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
Okay kiddos, it's your boy Tony DeLuca, this is Barely Possible, and we've got a plate full of stuff today that's less shiny than usual and a lot more useful. Grab your coffee, let's have at it.
I want to start with a router. Not a model or a chip or a trillion-dollar IPO. A router, sitting at what the Transportation Secretary called "a very remote location," that helped ground every flight in the United States on a Friday afternoon.
Here's what we know from Anthony Ha's piece at TechCrunch. Transportation Secretary Sean Duffy said on Saturday that the nationwide ground stop came out of what he called "routine maintenance." Copper wire at a remote site was being upgraded to fiber optic cable by L3Harris, a longtime contractor for the Federal Aviation Administration. After the upgrade, a router overloaded a critical communication system, and that led to the ground stop.
Duffy put it this way, and I'm quoting him: "It's pretty straightforward work that we were doing yesterday at, again, a very remote location." Then he said, "What we're working to determine was whether we had incompetence — meaning, was it user error or installator error from L3Harris? Or did they install a faulty router?"
Installator error. I'm going to let that one sit there on its own.
The more useful quote came from the FAA's acting CTO, Rebecca Guy. She said, "When we started to see the occurrence where our network was having instability, we started to look at what was the cause. And as such, we saw two places that were actually pinging our network at a very high rate — a thousand connections over every four to five seconds." She said the FAA disabled "both of those, right away and were able to stabilize our network," and that they're now doing a root cause analysis.
So walk through the chain. A crew swaps copper for fiber somewhere nobody's ever heard of. A piece of equipment at the end of that line starts knocking on the door of a central system, a thousand times every four or five seconds. The central system gets flooded and goes unstable. And the country stops flying.
The damage, per the piece: more than 8,200 flights canceled Friday evening, according to the aviation data firm Cirium, and at least another 3,000 canceled Saturday, after the ground stop had already been lifted. This was only the third nationwide ground stop of the twenty-first century. Third. Since 2001.
Then there's the timing question, which any of you who've ever run ops already have on your lips. Somebody asked Duffy why this work happened at 3 p.m. on a Friday instead of, say, midnight on a Saturday. His answer was that "switchovers" at bigger airports happen late at night, but remote switchovers happen during the day. And then he said, "There will be reevaluation of when we do this work."
I appreciate the honesty, but come on. Every engineering team listening has a version of the "no deploys on Friday afternoon" rule taped somewhere, and every one of them has broken it, and most of them have a story about why. The lesson is not that the FAA is uniquely sloppy. The lesson is that the risk policy was keyed to how important the site looked, not how important the connection was. Big airport, do it at night. Remote site, do it during business hours. But the remote site was plugged into the same critical network. The blast radius didn't care how small the building was.
And then there's the part of this story that tells you what moment we're living in. The ground stop came one day after the FAA announced the rollout of an AI tool for air traffic controllers, and the controllers' union said it wasn't consulted in designing that system. So of course the first question at every press conference was, did the AI do this?
FAA administrator Bryan Bedford said the agency sees "zero evidence of anything nefarious and zero evidence of anything AI-related." And he said, "We want to be clear that we have an equipment issue, and we're working through what this defective piece of equipment was."
And when Homeland Security Secretary Markwayne Mullin was asked earlier whether the ground stop was caused by AI, he answered, "We call it SI around here, remember? SI." Which, okay. I'll leave the branding department to sort that one out.
But notice what happened. Within hours, the public's first instinct was to blame the newest, scariest thing in the building. And the actual culprit was the oldest, dumbest category of failure there is: a device talking too much on a network that couldn't defend itself. Duffy, for his part, has been saying at press conference after press conference that Congress needs to fund aviation equipment upgrades. You can read this week as evidence for his point. You can also read it as evidence that the upgrade itself is where the danger lives. Probably both.
Now, here's why I'm opening the show with this, and it's not because I want to dunk on the FAA. It's because the rest of today keeps coming back to the same unglamorous layer. The old portals. The remote sites. The systems nobody bothered to build a proper front door for. That's where the action is this week, and that's where our big story lives.
So let's go from a router that knocked too hard to software that's learning to knock politely, which brings us to our deep dive.
