books. On the other end, the polished devices that work beautifully as long as you stay inside the fence. That's the trade in almost every consumer tech category right now. Openness costs you convenience, convenience costs you control. And the Flaming Hydra project is basically trying to add a third option for one specific buyer, libraries.
Now flip through the rest of the list with the same lens.
The Dumb Phone, twenty dollars plus a service plan. It's a TCL flip phone with software from a company called Dumb Co that lets you run WhatsApp, Spotify, Apple Music, Uber. And Silberling notes you can even get iMessage through a third-party app, then adds, quote, "though Apple could probably get that shut down… no one tell Apple, please!" She's joking, but she's also describing a business that runs partly on a feature somebody else could switch off. If you're a founder building on top of another company's walled garden, that parenthetical should make you a little nervous. Your best feature might be the one you have the least control over.
The Brick, fifty-nine dollars, and the newer Brick Night at a hundred twenty-nine. This is the one that made me laugh. Silberling describes it as "basically just a hunk of plastic with an NFC chip in it and a companion app." You tap your phone to it and it locks your social apps, and the Night version does it when you're in the same room during set hours. And she says she's embarrassed how much she recommends it, because her sleep has improved so much she does eight a.m. Pilates now. Who am I, she asks.
From a product point of view, that is a fantastic lesson. The hardware is almost nothing. The value is the ritual. A physical object you have to walk over to is the friction that software alone can't create, because you can always just tap "ignore limit" on your phone. Sometimes the smartest product decision is to add friction on purpose, in exactly the right place.
The Oura Ring 5, three ninety-nine to four ninety-nine, plus a five ninety-nine monthly subscription. Aisha Malik named it the winner after testing several rings and bands. Thinner, more durable, better battery, more premium-looking. Silberling tried a competitor, the RingConn 3, and returned it. But again, read the price tag. Four hundred dollars of hardware and then a monthly fee. You buy the ring, you rent the insights. That's the e-book model moved onto your finger.
And then the one that's actually in our lane. In the section titled, and I'm quoting, "For the AI-pilled vibe-coder," the guide says what they really want is a Mac mini or a Claude subscription, but if that's out of budget, try the Codex Micro, two hundred thirty dollars, described as OpenAI's first attempt at hardware. It was the favorite product reviewer Lucas Ropek tested this year. But the guide also notes that some hard-core coders, the very people it's aimed at, have found it "less than pleasing." Best case, Silberling writes, it's a productivity game-changer for your friend with a team of AI agents inside the OpenAI ecosystem. Worst case, "this will be a great piece of tech history in a few years."
That's a pretty honest gift recommendation. It's a nice gadget if you're all-in on one vendor's stack. And notice how the guide frames the whole category: the thing a vibe-coder actually wants is a subscription or a box to run the subscription on. The gift is access. Same as the Oura. Same as the Kindle library. The object is just a doorway into somebody's service.
I'll mention the kitchen stuff quickly because it's fun and because nobody gets hurt. The Ember Mug 3, ninety-nine dollars, from reviewer Lauren Forristal, who liked that you no longer need the phone app to change the temperature. Which, I'm sorry, needing an app to adjust your coffee was always a little much. The Ninja Crispi microwave and air fryer combo at four forty-nine ninety-nine, which Lauren said made "a pretty convincing case for ditching two appliances and making room for one." And the GoveeLife nugget ice maker at four ninety-nine ninety-nine, which Aisha says is for "those who crave what's called 'the good ice.'" If you know, you know. I'm not paying five hundred bucks for ice, but I respect a product that knows exactly who it's for.
And that's actually the takeaway from the gift guide for builders. The products that won were the ones that knew their one customer and their one job. A tiny e-reader for the coffee line. A plastic brick for the doomscroller. An ice maker for the ice obsessive. None of them tried to be everything. And the ones with real staying power have to answer the Bustillos question sooner or later: when the customer pays, what do they actually own, and who can take it away.
Now let's take that same question and scale it up from a flip phone to an entire national industry, because that's the story out of TechCrunch's mobility newsletter this weekend.
