Subscribe
Share
Share
Embed
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, I'm your boy Tony DeLuca, and Barely Possible is back on the air. Fresh pot of coffee, fresh stack of stories, and today's got a little bit of everything: surveillance networks eating themselves, Amazon feeding rare books into a shredder, and a poor guy who took a boulder to the back seven hundred years ago and just got his day in the science pages. Buckle up. Let's have at it.
Let me start where I want to spend the most time, because it's the story that actually teaches you something about building a business, not just running one. It's a piece from Nate Anderson about Flock, the license plate camera company, and how a whole run of Wisconsin towns have been pulling the plug. Now the headline sounds like a privacy story, and it is one, but underneath it there's a business lesson so clean it belongs on a whiteboard. So that's our deep dive today. And I want to be straight with you about the timing: this reporting walks through decisions that started back in April and rolled through the summer. So this isn't breaking-this-morning news. It's a recent piece that stitches the whole arc together, and the arc is what matters.
Here's the setup. Flock makes automated license plate readers. Cameras on poles, solar panels, the whole deal. They snap your plate, they log where you were and when, and — this is the important part — they share that data across a giant network of law enforcement agencies. That sharing is the product. That's the magic. A cop in one little Wisconsin town isn't just seeing his own cameras, he's tapping into a network that, in the case of Dane County, reached a hundred and forty other agencies. Cops in Missouri, New York, Tennessee. Your local plate reader becomes valuable precisely because it's plugged into everybody else's.
That's a network effect. Classic stuff. Every new town that joins makes the whole thing more useful for every town already in it. It's the same math that makes a phone network or a marketplace worth something. More nodes, more value, and the value compounds. Investors love it, founders build slide decks around it, and for a while it looks like a one-way ratchet. It only goes up.
Except. And here's the whole point of the story. Nate Anderson's framing is right there in the subhead: the network effect can run in reverse.
Watch how it actually played out, because the dominoes are beautiful and a little terrifying if you're the one selling the cameras. Dane County — that's the one with Madison in it, one of the most populous counties in the state — the Board of Supervisors votes on April 16th to cut eighty thousand dollars in funding and ban further spending on the system. County Board Chair Patrick Miles didn't mince words. He called the company "a proven bad actor" and talked about Fourth Amendment concerns, unreasonable searches, the whole civil liberties argument. And the cameras went dark.
Now here's where it stops being a privacy story and becomes a business story. When Dane County went dark, it didn't just lose its own cameras. It pulled that data out of the shared network. And the neighbors noticed immediately. The town of McFarland, right next to Monona, had been planning to deploy Flock cameras. They looked at the map and backed out. And listen to exactly why the McFarland police chief, Brian Redman, said they backed out. Quote: "Once Dane County lost their funding for Flock, we lost about half of the potential camera network within Dane County. I had to ask, 'Were we still going to get the return on our investment with that loss?' With the network being chipped away, that wasn't going to be the case." End quote.
You hear that? That's not a privacy objection. That chief still likes the technology. He did the math. Half the network evaporated, so the ROI didn't pencil out anymore. He walked because the product got worse, and the product got worse because somebody else left.
And then it cascades. Monona puts its contract "under review." Fitchburg votes to discontinue in late May. The University of Wisconsin police in Madison decline to renew in July. Stoughton cancels in late July, even paid a twelve-thousand-five-hundred-dollar early buyout fee to get out. And up in the Fox Valley, same movie: Oshkosh out in April, Appleton out in May, Grand Chute switches to a competitor, Axon, in June. And then Kaukauna. Kaukauna spelled it out even more plainly than McFarland did. They said the whole appeal was the ability to work with the surrounding communities to track suspects crossing city lines. With everybody else bailing, their police chief Jamie Graff said it was, quote, "not fiscally responsible" to keep investing. The reason to be in the network was the network. Once it hollowed out, there was no reason left.
Now here's why I'm making a founder sit through medieval-town-council minutes from Wisconsin. Because the thing that makes network-effect businesses so valuable on the way up is the exact same thing that makes them fragile on the way down, and almost nobody prices that in. When you're pitching, the network effect is your moat, your defensibility, your reason the valuation deserves a fat multiple. But defensibility and fragility are the same coin. A business whose value comes from everybody being in it is a business that can unravel fast when people start leaving — because each departure doesn't just cost you one customer, it degrades the product for everyone who stayed, which gives the next customer a fresh reason to go. It's a bank run with cameras on poles.
