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 the menu today is a little different than usual. We've got a surveillance guy who can't keep his story straight for two interviews in a row, a database startup out of Munich that might tell you something real about where enterprise software is heading, a cyberattack that cooked the groceries in seven hundred French supermarkets, and a batch of AI announcements that all say the same three words over and over again. Buckle up, let's have at it.
Let me start where I keep ending up lately, because the universe keeps handing me this story. If you've been riding along with me this month, you know Flock Safety has been a recurring character on this show. Back at the start of October we talked about those unexplained, unpermitted license-plate cameras showing up in St. Lucie County, Florida. A few days after that, we covered Judge Hill ruling that a warrantless Flock search violated the Fourth Amendment. This company is a flashpoint, and it's not going away. So when Flock's CEO shows up on stage at a big tech conference this week, I pay attention. And what I heard was one of the more revealing performances I've seen from a founder under pressure.
Here's the setup. Garrett Langley, the CEO of Flock, sat down for an interview with TechCrunch's editor-in-chief Connie Loizos. Flock, for anybody just tuning in, makes license plate reader cameras. They've got more than a hundred thousand of them across the United States, and they let local police departments and federal agencies track where vehicles go. And because departments can look up plates outside their own jurisdiction, some cops can effectively track a car across the whole country through Flock's central database. That's the product. That's the whole thing.
Now, the reason Langley was on defense is that there's been stream after stream of reporting about police officers abusing this system. Dozens of officers arrested, fired, or told to resign after getting caught misusing access to that nationwide database, oftentimes to stalk people without permission. So Loizos asks the obvious question. Have you ever fired a customer, meaning a police department, over this abuse? And Langley says, no. Not once. He says the company had the debate internally, but hasn't pulled the product from any agency. His exact framing was, quote, if a customer knowingly broke the law, we would fire them immediately. That's very easy. End quote. And then he pivots to, quote, most of the debate is nuanced into what if two officers out of fifty were caught. End quote.
Now listen. I'm a reasonable guy. I understand the two-out-of-fifty logic. You don't necessarily punish a whole precinct for a couple of bad apples. But sit with the structure of what he just said. He said we'd fire a customer immediately if they knowingly broke the law. And simultaneously, he's confirming that dozens of individual officers have been arrested or fired for breaking the law using his product, and no department has lost access. So which is it? The law got broken. Over and over. People got stalked. The answer he's giving is that it's complicated. And maybe it is complicated. But "we'd fire them immediately for breaking the law" and "we've never fired anyone despite all the law-breaking" cannot both be your clean talking point. One of them is doing a lot of heavy lifting.
And while he's up there saying this, there are protesters outside the venue. People dressed in the red robes from The Handmaid's Tale, cardboard Flock cameras on their heads, clips from the movie 1984 playing on a loudspeaker. Signs saying "Flock off" and "Flock equals ICE." Now, for the record, Flock doesn't have a direct contract with Immigration and Customs Enforcement. But the agency has reportedly gotten access to Flock data in the past by working with local police departments that use the cameras. So the "Flock equals ICE" sign isn't pulled out of thin air. It's a reasonable read of how the data flows.
What I found most interesting was a quote from one of the organizers, a woman named Ashley Ortiz from a group called The Wolves. She said something that I think founders in the surveillance space should tattoo on the inside of their eyelids. She said, quote, I think Flock is going to be the start of a lot more community building, because we can come together around something as bipartisan as not enjoying mass surveillance. We have Republicans who join us, conservatives who agree, we have sovereign citizens cutting down Flock cameras, so it's not just the left who doesn't want AI-powered mass surveillance. End quote.
That's the part the industry keeps missing. Opposition to mass surveillance is one of the last genuinely bipartisan positions in American life. The civil libertarian left and the leave-me-alone right land in exactly the same place here. And when your product manages to unite those two groups against you, that's not a messaging problem you solve with a better PR agency. That's a product problem.