I know. We've been chewing on Nvidia all week. On this show we've talked about the petition from Elon Musk and other execs asking the board to bring Jensen Huang back, about the hole his exit punched in AI startup funding, about the reporting that export-banned chips were behind China's top models. I'm not going to re-run any of that. But there's a genuinely new angle here, and it comes from the other side of the table, from a Chinese company that depends on that hardware.
Kirsten Korosec wrote this edition of TechCrunch Mobility after Web Summit. She sets the scene: the conference leaned heavily into Chinese founders and companies this year. Web Summit co-founder Paddy Cosgrave said in his opening remarks that the superstars of the future are coming from China, not California, and that Chinese companies are building more of our shared future than anyone else.
Then, a few days later, word spread through the conference that Jensen Huang had resigned from Nvidia after admitting to an affair with an alleged spy. Korosec points out that Huang has been one of the strongest proponents of open source in the United States, and that anybody following autonomous vehicles knows how central Nvidia is to that whole world.
And the reliance runs at different depths. Some companies use Nvidia's automotive processors to actually run the driving software in the car. Others use Nvidia GPUs to train their driving models back at the data center. And some go full stack with Nvidia, which the piece says includes Mercedes-Benz, Zoox, and Wayve.
So Korosec talked to Leo Wang, the CFO of the Chinese autonomous vehicle company Pony.ai, and asked him straight: what happens if the fallout from Huang's resignation leads to even stricter export controls on Nvidia chips going to China?
His first answer is that the squeeze already exists. Demand is far greater than supply, he said, and it's one of Pony.ai's pain points. So their strategy has two parts.
Part one: efficiency. Engineers are told to be as efficient as possible when deploying training models. That can mean using mid-tier compute instead of high-end. In other words, they're designing the work to fit the hardware they can actually get, not the hardware they wish they had.
Part two: domestic chips. "If you look at the whole ecosystem, especially in recent years, definitely domestic chips in China also get lots of attention and development," Wang said, adding that those chips give them another option for AI training. "I think that gives us some flexibility."
And then the quote that I think is the real headline: "If you look at the whole situation, this has been developing for already a couple of years. Everyone, I would say, is also practicing kind of like a Plan B for a few years, so at least if you think about that, I would say there are alternatives. Although cost-effectiveness-wise, probably it's not the best choice yet. But still, it's an alternative way that already exists and which is more mature than a few years ago."
Let me tell you why I like that quote. It's a CFO talking, and CFOs don't do pep talks. He's not saying the domestic chips are just as good. He says flat out, cost-effectiveness-wise, probably not the best choice yet. That's an honest number guy telling you the backup plan costs more. But he's also telling you the backup plan exists, it's been rehearsed for years, and it's more mature than it was.
Here's the part that matters for Washington, and for anyone building on top of that supply chain. Export controls work by making the alternative expensive and painful. They don't work by making the alternative impossible, not after people have had a few years to practice. Every round of tightening is also, in a way, a round of motivation for the other side to finish building the Plan B. Wang is basically saying: you tighten again, we eat a cost penalty, but we're not stopped. Whether that cost penalty is big enough to matter strategically is the whole policy argument, and smart people land on both sides. But you should hear the buyer's version of it, because that's the side you rarely get on the record.
Now the founder lesson, and it's not about geopolitics, it's about your own stack. Pony.ai's two moves are the two moves every company should be making with any critical supplier, and especially with model providers right now.
Move one: get efficient enough that you can run on the second-best option. If your product only works on the single most expensive, most constrained resource on the market, you don't have a product, you have a hostage situation. Pony.ai is telling its engineers to make things work on mid-tier compute. For you, that might mean making sure your app still does something useful on a cheaper or smaller model, so a price hike or a rate limit doesn't take you out.
Move two: keep a second source warm. Not on paper. Actually practiced. Wang said "practicing" a Plan B, and that word is doing a lot of work. A backup you've never run is a hope, not a plan. If you've got one model provider wired through your whole product and you've never once swapped it out in staging, you don't know what your Plan B costs, and you don't know if it works.
And notice the honesty about the price. The backup is more expensive. That's normal. Insurance costs money. The question isn't whether redundancy is free, it's whether you can afford to be without it on the day the main supplier changes the terms. Which, again, is the Bustillos question. Who can shut it off.
Let me run through the rest of that mobility newsletter, because there were a handful of deals and tidbits worth your time, and a couple of them connect back to things we've been tracking.
