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 open for business. Grab your coffee, pull up a chair at the counter, because we've got a menu today that runs from the FTC poking at your grocery loyalty card, all the way out to a spent rocket punching a hole in the Moon. Buckle up, let's have at it.
Let me start where I think the real meat is for you if you're building a company, and that's a small, unglamorous story that ought to make everybody who's shopping for milk sit up. The Federal Trade Commission is going after what it's calling personalized pricing. This is a current piece from Ashley Belanger, and the headline quote basically writes itself. Personalized pricing is, and I'm quoting one of the public commenters here, "abhorrent." But the twist, the reason this is worth your time and not just outrage bait, is that the critics on both sides think the FTC might be about to make things worse.
So let me lay it out plain. Personalized pricing, the way the FTC defines it, is when a business uses your personal data to figure out the highest price you'd personally be willing to pay. Not the market price. Your price. And the FTC chair, Andrew Ferguson, is being honest about the limits here. He said, and I'm quoting, "The FTC does not have the legal authority to ban personalized pricing in all circumstances, but businesses that fail to tell consumers how their personal data is being used to set a price may be in violation of the FTC Act." So they can't ban it. What they can do is come after you for not disclosing it.
Now the examples the FTC put in its own policy statement are the kind of thing that makes your stomach turn, and they're worth hearing because they're specific. A food delivery service raising prices because the data suggests you can't leave your home. A grocery chain charging a family more for milk because the data says they've got more kids. A hotel quoting you a higher rate after it figures out you're traveling for a funeral and you've got nowhere else to go. And here's the one that lands for anybody who's built an app: it would be deceptive, the FTC says, if a rideshare company charged you more after noticing no competing rideshare apps were installed on your phone. That's not hypothetical pricing theory. That's reading the software on your device to figure out how trapped you are.
And most of the commenters, dozens of regular people who wrote in, they hate it. They called it discriminatory. They pointed out it hits low-income folks hardest, the people with, quote, "less time, technical literacy, or resources to detect and circumvent these practices." That's from a commenter named Sarah Burdell, who also made the sharp point that these pricing algorithms lean on data correlated with race, gender, age, geography, so you can end up reproducing discriminatory outcomes even when nobody typed in a racist rule. The machine just finds the proxy.
So far, so simple, right? Bad practice, regulator steps in, hooray. Except here's why I'm giving this the deep dive slot instead of just a quick hit. The smartest criticism in the whole piece isn't from the industry. It's from a concerned citizen named Jessie Shettleroe, and this person is not defending the surcharges. Listen to this. Shettleroe wrote that the policy statement "describes conduct that outrages people, then offers a remedy that permits it with a notice attached. I do not want a disclosure. I want the practice prohibited where it is clearly exploitative." And then the kicker, the part that should matter to every founder listening: "Nowhere does the Commission define personalized pricing. The operative phrase is prices that vary based on their personal data. That covers my grocery loyalty card, emailed coupons, and app-only prices. Those save me money."
You hear that? That's the whole problem in two sentences. The FTC wrote a rule so broad it covers the surcharge that screws you and the coupon that helps you, and it didn't tell you which is which. And Shettleroe's prediction is the one I'd bet on: if a business can't tell whether its practice is covered, the safe move is to stop personalizing altogether, and the first thing to disappear won't be the surcharge, it'll be the discount. Because the discount is the optional nice-to-have. The surcharge is where the money is.
Now let me connect this to you specifically, the builder. If you run anything with dynamic pricing, loyalty tiers, promo codes, A-B price testing, this is your world. Another critic in the piece, Deymond Lashley, made an argument that cuts the other way, that if the FTC restricts personalized price discrimination but leaves loyalty programs and coupons alone, you could actually entrench the incumbents, because the big players run the loyalty machines and the little guy trying to compete on a clever pricing model gets clipped. So depending on how the scope lands, this rule could either protect small players or lock them out. Nobody knows yet, because the scope isn't defined.
There is a legislative version floating too. A bill in the House called the Stop AI Price Gouging and Wage Fixing Act, which would actually let Congress ban certain algorithmic pricing, and importantly it carves out the good stuff explicitly, discounts for teachers, veterans, seniors, students, and loyalty programs people actually signed up for. That's the version with the guardrails drawn in ink. The FTC version is the one drawn in fog.
My take, and I want to be careful here because I'm not going to pretend I know how the rulemaking shakes out. What I'd watch for is whether the final language names names. Does it explicitly say loyalty pricing, coupons, promo codes, and randomized price testing are outside the scope? Because that's the difference between a rule that kills the funeral surcharge and a rule that just makes every checkout page a little more expensive for everybody. And there's a wage angle buried in here too. One of the commenters, a data privacy attorney named Blake Hunter Yagman, who's litigated both surveillance pricing and surveillance wage cases, wanted the FTC to go after the wage side as well, the same extractive logic pointed at what they'll pay you instead of what they'll charge you. That's the version of this story that comes back around in a year, mark it down.