Now here's where it gets worse, and this is the second half of the Flock story, because the same CEO did a second thing in the same appearance that I want you to hear, because it's a masterclass in how not to talk about your own product.
Flock has an audio product. It's called Gunshot Detection. The pitch is these microphones listen for gunshots across neighborhoods to give police earlier awareness of a shooting. Fine. Reasonable use case on its face. Langley said on stage that the microphones detect, quote, just gunshots. End quote. When the interviewer pushed him on it, he said it again. Just gunshots.
The problem is Flock's own website says something different. The website says the audio systems can pinpoint gunfire, sure, but also, quote, detect street takeovers early to stop chaos before it escalates, and identify severe crashes in real time. It says they can expand awareness beyond gunfire with alerts for car crashes, fireworks, and sideshows. That's the company's own marketing copy. So when the reporter asks Langley directly about the contradiction between "just gunshots" and his website listing crashes and fireworks and sideshows, his answer was, and I'm quoting the man exactly, "It's just gunshots."
And then, predictably, a Flock spokesperson had to clean it up afterward. The spokesperson, Paris Lewbel, told TechCrunch that the technology is, quote, primarily used to detect gunfire but is also capable of detecting fireworks, sideshows, and potential car crashes. End quote. So the official company statement directly contradicts what the CEO said on stage twice. Primarily gunfire, but also these other four things. That is not "just gunshots." That's just not what those words mean.
And there's history here that makes it uglier. When Flock first announced this gunshot feature last year, they said the tech would listen for the sounds of screaming. Privacy groups came down on them hard for listening for human voices. So Flock amended the website to say "sounds of human distress" instead. Then they said they pulled the distress detection altogether from the customers who were testing it. So the arc of this product is: we listen for screaming, no wait, human distress, no wait, we pulled that, no wait, it's just gunshots. And the FAQ on the site still says the detection library is, quote, continually expanding to include emerging threats. End quote.
Here's why I'm spending real time on this, because I know some of you are builders and you might think, Tony, this is a surveillance company, not my problem. It's your problem in this sense. When you build a product that collects ambient data about people who never consented, you are going to end up in exactly this spot. On a stage. Being asked a direct question. And every time your public statement doesn't match your own documentation, you lose the one thing that actually matters in a trust business, which is trust. Langley could have said, look, the microphones are tuned primarily for gunfire, and yes, the system can flag a few other loud events, and here's exactly where the line is. That would have been true and defensible. Instead he said "just gunshots" to a journalist who had his website open. That's not a surveillance problem. That's a candor problem. And candor problems compound.
One more data point to close the loop on Flock, because it matters for anybody watching this company. Texas and Florida have both moved to end their relationships with Flock. Amazon's Ring has ended its partnership with them. Some communities have swapped Flock out for rivals like Axon, the Taser and bodycam company. When Langley was asked if Flock would sell to Axon, he said no, no desire to sell. Fine. But the trend line here is customers leaving, states backing away, and a CEO who can't give a straight answer about what his microphones hear. That's a company under real pressure, no matter how many cameras are in the ground.
Alright, let me shift from a company burning trust to a company that's quietly earned some. Let's talk about databases. And I know, I know, I hear you groaning in the car. Tony, databases? But stay with me, because this one actually tells you something about where enterprise software is going, and it's our deep dive today.
There's a Munich-based outfit called CedarDB. This is a recent piece from TechCrunch's Anna Heim. Now, the core idea is wonky but the implication is not. For basically the entire history of business computing, databases have been built for one of two jobs. You've got your analytical databases, the ones that help you spot trends and answer big questions about your business. And you've got your operational databases, the ones that store the boring day-to-day stuff, the orders, the payments, the transactions. Two different tools for two different jobs. That split has been gospel for decades.
CedarDB's whole pitch is that the split doesn't need to exist. One engine does both. And here's the line from the piece that made me sit up, from cofounder and CEO Moritz Sichert. He says the approach has, quote, gained relevance with the rise of agentic AI. End quote. Now, normally when I hear a founder bolt the words "agentic AI" onto a decade-old database project, my skeptic light starts blinking. But let me walk you through why this one actually holds together, because the reasoning is the interesting part.