On deals. May Mobility plans to use the Volkswagen all-electric ID. Buzz for its global ride-hailing platform, and says it's targeting more than ten thousand autonomous vehicles through 2031. Korosec was careful to note this does not mean VW's MOIA group is ditching Mobileye. VW told her it's offering an automated-driving-ready base vehicle to other companies that build their own self-driving systems. That's a smart platform move from VW. Instead of betting everything on one autonomy partner, sell the body that anybody's software can drive. It's the same lesson in reverse. Pony.ai is reducing its dependence on one supplier, and VW is making itself the supplier that doesn't care which software wins.
Newmo, a Japanese robotaxi startup, raised twenty point eight billion yen, roughly a hundred thirty-five million dollars, in a round led by KDDI. ParkBee, a Dutch parking technology company, acquired Mobian's city and parking operations. And Port, an urban logistics company, raised fourteen and a half million euros in a round led by the European Bank for Reconstruction and Development and Balnord.
On the tidbits. Lucid reported third-quarter revenue of three hundred fifty million dollars, below Wall Street expectations, and it's still losing money, though the net loss narrowed to eight hundred thirty point nine million. More telling: Lucid increased its delayed draw term loan from its largest shareholder, Saudi Arabia's Public Investment Fund, to one point seven five billion. When the biggest shareholder is also the lender of last resort, you know where the oxygen is coming from. That's a company whose survival depends on one very patient backer. Another off switch, just a friendlier-looking one.
The National Highway Traffic Safety Administration is investigating a Waymo software recall from about a year ago. That recall was issued after numerous reports of Waymo robotaxis illegally passing school buses. And here's the uncomfortable part: the agency said a majority of the reported incidents happened after Waymo deployed the software update meant to fix the problem. Korosec says sources familiar with Waymo told her there haven't been school-bus incidents since March, but she couldn't independently verify that with the Austin Independent School District. So, unconfirmed, and she said so, which I appreciate.
For builders, that one's a gut check. A fix that ships is not a fix that works. If your agent or your software does something wrong, pushing the patch is step one. Proving the behavior actually changed in the field is step two, and that's the step regulators are now looking at.
Tesla's Cybertruck will keep going through next summer, but production will ultimately end, according to Elon Musk, who defended the truck. Tesla is also expanding its robotaxi service to San Antonio, Texas. Uber is launching a service that lets customers reserve and pay for parking in major U.S. cities, about nine months after it bought the parking app SpotHero.
And the one I want to flag hardest: United Airlines will restrict access to AI agents that try to book flights and manage reservations on users' behalf. The newsletter gives it one line, so I won't pretend to know the details. But remember two days ago, when we covered OpenAI's ChatGPT Wallet, and how OpenAI is asking merchants to let signed, verified agents past their firewalls so those agents can shop and check out for you. We talked about Amazon being the biggest retailer blocking AI shopping agents outright. Now you've got a major airline saying, at least for now, not so fast.
This is going to be the fight of the next year in agent commerce. The agent builders want an open door. The merchants want to own the customer relationship, the upsell, the seat selection fee, the loyalty program. Every time an agent books on your behalf, the airline loses a chance to sell you something in the app. So expect a lot of "restrict," "verify," and "partner only" before you see wide-open doors. If you're building an agent that transacts, your roadmap needs a column for "who can block us," and it might be longer than the column for "what can we do."
Which takes us neatly to the deep dive. Because the most consequential story for builders today isn't about who can block you. It's about what happens when you go all-in on one vendor and it works.
OpenAI published a customer story on November 12th about a startup called Gondola. Now, you know my rule with these. This is a vendor writing about its own customer. It's marketing. Every number in it was picked to make the product look good. We've seen a run of these lately; on this show we've talked through Holiday Extras rethinking what it builds with Codex, and just yesterday Shield handing off legacy portal work. So I'm not going to read you the brochure. I'm going to read it like a skeptical investor reads a pitch deck, and pull out what's actually useful and what should make you squint.
First, the business. Jared Kleinstein was running a creative production agency called Fresh Tape Media and noticed creative work for social media was getting made, but the makers weren't getting proper credit. So he built Gondola, a platform to help creatives show what they made, prove its impact, and turn that work into the next gig.