Alright. That's the one to sit with. Now let me pivot to a different flavor of the same underlying idea, because it's a nice hinge. The pricing story is about a machine figuring out how much you'll tolerate. This next one is about a machine that tells you a comforting number that may not be true at all.
Oura, the smart ring people, they're facing a proposed class action, filed by the Clarkson Law Firm out of San Francisco, and the accusation is a beauty. The complaint alleges the ring can't actually measure any of the physiological signals you'd need to assess sleep quality or sleep stages, and instead leans on AI-generated estimates that have, quote, "a coin flip's chance of being correct." A coin flip. And the line in the complaint I keep coming back to, because it's almost poetic: "sleep happens in the brain, not on one's finger."
Now here's where it stings. The complaint says Oura marketed the ring as "built for accuracy," told customers it could hit 79 percent accuracy, and more recently claimed 95 percent sleep-staging accuracy compared to a clinical sleep lab. And the argument is that to actually stage sleep the way a lab does, you need electrodes on the scalp and sensors on the eyes. You're not getting that from a band around your finger. Ryan Clarkson, the managing partner, put the harm this way, and this is the part that matters for anybody shipping a consumer AI product: "People structure their days, interpret the way they feel, and design their lives around inaccurate figures spit out by these devices."
That's the whole hazard of confident AI output in one sentence. It's not that the number's wrong. It's that the number is presented cleanly, precisely, with a decimal point, and people rearrange their lives around it. Oura didn't respond to a request for comment, and look, a lawsuit is an allegation, not a verdict. But if you're building anything that hands a user a number, a health score, a productivity score, a risk score, the lesson here is the gap between "estimate" and "measurement." Market the estimate as a measurement and eventually somebody's lawyer reads your landing page copy back to you in a courtroom.
Now let me shift from what AI claims to measure over to what actually makes AI agents work, because there's a piece here about the guts of agent systems that's relevant even though I want to flag something up front. This one from Julie Bort is a recent report, but the underlying Nvidia research and the supporting studies it cites go back a few months, so I'm not going to sell it as breaking news. I'll give you the argument and move on, because I know you builders live in this stuff.
The claim is that Nvidia put out research showing the harness, the software wrapper around a model, the memory management, the tools, the rules, matters more than the model itself for long, multi-step tasks. They took Claude Opus 5, ran it on an interactive reasoning benchmark called ARC-AGI-3, which is a bunch of 2D games with no instructions where the thing has to figure out how to play. Naked model scored 30 percent. With a custom harness that handled memory well and added a supervisor component, a boss agent that nudges the worker agent when it wanders off, they hit 100 percent. Human-level on that benchmark.
Now I'm not going to camp out on the mechanics here, and I'll be straight with you, I'm a little allergic to the "it's the harness not the model" framing because it's become a bit of a industry catechism at this point. So let me pull the one genuinely useful thing for you out of it and leave the rest. The useful part isn't the accuracy number, it's the cost number. There's a bit in here from Databricks CEO Ali Ghodsi, who said, quote, "You can pick the same model but different harnesses, and you get significantly more cost if you use the wrong harness. That itself can 2x your cost." That's the line that should stick. Same model, wrong wrapper, double the bill. When you're evaluating why your agent is expensive, the answer might not be "switch to the cheaper model." It might be "your scaffolding is making forty calls where it needs eight." That's a founder problem, that's a burn-rate problem, and it's more actionable than the benchmark bragging.
And there's a real-world version of that same lesson sitting right next to it, from Mistral, which put out its Agentic Search product this week. I'll keep this tight because it's a product announcement and I'm not here to read you their press release. But the numbers are the interesting part. They built a retrieval layer that lets a model navigate inside long, dense documents, financial filings, contracts, scanned government PDFs, using five basic tools that look like file operations, search, open, navigate, read, grep. And on a financial-filings benchmark they took correctness from 26 percent up to 86 percent. On a harder benchmark of scanned Treasury bulletins, from 6 percent to 52 percent.
The reason I mention it next to the harness stuff is the mechanism is the same idea. Old-school retrieval, the one-shot approach, grabs a fixed set of text chunks and asks the model to answer in one pass. Works fine if the answer's sitting in the top result. Falls on its face when the answer's buried in a footnote or spread across three documents. The agentic version lets the model poke around, refine, go back, verify before it answers. And crucially, Mistral's own data showed adding the navigation tools improved accuracy while cutting token use by a third and dropping latency. More right, cheaper, faster, all at once. So if your product's doing enterprise document Q-and-A and it's been giving you garbage, the finding here is that the fix might be letting the model read like a person reads, going back to the table instead of trusting the first chunk it grabbed. That's a build decision, and it's a cheap one to test.
