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, pour the coffee, grab a seat, I'm your boy Tony DeLuca and this is Barely Possible, the show where we take the day's big pile of tech noise and pull out the parts that actually matter to you, the person who's building something. Today we've got a documentary premiere that felt like a heist, a jaw-dropping number about what a working researcher actually spends on AI in a single day, a couple of consumer AI features that tell you more about the industry's mood than the labs would like, a car with no child seat anchors, and a court telling us who really owns a word. Buckle up. Let's have at it.
Let me start with the one that stuck in my teeth, because it's the number that says the most about where this whole thing is headed, and it's not from a press release. It's a small thing, a line pulled out of a piece and posted by Simon Willison. And here's the line, word for word, from the article he was quoting: "the median researcher was integrating agents daily into their work, using more than $600 per day of inference at API prices."
Six hundred dollars a day. Per researcher. The median researcher. Not the one hotshot burning compute for a stunt. The middle of the pack.
Now I want you to sit with that for a second, because if you've been listening to this show you know I've spent a lot of time being the guy at the diner counter waving off the hype. And I'm still that guy. But this is not a hype number. This is a spend number. This is a receipt. And receipts are the thing I actually trust.
Let's do the math the way you'd do it on a napkin. Six hundred bucks a day, call it twenty working days a month, that's twelve thousand dollars a month. In inference. For one person. That's a hundred and forty-something thousand a year, per head, just in the tokens they're feeding these coding agents. That is not a per-seat SaaS line item. That's not twenty bucks a month for a productivity tool. That's a number that looks like a second salary. And Simon, in the surrounding chatter around this, was poking at something worth poking at, which was a big uptick in token spend around mid-July, wondering whether that lined up with a new model landing in the hands of employees. I'm not going to state that connection as fact, because the post is speculation, not confirmation. But the shape of it matters. When a stronger tool shows up, the spend doesn't creep. It jumps.
Here's why I'm leading the show with a single tweet quoting a single line. Because if you're a founder, this reframes your entire cost model, and it reframes it in a way most of the pitch decks I imagine floating around right now have not caught up to yet.
For years the mental model for software was: a seat is a seat. You sell somebody a login, they pay you a monthly fee, and your job is to keep your own costs below that fee. The whole SaaS religion is built on that. Gross margins in the eighties, low marginal cost per user, scale to the moon. That's the machine everybody learned to build.
Agents break that machine. Because now the person sitting in the seat isn't just clicking around your interface. They're spinning up work that consumes real compute, all day, continuously, and that compute has a real dollar cost that scales with how much they use it. The heavy user isn't your best margin anymore. The heavy user is potentially your worst margin, because the heavy user is the one running the loops overnight and burning six hundred bucks of somebody's money doing it.
And I want to be careful here about who "somebody" is. In this specific case, that six hundred a day is at API prices, and these are researchers inside a lab, so a lot of that is internal cost, not something a customer is cutting a check for. But the principle travels. Because the enterprise version of this is coming for your product. If your customer's employees can run agents through your tool, the question stops being "how many seats can I sell" and starts being "how do I price something whose cost floor moves around based on how ambitious my user gets on a Tuesday."
I've talked before on this show about the revenge of the CFOs, this idea that after everybody got drunk on agentic experimentation, the finance people walked in, looked at the invoices, and started asking hard questions about cost per task. This six hundred dollar number is the thing the CFO is staring at. And the smart builders right now, the ones who are going to still be standing in two years, are the ones designing for that reality instead of pretending the old per-seat math still holds.
What does designing for it look like? A few plain things. One, you meter. You have to know, per customer, per workflow, what a task actually costs you to run, or you are flying blind into a wall. Two, you route. Not every job needs the biggest, most expensive model. A quick database lookup and a full codebase refactor are not the same amount of thinking, and if you're paying frontier prices for both, you're lighting money on fire. Three, you either pass the cost through honestly, usage-based, or you cap it, so one power user doesn't eat the whole quarter. There's no clever fourth option where you eat unlimited compute at a flat monthly price and grow forever. That's not a business, that's a charity with a runway.
