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 back with a plate full of stuff that actually matters if you're trying to build something. Grab your coffee, pull up a chair, let's have at it.
Here's the thing that jumped out at me today, and I want to start here because it's the one that most people are gonna scroll past thinking it's a lawyer story. It's not. It's a founder story. There's a piece from TechCrunch by Amanda Silberling, called Is it legal to train AI models on copyrighted books? It's complicated. And I know, I know, the second somebody says "it's complicated" you want to change the channel. But stick with me, because buried in here is one of the clearest maps I've seen of the actual rules starting to form around what you can and can't build on top of other people's work. And if you're a founder, those rules are the ground under your feet whether you like it or not.
Let me set it up. You already know the big AI models — the ones behind ChatGPT, Gemini, Claude — got trained on basically everything. Hundreds of millions of books, articles, papers, whatever's on the internet. Most authors never knew, never said yes. Feels illegal, right? Feels like theft. And that's where most people stop. But the courts have not stopped there, and the distinctions they're drawing are the whole ballgame.
So here's the anchor case. Last year, Judge William Alsup ordered Anthropic to pay a $1.5 billion copyright settlement to a group of writers. On the surface, huge win for the authors. Money's on the board. But read what the judge actually said. He ruled that Anthropic's AI training itself was lawful. Let me say that again, because everybody misreads it. The training was fine. What Anthropic got hit for was pirating the books — pulling them off illegal shadow libraries. The judge wrote, and I'm quoting the ruling here, "Like any reader aspiring to be a writer, Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different." He compared a model ingesting trillions of words to a writer studying literature.
Now, Cathy Gellis — she's an attorney who works in IP and copyright and technology — she told TechCrunch that ruling is actually good news for the AI companies. And her reasoning is the part builders should tattoo on their arm. She said, "Copyright law hinges on copying, but it doesn't hinge on using the work or experiencing the work, consuming the work, reading the work." That's the crack in the door. The law is about the copy, not the learning. And she points out the fine is basically a rounding error — what's a billion and a half to a company projecting two hundred billion in annual revenue by 2028? It stings, it doesn't stop.
But here's where it gets sharp for you, and this is the sentence I want every founder to hear. Another attorney, Jason Henderson, who runs the IP and Media Practice at JWL International, laid out the actual test the courts keep landing on. He said the courts are all over the place, but "what's tending to win is if what you're doing is you're training on somebody's property because your purpose is to directly compete, then the courts will frown on it. If what you're doing is not going to compete, then the courts are tending to find ways that it will be okay."
You hear that? Compete versus don't compete. That's the line. And there's a case that proves it. Thomson Reuters — the big media and legal information company — sued a research firm called Ross Intelligence for copying Reuters content to build a competing AI legal platform. Judge Stephanos Bibas ruled against Ross. He wrote that Ross's use "is not transformative because it does not have a 'further purpose or different character' than Thomson Reuters's." Translation: you took my stuff to build the exact thing I sell. That's not learning, that's cloning my business. Not fair use.
So put those two rulings side by side and you get the shape of the thing. Train a general model that learns from a book and goes off to write something new — the courts are leaning toward okay. Train specifically on a competitor's proprietary data to replace that competitor in their own market — the courts are leaning toward no. If you're building a vertical AI product right now, and a lot of you are, that distinction is not academic. If your training data is the crown jewels of the incumbent you're trying to eat, you are walking into the Ross Intelligence buzzsaw. If you're building something genuinely new on broad data, you've got the Alsup ruling as cover, at least for now.
And I want to flag the "for now," because Gellis is honest about it. Copyright law hasn't been updated since 1976. Nineteen seventy-six. That's judges reading fifty-year-old guidelines and trying to figure out what they mean for a technology nobody could've dreamed of. And she says these early rulings are "the initial opening volleys" — influential, but they could get undone if other courts decide differently. Her closing line is the one I'd write on the whiteboard: "It would be kind of foolish for the AI companies to ignore them." Right. And it'd be foolish for you to ignore them either.
