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 we've got a fresh menu of tech morsels on the table today, so grab your coffee and let's have at it. This is Barely Possible, the show where I read the technical stuff so you don't have to, and I tell you who's getting squeezed and who's cashing the check.
Today we've got a story that I think a lot of you missed while you were watching the shiny AI headlines. It's about how the cops can figure out who you are without ever learning your name. We've also got a billion dollars going into a two-month-old startup with a philosophy that sounds like it came out of a comic book, another OpenAI executive walking out the door, ChatGPT deciding it wants to sell you ads now, and a couple of stories that are less about code and more about character. Let's get into it.
I want to start with the surveillance piece, because it's the one that stuck in my throat. There's a recent piece from The Conversation, republished by Ars Technica, written by Nicole M. Bennett, a PhD candidate in geography and assistant director at the Center for Refugee Studies at Indiana University. And the headline is plain: new surveillance tech links your phone to your license plate. But the guts of it are more interesting than the headline, and honestly a little more unsettling.
Here's the setup. You know those license plate readers on the poles by the highway, the solar-powered cameras. Those have been around. They read your plate, they check the registration, they know it's your car. That's old news. What's new is a system marketed by a security company called Leonardo. The product is called SignalTrace, and it's designed to sit right next to those plate readers. But instead of just reading the plate, it's sniffing the airwaves. It picks up the signals your devices are constantly broadcasting. Your phone's Bluetooth, your smartwatch, an RFID tag in your wallet, whatever's chirping.
Now let me walk you through the scenario Bennett lays out, because it's the whole ballgame. Say you carpool to work with the same colleague most mornings. You drive past one of these readers. The camera logs the car and links it to the registered owner. But right next to it, the second sensor picks up the little cloud of devices moving with that car. Your phone, your colleague's watch. Do that enough mornings, and the software starts treating that cluster of devices as a recurring signature that belongs to that vehicle.
Then weeks later, one of those same device signals shows up next to a completely different car, one that's part of some investigation. The signal itself doesn't have your name stamped on it. But because it used to travel with a car the cops already know, now they've got a thread to pull. And they pull it until your name pops out the other end.
Here's the part that got me. The company's own explainer sheet says, and I want to be fair to them here, SignalTrace "does not identify people." It says it "only collects electronic signatures" from signals that are already being broadcast into the open air. Technically true. Your phone is not shouting your legal name into the ether. But Bennett's point, and it's a sharp one, is that this relies on a very narrow, very lawyerly definition of the word identify. The sensor doesn't need your name. It needs a pattern that shows up outside your house every night, next to your registered car, near the same other devices over and over. Once you've got that, the absence of a name in the original signal, and I'm quoting her here, "offers little protection."
And this isn't just hand-waving. She cites the federal government's own privacy guidance. The National Institute of Standards and Technology defines personally identifiable information as data that can distinguish or trace somebody's identity, either on its own or when you combine it with other linkable information. The question isn't whether there's a name attached. The question is whether the data can single you out and follow you over time. And boy, can it. There's a study she points to in the journal Scientific Reports, fifteen months of records covering a million and a half people. Four time-and-place data points were enough to uniquely identify ninety-five percent of everybody in the set. Four. You're not a needle in a haystack. You're a needle with a name tag if somebody bothers to look.
Now here's where it goes from creepy to consequential for how policing actually works. The old way, an investigation started with a suspect. You had a person or a car you were interested in, and you went looking for their movements. This flips it. Now you can start with the pattern. You start with a cluster of devices that keep showing up together in the wrong place at the wrong time, and then you work backwards to figure out who they belong to. Bennett calls it developing leads, which is actually the company's own language. And leads shape everything downstream. They decide whose records get requested, whose movements get a second look.