Let me set it up. Yesterday on the show, we spent our time on ChatGPT Wallet, OpenAI's setup that lets agents pay merchants using card network tokens without the agent or the merchant ever touching your real card number. That was about agents getting through the checkout counter. Today's story is about agents getting through a different door, the login screen of a crusty old web portal that was never designed for anybody but a tired human with a password manager.
The piece is an OpenAI customer story, published on OpenAI's business site, titled "Shield helps technicians hand off legacy portal work with OpenAI." I'll say up front, this is a vendor case study. It's marketing. It is written to make GPT-6 Astra look good. So we're going to read it the way you read a press release from a guy trying to sell you a used car: interested, polite, hand on your wallet.
Here's the setup. A lot of small and medium-sized businesses outsource their IT. The piece describes the range nicely: everything from "an employee locked out of their laptop to a misconfigured firewall that took the office offline." The people doing that work are IT services providers, and they live inside a pile of portals. Ticketing portals, device management portals, software license portals.
Shield built a product called Forge, which the piece describes as "an AI operating system for IT services providers." The idea is it pulls together business data and customer context so providers can complete IT support tasks end to end. And here's the catch they ran into: if your agent is supposed to finish the job, it has to work inside the same old portals the technicians use every day.
The money quote in the whole piece comes from Mayank Jain, Shield's Head of Technology: "The only time you should have to bring in a human technician is if you need their judgment, not because you're blocked on access."
I want you to chew on that one, because it's the most precise description of the agent bottleneck I've heard in a while. For two years the conversation's been about whether the model is smart enough. Can it reason, can it plan, can it write the fix. Jain's saying the model was often smart enough, and the agent still got stuck, because there was a login screen in the way and no API behind it. The human wasn't being brought in for brains. The human was being brought in for hands.
He describes the stakes plainly. "If a customer is locked out of their computer or there's a phishing attack, speed is paramount." And then, "Whether you're going to solve their issue or you're going to escalate to a human, you want to do either as fast as possible."
That second line is underrated. A lot of agent pitches only count the wins, the tickets the AI closed. Jain's framing counts the escalation too. If the agent is going to fail, fail fast and hand off clean. Anybody building agents should steal that as a design principle. An agent that spends eleven minutes flailing before it gives up is worse than an agent that gives up in thirty seconds with a good summary.
Before GPT-6 Astra, the piece says, Forge's agents "could get stuck at systems they could not access," and customers had to wait for a technician to finish that step. Shield had tried computer use with earlier models, computer use meaning the model reads the screen, clicks, and types, and Jain says they weren't reliable or fast enough.
Then he gives you the line that every single person who's ever used an enterprise vendor portal will feel in their bones: "Usually, a system that doesn't have an API is an old-school system that also doesn't look very nice and is very hard to use, even by a human."
That's the truth. The systems with no API are never the beautiful ones. They're the ones with the nested frames, the dropdown that reloads the whole page, the session that times out while you're reading. So the hardest UI work landed on the systems least designed for automation. That's exactly why earlier computer-use attempts stuttered there.
Here's how the workflow actually runs, per the piece. Forge has a component they call the AI Engineer, which assigns work to a computer-use agent. GPT-6 Astra navigates the interface to complete supported tasks. And, importantly, "Technicians decide which actions it can take on its own and where they need to review or approve the work."
Shield says these browser workflows are live, with "thousands of runs each month across its IT providers." And it's honest about the limits: phone calls, on-site work, and remote-desktop tasks still need human involvement.
The quote the OpenAI marketing team clearly loved: "What shocked me the most was how natural GPT-6 Astra was with computer use. We'd seen models perform computer use before, but you could clearly see them stutter as they try to figure out how to navigate the UI. With GPT-6 Astra, it feels like you're just watching a human click around."
And the numbers. Jain says GPT-6 Astra was 35 to 40 percent faster on average than the legacy models Shield had used for computer-use tasks. He says customer satisfaction is higher on tickets where AI is used, with speed of resolution being a critical factor. Shield reports faster execution, fewer failures, and fewer timeouts, and its teams spend less time writing detailed instructions or nursing the agent through a task.
What's next, per the piece: Shield is testing remote-desktop workflows that would let agents do more complex actions directly on a computer's desktop. And Jain closes with, "With GPT-6 Astra, Forge is now truly an extension of our human technicians."
Okay. That's what the piece says. Now let me tell you what I think it means, and where I'd poke at it.