There's a second layer here that's a little sci-fi but worth thirty seconds. Separate from training, there's the question of who owns AI-generated output. A case called Thaler v. Perlmutter held that if a work is a hundred percent AI-generated, it's not copyrightable. Nobody owns it. And Gellis makes a great point about how slippery that gets. She says if you write your novel in Word and run spell check, we're all comfortable saying Word doesn't own your novel. But AI is forcing us to draw a line we've been ignoring — how much AI assistance is too much before your work stops being yours? Nobody knows. If your product spits out content your customers want to own and protect, that's an open question sitting inside your business model. Watch it.
Alright, that copyright fight is really a fight about where a computer's decision-making crosses a legal line. And that same tension — a computer making a call that has real consequences for a real person — is exactly what's playing out over in Europe right now, so let's shift over there.
The Dutch Data Protection Authority just fined Uber 825 million euros. That's about 966 million dollars — call it a billion. It's the second-largest penalty ever handed out under Europe's GDPR. This one's current, reported by Anthony Ha at TechCrunch, off Reuters. The charge: Uber deactivated driver accounts through an automated process without enough warning or human oversight. The deputy chair of the regulator, Monique Verdier, said Uber "committed serious infringements," and she said, quote, "A computer should not make decisions on its own that have such major consequences."
Uber, naturally, is fighting it. They say most suspensions are brief, no permanent deactivations happen without human review, drivers can appeal. The regulator says some drivers did get permanently cut off with no human in the loop, which Uber disputes. Uber's appealing.
Now here's why I like this story for builders, and it's not the Uber-versus-Europe soap opera. There's a genuinely good argument on both sides, and you're going to have to pick a lane on it eventually if you build anything with automation in the enterprise. John Gruber of Daring Fireball — he pushed back on the whole thing. His worry: this fine could make it "unlawful in the EU for Uber to monitor its drivers for pulling scams against customers, or just never picking riders up, leaving them stranded." And he took a shot at Verdier's framing. He said saying "a computer" made the decision is like saying when a company fires a habitually late employee, "the time clock" made the decision. The managers set the policy. The device just measures compliance. Fair point, right? The tool didn't decide anything. A human wrote the rule.
But then Paul-Olivier Dehaye — he runs a Swiss digital-rights nonprofit that helped the drivers bring the case — he fires back, and I think his answer is the one founders need to sit with. He said Uber is free to use humans to punish drivers who scam. But then, quote, Uber "has to take responsibility for this decision making — like being an employer, not being a marketplace." That's the whole thing right there. That's the trap. Uber wants to be a marketplace when it comes to responsibility and an employer when it comes to control. You can't have both. If your automated system is exercising employer-level power over someone's livelihood, the law is starting to say you carry employer-level accountability for it.
And this is the third fine the same group of drivers has pulled off Uber. There was a 290 million euro fine before, a 10 million euro fine before that, all from complaints by the same crew. Dehaye's now starting a company called StartClaims to fund class actions — first against Uber, then gig-economy cases broadly, then adtech. So this isn't a one-off. This is somebody building a litigation machine aimed at exactly the kind of automated decision-making that half the products in your pipeline probably contain. If you're deploying agents that suspend accounts, deny claims, flag fraud, cut people off — in Europe, "the algorithm did it" is not a defense. It's an admission. Build the human review in on purpose, not as an afterthought, and document it.
Now, staying on the theme of a surveillance tool colliding with public backlash — let's talk about Flock Safety, because there's a fresh wrinkle here and I want to be careful with it. We covered Flock's troubles about a week back, when Wisconsin towns started canceling contracts over privacy and the network-effect problem. So this is a callback, not a rerun — but there's a new development worth the time.
Flock's CEO, Garrett Langley, has gone on a media tour, and the framing he's pushing is "compromise." On Fox News he said, quote, "When people talk about just one of these, privacy or safety, they're prioritizing the wrong thing, and what we have to prioritize as a country is compromise." Okay. But here's what's underneath the PR. The Washington Post identified 46 cases where police officers were accused of using Flock's tech for unauthorized stuff — stalking wives, girlfriends, exes. Forty-six. After hearing an interview with one of the victims, Langley said on CBS, "I apologize. It kills me that she went through that." And then, same breath, "I don't think that Flock created police abuse. I think we're the first company to ever shine a light on it and build the tools to find it."
I'll let you sit with that one. We built the panopticon, but hey, we also built the flashlight.
