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 yourself something hot, I'm your boy Tony DeLuca and this is Barely Possible, your daily plate of AI news served without the salesman's smile. We've got a menu today that runs from a newspaper that took money from the same companies it's now suing, to a settlement where authors are getting robbed by their own publishers, to a fresh set of numbers from OpenAI about AI building AI. Buckle up. Let's have at it.
Let me start with the one that made me put my coffee down, because it's got a wrinkle you don't see every day in these copyright fights.
The Seattle Times and Newsday have filed suit against OpenAI and Microsoft. This is a recent report from TechCrunch, published September 5th, written by Anthony Ha. And on the surface, okay, another day, another newspaper suing an AI company over training data. We've all seen the New York Times case grinding along since 2023. Other publications have piled on since then. This is the pattern now. If you produce words for a living, and somebody scraped those words to build a machine that competes with you, you go to court. Fine.
But here's the part that separates this one from the pack. The Seattle Times took money from Microsoft and OpenAI. Those two companies funded some of the paper's journalism projects and fellowships. So you've got a situation where a news organization accepted checks from the very companies it is now dragging into a courtroom. Now, I'm not going to sit here and tell you that's hypocrisy, because it isn't, not really. You can take grant money for a fellowship program and still believe somebody's core business model is eating you alive. Those two things can both be true. But it tells you something about how sour this relationship has gotten. When the paper you helped bankroll turns around and sues you, the honeymoon, whatever honeymoon there was, is long over.
And the language in this lawsuit, I want to read you the actual words, because it's more colorful than the usual legal boilerplate. They described generative AI as, quote, a snake eating its own tail, that could, quote, destroy the very organizations that produce the content it's trained on. And then this line, which is the whole argument in one sentence: quote, AI products like ChatGPT and CoPilot are touted as producers of content, but in fact they are rapacious consumers, devouring human-authored content and delivering back to the world copies and derivative imitations of that same original content they consumed to achieve their commercial objectives.
Rapacious consumers. That's a hell of a phrase to put in a legal filing. And here's why it matters for you, if you're building anything on top of these models. The industry has been operating on a bet. The bet is that training on this stuff is fair use, and the courts will bless it after the fact, and the licensing deals they do sign will be cheap enough and voluntary enough that the whole thing stays manageable. We've talked before about the Anthropic ruling, where the judge drew that line, training on copyrighted material is legal under fair use, but pirating the material to get it in the first place, that's not. That distinction is now the fault line for the entire industry. And every new lawsuit is another test of exactly where that line sits.
Microsoft's response? A spokesperson told GeekWire the company is, quote, surprised by the lawsuit, but is, quote, always happy to sit down and explore solutions to this type of dispute. Surprised. Come on. Nobody at Microsoft is surprised. This is the ninth or tenth publication to file. The surprise is theater. What "explore solutions" actually means is: we'd rather write you a licensing check than let a jury decide what our training data is worth. Because that's the real risk hanging over all of this. It's not any one lawsuit. It's the possibility that a court someday puts a dollar figure on the whole corpus, and that figure turns out to be enormous.
So if you're a founder, the takeaway isn't "avoid AI." The takeaway is: understand that the data layer underneath your product has unsettled legal ownership, and the bill for that hasn't come due yet. When it does, it flows downhill.
And speaking of bills coming due and money getting split up in ways nobody expected, let me take you to a story that's the exact same theme from a completely different angle. Because this one is the copyright fight after the copyright fight.
This is about the Anthropic settlement, and I want to be careful here, because this is an older story resurfaced. Anthropic settled that copyright class action last year. The deal got final approval back in July. So the settlement itself is not news. What's new, and what TechCrunch's Anthony Ha reported on, is the mess that's happening now as the money actually gets distributed. And it is a mess.
Here's the setup. Under the settlement, the authors of nearly 500,000 titles get paid $3,000 for each pirated work. Simple enough. But then the split rules kick in. If the book is still in print with a traditional publisher, the money gets split fifty-fifty between the author and the publisher. If the book was self-published, or if the publisher let the rights revert because the book went out of print, then the author should get the whole thing.