And here's the part that should actually excite you if you're building, because I don't want to leave you at the bar crying into a napkin. The flip side of "this stuff is expensive" is "this stuff is worth a lot." Nobody spends six hundred dollars a day on something that isn't paying for itself. If a researcher, or an engineer, or a lawyer, or an analyst is comfortably burning that kind of money on inference, it's because the output is worth multiples of that. The total amount a company can usefully spend on this went from twenty bucks a head to potentially thousands a head. That means the pie you're selling into is enormous. It just isn't shaped like a SaaS seat anymore. It's shaped like consumption. And the builders who internalize that early get to price like a utility instead of like a to-do list app.
So that's my one big number for the day, and it's the thing I'd write on the whiteboard if I ran your Monday standup. The unit of cost has changed. Build like it has.
Now let me connect that to something happening on the other end of the pipe, because if the researchers are the ones spending six hundred a day, the everyday consumer is the one the industry is still trying to convince to care at all.
Let's shift from what power users spend to what regular people won't. There's a recent piece about Google's Gemini Spark now being able to manage your Google Photos library. And on paper it sounds fine. You ask your personal agent to edit images, curate albums, build a shared album of your best shots, turn a photo of a concert flyer into a calendar appointment. Google Photos lead Shimrit Ben-Yair posted about it, talking about finally having a power agent to help get the most out of a hundred-and-forty-three-thousand-photo library. It's rolling out over the coming weeks to eligible Gemini AI Pro and Ultra subscribers in the U.S., in English, no word yet on other markets.
And look, I'm not going to sit here and tell you organizing a hundred and forty thousand photos is nothing. If you've got a photo library that looks like a hoarder's garage, sure, an agent that tidies it up has some value. But the piece itself makes the sharper point, and I think it's the honest one: on its own, this doesn't feel revolutionary or even, really, necessary. It's not that hard to make a photo album.
And here's where it ties into the mood of the whole industry right now. In that same coverage there's a reference to OpenAI's Sam Altman telling Bloomberg this week that the industry has, quote, done a terrible job communicating the benefits of the technology, which has led to backlash from communities around the world. Now I want to be precise, because I'm not going to overbuild that quote into a thesis it can't hold. He said the industry did a terrible job explaining the upside. Fine. But the delicious irony that the TechCrunch piece points at is that announcements like "your photo agent can now make an album" are part of the problem, not the solution. Because the competitive pressure is so intense that every company feels like it has to trumpet every tiny AI feature, no matter how small, instead of shutting up and waiting until they've got a story worth telling.
Think about the split we just walked through. On one side, researchers burning six hundred bucks a day because agents are genuinely transforming deep, hard, valuable work. On the other side, the consumer pitch is "hey, we'll sort your photos." That gap is the whole ballgame. The value is real and enormous at the frontier of professional work, and it's thin and unconvincing at the kitchen table. And the labs know it. That's what Altman's "terrible job communicating" comment is really admitting, whether he means to or not. It's not that they failed to communicate the benefits. It's that for the average person, the benefits they can point to right now are kind of small.
For you as a builder, there's a lesson buried in there, and it's an old one that my show keeps coming back to because it keeps being true. The money is where the work is hard and expensive and high-stakes. It is not, yet, in shaving thirty seconds off making a photo album. If you're building consumer AI, you're fighting for attention in a space where even Google's own coverage admits the killer feature isn't killer. If you're building for professionals doing expensive, verifiable work, you're playing in the pond where people are already comfortable spending real money. Pick your pond knowing which one has the fish.
Alright, let me pivot to the story that had the most pure showmanship in it this week, because after all this talk of tokens and margins, you deserve a little drama.
There's a recent report out of the Telluride Film Festival, and I want to be clear about the timeline before I get into it, because the underlying subject here is not new. This is the Elizabeth Holmes story, Theranos, the whole saga that landed her an eleven-year sentence back in May of 2023 for defrauding investors about her blood-testing technology. That conviction is old news, that's history. What's new, per the recent reporting, is the movie.
And the way this thing dropped is genuinely wild. Telluride pulled off one of its best-kept secrets in years. Sunday night, festivalgoers filed into a six-hundred-and-fifty-seat theater, phones locked away, no idea what they were about to watch. All they were told was that it ran nearly three hours, it had never been shown anywhere, and its principal creators would be sitting in the audience with them. And then the lights go down and it's a brand-new documentary about Elizabeth Holmes, from co-directors Nathan Fielder and Lance Oppenheim. It's called You Can See Everything.