Now let me get to the story that I think deserves a good long look, because it's the kind of quiet failure that carries a big lesson: Tesla's solar roof is dead. Tim De Chant reported it, Tesla scrubbed the product off its website, all the pages now redirect to the generic solar landing page. And I want to spend a minute here because this is a founder story dressed up as a climate story.
Nearly a decade ago Tesla launched this thing, a roof that generated power but looked like fancy shingles, terra-cotta, slate. Beautiful pitch. Elon promised a thousand installations a week. You know what they got to, after years of, quote, "refinement and process improvement"? Somewhere between twenty and forty a week. As of 2022. Twenty to forty against a target of a thousand. That's not a shortfall, that's a different universe.
And the autopsy is instructive, so let me walk it. First problem, it was a luxury product that only got more expensive. People reported quotes of two hundred thousand dollars to install one of these. Second problem, and this is the killer for anybody who thinks hardware, regular solar panels got cheap because the whole world makes billions of standardized cells and slaps them into standardized panels. Economies of scale you didn't have to build. Tesla's tiles were unique to the system. So Tesla had to design and buy its own manufacturing equipment, and when the sales didn't show up, the per-unit cost of all that custom equipment went up, not down. You want to hear the trap in that? You need volume to make the custom line cheap, but the thing's too expensive to hit volume, because the line is custom. That's a snake eating its own tail.
And then physics showed up to kick it while it was down. There were reports of the non-solar parts warping, roof underlayment melting under full sun, and worse for Tesla, underproduction, the system just not making the electricity they said it would. Likely culprit, heat. Solar cells work better cold, voltage drops about half a percent for every degree Celsius, and regular panels leave an air gap under them to cool off. Tesla's gap was smaller. So the thing cooked itself and lost efficiency.
Here's why I'm giving a dead roof this much airtime. The takeaway isn't "solar roofs are dumb." Other companies, GAF, Merlin Solar, are still in that market. The takeaway is about the seduction of the integrated, beautiful, custom product versus the boring, standardized, ugly one that wins on cost. Tesla bet that people would pay a premium for the thing that looks nice, and enough of the market said, I'll take the ugly panels and the cheaper bill. And as the piece put it, if one of the most valuable companies on Earth can't make it work, that's either an indictment of the whole concept or a challenge to some founder stubborn enough to prove Musk wrong. I'd frame it slightly differently for you: any time your product's economics depend on being both premium and high-volume at the same time, sit with that tension a long, long time before you tool up the factory.
Let me stay in the hardware and jobs lane for a second, because Apple gave us a companion piece. Bloomberg reported, and TechCrunch picked up, that Apple's cutting over 200 jobs, roughly a hundred off the Vision Pro team, its long-suffering headset, and another hundred or so from Siri and a group called Intelligent Systems Experience, which handles AI integration into devices. Apple confirmed the cuts, gave the standard corporate line about evolving the business, said it'll create new roles even as it eliminates some. The context that matters here: Apple's in a genuinely bumpy stretch. It's trying to fix its AI position, it's dealing with a memory shortage driven by the whole industry's AI buildout that's jacked up the cost of making its flagship products, it raised prices on Macs and iPads this summer, and it even rolled out a leasing plan for its hardware. And oh, by the way, it's suing OpenAI, accused them of trade theft. So the Vision Pro headcount coming down while the AI headcount gets reshuffled tells you where Apple thinks the next decade is, and it isn't strapped to your face.
Now let me swing over to the geopolitical corner, because there are two stories here that are current and consequential, and then a resurfaced one I want to handle carefully.
The current one first. Romania, this week, accused Russia of sending a drone boat loaded with explosives at a European natural gas platform in the Black Sea. This is the Neptun Deep project, and it matters because once it's fully running in 2027 it doubles Romania's gas production and helps supply Germany and Moldova. A Romanian F-16 used its cannons to disable the surface drone, then ordnance experts blew it up in a controlled detonation. Several hundred workers were on that platform doing the installation. Romania's president condemned what he called the intensification of these irresponsible incidents from the Russian Federation.
And here's the detail that made me put down my coffee. A naval expert who runs a site called Covert Shores noted this drone boat was a type Russia hadn't deployed before, with a large hinged hangar in the forward hull that might've let it carry missiles or aerial drones, plus two flat antennas that may have allowed satellite comms through services like Starlink. Think about that for a second. An uncrewed attack boat that might be phoning home over commercial satellite internet. That's the war we're in now. This is the second incident near that gas site this month, and both times Romania checked with Ukrainian partners to confirm the drones weren't Ukraine's. The Black Sea's become a proving ground for cheap, autonomous, expendable naval weapons, and if you build anything in the maritime or defense-tech space, that hinged-hangar drone boat is the shape of the threat.