Here's the part that matters for anybody building in this space, though, and it's the political turn. Flock has become radioactive across the whole spectrum. On the left, a Democratic Senate nominee in Michigan is hammering his opponent over Flock cameras "watching your every move." Bernie Sanders posted, all caps, "STOP AI MASS SURVEILLANCE. STOP FLOCK." And on the right — this is the interesting part — three House Republicans introduced a bill to ban the federal government from buying automated surveillance systems using facial recognition, biometric IDs, or license plate readers, and the bill literally names "a Flock Safety camera." When you get named by name in federal legislation, that's not a PR problem anymore, that's an existential one.
Flock's made some changes — cut default data retention from 30 days to seven, added a case-code requirement before accessing data. But both can be overridden with a setting called "Evidence Mode." The ACLU's response was basically, nice, but you didn't go to our recommended 48 hours, and whether this is real or just PR depends entirely on how Evidence Mode actually works. The lesson for founders: if your defaults can be overridden by the same people you're supposedly protecting against, you don't have a safeguard. You have a marketing line. And regulators, left and right, are getting real good at spotting the difference.
Alright, let me pivot to something with a completely different flavor, because not everything today is about a computer making a decision that gets somebody in trouble. Some of it's just a genuinely wild business idea. Let me tell you about Michael Polansky and a company called Outer Biosciences. This one ran as a recent profile by Connie Loizos at TechCrunch, and the underlying reporting is from a few weeks back, so I'm not calling it breaking — but the idea is too good not to unpack.
Polansky is training an AI model on human skin that's still alive. Literally. Living human tissue, kept alive outside the body for up to a month. And the wrapper on this — Polansky's the partner of Lady Gaga, he came up through Sean Parker's world and Founders Fund — that's the headline candy. Ignore it. The business underneath is the thing.
Here's the actual model, and this is where I want you builders leaning in. Outer Biosciences sources human skin that would otherwise get thrown out after surgery — mostly plastic surgery — through vetted biobanks with donor consent, all de-identified upstream. Then they've got a proprietary system that feeds the tissue nutrients and clears out waste, keeping it alive way past the normal industry standard. And that standard, by the way, is a matter of days. Outer keeps it alive for about 30 days. Why does that matter? Because a few days is enough to test for acute toxicity, but it's not enough to watch the slow stuff — collagen remodeling, pigmentation change, barrier repair. That takes weeks. They can sunburn a piece of living skin with UVB and then watch the whole inflammation and recovery process unfold over time.
But the tissue's not the product. Here's the product. An AI model predicts which untested chemicals will have a beneficial effect on some skin function. Those chemicals get run through the living tissue. The results — right or wrong — get fed back into the model, and the next round of guesses gets better. That loop is what they're selling. Not a cream. Not a serum. A faster way to find one.
And the numbers on that loop are the whole pitch. Early on, brute-force literature mining got them a couple of leads in eighteen months. With AI layered into the loop, they're now generating a new candidate roughly every six weeks. Six leads active, several dozen additional hits logged. And the market they're hunting is absurdly small — Polansky says across all FDA over-the-counter skin drug categories, there's only about 120 to 130 approved active ingredients. Add cosmetic ingredients with real research behind them, maybe 200 total. Two hundred. That's the entire known universe. If your loop can expand that catalog, every new ingredient is worth a fortune.
Now here's the part I really want founders to hear, because it's a defensibility lesson dressed up as a skincare story. Polansky said, and I'm quoting, "Unlike AI companies training on scraping the internet, there is no biology internet to scrape." Think about that. Every LLM company is fighting over the same pile of scraped text. Outer's data doesn't exist anywhere else — they generate it themselves, in-house, on their own living tissue. It's slower and more expensive to build than a software startup. But it's a moat you can't just copy by scraping harder. And get this — all their AI runs on-premise, not in the cloud, because, Polansky says, "we don't want the data in the cloud." When your data is the whole company, you don't hand it to somebody else's servers.
The takeaway for you, whatever you're building: everybody's got the same models now. Same weights, same APIs. The edge isn't the model. The edge is having data nobody else has and can't easily get. Outer built a physical machine to generate proprietary data. That's harder than a wrapper on GPT, and that's exactly why it might actually last.