The constraint, according to the story, was the classic small-team problem. Every small experiment competed with core product work for engineering time. And finding information or drafting anything meant digging through a pile of tools.
Here are the headline numbers. Monthly engineering output up twenty-five percent across the team, up fifty percent for the lead engineer. Company revenue more than doubled year over year. And Kleinstein's quote: "Using OpenAI has increased the ambition of what we think a company our size can build."
Okay. Let's take those numbers apart one by one, because how they're measured matters more than the number itself.
Engineering output. The story says, and this is important, the lead engineer's fifty percent increase is "measured in merged changes." Merged changes. Not features. Not customer outcomes. Pull requests that got merged into the codebase. Now, to their credit, the story also says the team is shipping larger changes and that pull request review time has held steady. That's two good signs. If they'd said "fifty percent more PRs" and the PRs got tiny, I'd call it gaming the metric. If review time had ballooned, I'd say the work just moved from writing to reviewing. Neither of those appears to have happened, at least by their account.
But here's what's missing. Revert rates. Bugs in production. Incidents. How much of that merged code had to be pulled back out a week later. The story does claim that two launches, which we'll get to, shipped with fewer bugs and more polish than they would have without Codex. But that's Kleinstein's belief and estimate, not a measured comparison. You can't A/B test your own product launch against the version you didn't build. So treat "fewer bugs" as a founder's impression, not data.
If you're measuring AI coding productivity at your own company, please don't stop at merged changes. Measure what breaks. Measure what gets reverted. Measure time from idea to a customer actually using the thing. Merged changes are the easiest number to move and the least meaningful on their own.
Second number. Revenue more than doubled year over year, "with no added payroll costs or significant overhead." That's the one everybody's going to screenshot. But be careful about the causality. The story does not say Codex doubled their revenue. It says revenue doubled during the same period they adopted Codex. A creator marketplace can double revenue for a dozen reasons: a good sales hire last year, a few big brand clients, a hot market for creator content, a World Cup summer. Some of that growth is surely the team shipping more product. How much? The story can't tell you and doesn't really try.
What I do find credible is the shape of the claim: revenue went up and headcount didn't. That's the thing every investor in the room cares about right now. Revenue per employee. And if a small team can grow the top line without growing payroll, the tools are at least not getting in the way, and probably helping.
The story also says Gondola helped brands hire two and a half times as many creators year over year. That's an outcome for the customer, and frankly it's the most interesting number in the piece, because it's about the marketplace actually working, not about code volume.
Now let's get to the part I think actually matters for how you run a company. Who writes code.
Before Codex, the story says, Gondola's ideas often came from the business development and leadership side, from customer conversations and operational headaches. And the process was the old one: the business person defines the need, writes requirements, then waits for an engineer to have time to build a first version.
Now, the same business folks use Codex Cloud to explore the codebase, prototype an interface or a workflow, test their assumptions, and often get to a functional first version before an engineer needs to step in. And both Kleinstein and Jojo Hesketh, whose title is Creator and Client Partnerships Manager, have merged their first-ever code contributions into the main codebase.
Kleinstein's line: "A product manager doesn't need to stop at a written product requirement anymore. They can bring engineering a working prototype."
That's a real shift, and I want to give it its due. For as long as I've been watching software get made, the most expensive, error-prone step has been translation. The person who talked to the customer writes down what they think the customer wants. The engineer reads it and builds what they think the document means. Then everybody sits in a meeting going, that's not what I meant. A working prototype from the person who actually heard the customer cuts out a whole layer of that telephone game.
Here's the concrete example. A tool called Programs, which helps companies find creatives, review applicants, manage projects and track budgets, with plans to handle contracts and payments. Hesketh sponsored it and used Codex to prototype the workflows before handing them to engineers to refine and release. Kleinstein estimates it saved days, if not weeks. And his quote: "The Programs feature benefited from our ability to rapidly prototype, throw away things that didn't work, and keep the things that did."
That last part is the real value, by the way. Not "build faster." "Throw away faster." When a prototype costs an afternoon instead of a sprint, you stop falling in love with your first idea. You try three versions and kill two. That's how you find the thing customers actually want.