First, the skeptic's corner. "Thousands of runs each month across its IT providers" is real, but it's not enormous. That's production, not scale. We don't get a success rate. We don't get a failure rate. We get "fewer failures" and "fewer timeouts," which are relative to a baseline we never see. Thirty-five to forty percent faster than legacy models is a speed claim, not a correctness claim. And "it feels like you're just watching a human click around" is a vibe. A good vibe, maybe an important vibe, but a vibe. Also, customer satisfaction being higher on AI-assisted tickets could mean the AI is great, or it could mean the AI gets assigned the easier, more routine tickets in the first place. The piece doesn't say, so I won't assume either way. I'd just like to see it.
Now, with all that said, here's why this is the story I'd want a founder to sit with today.
Point one. The no-API long tail just got addressable. For twenty years there's been an unspoken deal in enterprise software. If you didn't ship an API, you were protected from automation, for better or worse. Integrations had to go through you, or through some screen-scraping RPA script that broke every time someone moved a button. If computer use is actually getting reliable on ugly portals, that protection is eroding. Your user interface is becoming an API whether you designed it as one or not.
If you're the founder of the agent company, that's opportunity. The whole back-office world of small business runs on portals nobody's going to modernize. Insurance carrier portals, payroll portals, state licensing sites, distributor ordering systems. Jain's line, "performing tasks that only a human could six months ago," is the pitch for a hundred vertical startups.
If you're the founder of the portal company, it's a different kind of wake-up call. Somebody else's agent is going to log into your product and do your customer's work, and the customer is going to judge you partly on how well that goes. You can fight it, you can ignore it, or you can get ahead of it by building the API you should've built years ago, and charging for it.
Point two. The permission matrix is the product. Read that line again: "Technicians decide which actions it can take on its own and where they need to review or approve the work." That sounds like a settings page. It's not. That's the trust architecture of the whole business. An MSP that lets an agent reset a password on its own, but requires approval to change a firewall rule, is encoding years of professional judgment about what's reversible and what's not. The companies that win in agent land are going to be the ones that make that matrix easy to set, easy to audit, and easy to tighten after something goes wrong. If you're building agents and you've got a sophisticated model with a crappy approval system, flip your priorities.
Point three, and this is where I want to tie it back to the FAA. Think about what an agent logging into legacy portals actually looks like from the other side of the wire. It's a new client, hitting an old system, at machine speed, possibly in parallel across dozens of customers. Rebecca Guy described the FAA problem as two places "pinging our network at a very high rate — a thousand connections over every four to five seconds." Now picture a fleet of well-meaning computer-use agents all logging into the same creaky ticketing vendor at nine on a Monday morning because a phishing wave hit a bunch of clients at once. That vendor was built for humans clicking at human speed. It may not have rate limiting worth the name. The agent could become the router.
I'm not saying Shield is doing this. Nothing in the piece suggests it. I'm saying if you're building in this space, you need to be the good neighbor. Throttle yourself. Back off when you see errors. Identify yourself where the system allows it. Yesterday we talked about OpenAI asking merchants to let signed agents past their firewalls. That conversation is going to spread from checkout pages to every portal on earth, and the agent builders who behave well are going to get let in, and the ones who hammer are going to get blocked, and they're going to deserve it.
Point four is credentials. For an agent to log into a portal, somebody has to give it a login. That means the agent platform is now holding the keys to a lot of other companies' systems. For an IT services provider, that's the keys to their clients' devices, licenses, and tickets. That's a juicy target. The piece doesn't get into how Shield handles that, and to be fair, a customer story isn't where you'd expect a security architecture write-up. But if you're evaluating any agent vendor that does computer use, credential custody is question one, not question ten.
Point five, the money. Who benefits, who gets squeezed. The obvious winner is the small business customer with the locked-out laptop, who gets helped faster. The IT provider wins if it can handle more clients per technician, and the piece is pretty clearly framing the agent as making each tech go further. Jain calls it "an extension of our human technicians," and I take him at his word that that's how they're using it right now. But let's be grown-ups. The tasks being handed off first are the access tasks, the logging in and clicking through, and in a lot of IT shops that's exactly what the junior tier-one person spends the day doing. That's also how a junior person learns the trade. If the agent eats that layer, the industry has to figure out how anybody becomes a senior technician. That's not a reason to stop. It's a reason to think about it before it bites.