Here's the part that made me put my coffee down. The film reportedly captures Holmes and her husband, Billy Evans, in the thirty-four days before she reported to federal prison. And the access, by all accounts, was total. Nathan Fielder apparently lived in the couple's guest house during filming. The guest house. And there's more buzz tied to an A-list actress who apparently becomes part of the story itself. A24 is set to release the film in October.
Now why am I telling you, a founder, about a documentary? Two reasons.
One, because this is the culture completing its digestion of a founder-fraud story, and that matters for the whole ecosystem you operate in. The Holmes saga became the load-bearing myth for "fake it till you make it gone wrong." Every skeptical journalist, every cautious investor, every regulator eyeing an AI startup's revenue claims, they've got that story in the back of their head. And I'll connect a thread from earlier this week's episodes here for continuity: we've been talking a lot lately about AI-generated ARR and the fragility of startup revenue moats, about whether the numbers in these decks are real. The Holmes story is the cautionary tale that lives underneath all of that scrutiny. A big cultural artifact re-litigating it, with that kind of intimate access, keeps that skepticism sharp. Which, honestly, is probably healthy for everyone building something real.
Two, and this is the part I find genuinely interesting as a media guy at heart, is the access itself. A convicted fraudster and her husband invited a documentary crew to live with them in the final month before prison. That's not an accident. That's a narrative play. Somebody made a decision that controlling the story, or at least being present in it, was worth the risk. And I don't know how it plays out until the thing hits theaters in October, and I'm not going to pretend I've seen it. But as a lesson in how founders think about narrative, even at the absolute bottom, it's a heck of a case study. When the walls are closing in, the instinct of a certain kind of founder is still to shape the story. File that away.
Let me leave Hollywood and get back to the machinery, because there's a legal ruling this week that's a genuinely useful lesson in how intangible assets work, and it's got Elon Musk in it, which never hurts for entertainment value.
A federal court in Delaware handed down a split ruling in a recent report on a trademark fight between Musk's X and a startup called Operation Bluebird. And this one's a little bit delicious. The startup originally launched under the name Twitter.now. The court told them no, you can't use that, X still holds the "Twitter" trademark. Fine. But then the court found something more interesting. It found that X had likely abandoned two other related trademarks: the word "tweet," and the old Twitter bird logo. Abandoned them. Left them on the sidewalk.
So Operation Bluebird did exactly what you'd expect a couple of sharp lawyers to do. They rebranded to Tweet.app and launched to the public using the freed-up name. And let's be honest about what's going on here, because the coverage is refreshingly blunt about it. This isn't really a company trying to build the next great social network. This is a company trying to grab valuable trademarks that X threw away. It's headed up by two lawyers, founder Michael Peroff and Stephen Coates, who was previously a trademark lawyer at Twitter. These are not naive kids with a dream. These are people who read the fine print for a living and spotted an opening.
And here's the judge's actual reasoning, from the opinion, which is worth hearing because it's the whole principle in one sentence. He wrote that Bluebird was, quote, likely to succeed in proving both that X Corp. discontinued the bona fide use of the Tweet mark and Bird logo and that it intends not to resume the use of the marks.
That's the lesson right there, and it's one every founder should tattoo somewhere. A trademark is not a deed. It's not something you own forever just because you registered it once. It's a use-it-or-lose-it right. It exists to identify a source of goods and services in commerce. The moment you stop using it, and signal you don't intend to go back, you're at risk of it going up for grabs. Musk renamed the town square to X. He, in the words of Bluebird's own homepage, threw the bird away on his way out. And the law, at least in this preliminary ruling, is saying: okay, then it's not yours anymore.
Coates put it in a way that I have to admit is pretty good rhetoric. He said, quote, they kept the word. They let go of the bird, and they let go of the tweet. And then, quote, a tweet was never a corporation. It's one person saying something. That word survived three years of a company trying to replace it, because the public declined to stop using it. We think that tells you who it belongs to.
Now I want to add a layer of Beacon skepticism to the halo of that quote, because it sounds noble and it's mostly self-serving. The startup is charging users twenty bucks to reserve a handle, and the coverage flatly notes that probably helps cover the legal bills. They said more than a hundred and seventy-two thousand people requested a handle before launch, and that's almost certainly just leftover affinity for the Twitter brand, a name the company literally can't use anymore. And the ruling is not final. This is a preliminary injunction fight. The case proceeds to figure out whether X ultimately keeps any of these marks. So don't go treating Tweet.app as a done deal or a business model, because it's neither yet.