And that's where it's falling apart. Writers have been posting that publishers are claiming more than their fair share. I'll give you the sharpest example, from mystery and thriller author April Henry, who asked, quote, WTF is HarperCollins playing at? They claimed one of my books on the Anthropic Settlement that reverted back at least 17 years ago AND on the same day I got a credit alert saying they had been added as my employer! Which they never were.
Seventeen years ago the rights reverted. And they're still putting in a claim. And somehow her credit report now lists HarperCollins as her employer, which is its own separate weirdness that nobody's fully explained.
Over at the blog Writers Beware, Victoria Strauss has been cataloging complaints, and she says they fall into two buckets. One, publishers seeking payment on works where the rights already reverted, so they've got no legitimate claim. Two, publishers claiming a full 100% when they're only entitled to 50. Now, Strauss is being fair about it. She wrote that she's, quote, reluctant to attribute to malice what can be plausibly explained by poor recordkeeping. And the Authors Guild CEO, Mary Rasenberger, told the New York Times she doesn't see this as, quote, a grab by the publishers, and doesn't think they're, quote, specifically trying to screw any author over. She chalks it up to bad record-keeping and a confusing process.
Maybe. But Strauss added a line that stuck with me. She said what she's seen is, quote, a peek through a small crack in a massive wall. And she wrote that the volume of reports, and the fact that authors are reporting the exact same errors over and over, suggests to her these aren't routine glitches, but, quote, something much more widespread and systemic.
Here's my read, and it's the DeLuca kitchen-table read. When the same mistake happens over and over, and every single one of those mistakes happens to favor the bigger party, at some point "sloppy recordkeeping" stops being an innocent explanation and starts being a business model. I'm not saying anybody committed fraud. I'm saying that a system where the default error always benefits the publisher is not a neutral system. It's a system that's been designed, on purpose or by neglect, to make the little guy fight for what's already his.
And then it gets richer. Because publishers aren't the only ones with their hands out. Strauss says literary agencies are also filing claims. Which is bizarre, because agents aren't rightsholders. They sell the book, they take their cut of the sale, but they don't own the copyright. Author Courtney Milan, who writes under that pen name but is a former law clerk and law professor, put it bluntly on Bluesky, and I'll clean this up a touch for the airwaves: quote, Apparently some agents are trying to claim percentages on the Anthropic settlement, and I do not REMOTELY think they should do this, what the... stop that.
I'm with Courtney on this one.
Now why does this matter to you, the builder? Because this is a preview of every AI settlement to come. When the copyright dust finally settles across this whole industry, and it will settle, there's going to be a giant pile of money that needs to get distributed to the actual humans whose work got used. And the plumbing for distributing that money doesn't exist yet. It's ad hoc. It's built on top of decades-old publishing contracts that nobody kept clean records on. The lesson here, if you're building anything that involves paying creators, licensing content, or splitting revenue with rightsholders: the accounting layer is the whole ballgame. Get that wrong and you don't just have an operational problem, you have a trust problem that outlives every check you cut. One tricky detail from the filing, for the writers listening who might be affected: to make a 100% claim on a reverted book, the rights reversion had to happen before August 10th, 2022, which is the download date in the settlement. So if you're an author in this pool, that date is the one to know.
Now let's shift from who owns the words to who's using them to build the next generation of the machines themselves. Because OpenAI put out something this week that's worth sitting with, and it ties directly into the biggest open question hanging over this entire industry.
OpenAI published a piece called Research acceleration: The view inside OpenAI. And the framing is exactly what it sounds like. It's early data on how coding agents are reshaping AI research inside the company. Agent usage, experiment velocity, task complexity, research acceleration. And Sam Altman amplified it on X, reposting a colleague who wrote, quote, Today we're releasing data on models accelerating research at OpenAI. Recursive self-improvement could be the most important, and the post trails off there.
Recursive self-improvement. Let's talk about what that phrase actually means, because it's the whole reason this little blog post matters more than a little blog post should.