Now, this pattern isn't just Shield. There was another short recent item on OpenAI's site, from earlier in the week, about a company called Cogent. The description is thin, one line: Cogent "combines OpenAI models with enterprise context to help automatically remediate vulnerabilities with agents." The headline calls it building "cyber defense factories for enterprises." I don't have much more detail than that, so I'm not going to pretend I do. But look at the shape of it, same as Shield. Model plus the customer's context plus agents that don't just tell you about the problem, they go fix it. The thing being sold isn't intelligence. It's the last mile of action.
So if you take one thing from the deep dive, take Jain's sentence and pin it up: bring in the human for judgment, not for access. Then ask yourself, in your own product, how many of the times a person gets pulled in are really about judgment, and how many are about a login screen. That ratio is about to change, and it's going to change faster in the ugly corners than in the pretty ones.
Alright. Let's shift from agents to the biggest corporate soap opera in tech, because there's genuinely new material on the Nvidia front.
We've been following the Jensen Huang saga on this show; back on the thirteenth we spent time on whether his conduct was more common in the C-suite than people admit, and on the deals Nvidia struck under him. I'm not going to rehash all that. Two new pieces of reporting moved things this weekend, so let's take them in order.
First, the petition. Anthony Ha at TechCrunch again. A petition calling on Nvidia to reinstate Huang appears to have attracted more than 115,000 signatures as of Saturday afternoon. Note the word "appears." These are the petition website's own numbers. According to that site, the signers include 5,965 current Nvidia employees, 1,108 former employees, and 63,063 shareholders.
The notable names: Tesla and SpaceX CEO Elon Musk, Replit CEO Amjad Masad, and Epic Games CEO Tim Sweeney.
The petition says its signers are "concerned that Jensen was removed without meaningful due process," and demands that the board "bring him back while an independent investigation establishes the facts." It also asks for a fair investigation and more public information. And the line that'll end up on the protest signs: "Jensen deserves due process, not mob justice."
For the facts as TechCrunch lays them out: Huang resigned on Thursday, several days after confessing to an affair with a chip executive whom U.S. intelligence identified as a suspected Chinese spy. News reports have also suggested multiple extramarital affairs. Nvidia appointed its CFO, Colette Kress, as interim CEO while it searches for a permanent replacement. Nvidia's stock fell 2 percent on Friday. And TechCrunch notes his departure is expected to leave a big gap in AI industry funding.
Now let me give you my read on the petition, because I've been watching neighborhood petitions since before I could vote, and they all have the same tell.
The number that matters isn't 115,000. Anybody can sign anything on the internet. The number that matters is 5,965 current employees, if that figure holds up. That's people putting their name on a public document asking their own board to reverse a decision. That's not nothing. Even discounting for the petition site's self-reporting, that's a workforce telling its board where its loyalty sits.
And here's the uncomfortable part for the board. There's a real difference between "he resigned" and "he was removed without meaningful due process." The petition is betting the second framing sticks. TechCrunch's headline calls him "recently ousted," while the body says he resigned. That wobble in language is the whole fight. If it was a voluntary resignation, there's no due process to argue about. If it was a push, everybody wants to know who pushed and on what evidence.
Now, a little skepticism the other way. Shareholders signing a petition to bring back the founder of a company whose stock they own is not a profound act of conscience. It's a position. And the famous names on the list are founders and CEOs who have their own reasons to dislike the idea that a board, or Washington, can take out a founder. I'm not saying they're wrong. I'm saying everybody in this story is talking their book, and that includes the petitioners.
For founders listening, the governance lesson writes itself and I'll say it once. If your company's identity, culture, and partner relationships run through one person, then that person's private life is a business risk whether you like it or not. That's not a moral statement. It's a balance-sheet statement. The time to build the bench is before you need it.
And that's where the second piece comes in, because the partner relationships are real and they're international.
Kate Park at TechCrunch put together a roundup of how governments across Asia are responding to Huang's exit, and a fair amount of what's in it is older material that's resurfacing now because of the scandal. Let me separate the threads.
The older thread first, Singapore. According to Singapore's Ministry of Home Affairs, authorities seized the passport of Alice Huang, the former CEO of Singapore-based cloud computing company Megaspeed, back in June 2025, as part of an investigation into possible false declarations about server purchases. The ministry said police took the passport to make sure she stayed available for questioning, since she wasn't based in Singapore. The investigation looked at whether servers shipped from Singapore to Malaysia were then sent on to other countries without proper declaration.