But the underlying principle for you as a builder is rock solid and it's worth money. Your brand assets are living things. If you've got a product name, a logo, a mark you're not actively using, understand that the protection you think you have may be quietly evaporating. And on the flip side, if you're eyeing a name a bigger company clearly walked away from, there may genuinely be an opening there. It's not glamorous. It's not a moat in the "we have the best model" sense. But trademarks, domains, brand equity, these are the intangible assets that outlast the tech, and most founders treat them as an afterthought until a lawyer with better attention to detail eats their lunch.
Now shift from what a company let go of to what a company built in, because there's a product-decisions story in the Tesla Cybercab coverage that I want to hit carefully, since we touched on the launch itself earlier this week and I'm not going to rehash the whole thing.
We covered the Cybercab launch and the NHTSA investigation in the last couple of episodes, the self-certification questions, the small size of the initial Austin fleet, all of that. I'm not repeating it. But there was a recent report digging into the actual fine print of the vehicle, the terms of service and the released documents, and a couple of the details are genuinely interesting as case studies in product decisions, so let me pull just those threads.
The one everybody fixated on: no little kids. Tesla does not allow minors under thirteen in the Cybercab, at least "at this time." And here's the detail that raised my eyebrow as somebody who thinks about how products get built. The Cybercab does not have the standard LATCH anchors for child seats. The anchors that are standard equipment in basically every car sold in America. You can only attach a child seat with the seat belt. Tesla's whole philosophy, stated openly by their executives, is stripping out unnecessary parts to cut cost. And apparently child seat anchors made the "unnecessary" list.
Now I'm not here to do vehicle safety commentary, that's not my lane and it's not this show's job. But as a lesson in product tradeoffs, this is a clean one. When your core design principle is "remove everything you can to cut cost," you will eventually remove something that maps to an actual customer need, and then you have to paper over it with a policy. That's what "no kids under thirteen at this time" is. It's a policy patch over a hardware decision. And there's a broader pattern in the documents, brake-by-wire instead of hydraulic, again to avoid the complexity and cost of plumbing, windows that "cannot be fully opened at this time" with no explanation given. Some of these are legitimately clever engineering. Some are cost-cutting that shows up later as a limitation the user has to live with.
The builder takeaway, and then I'll move on, because I promised I wouldn't relitigate the whole launch: minimalism is a virtue right up until the moment it isn't. "Delete the part" is a great instinct in software and in hardware. But every part you delete either wasn't needed, or it was, and if it was, the need doesn't disappear, it just resurfaces as a constraint, a policy, or a support ticket. Know which deletions are which before you ship.
Alright, let me do a quick lightning round on a couple of smaller items, because they each carry one clean idea worth thirty seconds of your attention.
First, an older story that resurfaced this week, and I want to frame it correctly: this is not a fresh event. The reporting notes the underlying changes rolled out earlier this year, with the Air Force implementing its piece more recently in July. The story is that the U.S. Department of Defense disabled advertising tracking on troops' devices, across the Army, Air Force, Navy, Marine Corps, and Special Operations Command, on government-issued iPhones, Android phones, and Windows computers. This came out through a letter shared with Senator Ron Wyden, and Reuters first reported it. The reason is chilling and simple: foreign adversaries were using commercially bought location data, the same ad-tracking data that flows to data brokers, to target U.S. troops in the Middle East. Turning off the advertising ID makes a device blend into the crowd of everyone else who's disabled theirs.
Here's the builder-and-citizen point, and it's the one Wyden himself makes. The exact same data economy that powers your growth funnel, the ad IDs, the location data, the third-party brokers, is a national security liability when it's pointed at a soldier. And in the ultimate bit of institutional hypocrisy, the piece notes that the U.S. intelligence community and the FBI have confirmed they buy that same commercial data for surveillance without a warrant. So the government is simultaneously protecting its own troops from the data broker economy and buying from it. If you build products that touch location or advertising identifiers, understand you're operating inside a supply chain that governments treat as both a weapon and a threat. That's not a compliance footnote. That's the whole moral weather of the space you're building in.