Now let's shift from a company that owns its data to a mystery about who owns a model at all. There's a fun little one out of TechCrunch — a stealth model called Ox Alpha showed up on OpenRouter on Thursday, described as a reasoning model built for coding, sustained agentic work, and production workloads. Free. And the whole internet is losing its mind trying to figure out who made it. The listing just says it's from "a third-party provider who has chosen to remain anonymous during this preview." Stripe's Patrick Collison — and remember, Stripe is acquiring OpenRouter, which we covered when that deal broke — he called it "very impressive."
The speculation, predictably, has swung toward China. One AI analyst said the early guessing pointed at the GLM models from the Chinese company Z.ai, but then, quote, "this morning people seem less sure of anything." Wccftech floated that it could be an unreleased version of Microsoft's MAI. Reddit's got one post insisting it "can't be the Chinese" and another with "high confidence" that it absolutely is. So — nobody knows.
Why do I even bother mentioning it? Because the fact that a top-tier coding model can drop anonymously and have serious people impressed before anyone knows the source — that tells you something real about where we are. The model itself is becoming a commodity. Provenance is a guessing game. If you're building on top of these things, your loyalty to any one lab is worth about as much as the next benchmark. Route to what works. Watch OpenRouter — that's where this stuff surfaces first now, which is exactly why Stripe paid up for it.
Okay, let me get out of pure AI for a minute and give you a couple from the wider world, because builders don't live in a silo.
Quick one from India that's a masterclass in distribution. Walmart's Flipkart — this is an older data point resurfaced, the service launched back in 2024, so I'm framing it as where things stand now, not breaking news. Flipkart Minutes, their quick-commerce play, groceries in minutes, is now doing 1.1 to 1.2 million orders a day. In November it was around 390,000. That's roughly triple in well under a year. It's closing in on Swiggy's Instamart at about 1.4 million a day. The market leaders, Blinkit and Zepto, are way ahead at three-and-a-half and two-and-a-half million. But here's the founder lesson from an adviser named Satish Meena: Flipkart came late, but it had something the pioneers had to buy — an enormous existing customer base it already spent years and billions acquiring. He said, "Once you open 1,000 dark stores and are doing a million orders per day, it's serious enough." Distribution beats first-mover. Again. If you've got the customers, you can arrive late and still catch the leaders. And the kicker line from Meena on why everybody's forced to play: "Can you go back to scheduled delivery now in grocery? No. You will not go back." Once you retrain a customer's expectations, you can't un-ring that bell. That's a moat and a treadmill at the same time.
And one from space, because it's a nice reminder that even superpowers get beaten by the weather. China delayed its Chang'e 7 Moon mission — this was reported by Eric Berger at Ars Technica, and the underlying delay traces to conditions earlier in the year, so I'm calling it what it is, a resurfaced report on an ongoing situation. China put out this beautifully bureaucratic statement about "ensuring absolute success" and how the mission "does not meet the conditions for launch." Translation, most likely: there's a typhoon named Narra parked near the launch site. The next good window might not open until February or March. And here's the geopolitics for you — Chang'e 7 is aimed at the lunar South Pole, near Shackleton Crater, where there might be water ice in the permanently shadowed cold traps. That's the most valuable real estate on the Moon, and it's a tight race between China and the US to get there first. The delay actually cracks open a window for NASA to arrive first via Blue Origin's Blue Moon lander — if they can get New Glenn flying again after losing a rocket and a launch pad back in the spring. Nobody's ever softly landed at a lunar pole. Whoever does it first plants a very expensive flag.
Alright, let me bring it home. If there's a thread running through today, it's not "regulation can't keep up" — I promised you I wouldn't hand you that stale line. It's sharper than that. It's about where accountability lands when a machine makes the call. The copyright courts are drawing a line around whether your model competes with what it learned from. Europe just told Uber that "the algorithm did it" makes you the employer, not the marketplace. Flock's finding out that an override switch isn't a safeguard. Different stories, same question — when the software decides, who's holding the bag? And the answer, more and more, is: you are. The builder. So build like you know that.
The bright spots are the ones that sidestep the whole fight — Outer Biosciences generating data that doesn't exist anywhere else, Flipkart winning on distribution it already owned. The durable edges aren't the model. Everybody's got the model. It's the proprietary data, the customer base, the thing that's hard to copy.
That's the menu for today. Chew on it. This has been Barely Possible — I'm Tony DeLuca, thanks for spending a little of your day with me. Go build something you'd be willing to put your name on. Catch you next time.