The dream, or the nightmare, depending on where you sit, is a feedback loop. You build an AI that's good at AI research. That AI helps you build a better AI. The better AI is even better at AI research. And around and around it goes, each turn of the wheel faster than the last, until the thing is improving itself faster than any human team ever could. That's recursive self-improvement. It's been a theoretical concept in these circles for years. And what OpenAI is now saying, with data, is: it's not theoretical anymore. The coding agents are already inside the research loop, accelerating the experiments.
Now, I want to put my skeptic's hat on, which is where I keep it most of the time. This is a company blog post. OpenAI has every incentive in the world to tell you that its models are so good they're accelerating the invention of even better models. That's the entire investment thesis for a company chasing a multi-tens-of-billions run rate. "Our AI makes our AI better" is the most bullish sentence a frontier lab can possibly utter, and you should always squint a little when the party selling the thing also happens to be the party measuring the thing. The details we didn't get here, and the piece wouldn't fully load for me, are the ones that matter: accelerating research on what, exactly? Grunt-work coding tasks that a junior engineer would've done anyway? Or genuine research direction, the hard creative part? Those are wildly different claims, and "experiment velocity went up" can mean either one.
But here's why I'm not dismissing it either. It lines up with something we covered in this show a couple days back, the loops discussion, the idea that the frontier of AI work has shifted from prompting a model to designing systems where agents run in loops until they hit a verifiable goal. Coding has that verifiability built in. The code compiles or it doesn't. The test passes or it fails. So of all the domains where you'd expect agents to genuinely accelerate real work, AI research and software engineering are exactly the two where the finish line is cleanest. It's the one place where "put it in a loop and let it run" actually works, because the machine can check its own homework. So the claim is at least plausible in the one domain where you'd most expect it to be true.
And I'll pair this with the other thing OpenAI dropped the same day, an essay from Jakub Pachocki called An Alien Mind. The gist, and this is a short summary so I'm not going to over-read it, is Pachocki reflecting on increasingly capable AI and the challenge of keeping it aligned, calling for stronger safeguards and international coordination.
Now think about the timing of those two together. On the same day, the same company tells you two things. One: our AI is now good enough to accelerate the building of better AI. Two: this stuff is starting to feel like an alien mind and we need stronger safeguards and international coordination. Read those side by side and the message is, we're pressing the accelerator and asking somebody, anybody, to help us think about the brakes. I'm not going to tell you whether that's genuine concern or savvy positioning. Probably it's both, because in this industry it usually is. But if you're a founder building on OpenAI's models, the operational takeaway is simple: the pace underneath your foundation is going to keep increasing, and the ground rules for how these models get released are getting more political, not less. Plan for both.
Alright, let me clear the palette with a couple of quicker bites before we get to the meat of today's show, because there's a real builder story I want to spend time on.
Travis Kalanick's company, Atoms, might be getting into the robotaxi business. You remember Kalanick, the Uber founder. Earlier this summer his company Atoms announced a $1.7 billion funding round led by Andreessen Horowitz, and Kalanick was cagey about what he was actually trying to do with all that money. Now the Financial Times has more, reported through TechCrunch by Anthony Ha. Atoms is reportedly prepping a hiring spree and acquisitions that could make it a major player in autonomous vehicles. And here's the kicker, the FT says Atoms has talked to Uber, Kalanick's old company, about how Uber could use the startup's robotaxi tech. Uber has also invested $100 million in Atoms.
So the guy who got pushed out of Uber is now potentially selling robotaxi technology back to Uber. That's a hell of a full circle. Kalanick called the round, quote, unfinished business. And it fits, because Atoms already acquired Pronto, an autonomous mining startup run by Anthony Levandowski, Uber's former self-driving chief. Levandowski, if you don't remember, was convicted of stealing trade secrets, sentenced to prison, then pardoned by President Trump. So the band's getting back together, and it's the same band that gave us the messiest self-driving saga in Silicon Valley history. I'm watching this one purely for the drama, but there's a real business point underneath it: the robotaxi space is consolidating fast, and the money's flowing to founders with scar tissue, not just clean résumés.