The second example is the World Cup Creator Map. This summer, Gondola built a map that let brands, teams and agencies find photographers and videographers in World Cup host cities. It used data Gondola already had. Hesketh, again, built a fully working prototype with Codex, and engineers polished it for launch shortly after. Kleinstein called the build "incredibly quick," and the story is honest about why: small, self-contained scope, hard deadline.
Pay attention to that caveat, because it's the most useful sentence in the whole piece. Small. Self-contained. Hard deadline. That's the sweet spot. A feature that doesn't touch the guts of your billing system, uses data you already have, and is worthless if it ships a month late. That's exactly where you let the business side run with an AI coding tool. It's a very different thing from letting your partnerships manager refactor your payment flow.
And the story, to its credit, draws that line. Quote: "Engineers remain responsible for architecture, security, data handling, and production readiness."
I want to underline that, because of what we covered earlier this week. Remember the Grok Build mess, where vibe-coded apps were exposing secrets through subdomain URLs, including sales data belonging to SpaceXSI itself. That's what it looks like when nobody owns security and the prototype just becomes the product. Gondola's setup, at least as described, has a gate: the business person brings the prototype, the engineer decides whether it's safe and solid enough to ship. If you take one structural idea from this story, take that one. Anybody can propose code. Somebody with the right job title is accountable for what gets merged.
How the engineers use it, quickly: Codex through the command line and editor tools for implementation, and in GitHub to review pull requests. And Gondola connects Codex to its repositories, development environments, and Slack, so product discussions can inform the work.
The operations side gets a section too. The team uses ChatGPT to prepare proposals, manage schedules, find information across company systems, and automate tasks. They built a daily report template employees use first thing in the morning, which pulls from Google Workspace, meaning Gmail, Calendar, Sheets and Drive, plus Asana, Slack, QuickBooks, the weather, and other sources. Kleinstein says doing that by hand would take "hours and hours of work."
And again, a guardrail: a person reviews consequential work before it goes out externally or gets used for a business decision.
Now, the morning briefing is genuinely useful and genuinely stealable. Any small company can do some version of that. But let me point out what it means. To make that briefing work, you have wired your email, calendar, spreadsheets, documents, project tracker, chat, and accounting system into one vendor's assistant. That's a lot of keys on one keyring.
Which brings me to the detail in this story that I think most people will skim right past, and it's the one that ties to everything else we talked about today.
The model churn. The story says, about building Programs: "The team started with GPT-5.4, did most of the work with GPT-5.5, and used GPT-6.1 for later features, optimizations, and bug fixes."
Three different models, on one feature, inside one year. OpenAI puts that in as a brag, and fair enough, it shows the tools keep getting better. But read it as an operator. The engine under your workflow got swapped twice while you were building. Your prompts, your expectations of what it'd catch, how it behaved in code review, all of that shifted under your feet. This time, every swap was presumably an upgrade. But the point is, you don't control the schedule, and you don't control the direction. The vendor does.
That's not an argument against using Codex. Gondola's results, even discounted for marketing, look real enough. It's an argument for knowing exactly where your dependencies are. Gondola has wired its code, its Slack, its Google Workspace, its accounting and its project management into one provider. That's efficient. It's also a single point of failure, and a single point of pricing power. Think of the Pony.ai CFO: practice your Plan B before you need it. Think of Bustillos: if somebody can shut it off, you don't own it.
So let me give you the founder's takeaways from Gondola, straight.
One. The real unlock isn't engineers typing faster. It's non-engineers bringing working prototypes instead of requirement documents. That changes who in your company gets to try ideas. If you're not letting your customer-facing people prototype, you're leaving the best ideas stuck in a doc.
Two. Pick the right jobs for that. Small, self-contained, time-sensitive, using data you already have. Keep anybody-can-prototype away from the core systems that move money or hold secrets.
Three. Keep an accountable human at the merge button. Engineers own architecture, security, data, production readiness. Write that down. Make it a rule, not a vibe.
Four. Measure what breaks, not just what ships. Merged changes is a starting point. Reverts, incidents, and customer outcomes are the real scoreboard. And Gondola's best number isn't code at all, it's two and a half times as many creators hired for brands.