The company is a spinout from the Technical University of Munich, built on a research project called Umbra. Five of the Umbra researchers are the cofounders. The original goal of Umbra was pure efficiency, to beat legacy database engines at raw processing speed by squeezing modern hardware harder. The promise is faster performance without you having to buy more CPUs or memory. And critically, they made it compatible with PostgreSQL, which is the popular open-source database a huge chunk of the world already runs on. Compatibility keeps your switching costs low, because you don't have to rebuild your apps to move over.
Now here's the clever bit, the part that connects to AI. Sichert points out that because they follow PostgreSQL's logic, they also follow the logic that large language models are most familiar with, because the models were trained on mountains of PostgreSQL. So when you've got an AI agent writing queries against your database in natural language, an agent that was trained on all that Postgres-flavored SQL, it's going to be more fluent against a Postgres-compatible engine. Sichert, to his credit, called that one luck in retrospect. They didn't build it for the AI agents. The AI agents showed up and happened to speak the dialect they'd already chosen.
But here's the real reason this story matters for you, the builder, and it's not the Munich startup specifically. It's what's happening around them. The piece notes that Databricks, one of the giants of the data world, reached the same conclusion. At their big summit in June, Databricks cited, quote, the agentic era, end quote, as the reason it was launching a new data architecture to unify transactions, analytics, streaming, and operational data on a single copy of storage. So you've got a fourteen-person startup out of a German university and a multibillion-dollar American data company both deciding, independently, that the old wall between transactions and analytics needs to come down, and both pointing at AI agents as the reason.
Why would agents change this? Think about what an agent actually does when it's turned loose on your business data. It's not a human analyst who runs a scheduled report once a week against your nice clean analytical warehouse. An agent is asking questions on the fly, in natural language, and it wants to ask about what's happening right now, not what happened last quarter after the data got shoveled from your operational system into your analytical one. The traditional setup has a lag built into it. You run transactions over here, then you pipe the data over there for analysis, and there's always a gap. An agent asking "what's the state of this customer's orders and what's the trend across similar customers" wants both the live transaction data and the analytical rollup in one shot, with no lag. If your architecture forces the agent to wait for data to migrate from one system to another, the agent is working from stale information, and the whole promise of real-time autonomy falls apart.
So the deeper point is this. We spent a lot of this month talking about agents at the model layer, at the harness layer. This is the data layer waking up to the same pressure. The agent doesn't care about your org chart, and it doesn't care that your company decided in 2014 to keep operational and analytical data in separate buildings. It just wants an answer, fast, from one place. And the companies building the plumbing underneath are starting to redesign for a consumer that never sleeps and never waits.
Now the honest counterweight, because I'm not here to sell you a database. The on-premise and free versions of CedarDB launched back in May of last year with a roughly six-million-dollar seed round. The cloud version just hit limited availability. It's early. Fourteen people. The founder himself says they'll need a bigger team and another funding round once they get traction. And "one database to rule them all" is a pitch the industry has heard many times before and been burned by, because the reason the two-database split existed is that the two jobs genuinely have different performance profiles. Maybe the hardware has finally gotten good enough to erase that tradeoff. Maybe not. But the fact that both the scrappy startup and the incumbent are moving the same direction, for the same stated reason, tells me the demand signal is real even if the winner isn't decided. If you're building anything agentic on top of enterprise data, this is the layer to watch. Not because of who wins, but because the shape of the question is changing underneath you.
Let me connect that to the AI announcements that came across the desk this week, because they're all circling the same word, and it's the same word CedarDB and Databricks are using. Agentic.