And notice the trigger here. The thing that started the exodus wasn't a better competitor or a price war. It was trust. It was public resistance, citizens showing up to county board meetings, privacy experts making noise, that word "bad actor" getting said out loud by an elected official. The reputational hit is what knocked the first domino. And in a network-effect business, a reputational hit doesn't stay contained to one account — it propagates through the whole graph because of how the value is wired together.
Flock, to their credit or maybe just their survival instinct, saw the fire. The piece notes that the backlash got strong enough that the company announced new mandatory audit controls and lower default data retention periods for US users. That's a company trying to patch the trust hole before more nodes drop off. Whether it's enough, we'll see. But the lesson for you, whether you're building a marketplace, a payments network, a developer platform, a social product, anything where your value scales with participation: your churn model is probably wrong. If you model each departing customer as costing you just that customer's revenue, you're missing the second-order effect. In a real network business, a departure lowers the value delivered to everyone else, and that's where the cascade lives. Model the compounding on the way down the same way you brag about it on the way up.
And one more thing, because it's the part builders always want to skip. These towns didn't reject the technology. Read the quotes again — half of them said they still believe in license plate readers, they just don't trust this vendor. Grand Chute went to Axon. UW-Madison said they'd find a vendor that "better meets our needs." The market didn't disappear. The trust did. Which means the moat was never really the cameras or the software. It was being the trusted operator of the shared thing. Lose that, and the network you built becomes the mechanism of your own unwinding.
That belief-in-the-tool-but-not-the-vendor split is a good bridge to the next one, because it shows up again in a completely different corner of the industry.
Let's talk about Amazon and the books. TechCrunch's Amanda Silberling, building on reporting from 404 Media, and it's a doozy. Amazon — the company that literally started life as an online bookstore, that's the whole origin myth, Bezos and the cardboard boxes — is now buying up rare books, cutting the spines off, and scanning the pages to train AI models. How do we know? 404 Media put a tracking device inside a rare book and watched it travel to an Amazon facility in Las Vegas. A facility that, and I am not making this up, marks itself with a logo of a dinosaur holding a book in its claws. Somebody in a branding meeting thought that was a good look. A dinosaur. Holding a book. You cannot write it better.
Amazon's statement was the corporate equivalent of a shrug: it "purchases books through commercial channels to improve the products and services customers use." Okay. But here's the why, and the why is genuinely interesting for anybody thinking about where AI goes next. These companies have basically eaten the internet. They've ingested everything they can scrape off the open web. So now they need text that's rare, out of print, impossible to find online — and here's the kicker — text that was written before 2022. Why does the date matter? Because anything written after the chatbots showed up might itself be AI-generated, and when you train a model on AI slop, you get what researchers call model collapse. The quality degrades. It's a snake eating its own tail. So pre-2022 human writing has become a kind of clean, uncontaminated resource. And rare physical books are a vault of it that hasn't been strip-mined yet.
So think about what that means. We've hit a point where the raw material — genuine, un-poisoned human text — is scarce enough that it's worth buying rare books just to shred them for the pages. The article even reminds us that in Anthropic's case, some of that book-hunger went the illegal route, with pirated books. Amazon at least is buying them. But the destruction is the part that sticks. These are physical objects, some of them irreplaceable, and they're being fed into a scanner spine-first and then, presumably, into a bin. There's something about the bookstore company killing books for parts that just lands wrong, and I don't think you have to be a sentimentalist to feel it.
For a builder, the takeaway underneath the outrage is this: data provenance is becoming a real competitive axis. Clean, dated, verifiably-human training data is turning into an asset class. If you're building anything that depends on model quality downstream, the pipeline your provider trains on actually matters, and "we scraped the web" is starting to be a liability, not a flex.
Now let's shift from what's feeding the models to who's getting hit by the fancy stuff built on top of them. There's a cybersecurity story from TechCrunch's Lorenzo Franceschi-Bicchierai that's worth your attention. Apple sent out a fresh wave of spyware notifications on Friday — these are the alerts Apple sends when it believes you've been targeted by what they call "mercenary spyware," the government-grade stuff. And investigators are saying this batch looks like the biggest yet. Apple hit customers in a hundred and ten countries.