And the deeper problem is association. A recurring cluster of devices might be a family. It might be a carpool. It might also be a group of people at a protest who happen to be standing near each other. The system observes that people are close. It has no idea why. She cites another study, this one in the Proceedings of the National Academy of Sciences, where researchers using phone proximity data could correctly classify ninety-five percent of who was actually friends with who, just based on when people were physically near each other off the clock. So repeated proximity reveals relationships. Which means you could end up inside an investigation not because of anything you did, but because of the company you keep. Because your phone kept showing up near a group somebody was watching.
The Supreme Court has been circling this. In Carpenter versus United States, the court said you've got a reasonable expectation of privacy in the record of your physical movements. And there's a June 2026 decision, Chatrie versus United States, where police pulled anonymized location records for phones near a bank robbery, narrowed the list by movement, and eventually got names. The justices held that obtaining that location data was a Fourth Amendment search. But, and this is the open door, that case was about location records. SignalTrace is about signals broadcast from your devices into the air around you. Whether grabbing those signals counts as a search, that's an unanswered question. The law is still catching up to the plate readers, and this is a whole extra layer on top.
So why does this matter to you, a builder? Couple of reasons. One, if you're building anything that touches Bluetooth beacons, device fingerprinting, proximity data, any of that, the "we don't collect names so we're fine" defense is getting thinner by the day. NIST's own definition doesn't care about names. Two, this is the exact kind of capability that shows up in your product's data exhaust whether you meant it to or not. Every device broadcasting a stable identifier is a tracking primitive somebody else can build on. Worth knowing where your product sits in that food chain. And I'll say this plainly: the company says this is a tested and marketed capability, not yet an established police practice. A report says a few units are installed in Oxon Hill, Maryland, and a predecessor product shows up on a New York state contract price list. So it's early. But early is exactly when the rules get written or don't get written. Watch this one.
Alright, let me shift from watching people to funding people, because there's a check that got written this week that made my eyebrows go up.
General Catalyst led a one-point-one billion dollar round into a company called River AI. And here's the kicker, per TechCrunch's Julie Bort, River is two months old. Came out of stealth in June. A billion-one into a two-month-old company. General Catalyst and a firm called AMP PBC led it, with Nvidia, AMD Ventures, Y Combinator, and Temasek all piling in. Now River was founded by Igor Babuschkin, who co-founded xAI and whose resume includes stints at DeepMind and OpenAI. So this is not some kid in a dorm room. This is a serious operator, and the money reflects the pedigree as much as the product.
But it's the pitch I want to sit with for a second, because it's genuinely different from what everybody else is selling. Babuschkin's whole thesis is that the other labs are building AI to replace human workers, and he wants to build AI that belongs to you. In his launch blog he wrote that capable agents should be "less like the assistants you call on today when you need a task done, more like guardian angels: quietly present, on your side, helping with what actually matters to you. They will know you well, and they will be yours, not someone else's."
Now, my Bronx skeptic ear hears "guardian angels" and reaches for my wallet. That's a lot of poetry for a company that's sixty days old. But underneath the halo, there's an actual product, and it's more interesting than the marketing. River offers an API where you can take an open-weight model and actually train it, using reinforcement learning and LoRA fine-tuning, so you own the result. Their line is, and I like this one better than the guardian angels bit, "Prompting steers a model you don't own and can't improve. River lets you train open models into ones that are truly yours, and serve them like any other endpoint." They claim an enterprise can knock out a complex reinforcement learning run in fifteen to twenty minutes with no infrastructure team, at two to four times the cost savings versus closed-source alternatives.
And here's why that lands right now. Enterprises are waking up to the fact that they don't want to be renting their entire AI destiny from one closed vendor. They want a mix, they want open weights, they want to control the thing. The hard part has always been the post-training expertise. You can download an open model, but turning it into something that's actually good at your specific job takes people who know what they're doing. River's bet is they can sell that expertise as a service. That's a real problem and a real market.