Which brings me, cleanly, to the money side of that same trend. Castelion, a missile startup founded by ex-SpaceX people, raised a billion-dollar Series C at a 13 billion dollar valuation, co-led by Andreessen Horowitz, Carlyle, and JPMorgan Chase. Founded in 2022, they've already pulled more than 500 million in US military contracts. Their whole pitch is manufacturing hypersonic weapons faster and cheaper than the traditional defense primes, and the timing's pointed, because the Pentagon's stockpile of hypersonics, the ones going above Mach 5, hasn't kept pace with China's. They're building missiles they named Blackbeard, after the pirate, out of a New Mexico facility. Whatever you think about the defense-tech gold rush, that drone boat in the Black Sea and this 13 billion dollar valuation are two ends of the same wire. The cheap autonomous stuff is proliferating on the battlefield, and the capital is stampeding into the companies that mass-produce it.
Now the resurfaced one, and I'm flagging the timing on purpose. There's a piece that came around on Chinese lidar sensors and whether they're a security risk on American cars, being probed by the Idaho National Laboratory. I want to be honest with you, the underlying reporting on that review dates back to the early summer, so I'm not going to hand it to you as fresh news. But the tension in it is durable and worth one clean beat: Chinese lidar suppliers like Hesai and RoboSense drove the price of a sensor from as much as seventy-five grand a decade ago down to a few hundred bucks, US lidar makers got crushed, Luminar went bankrupt, and now American automakers are tempted to use the cheap Chinese sensors while lawmakers panic about back doors. It's the same movie as the drone boat and the milk pricing, honestly. The cheap capable thing wins on cost, and everybody argues about the risk after it's already everywhere. I'll leave it there.
Let me do a quick lap through the policy and consumer stuff before I take you to the Moon, because there's a settlement worth naming. TikTok and ByteDance reached a 400 million dollar settlement with the Department of Justice over children's privacy, resolving a case first filed back in 2024. The allegation was that TikTok let millions of kids under 13 onto the platform and collected their data without parental consent, violating COPPA. No admission of wrongdoing, but they're paying the 400 million and adding stronger age controls and parental oversight. And the detail that gives it teeth: this is the second time around, the predecessor company Musical.ly paid 5.7 million back in 2019 and promised to keep under-13s out, and the DOJ says they just kept struggling to do it. And it lands right after a Bloomberg report that TikTok deliberately switched off a content safeguard for about 10 percent of US users as an experiment, which got two senators, Blackburn and Blumenthal, writing angry letters. So, 400 million, and the age-verification fight for social platforms grinds on.
Alright. I promised you the Moon, and I want to close there, partly because it's a callback. We were on Ariane 6 and the launch economics yesterday, so let me tie the thread. There's a spent Falcon 9 upper stage that got left drifting in deep space after launching a commercial Moon lander back in 2025, and on August 5th it did what abandoned rockets occasionally do. It hit the Moon. NASA's Lunar Reconnaissance Orbiter got images of the fresh scar near Einstein Crater, a new crater about sixty feet across and less than ten feet deep. And the genuinely cool part, the accidental science: the impact dug up pristine material from below the surface, stuff that's escaped weathering from solar wind and cosmic rays, and it gave the asteroid-tracking folks a chance to test their prediction methods. Turns out they nailed the impact site to within about six-tenths of a mile. So a piece of space junk became a free experiment.
And in the same report, two things I want you to hold onto because they're the real story under the pretty crater. One, the launch market is being reshaped hard. New research out of the UK and Italy quantified it: the US now sends a kilogram to orbit for about 3,225 dollars. Europe? 9,897. India nearly 15,000. That's not a gap, that's a chasm, and it's almost entirely reuse. Two, and this connects straight back to yesterday, there's a brewing launch crisis, because SpaceX wants to retire the workhorse Falcon 9 for the still-unproven Starship, and it's told customers it intends to end commercial launches after 2028. Satellite operators are in a panic because they can't plan around the one reliable ride anymore. Meanwhile China just landed a second reusable booster, from an eleven-year-old company called LandSpace, and it's about to send its Chang'e 7 mission to the lunar south pole, to Shackleton Crater, where the permanently shadowed cold traps might hold water ice. That's the most coveted real estate on the Moon, and China's going for it while the American commercial launch market's got a question mark hanging over its head.
That's the through-line for the whole episode, if you want one. From the FTC fog to the dead solar roof to the launch economics, it's the same question over and over: who benefits, who gets squeezed, and does the cheap capable thing win before anybody's figured out the rules. Watch the FTC scope language, watch whether your agent harness is doubling your bill, and watch 2028 on the launch calendar.
That's the menu, kiddos. Plates are clean. I'm Tony DeLuca, thanks for spending a little of your day at the counter with me, and I'll see you next time on Barely Possible. Take care of each other out there.