The group Access Now, which runs a helpline for spyware victims, said they got a record number of people reaching out — thirty to forty percent above what they normally see after one of these waves. And here's the detail that stuck with me. One of the people who got flagged was a Ukrainian soldier, fighting the war against Russia. He first thought it was a scam, verified it with Apple, and then said, quote, "I wouldn't have thought I was important enough for them to target me like this. I am flattered though." A little gallows humor from a guy who just found out a government thinks he's worth hacking.
Now there's a genuinely important nuance here, and John Scott-Railton at Citizen Lab and the Access Now folks both flagged it. Part of why the numbers look so big is that Apple changed how it notifies people. Starting this year, they hit you on the lock screen, in Settings, by email, and when you log into your account on the web. Harder to miss, harder to ignore. So some of the spike is real growth in attacks, and some of it is just that Apple got louder about telling people. Both things can be true. But as Scott-Railton put it, for every public notification, there's a huge iceberg the public never sees. The practical advice, if you or someone on your team ever gets one of these: take it seriously, and turn on Lockdown Mode. Apple says it's not aware of anyone with Lockdown Mode enabled getting hacked. That's not nothing.
Now let's move from security to the churn of the startup world, and there's a cluster of items here that tell a story when you put them side by side. First, Relay. This is an AI workflow automation startup, launched back in 2021, and its whole ambition was to become the new Zapier. It's shutting down. Paying customers lose access September 14th. And the founder and CEO, Jacob Bank, is going back to Google — he's rejoining as VP of Product for Google Chrome. Now, I want to be careful here, because the shutdown itself was actually announced back in July; this is the follow-up with where the people landed. Bank spent six years at Google before this, was product lead on Gmail and Calendar, so it's a homecoming. And his line is telling. He called Chrome "a perfect place to collaborate with agents." Which tells you exactly where the wind is blowing. The independent automation startup couldn't hold the ground, and the guy who built it decided the browser — with a billion Gemini users flowing through Google's front door — is where the agent action is going to be.
Put that next to Higgsfield, which sits at the opposite end of the same market. Higgsfield does AI image and video generation, and they just raised a four hundred million dollar Series B at a five-point-four billion dollar valuation. That's a quadrupling of their valuation in eight months. Eight months. They're claiming seven hundred million in annualized revenue, thirty million users, and working with three hundred ninety of the Fortune 500. Founder Alex Mashrabov, former Snap exec, said something that connects right back to our Flock conversation about second-order costs. He said video is one of the most compute-intensive domains in AI — one minute of video is like processing sixty thousand words. So a big chunk of that four hundred million isn't going to hiring, it's going to buying compute just to stay competitive with Runway and Synthesia. That's the tax. In the video generation game, your gross margin lives and dies on your compute deal.
So look at the two together. One AI startup folds and its talent gets absorbed into a trillion-dollar platform. Another AI startup quadruples in eight months but has to spend a fortune on compute to keep the lights on. That's the barbell of this moment. Either you get big enough fast enough to justify the compute bill, or you get acqui-absorbed into the mothership. The comfortable middle is getting squeezed out, and Relay is exhibit A.
Let me hit a couple of platform-mechanics stories quickly, because they matter if you make or measure content. YouTube announced it's changing how it counts a view. Used to be, roughly, you had to watch about thirty seconds. Now, as of August 24th, a view counts the instant the video starts playing. Same thing TikTok and Instagram already do. And here's the honest consequence: view counts are about to inflate, especially on videos that grab you for two seconds and lose you. So the public view number becomes a much worse signal of whether anything's actually popular. YouTube's keeping the old metric, renaming it "Engaged views," and tucking it into the analytics dashboard for creators. Read that as: the vanity number goes up for everybody, and the real number goes private. If you're an advertiser or a builder relying on public view counts to judge reach, recalibrate. And this lands the same week YouTube raised the bar for new creators to start earning — eight thousand watch hours now, up from four thousand. So it's easier to rack up "views" and harder to get paid. Funny how that works.
And a quick companion note that'll sting for some of you: Feedly, the biggest standalone RSS reader out there, fifteen million users, has been running unusably slow for over a week, and paying customers say their support tickets got ignored. The CEO says it's a bug related to "Mark as Read" on accounts with lots of folders, not their pivot to AI cyber-threat intelligence. Maybe. But the piece raises the fair question of whether a company that now advertises the AI product on its homepage still has its eye on the RSS ball. It's the read-it-later graveyard all over again — and speaking of which, we also saw the resurfaced rundown of Pocket alternatives, since Mozilla shut Pocket down. If you're one of the people still living in RSS, keep an export handy. When the core product isn't the growth story anymore, the core product tends to rot quietly.