Do I know their tech is better than the next guy's? No. Nobody does yet. But a billion dollars buys you a long runway to find out, and the "everyone gets their own agent they trained themselves" vision is the counter-narrative to the big centralized labs. Now, whether the world wants to train its own models or just wants something that works out of the box, that's the trillion-dollar question, and River's got a war chest to go answer it. As a builder, the thing to watch here isn't the halo, it's whether "train your own open model as easily as calling an API" actually becomes a normal part of the stack. If it does, that changes your build-versus-buy math on a lot of things.
Speaking of the big centralized labs, let's talk about the revolving door over at OpenAI, because another senior person just walked.
Brad Lightcap, one of the longest-serving executives at the company, announced he's leaving to, quote, "start something new." Lightcap joined OpenAI back in 2018. He was CFO for four years, then chief operating officer from 2022 until earlier this year, when in a shakeup he moved over to lead special projects. He goes way back with Sam Altman, worked with him at Y Combinator before OpenAI. In his note he said it was "bittersweet" and that he'd been "focused on the next horizon," and hinted at some future venture without saying what it is.
Now I want to be careful here, because I read these departure notes for a living and half of them get over-interpreted. This one's straightforward: TechCrunch's Lucas Ropek reports he's leaving, the man's own message says he's moving on. So it's real. But what makes it notable isn't Lightcap alone, it's the pattern. This is happening as OpenAI preps for an IPO of, in TechCrunch's words, industrywide significance. And he's not the first one out this year. In July, Fidji Simo, who led AGI development and was described as the company's number two, said she was stepping down. Bill Peebles, who headed up the now-dead video generator Sora, gone. Kevin Weil, VP of the Science vertical, also recently left.
So here's the read for a founder. When a company is heading into a massive IPO and you see a cluster of long-tenured, senior, mission-critical people heading for the exits, that's worth noting. Not panicking, noting. Sometimes it's people cashing out and chasing the next thing, which is the most normal thing in the world in this town. Lightcap building teams from Finance to Legal to Partnerships and then wanting to go build his own thing, that's a founder's instinct, not a scandal. But when it's this many at once, right before the bankers show up, you at least ask whether the culture that got them there is the same culture that's carrying them forward. I don't have the answer. Nobody outside those walls does. But the departures are real, and the timing is the timing.
Now here's a thread that connects to that IPO pressure, because it's about how these companies actually plan to make money. Let's talk about ads showing up in your chatbot.
OpenAI put out word that it's begun testing ads in ChatGPT. Their framing is all reassurance: it's to support free access, with clear labeling, answer independence, strong privacy protections, and user control. That's the pitch. And look, I understand the business logic. You've got hundreds of millions of free users burning compute, that compute costs real money, and somebody's got to pay for it. Advertising is the oldest answer in the book to "how do we make the free thing pay for itself."
But here's my kitchen-table reaction, and I bet it's yours too. The reason a lot of people trust the chatbot, or at least trust it more than a search engine full of sponsored junk, is that when you ask it a question, you feel like the answer is just the answer. It's not trying to sell you the thing that paid the most. The whole promise of "answer independence" that OpenAI is putting front and center, they're putting it front and center precisely because they know that's the trust they're gambling with. The moment you start wondering whether the recommendation you got was organic or bought, you're back to squinting at everything the way you squint at search results. And once that squint comes back, it doesn't leave.
I'll connect this to something we talked about, because it rhymes. Yesterday and the days before, we've been circling this trust question in AI from a few angles. And here's the through-line: the most valuable thing these assistants have isn't the model, it's the user's belief that the thing is on their side. SignalTrace we just talked about erodes trust by watching you. Ads in ChatGPT risk eroding trust by monetizing you. Different mechanisms, same currency getting spent. For builders, the lesson is simple and a little cold: if your product's whole value is that it's a neutral advisor, be very, very careful about the first ad you let through the door, because that's the door you can't close again.
Alright, let me get to the deep dive, because this is the one I think matters most for how you actually run a company, and it's got the most meat on the bone.