And staying on the theme of executives making moves, there's an Apple story that's actually a strategy story in disguise. Phil Schiller, the longtime Apple executive who runs the App Store, is stepping down from that role. He'll stay on as an Apple Fellow. And per Bloomberg's Mark Gurman, reported through TechCrunch, the official reason is family and philanthropy. Fine. But there's a second factor, and this is the one that matters. New CEO John Ternus and services chief Eddy Cue want to improve App Store margins and bring in more recurring revenue. And Schiller, according to Gurman's telling, believed that squeezing more profit out of the App Store would only increase conflict with governments and developers. So rather than fight that battle, he stepped aside.
Read that carefully, because it's a signal. Apple's App Store has been the subject of lawsuits and regulatory heat for years, all over its cut of payments and its control over developers. And here you've got a veteran executive, the guy who ran the thing, essentially saying: I don't want my name on what comes next. If you build apps, if you sell through that store, that's your early warning. The direction of travel is toward more aggressive monetization, and the person who understood the political risk of that just quietly excused himself from the room. Watch your margins.
And one palate-cleanser from way outside the AI bubble, because I like to remind you there's a whole universe out there. Europe's BepiColombo mission is on final approach to Mercury. This is a resurfaced report from Ars Technica by Stephen Clark, describing a milestone from last fall, so I'm framing it as background, not breaking news. But it's a beautiful story. Eight years, six billion miles traveled, nine planetary flybys, a two-billion-dollar spacecraft that's actually three spacecraft stacked together, threading its way to the hardest-to-reach rocky planet in the solar system. It jettisoned its propulsion module 39 million miles from the sun, in temperatures and radiation that would cook anything, and everything worked. It's scheduled to slip into orbit around Mercury on November 21st. I bring it up because in a week full of lawsuits and margin squeezes and recursive self-improvement, it's worth remembering that somewhere, quietly, engineers spent eight years on a problem with no shortcut, no loop, no agent, just patience and math. Sometimes the barely possible is just people being very good at their jobs for a very long time.
Okay. Now let's dig into the story I think matters most for the builders and founders listening, and it's not a headline story. It's a quiet one, three little posts from a developer named Simon Willison, and taken together they say something about the labor economics of AI that nobody's putting on the front page.
Here's the deep dive. Over the last day, Simon Willison, who's one of the more grounded voices on how these tools actually get used day to day, put out a short thread that's really about who gets automated and when. And I want to walk through it because it captures a shift that's happening in real time, and it has direct implications for anyone building a company.
The first post is a small, practical one. He wrote, quote, This is why I always generate my pelican SVGs with my LLM CLI tool, it calls their API directly and keeps me 100% in control of the context visible to the model. Now, on the surface that's a nerdy little workflow note. He's got a command-line tool that talks straight to the model's API, and he likes it because he controls exactly what context the model sees. But there's a principle buried in there that every builder should internalize. Control of context is the whole game. When you go through somebody else's app, somebody else's chat interface, you're getting their idea of what the model should know, their system prompt, their guardrails, their invisible instructions. When you call the API directly, you own that. For a founder, that's the difference between building on a platform and building on a foundation. The closer you are to the raw model, the more of the product is actually yours. That's not a small thing. That's the difference between having a business and being a feature of somebody else's business.
But it's the next two posts where it gets sharp. He wrote, quote, Professional 3D modelers are just hitting that point now. That point. Meaning the point where the AI tools have gotten good enough that professional 3D modelers, skilled people, people who spent years learning a craft, are feeling the ground move under them. Just now. Not two years ago. Now.
And then the question he asks, which is the whole ballgame, quote, If the cost of producing great animated movies and TV shows drops, will we see more of them? Would that be enough to offset the change in labor demand? And will the market for 3D models outside of animation dramatically expand now that they're faster to build?
Now stop and appreciate what he's actually doing there, because it's the smartest way to think about this whole disruption, and it's the opposite of the panic you usually hear. He's not saying "3D modelers are finished." And he's not saying "everything's fine, don't worry." He's asking the two-sided question that actually determines the outcome. When the cost of making a thing collapses, two things happen at the same time, and they pull in opposite directions. On one side, you need fewer people to make each individual thing. That's the labor demand shrinking. That's the scary side. But on the other side, when something gets cheap, people make a lot more of it. Demand that was never economical before suddenly becomes economical. Markets that couldn't afford custom 3D models, small businesses, indie games, one-off projects, product visualizations, they can afford it now. That's the market expanding. And whether the net effect on jobs is good or bad depends entirely on which of those two forces is bigger.