OpenAI put out a cluster of posts. There's a piece on an outfit called Third Moment Research running ten thousand agents for thirty-six-plus hours. There's SentinelOne using OpenAI's security tooling to cut malware analysis from weeks down to hours. There's a developer tool called Amp pushing agents beyond just writing code. And there's a broader OpenAI analysis titled, roughly, the EU's next chapter in AI is agentic, arguing agentic adoption is growing across European enterprises. Now I want to be straight with you about these. These are company announcement posts. They're marketing. The detailed bodies didn't fully come through on my end, so I'm not going to pretend I can give you an independent audit of whether SentinelOne really went from weeks to hours. What I can tell you is the theme. Every frontier lab, every enterprise case study, every database startup, is now organizing their entire pitch around agents that run for a long time, with some autonomy, and improve with use.
The one concrete number that's worth holding onto is that malware-analysis claim from SentinelOne, weeks to hours, because security is one of the places where the agent story is most plausible. Malware analysis is tedious, pattern-heavy, high-volume work, exactly the kind of thing where a tireless agent that can grind through samples creates real capacity. If it holds up, it means the security analysts get to spend their time on the hard, weird threats instead of the grunt triage. That's a believable win. I'm just flagging that it's a vendor telling you about a vendor, so hold it loosely until the independent numbers show up.
Now, the counterweight to all this agentic optimism came, interestingly, from Anthropic, in a recent piece on how they prepared for the 2026 US midterms. And buried in there is a finding that every agent enthusiast should chew on. Anthropic analyzed a sample of a quarter million conversations and tested their election safeguards. They found that Claude reliably refused a single harmful request. Ask it to do one obviously bad thing, it says no. But, quote, could miss a malicious campaign when it was spread across many individual requests that looked harmless on their own. End quote.
Sit with that, because it's the whole agent safety problem in one sentence. The model is good at catching the one bad step. It's much worse at catching the bad plan that's been chopped up into a hundred innocent-looking steps. An agent that runs for thirty-six hours across ten thousand tasks is, by definition, a system taking many small steps where no single step looks alarming. That's the exact blind spot. The same architecture that makes long-running agents powerful is the architecture that makes them hard to supervise, because harm doesn't live in any one action anymore. It lives in the shape of the whole run. Anthropic says they've added new classifiers to catch multi-step misuse. Good. But that's a cat-and-mouse game, and right now the mouse has the structural advantage.
A couple of the other numbers from that Anthropic piece are worth a quick mention because they cut against a lot of the panic you hear. People, it turns out, rarely ask the chatbot who to vote for. Requests asking Claude whom or what to vote for showed up in about one in ten thousand conversations. And requests that actually violated their election integrity policies were rarer still, about two in a hundred thousand. So the dominant use isn't people outsourcing their vote to a robot. It's people asking where a candidate stands on an issue so they can make up their own minds. That's a more boring and more reassuring picture than the doom headlines. The real threat, per Anthropic's own findings, isn't the confused voter. It's the coordinated operator slicing a propaganda campaign into harmless-looking pieces. One of their cases was foreign operators running a fake news site using Claude to produce midterm coverage pushing Russian, Iranian, and North Korean narratives. That's the distributed-harm problem again, just wearing a different coat.
Let me stay in the enterprise-AI lane for one more, because Anthropic also made a move this week that's more about the map than the model. They opened an office in Singapore, their fifth in the Asia-Pacific region, and signed a memorandum of understanding with the Singapore government. And the numbers in that announcement are the actual story. In Anthropic's own Economic Index, Singapore ranks second out of a hundred and twenty-one countries in Claude usage relative to population. Nearly six times what the country's size alone would predict. Six times. Think about that. A city-state is punching six times above its weight in AI adoption.
And the customer list tells you why they bothered flying out there. GovTech Singapore has more than fifteen hundred engineers using Claude Code to modernize government systems. DBS, the biggest bank in Southeast Asia. GIC, the sovereign wealth fund, rolling it across investment teams. Grab using it to power a sales assistant that the company says saves its teams about forty thousand hours a quarter. For a founder, the lesson isn't about Anthropic. It's about where the adoption is actually dense. The conventional wisdom puts the AI center of gravity in San Francisco. But the usage data says a disciplined, regulation-friendly, government-coordinated market like Singapore is absorbing this stuff faster per capita than almost anywhere. If you're thinking about where to expand, the heat map doesn't always match the hype map.