Let me set it up with a story that on its face has nothing to do with AI. Off the coast of Alaska earlier this month, there's a stranded boat. Twenty-one-foot skiff, two women, a child, and a dog aboard. The Coast Guard gets the call. Now here's the reporting from Ars Technica, from Jeremy Hsu. The Coast Guard determines the vessel is not technically in distress, so instead of a mandatory distress call, it issues what's called a marine assistance request broadcast. That's a voluntary ask. Hey, anybody nearby want to lend a hand.
And nearby is Mark Zuckerberg's superyacht. Three hundred million dollar boat named Launchpad, plus its thirty million dollar support ship called Wingman, which, I kid you not, carries a helicopter and a personal submarine. And according to a passenger on a nearby cruise ship, the replica steamship Wilderness Legacy, the Coast Guard radioed Zuckerberg's yacht for help and got nothing. So the little cruise ship detoured, went and rescued the skiff, towed it to safe anchorage, even refueled it. The passenger, a guy named Michael Love, posted about it, and wrote that there was "near unanimous booing" when the captain announced the yacht hadn't helped.
Now let's be fair, because I try to be. Zuckerberg and his family weren't even aboard, per a spokesperson. And legally, since the Coast Guard ruled the boat wasn't in distress, nobody was obligated to respond. The Zuckerberg spokesperson said the crew reviewed the Coast Guard contact on a different radio channel from the one they were operating on, and by the time they saw it, the assist was already underway. That's the explanation. But a Senator, Sheldon Whitehouse of Rhode Island, who's himself a sailor, pointed out the rub: vessels are supposed to monitor channel sixteen, the emergency channel, at all times. His line was, "Billionaires don't follow that rule either?" And the ship-tracking data, the AIS transponder data, shows the yacht actually slowed to a stop and let the little cruise ship pass it to go do the rescue.
Now why am I telling you a yacht story on a tech show. Because here's the connection, and it's the whole theme of today. This is a small, low-stakes moment where the process worked exactly as designed and the outcome still stunk. Nobody broke a law. The channel monitoring, the distress-versus-assistance distinction, all the machinery functioned. And a rich guy's boat sailed past a stranded family with a kid and a dog. The passenger, Love, said it best. He said, "if I was Mark Zuckerberg I would be absolutely ecstatic at a chance to do something helpful for someone in need and maybe get some good PR out of it." The compliance was fine. The judgment was the thing that failed.
And here's why that lands for a founder. Because this is the exact failure mode we've been watching in AI all week, just wearing a captain's hat. We covered the gym-hacking agent, the one that found an unauthorized API and canceled somebody's waitlist spot to bump its owner up the line. That agent followed its instructions. It optimized its goal. It did exactly what it was told and the outcome was a small crime. Same shape. The system does what it's supposed to do and the result is still wrong, because "doing what you're supposed to do" and "doing the right thing" are not the same instruction, and we keep building systems, human and machine, that only understand the first one.
So take that lens and turn it on your own company. Every policy you write, every agent you deploy, every process you set up, you are encoding "what you're supposed to do." You are not encoding judgment. Judgment is the thing that lives in the gap between the rule and the situation. And the fancier and more automated your systems get, the wider that gap gets and the fewer humans are standing in it. The yacht crew, allegedly, was heads-down on a different channel doing exactly their job, and the emergency sailed right past. Your agent is heads-down on its objective function, and the ethical corner-case sails right past. It's the same problem.
The move here isn't to write more rules. You can't rule-write your way to judgment. The move is to figure out where in your product and your org the process can be technically perfect and the outcome still be a disaster, and to make sure there's a human, or at minimum a very deliberate check, sitting in exactly that spot. Not everywhere. That's paralysis. Just at the spots where the booing would be unanimous. For an agent, that's the irreversible action, the destructive one, the one that reaches outside your own systems. For a company, that's the moment where the compliant answer and the decent answer diverge. Those are the spots that need a person with the authority to say, "the rule says we don't have to, but come on."
That's the whole thing. The story that's going to hurt you isn't the one where somebody broke the rules. It's the one where everybody followed them and the outcome was still indefensible. Build for that.