This is the induced demand question, and it's the single most important framework for any founder trying to figure out where the opportunity is. Here's the history lesson. When spreadsheets came along, everybody thought accountants and bookkeepers were done. Software could do the arithmetic instantly. And in the short term, yeah, a certain kind of manual number-crunching job did shrink. But the number of people working with financial data exploded, because suddenly it was cheap to run the numbers, so everybody ran the numbers, on everything, all the time. The pie got so much bigger that even a smaller slice of the work per person added up to more total work. That's induced demand. Cheap doesn't mean less. Cheap very often means more.
So when Willison asks whether cheaper animation means more animation, whether the market for 3D models outside of animation expands now that they're faster to build, he's asking the question that separates the founders who are going to win from the ones who are going to get flattened. Because here's the thing. If you're a 3D modeler thinking of yourself as a person who moves polygons around, the tools are coming for you, and that's real, and I'm not going to sugarcoat it, that professional he's talking about is feeling genuine pain right now. But if you're a founder looking at that same collapse in cost, the question is: what becomes possible now that was impossible last year? Who needs 3D models who could never afford them before? What product can I build for a market that didn't exist because the input was too expensive?
That's where the companies get built. Not in defending the old expensive way of doing things, but in going and finding the demand that was sitting there, priced out, waiting for the cost to fall.
And here's the part that connects back to everything else we talked about today. Remember the OpenAI recursive self-improvement story? The reason that one matters is the same reason this one does. The cost of producing intelligence-adjacent work, code, research, models, art, is falling on a curve that keeps getting steeper. Every story on today's menu is a version of the same underlying event. The newspapers suing over training data are fighting over what happens when the cost of producing content collapses. The authors fighting over the settlement are fighting over how to split up the value of work that got fed into that cost collapse. The 3D modelers are living through it in real time. It's all one story. The cost of making things is falling, and the fight is over who captures the value that gets freed up when it does.
And I'll give you the DeLuca bottom line, the thing I'd tell you across the kitchen table. When a cost collapses, don't ask "who loses the old job." Ask "who gets to do the new thing they could never afford before." The losers are visible, they're loud, they're the professionals hitting that point now, and their pain is real and worth taking seriously. But the winners are invisible at first, because they're the markets that don't exist yet, the customers who were priced out, the products nobody bothered to build because the inputs cost too much. If you're a founder, your job is to see the invisible side before everybody else does. That's the whole game. That's been the whole game since the spreadsheet.
Now let me tie one more thread in before I let you go, because it's the counterweight to all this optimism. There's a real risk in the induced-demand story that the optimists gloss over. Induced demand doesn't happen instantly, and it doesn't happen evenly. When spreadsheets landed, there was a gap, a painful stretch, between the accountants losing the old work and the new financial-analysis economy showing up to absorb them. Some people never made that jump. The transition is real even when the long-run outcome is fine. So when Willison says professional 3D modelers are hitting that point now, the humane response isn't "don't worry, the market will expand." It's "the market probably will expand, and also some specific people are going to have a rough couple of years in the gap, and both of those things are true at once." Anybody selling you pure optimism or pure doom on this is selling you something. The truth lives in the messy middle, in the timing, in who bridges the gap and who falls into it.
Alright, that's the plate for today. Newspapers biting the hand that fed them, publishers pocketing money that isn't theirs, OpenAI telling us its machines are learning to build better machines while asking someone to please think about the brakes, and one clear-eyed developer asking the only question that actually matters when the cost of everything falls: what becomes possible now?
Keep your eyes on the invisible side of the ledger, folks. That's where the barely possible turns into the actually built.
This has been Barely Possible. I'm Tony DeLuca. Take care of each other out there, and I'll see you tomorrow.