And five. Know your off switches. Write down every system wired into your AI vendor. Know what you'd do if the price doubled, or a model changed behavior, or access went away for a week. You don't have to have a full backup. But you should know what the backup would cost you. Pony.ai knows its Plan B is more expensive. That knowledge is worth a lot.
Kleinstein's closing line in the story is: "OpenAI gives product people more ability to build, engineers more time for difficult engineering, and our business team more time to spend with customers, partners, and creatives." If that's true, great. That's the dream. Just remember who wrote the sentence down and who published it. Take the lessons, keep your wallet in your front pocket.
Alright. Let's end the main stretch on something lighter, though it's got a little edge to it, because the culture caught up with the tech news this weekend in a way that's worth noticing.
Anthony Ha at TechCrunch wrote up Saturday Night Live's sketch, which put basically the whole tech elite in one room. John Mulaney hosted and played a CIA agent. James Austin Johnson played Peter Thiel. Ben Marshall played Sam Altman. Jane Wickline played Dario Amodei. Jeremy Culhane played Elon Musk. Andrew Dismukes played Mark Zuckerberg. And eventually Bowen Yang showed up as Jensen Huang.
The setup: Mulaney's CIA agent tries, very patiently, to explain to these executives why they should be suspicious when somebody attractive suddenly takes a romantic interest in them. Ha notes Mulaney didn't seem entirely sure how to pronounce "Palantir" or "Nvidia," which, honestly, relatable.
The bit that got me: Johnson's Thiel pushes back and says tech tycoons deserve more respect for everything they've done for the country. "Name one time technology has let us down." And Mulaney reminds everybody about flights getting halted all across the United States on Friday night. Which, if you listened yesterday, you know was traced to a router getting overloaded during a copper-to-fiber upgrade at a remote site, with zero evidence AI was involved. Then Thiel says, "Okay, name one more time," and Mulaney brings up Microsoft Outlook going down on Thursday. Asked for a third, Mulaney says, "No, I'm done naming times."
And then the closer: "I can't imagine how disorienting this has been for you, to go from being bug-eyed weirdos to being gods."
Now, why am I spending time on a comedy sketch? Because when the late-night shows can put Altman, Amodei, Musk, Zuckerberg and Thiel in one sketch and the audience knows every one of them, these people aren't industry figures anymore. They're public figures, like senators or ballplayers. And the jokes have turned. A few years ago the joke was "nerds are awkward." Now the joke is "these guys have all the power and they still fall for the oldest trick in the book." That's not a joke about technology. That's a joke about trust.
For founders, that shift matters. The public mood around tech leadership is skeptical, and it's getting baked into the jokes, which is usually where public opinion settles in for a long stay. When your customers, your city council, or your regulators think about AI, some piece of their gut reaction now comes from sketches like this one. You can't do anything about Saturday Night Live. But you can make sure your own company doesn't hand anyone the next punchline: own your security, be straight with your customers, and don't treat your outage as a "routine upgrade" if it isn't one.
So let's wrap it up.
Every story today came back to the same question. Bustillos put it best: "If you are at the mercy of somebody who can shut the book off, you do not own it." The library paying rent on The Diary of Anne Frank. The Kindle book that can't leave the Amazon store. The flip phone whose best feature works as long as Apple lets it. The ring that comes with a monthly bill. The Chinese robotaxi company practicing life without Nvidia. The airline deciding whether your AI agent gets to book a seat. And a small startup that got genuinely good results by putting its code, its email, its calendar and its books into one vendor's hands.
None of that is automatically bad. Renting is fine. Dependence is fine. Every business depends on somebody. What's not fine is not knowing it.
Here's what I'd watch this week. Whether more merchants and airlines follow United and start restricting AI agents, and whether the verified-agent approach from earlier this week starts winning deals. Whether the export control push turns into actual rules, and whether more Chinese companies talk as openly as Pony.ai about what their Plan B costs. Whether the Flaming Hydra Kickstarter hits its thirty thousand by November 27th, and whether any library consortium signs on, because that'd be the real signal. And in your own shop, take ten minutes, list every system you can't run your business without, and next to each one write down who can shut it off. If you don't like the list, now's a much cheaper time to fix it than later.
That's the show. I'm Tony DeLuca, this was Barely Possible, keep your receipts, keep a Plan B, and I'll see you next time.