Alright, let me get out of AI for a bit, because today's menu has some genuinely different flavors, and I promised you groceries. Let's talk about the French supermarket fridge hack, which is one of those stories that's funny until you realize it's a preview of coming attractions.
Now, one note on timing. This is a recent piece from Ars Technica's Dan Goodin walking through an incident, and the underlying warnings go back several months. The actual hack hit earlier this week, as the reporting frames it. So here's what happened. An estimated seven hundred supermarkets across France had their refrigeration systems hacked. Attackers got into the internet-connected controllers that let technicians remotely manage the fridges and freezers, and they cranked the defrosters until the units hit fifty degrees Celsius. That's a hundred and twenty-two Fahrenheit. In a refrigerator. Tons of food spoiled. And in some cases the electrical systems caught fire. This wasn't just France either. Spain, Portugal, Poland, Germany, and Belgium got hit too. The damage is estimated in the tens or hundreds of millions of euros.
And here's the kicker, the part that should make every person who builds connected devices wince. This was entirely preventable with basic hygiene. The controllers came from a few manufacturers, and they had critical known vulnerabilities. One of them, from a company called Danfoss, used a daily password. Sounds secure, right, changes every day. Except the password was generated by the publicly available firmware based on the date. So a hacker who reads the code can generate that day's password for any device that hasn't been updated. And the same daily password works across every one of those devices in hundreds or thousands of stores. That's not a lock. That's a lock where the key is printed on the calendar.
On top of that, the stores mostly didn't bother putting these controllers behind a VPN or restricting which addresses could log in. The researchers at Pen Test Partners spelled out the fix, and it's embarrassingly basic. Take the controller off the public internet. Put remote access behind a VPN and an allowlist. Change the default credentials. Patch to a current release. That's it. That's the whole defense, and nobody did it. Security firm Claroty warned about this last October. The US cybersecurity agency warned about it in February. The warnings were sitting right there.
The manufacturers, for their part, are doing the thing companies do. One of them, Copeland, acknowledged a vulnerability in what they called a discontinued, end-of-life generation of controllers. Which, sure, but those discontinued controllers are still sitting in seven hundred stores keeping the yogurt cold.
Here's the line from the piece that I want builders to really hear, because it's the whole lesson. Goodin writes that the incident underscores the tension between the finance department and the security department in most businesses. Companies hate spending real money on new equipment when there's no visible benefit. And then, quote, only after a devastating hack do they realize why the investment was worthwhile. End quote. That's the entire story of operational technology security in one sentence. The fridge works. The fridge has worked for years. Why would I spend money hardening the fridge? And then one Tuesday the fridge catches fire in seven hundred stores at once.
And I want to point at the scary possibility the piece raises, because it's not just vandalism. Nobody's claimed responsibility. That silence could mean this was mass vandalism for its own sake. Or it could mean a nation-state or some other serious actor testing whether they can cause continent-wide disruption through the most boring infrastructure imaginable. Not the power grid, not the banks. The grocery store fridges. Because that's the stuff nobody armored. The lesson for anyone building connected hardware is permanent: the attack surface isn't your flagship product. It's the dumb, cheap, forgotten controller in the back room running firmware from four years ago. That's where they get in.
Let me bring it down to earth with some shorter ones, because there's good stuff that doesn't need a full sermon.