Then in February 2026, per the ministry, police found insufficient evidence to establish she'd committed offenses on the suspected false declarations. But authorities said she had knowingly provided inaccurate information about her role at Megaspeed. She got a warning, was barred from re-entering Singapore, and her passport was returned that month. Singapore also said it got limited help from U.S. authorities despite repeated requests for information about U.S. server maker Supermicro. And the ministry says its investigation into Megaspeed and six other unnamed individuals remains ongoing.
So that's not breaking news in the strict sense. It's a year-and-a-half-old trail getting dragged back into the light. But it matters, because it tells you the server-routing questions were on at least one government's desk well before any of this hit the front pages.
Then there's how the region's governments are positioning, per the roundup.
China. The foreign ministry rejected, on November 10, the allegation that the businesswoman linked to Huang was a spy, calling it groundless and a smear. A spokesperson later reiterated opposition to export controls that disrupt supply chains, and said the ministry wasn't aware of reports linking restricted Nvidia chips to a Shanghai data center associated with Alice Huang. The piece also cites Reuters' review of procurement notices and corporate disclosures finding more than $3 billion in Chinese AI computing projects reliant on Nvidia's restricted processors now facing uncertainty.
South Korea. The Science Ministry said Nvidia reaffirmed its plan to supply 260,000 advanced GPUs to the country and continue joint AI projects. Separately, per local media, the government's AI strategy council held an emergency meeting to assess the impact.
Taiwan. Economy Minister Kung Ming-shin said the departure was unlikely to significantly change the growth of the global AI industry. Taipei's city government said Nvidia's planned overseas headquarters there remains on track, that the company has met its contractual obligations so far, and that officials will keep supporting it.
Japan. The Industry Ministry said it would closely monitor Nvidia, calling it an important player in Japan's AI semiconductor industry. Deputy chief cabinet secretary Masanao Ozaki said it was too early to assess the impact, while stressing Japan's commitment to stable semiconductor supply.
Put those side by side and you get a pretty clear picture. The allies are doing the same thing: get the supplier to say the commitments still stand, put it on the record, and say "too early to tell" about everything else. South Korea's 260,000 GPU reaffirmation is the tell. When a government feels it needs the vendor to re-promise a deal publicly because the CEO left, that tells you how much of the relationship was personal.
For builders, especially if you're buying compute or depending on a neocloud that depends on Nvidia's goodwill, here's what to watch. Not the petition. The contracts. Do the big allocation commitments get honored on the same terms under an interim CEO, or do they quietly get renegotiated? And does the "gap in AI industry funding" TechCrunch mentions show up as slower financing for the smaller cloud players you might be renting from? That's where this scandal stops being gossip and starts showing up in your infrastructure bill.
Which is a pretty good bridge into a quick one on data centers.
There's a recent TechCrunch video segment from earlier in the week, with Theresa Loconsolo, Rebecca Bellan, and Tim De Chant, with the title "Why the data center you dread might never even be built." The write-up is short, but the argument's clean. The data center boom is running into local opposition, permitting delays, and a scramble for electricity. But before any of that, there's a more basic problem: a lot of the projects being announced may never get built in the first place.
I only have the summary, not the full segment, so I'll stay in my lane. But the point lands on its own. An announcement is not a building. A press release about gigawatts is not gigawatts. And when the financing behind those announcements is shaky, or the guy who was lending his company's balance sheet to the whole sector just walked out the door, the list of announcements gets a lot less reliable.
For founders, the practical takeaway: when you're planning on compute that's "coming online next year," find out whether it's coming online from a hole in the ground with steel in it, or from a slide deck. Ask your provider which one. They'll know.
Now let's switch from things that might not get built to things that are getting sold before they really exist in the way you'd assume. Tesla opened Cybercab pre-orders.
Also from Anthony Ha. Tesla said Friday evening that "select customers" can now pre-order the autonomous, golden Cybercab "for personal use." The details came out as those customers looked at the order page. Tech reviewer Marques Brownlee posted that buyers "must accept delivery" in one of a handful of Texas cities, and that the Cybercab will only operate in certain geofenced areas. Estimated delivery is December 2026.