Second lightning item, on the lighter side, because I can't do a whole episode without something that made me laugh. There's a recent report on a new iPhone alarm app called Clucky. Its signature feature is that it wakes you up with a crowing rooster and then won't let you turn it off until you complete a mission, math problems, tapping tiles in order, doing push-ups, walking around, an AI camera challenge. It's in a category with apps like Alarmy and Awake, the whole idea being to make dismissing your alarm so annoying you have no choice but to actually wake up. There's a Rooster Report grading your week, a "Flock" feature so you and your family can hold each other accountable, and a pro subscription at just under forty bucks a year. It was built by a former Fanatics software engineer, Adrian Angelo Abelarde.
And honestly? As a business, I kind of respect it. This is not AI moonshot stuff. This is a solo builder finding a real, dumb, universal human problem, people hitting snooze, and shipping a product with a personality and a subscription. Forty bucks a year, no six-hundred-dollar-a-day inference bill, no frontier model, no regulatory investigation. There's something clarifying about that after a whole episode about tokens and trademarks. Sometimes the business is just: a rooster, a paywall, and people who can't get out of bed. Not everything has to be an agent swarm. File that under "the market is bigger and dumber and more human than your pitch deck assumes," and I mean that as the highest compliment.
Two more quick ones from OpenAI's side of the fence, and I'll keep these tight.
There's a current announcement that OpenAI, along with two partners, AIRPPU and WAN-IFRA, is launching a program to help Ukrainian news organizations strengthen innovation, resilience, and independent journalism. That's about the extent of the detail available, so I'm not going to inflate it. But I'll note the irony, gently, and let you hold both things in your head. The same set of companies that publishers around the world are suing over training data, and we've covered those suits, is also funding independent journalism in a country at war. Both things are true. Don't flatten it into a single villain-or-hero story. The relationship between AI labs and the news business right now is genuinely contradictory, cash flowing one direction while lawsuits fly the other, and pretending otherwise makes you a worse analyst.
And the last item, which I'll frame carefully because it's a recent report, not fresh-today news. OpenAI's most capable model, GPT-6 Astra, landed a few days back, positioned as their most intelligent and aligned model yet, with the pitch centered on computer use, coding, cybersecurity, and science. And there's a Sam Altman post from around the launch worth noting for its honesty, where he said, and I'll read it, "astra is a powerful model and we are working to make it generally available. we do not think it is a good strategy to keep powerful models to a chosen few. given its cyber capabilities, we need a little big longer to do this safely. but hopefully not too long."
I bring that up not to re-cover the model, we've touched Astra already this week, but to close the loop on the theme I opened with. Remember that mid-July spike in token spend Simon was wondering about? A more capable model showing up in employees' hands is exactly the kind of thing that drives spend up like that. Stronger model, more ambitious agent runs, bigger inference bill. The capability and the cost move together. And Altman openly saying they're holding a cyber-capable model back a little longer for safety reasons tells you the frontier is now a place where the gating factor isn't "can we build it," it's "can we release it without it becoming a problem." That's the world you're building in now. The most powerful tools show up on a delay, the spend jumps when they land, and the value has to justify a number that would've sounded insane two years ago.
So let me tie the whole menu together before I let you go, because there's a single thread running through all of it, and it's not "AI is fast," which is the laziest theme in this business and I try never to serve it.
The thread today is about what things actually cost and what things are actually worth, and how badly the old models measure both. Six hundred dollars a day for a working researcher blows up the per-seat pricing religion. A photo-organizing agent shows you the value is thin at the kitchen table even as it's enormous at the frontier. A trademark ruling shows you that an asset you thought you owned forever is worth exactly as much as your continued use of it. A car with no child seat anchors shows you that a deleted part isn't a saved cost, it's a deferred one. And a forty-dollar rooster app shows you that sometimes the whole business is just knowing a real human problem and charging for it.
The unit of value changed. Price for it, build for it, and don't let anybody sell you a SaaS-shaped story about a consumption-shaped world.
That's the menu, kiddos. If today's episode saved you from one bad pricing assumption or one lazy moat, then I did my job. This has been Barely Possible, I'm Tony DeLuca, thanks for spending part of your day at my counter. Go build something that's worth what it costs. See you next time.