Jio, the digital arm of Mukesh Ambani's Reliance Industries in India, is going public. They're looking to raise as much as three point two billion dollars, which would make it India's biggest IPO ever, overtaking Hyundai Motor India's listing from a couple years back. At the top of the price range, the whole thing values Jio at around a hundred and eight billion dollars. For context, back in 2020, Jio pulled in more than twenty billion from global investors, including Google and Meta, when it was valued at around sixty-five billion. So the value's gone up substantially. Jio launched in 2016, drew people in with free data and cheap plans, and now has more than five hundred twenty-four million subscribers. It's grown way past being a telco into broadband, cloud, all of it. The subscription window opens October twenty-first. And the spice in this story is that it's happening while Elon Musk is publicly attacking Ambani, accusing him of trying to keep Starlink out of India. Indian authorities have rejected Musk's claims. So you've got the biggest IPO in the country's history playing out against a billionaire-versus-billionaire sideshow. The thing to actually watch is whether public-market investors have the appetite for a company that soaked up twenty billion in private money. This is the test of whether all that private cash can find a public exit at a good number.
Next one, and this is a palate cleanser. Colossal Biosciences, the de-extinction company, announced a new AI platform called GenoMorph. The pitch is they're trying to bring back the dodo bird, and the problem is they've only got one usable dodo genome to work with. Compare that to the woolly mammoth project, where they've got a hundred genomes, because cold climates preserve DNA and the tropics where the dodo lived do not. Normally scientists compare many genomes to figure out which genetic differences create which traits. With one genome, you're flying blind. So GenoMorph was trained on developmental data from living birds, chickens, quail, ducks, pigeons, plus genetic info and 3D skull measurements from more than seventy bird species, to predict which genetic switches gave the dodo its look. CEO Ben Lamm called it a treasure map for the dodo genome, and they say they've already identified around twenty-three hundred promising DNA editing targets. Now, I file this under fascinating-but-watch-the-claims. Identifying a target is not the same as resurrecting a bird. Lamm himself was careful to say identifying a potential edit doesn't mean they immediately put it in an animal. But the broader signal is real: AI as a hypothesis-generator for biology, narrowing a giant search space down to the handful of things worth testing in the lab. That's a legitimately useful pattern, de-extinction drama aside.
Couple of quick hitters to round it out. US electric vehicle sales are showing signs of life. After tanking when the tax credits got cancelled last year, they've grown three quarters in a row, hitting six point one percent of the market, per a Cox Automotive report. Toyota, Rivian, and Subaru led the way, and interestingly, Tesla's own sales dropped twenty percent year over year even though it still sells over half of all EVs. The pattern mirrors Germany, which pulled its EV incentives back in 2023, watched sales crater, and then recovered. So the subsidy cliff isn't permanent. The market finds a floor and climbs back.
And one more from the hardware world, worth it for builders thinking about consumer devices. There's a recent report on a startup called Hiyd, founded by a former Apple manufacturing design engineer named Salah Batroukh. The product is a tiny health-tracking disc, under three millimeters thick, that sticks to the back of the watch you already own using something called microsuction, no magnets, no adhesive. The whole thesis is invisible health tracking, don't make people sacrifice their nice watch for a clunky fitness band. They've had over sixteen thousand people put down deposits. The founder's line about why he left Apple stuck with me. He said the products became, quote, more about being an Apple product rather than being something for the customer. End quote. Whether or not Hiyd makes it, that's a real tension every big-company builder eventually feels, and it's why startups keep getting born.
So let me tie the bow on today. The thread I keep pulling on this episode is candor and consequences. Flock's CEO couldn't give a straight answer about his own microphones, and states are walking away. Seven hundred French supermarkets ignored years of straight-up security warnings, and the fridges caught fire. And on the hopeful side, you've got CedarDB and Databricks and the whole agentic wave honestly redrawing the architecture underneath enterprise software to match a new kind of user, while Anthropic quietly admits the hardest part of that new world is catching harm that's been sliced too thin to see. The tools are getting more autonomous. The question that decides who comes out ahead isn't capability. It's whether you can look at your own product, your own data, your own fridge in the back room, and tell the truth about what it actually does.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible, and whatever you're building out there, keep the key off the calendar. Catch you next time.