The price: $29,990. Plus $299 a month to operate the vehicle.
Musk responded to Brownlee's post saying the Cybercab "will be available nationwide, as we receive regulatory approval." And then, interestingly, he said the monthly fee is "too high," that it "Should match the cost of Tesla Full Self-Driving and connectivity for other models, which is more like" about $115 a month.
Tesla's Ashok Elluswamy added detail on the geofence, saying Cybercab will work "unsupervised inside all the same geo locations as @robotaxi, just serving the owners' needs."
And the backdrop: after falling far short of projected sales, Tesla recently announced it'll stop Cybertruck production next year and use the freed-up factory capacity to build more Cybercabs.
Okay. Let me do this the kitchen-table way. You pay thirty grand for a car. You pay three hundred a month to operate it. And it only drives itself inside the same zones where Tesla's own robotaxi service already runs. So in practice you're buying a vehicle that's a private version of a taxi network you could also just hail.
That's a really strange product to own. And I think the strangeness is the whole story. A car you own has always been the thing that goes where you go. This car goes where the permission goes. The geofence isn't a feature limitation; it's the business model. It's a subscription to a service area, with a car attached.
And look at the fee. The order page says $299. Within hours, the CEO says publicly it's too high and should be more like $115. Either the order page was wrong, or the CEO is negotiating against his own pricing page in public, or the number is a placeholder for a price nobody's actually settled. Any of those would be fine for an early beta. They're less fine for something you're asking people to put down money on. If your monthly fee can drop by more than half because of a reply to a YouTuber, the fee isn't a price yet. It's a guess.
Builder lesson, and this one's cheap: pricing is a promise. When you change it in public, in a reply, you're telling your customers that the number they saw at checkout wasn't real. Sometimes that helps you, like here, where it's going down. But you're also teaching people to wait for the next reply.
And there's a bigger angle for anyone building in mobility or local services. The geofence is the real map of where autonomy is a business. Not the national map Musk promises "as we receive regulatory approval," but the actual polygons where the thing runs unsupervised today. If you're building anything that depends on autonomous delivery or rides, plan for the polygons, not the promise.
Next, a quick one on a story that's not exactly new but is getting some renewed attention. An Ars Technica report from Eric Berger, the site's senior space editor, on Starship flight 15 resurfaced in the pile this week, so treat this as an older story coming back around rather than breaking news. And I'll say plainly, I don't have the outcome of everything it was previewing, so I'm only going to tell you what the piece itself says.
Per Berger's report, Starship's first stage ascended cleanly and released the upper stage, which powered into what he called a nominal orbit. It was the second time SpaceX had flown Starship into low Earth orbit. The primary mission was deploying 26 Starlink V3 satellites, each with solar panels stretching 60 meters end to end when deployed, and all 26 made it into orbit. Berger's line: "This means that Starship is now an operational, orbital rocket."
But the booster, which SpaceX wanted to catch with the launch tower, was commanded to divert into the Gulf due to an unspecified issue, and it exploded. And the piece was setting up a potential first-ever catch of the upper stage, contingent on a series of health checks in orbit and on reentry, because the vehicle would be passing over parts of Mexico and the United States before returning. Berger flagged the heat shield as one of the biggest remaining hurdles, with SpaceX citing extra tile retention, fixes for plasma flow paths behind tiles, and a curved tile design meant to reduce heating in the gaps.
Here's why I bring it up on a founder show. Look at how Berger structures it. Primary mission, a success. Reuse, not yet. "The challenge for the company now is to make Starship reusable and dramatically increase its flight rate." That is the exact transition every hardware company, and honestly a lot of software companies, hits. Working once is a demo. Working cheaply, over and over, is a business. Getting the payload up proved the rocket. Getting it back proves the economics. And notice the health-check logic before any catch attempt: you don't bring the expensive, dangerous thing home over people's houses unless every gauge says it's safe, and you always have an earlier option to dump it in the ocean. That's a release process. It's the opposite of 3 p.m. on a Friday at a remote site.
Now, let me take you somewhere you might not expect on a tech show: vineyards in South Africa. Stick with me, because this one's got more to say about running a business under pressure than half the startup advice on the internet.
It's a piece republished at Ars Technica from The Conversation, by Hannah V. Herrero, Erin L. Bunting, Jane Southworth, Jason K. Blackburn, and Stephanie Insalaco-Wyner, under the title "Winemakers are learning that conservation can be good for business."