Now a couple of quick ones from the labs and the ad world, and then I want to close somewhere unexpected. OpenAI put out a run of posts — funding fourteen independent AI policy projects, joining a community investment project in Southern Ohio, and a cybersecurity piece called The Defender's Window about AI reshaping attack and defense. I'll be honest with you, these came through thin on detail, so I'm not going to pretend there's a big story where there isn't one. File it under: OpenAI is doing the community-investment, policy-friendly dance that companies do when they're getting very large and want the neighbors on their side. Which, funny enough, connects back to a debate we covered yesterday — Dario Amodei and Gavin Baker going back and forth about trust, regulation, and whether the labs have delivered on their promises. I won't relitigate it. But you can see the whole industry, OpenAI included, waking up to the idea that trust is the scarce resource. Same lesson Flock learned the hard way in Wisconsin.
And Google's out here signing five of the biggest football clubs on earth — Arsenal, Barcelona, Bayern, Liverpool, PSG — as partners for Gemini and Pixel. Match insights, behind-the-scenes content, and a real commitment to raising the visibility of the women's game, which I'll give them genuine credit for. But strip the jersey off it and it's a distribution play. Google wants Gemini to be the thing you pull up mid-match to settle an argument about a formation change. Get into the habit loop, get into the daily ritual. That's the whole game with consumer AI right now — not can it do the thing, but can it become the thing you reach for without thinking.
Alright. I promised you something unexpected to close, and I'm a man of my word. Let me tell you about Skeleton 150.
This is Kiona Smith's piece, and I want to flag up front — this is a recent write-up of research presented last week, but the events themselves are from the year 1304, so we're safely out of "breaking news" territory by about seven hundred and twenty years. During renovation work at Scotland's Stirling Castle back in 1997, they found nine bodies buried under the medieval chapel. Five of them clearly died violent deaths. But one, catalogued as Skeleton 150 — the researcher's notes just call him "trebuchet guy" — was in a category of his own. His skull was broken in sixty-one separate places. Another sixty bits of shattered bone across his ribs. Shoulder broken, leg above the knee basically pulverized. One single catastrophic impact.
University of Bradford paleopathologist Jo Buckberry and her team threw everything at it — X-rays, micro-CT scans, forensics — and the pattern kept pointing to one thing: something heavy and very, very fast. Buckberry said the closest comparable cases she could find in the modern record were car crashes and people hit by trains. And she said, quote, "What is large and moving very quickly in 1304?" The answer: a big rock flung by a trebuchet. She believes this is the only known trebuchet casualty in the entire archaeological record — the first physical evidence of trebuchet trauma ever documented.
Here's the historical kicker, and it's the part I can't stop thinking about. In 1304, King Edward I of England was besieging Stirling Castle, held by a small Scottish garrison. Thirteen siege engines pounded the walls around the clock for three months. And Edward was building the biggest trebuchet the world had ever seen — he named it the War Wolf. The Scottish commander, seeing this monster get assembled, sent word: actually, we surrender, let's call it a day. And Edward said no. He wanted to use his new toy. He made them wait so he could fire the War Wolf through the castle wall — witnesses said it was like watching an arrow pierce a paper target — purely as a flex. Skeleton 150 was probably already dead and buried by then, killed by one of the thirteen ordinary engines during those three grinding months. Just a guy in the wrong spot when a ninety-kilogram rock came down at somewhere near three hundred kilometers an hour.
And I'll tell you why a medieval skeleton belongs in a show for founders and builders. Because it's the oldest version of a lesson we keep relearning. The most powerful weapon on the field, the War Wolf, the thing everybody remembers, the thing with the terrifying name — it fired exactly one shot, on the last day, and probably didn't kill anybody who mattered. The thing that actually did the work was thirteen unglamorous machines grinding away for three months, one boring rock at a time. The headline weapon got the legend. The workhorses got the results. Next time somebody's selling you the War Wolf, ask what's actually doing the damage.
That's the menu for today. Surveillance networks that eat themselves, bookstores shredding books, and a Scottish defender who's been waiting seven centuries for someone to figure out what hit him.
I'm Tony DeLuca, this has been Barely Possible, and I appreciate you spending a little of your one and only life in here with me. Watch your trust, watch your compute bill, and I'll see you next time.