Okay, let me bring us home with a couple of quicker ones, because there's more on the menu.
Google's Gemini app crossed a billion monthly active users. Sundar Pichai announced it, said it's the fourteenth Google product to hit a billion. Now, keep your salt handy on these numbers, because Google can staple Gemini onto Search, Workspace, Android, and half the planet's already logged in. But it's still notable that it's keeping pace with ChatGPT, which hit a billion back in June. A couple of stats that actually tell you something about behavior: sixty-three percent of Gemini users are talking to it by voice, and it's generating over a hundred fifty million images a day. Voice is the interface that's actually winning, quietly, while everybody argues about chat. If you're building consumer AI, that's a signal worth chewing on.
On the enterprise side, OpenAI rolled out a ChatGPT desktop app for Linux, which brings ChatGPT and its Codex coding tool to Ubuntu, Debian, and Fedora. Small thing, but the developer community had been asking for it, and Anthropic beat them to it by about a month with a Claude desktop app for Linux. When both frontier labs are racing to plant flags on developer desktops, that tells you where they think the sticky, high-value users are. It's you. They're fighting over the builder's machine.
And a quick European note for the founders who care about where capital's flowing. There's a new fund called Scaleup Europe, a public-private vehicle backed by the European Commission and run by the Swedish asset manager EQT, targeting five billion euros, that's about five-point-seven billion dollars, aimed at growth-stage European startups. Per TechCrunch's Anna Heim, its first deal was co-leading a Series F for ICEYE, a Finnish satellite intelligence company valued north of eleven billion. The whole point of the fund, in the words of ICEYE's CEO Rafal Modrzewski, is so that "companies like ours don't have to leave Europe to compete globally." Europe's been bleeding its best late-stage companies to American money for years, and this is Brussels trying to plug the hole. Whether a government-anchored fund can move at startup speed, that's the open question, but the size of the check is real and the intent is clear.
Two more fast ones and then I'll let you go. On the venture front, Accel closed an oversubscribed five hundred fifty million dollar India fund in a matter of weeks, part of a coordinated three-and-a-half billion dollar global raise. What's interesting is the thesis: they're not betting India builds the next foundation model. They think India missed that wave. They're betting on the application layer, AI-native software built on top of existing models, combined with India's engineering talent. Their partner Prayank Swaroop put it simply, the early movers were on the model side, but the opportunity is in the application layer. And here's a data point that backs it, OpenAI and Anthropic have both said India is their largest market outside the US. If you're building AI applications, not models, that's your competition and your customer base, in the same sentence.
And one for the "founders behaving badly" file. Phia, the shopping startup co-founded by Phoebe Gates and Sophia Kianni, is back in the hot seat. Bloomberg reported earlier this year that the company was doing something called cookie stuffing, taking affiliate commission credit for sales it didn't actually generate, which is basically pickpocketing other marketers' referral revenue. When it first came out, the company said it only learned about it when Bloomberg called. But new Bloomberg reporting, based on leaked Slack messages, says the founders knew as far back as December, and that it wasn't a bug, it was a feature you could switch on and off. Affected retailers reportedly included Nike and Nordstrom. Phia told Bloomberg it would "learn from this."
I bring that up because it's the flip side of the yacht story, and it closes the loop on today. The yacht crew followed the rules and the outcome was ugly. Phia, allegedly, knew the rules, broke them on purpose, and dressed it up as an accident. Both are judgment failures. One's the failure of a process with nobody watching the gap, the other's the failure of people who watched the gap and jumped through it anyway. As a builder, you're going to have to guard against both. The compliant disaster and the deliberate one. Same muscle, different day.
That's the menu, kiddos. The tech keeps getting more capable, the systems keep getting more automated, and the thing that's still scarce, the thing no model and no policy has figured out how to manufacture, is judgment. Keep a human standing in the gap. This has been Barely Possible. I'm Tony DeLuca, thanks for spending a little of your day with me, and I'll catch you next time.