The setting is South Africa's Western Cape. More than 500 vineyards, producing about 90 percent of the country's wine, in one of the most biologically diverse landscapes on the planet. The pressures are brutal. The authors say temperatures there have risen 1.5 to 2 times faster than the global average over five decades. Storms are more intense. And the late 2010s brought three years of severe drought that put vineyards under intense stress and cut grape production.
The researchers talked with growers, conservation practitioners, and industry experts, and analyzed how 107 vineyards, about 20 percent of the total, were using land stewardship practices to adapt. And the headline finding is that many vineyard operators don't see conservation as land taken out of production. They see it as protecting the processes the business depends on.
The practices are concrete. Restoring native vegetation along waterways and removing invasive species to protect water supplies. Cover crops between vine rows: legumes, oats and rye, radish and mustard. Compost and mulching to build organic matter in the soil. Some vineyards using regenerative practices like letting cattle graze, so they fertilize the ground and work organic material in with their hooves. And on pests, monitoring populations closely and responding in a targeted way instead of spraying broadly, encouraging beneficial insects, managing habitat for natural predators, and using pruning to prevent outbreaks.
There's a quietly important detail: the industry's own certification system, the Integrated Production of Wine, dates back to 1998, with standards on pesticides, water, wastewater, and stewardship, plus independent auditing. Most South African vineyards participate. So this isn't a fad. It's a decades-old operating system that's getting more use as the pressure rises.
And the authors don't sugarcoat it. Cover crops take maintenance and can harbor pests. Chemical treatments are still part of many operations. Growers are facing rising costs across exports, labor, and chemicals. The area in vineyard production has actually shrunk, and some growers are diversifying or switching crops entirely to survive hotter, drier conditions.
So why's Tony talking about soil on a tech show? Because the authors put their finger on a management idea that applies straight across. Their phrase is that these approaches emphasize "prevention, monitoring and ecological balance," rather than "treating problems after they arise."
Go back to the top of the show. The FAA treated a remote site as low-risk because it looked small, and a single router flooded the whole system. The vineyards that are making it are the ones that stopped treating the edges of their land as wasted space and started treating them as the thing holding the water in. Same lesson, different dirt. The edges of your system, the remote site, the old portal, the strip of native brush along the creek, are not overhead. They're load-bearing.
There's also a business-model point I really like. Growers use sustainability certification to make their wine more marketable. That's the label. But the researchers found the deeper reason they stick with it is that it lowers their costs and risk, fewer chemicals bought, better water retention, less drought vulnerability. The label sells the bottle, but the practice keeps the farm alive. Founders, that's the difference between compliance theater and compliance that actually earns its keep. If the only reason you're doing the responsible thing is the badge on the website, you'll drop it the first bad quarter. If it's also lowering your cost of failure, you'll keep it.
Alright, let's lighten up and do a little hardware corner, because two consumer pieces landed that say something about how products get shaped.
First, Apple's iPhone Duo, the foldable. Sarah Perez at TechCrunch wrote up nine settings to make it easier to use, and the piece notes the Duo has been popular enough that Apple can't keep it in stock. I'm not going to read you a settings menu, but a couple of details are worth your time.
The Duo doesn't support Face ID. It's back to Touch ID on the side. And because you can hold it in multiple orientations and flip it around, Perez recommends registering a bunch of fingerprints, on both hands, wherever you find yourself naturally gripping it. Unlocking with your Apple Watch helps. There's a "rest finger to open" option, which is technically an accessibility feature, that lets you unlock without pressing.
The device is wide, described as shaped like a passport. So Reachability, which on a normal tall iPhone pulls the top of the screen down to your thumb, on the Duo lets you swipe the left or right edge to bring the far side into reach. The volume buttons are "fixed" by default, meaning up stays up no matter how you turn the phone, and a lot of owners turn that off so the buttons rotate with the device. And there's no built-in setting to lock the phone when you fold it closed, so you have to build a Shortcuts automation to do it.
Here's the builder angle. When a top-selling new device ships with its owners immediately reaching into accessibility settings and hand-built automations to make daily use feel right, that's a first-generation product telling you where the rough edges are. For app developers, the passport shape is the big one. If your app assumes tall and narrow, with key controls at the top, it's going to feel awkward on a wide device where the hard-to-reach spot is off to the side. If the Duo really is selling out, it's worth an afternoon to see what your app looks like on it.
Second, Amazon's refreshed Kindle lineup, and specifically the new Kindle Colorsoft, reviewed by Aisha Malik at TechCrunch. It starts at $290; the Signature Edition she tested is $320, with 32 gigabytes of storage and an aluminum back. Lighter and thinner than the first-gen Colorsoft, edge-to-edge display, 7-inch color E Ink screen, 300 pixels per inch for black-and-white and 150 for color. Colors are still muted compared to an iPad. Battery up to eight weeks, versus up to twelve for the Paperwhite. The power button finally moved off the bottom edge, where people kept hitting it by accident, up to the top right.
Her verdict is refreshingly blunt. If you already own the first-gen Colorsoft, there's little reason to upgrade. If you really want color for comics, graphic novels, or cookbooks, it's worth a look. For everyone else, the basic Kindle at $169.99 or the Paperwhite at $199.99 is probably the better buy. "A nice upgrade, but far from a must-have."
I mention it because there's a quiet product lesson in there. The best-received change in her review isn't the color. It's moving the power button. Nobody puts "we moved the button so you'd stop accidentally turning it off" in the keynote, but that's the kind of fix that makes people like a product. The flashy feature gets you the review. The boring fix gets you the second purchase.
One more before we wrap, and it's about distribution, not technology.
OpenAI published a recent post, around Veterans Day, saying more than 20,000 transitioning servicemembers and veterans have claimed a free year of ChatGPT Plus over the past year. Beyond the free access, OpenAI Academy and community partners have hosted more than a dozen trainings for nearly 1,000 members of the military community, with seven more scheduled before year-end. The post mentions a workshop in Reston, Virginia, with nearly 100 veterans and active-duty servicemembers, and more than 80 American Legion members attending three sessions at the Legion's national convention in Louisville. The use cases are practical: organizing a job search, rehearsing interviews, translating military experience into a civilian résumé. One participant, Ryan Kaono-Wedemeyer, learned to start by writing out a detailed account of his service and then asking the AI to help shape it into a résumé. An instructor, Erin Marie, helped participants build custom GPTs for their transitions.
My take is simple and not cynical. Translating years of military experience into language a civilian hiring manager understands is a real, painful problem, and if a tool helps, good. And for founders, notice the playbook. The free year gets the tool in the door. But the post spends most of its words on training, workshops, and partners, because handing someone a free account doesn't make them a user. Sitting with them in a room while they figure out how it helps their actual problem does. Twenty thousand claims is the headline. The thousand people in a room is the part that probably moves the needle. If you're selling to any audience that isn't already sold on AI, budget for the room, not just the free tier.
Alright, let's pull it together.
What ran through today was the unglamorous middle of the stack. A router at a site so remote the Transportation Secretary had to say "very remote" twice, flooding a network until the country stopped flying. A company that figured out its agents weren't held back by intelligence but by login screens, and found a model that could finally click through the ugly ones. A chip giant whose foreign partners are suddenly getting supply promises re-stated on the record because the relationship turned out to be more personal than anybody wanted to admit. A car you buy that only works inside somebody else's map. And South African farmers who figured out that the scrubby edges of their land were holding the whole operation together.
Here's what I'd watch this week.
On agents: watch whether more computer-use customer stories start showing real failure rates, not just speed. When vendors are confident, they publish the denominator. And watch how portal vendors react once agents start showing up at their login pages in volume. Rate limits, bot policies, and new APIs are all coming.
On the FAA: watch for the root cause analysis. Specifically, whether it lands on bad hardware, bad installation, or the more uncomfortable answer, which is a network that let two devices knock that hard without anything stopping them. And watch whether that "reevaluation of when we do this work" turns into an actual policy.
On Nvidia: skip the signature count and watch the contracts. Do the big allocation promises, like South Korea's 260,000 GPUs, hold on the same terms? And does the funding gap show up in the financing of the smaller clouds you might rent from?
On Tesla: watch the monthly fee. If it drops toward that $115 number Musk floated, the pricing page was a guess. If it doesn't, the CEO was. Either way, watch the geofence, because that's the real product.
That's the show. Treat the edges of your system like they hold the water in, because they do. Tony DeLuca, Barely Possible, see you next time.