Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and we've got a fresh menu of tech morsels sizzling on the griddle today. Grab your coffee, pull up a stool, and let's have at it.
I want to start with a story that on the surface looks like a video game curiosity, but if you're building anything, running anything, storing anybody's data anywhere, it should make the hair on your neck stand up a little. Over the weekend, somebody dumped twelve terabytes of Valve's old Steam content onto a BitTorrent tracker. Twelve terabytes. They're calling it the "teraleak," and it covers seemingly every game that was on Steam between 2003 and 2013. Cut content from Portal 2, dialogue from Cave Johnson's consciousness stuck in a cube saying "my life is torture, please kill me" — which, honestly, is how I feel scrolling through my inbox some mornings — and even fragments of that Half-Life 2: Episode 3 game they never shipped. The gaming crowd is losing their minds over the lost history, and that's the fun part.
But here's where I want you to lean in, because this is the part that actually matters for you as a builder. Kyle Orland's piece over at Ars Technica had a quote from a longtime Valve watcher named Gabe Follower, who said he "verified that everything in Steam2 Teraleak was obtained via a publicly accessible API endpoint. It's Valve's fault." And a Spanish-language analyst who goes by Scolcer put it even plainer, saying it was, and I'm reading this from the piece, "obtained from a site that was 100% accessible to the public. It was there for everyone to download. No passwords. Nothing. Hidden in plain sight, but with no protection whatsoever."
Now think about what that means. This wasn't some elite crew of hackers cracking a vault. This was a door somebody left unlocked for over a decade, and the internet finally noticed. And the reason it cuts off at 2013 is because that's when Valve moved off its old "Steam2" server system to the modern one. So this is old, forgotten infrastructure. The stuff nobody's looked at in years. The stuff nobody owns anymore because the person who built it left the company three reorgs ago.
Every one of you has a Steam2 somewhere. Some old S3 bucket, some legacy endpoint from a product you sunset, some internal API you spun up during a hackathon that never got locked down because it "wasn't important." And here's the kicker Ars pointed out — this isn't just Valve's problem. That archive contains a mountain of third-party content, early builds from other publishers who trusted Valve to hold their work. One Valve watcher, Tyler McVicker, wrote on social media, "Don't touch it, it's HOT. Take it from someone who got in trouble for beta stuff in the past." So now Valve's servers are what the piece called a single point of failure for a huge chunk of PC gaming history, and everybody who uploaded to them is exposed too.
The lesson for the builder isn't glamorous. It's boring. It's go audit your dead infrastructure. The breach that gets you probably won't be the shiny new system you're watching like a hawk. It'll be the thing you forgot you built. Okay, sermon over. Let it marinate.
Now let's shift from the door you left unlocked to the door somebody's slamming shut on purpose. And this one's a real business story for anybody building on top of these AI labs.
OpenAI put out a short note — a recent post, dated the 28th — titled, and I'll say it plain so nobody mistakes it for a command, an item called "Our decision on Cursor following its acquisition by SpaceX." The whole thing is basically one sentence: they're winding down the contract that provided OpenAI models to Cursor, following Cursor's acquisition by SpaceX. That's it. That's the announcement.
Now, I want to be careful here, because the source is thin — it's a headline and a line of text, and I'm not going to invent drama that isn't on the page. But even the bare fact is worth sitting with if you're a founder. Because what it tells you, cold and clear, is this: when you build your company on somebody else's model API, that supplier relationship is not neutral. It's political. It's competitive. Cursor, the coding tool, got bought by SpaceX — which, if you've been listening the last couple weeks, is now knee-deep in its own AI ambitions. And the second that happened, OpenAI looked at Cursor not as a customer but as an arm of a rival, and pulled the plug on the models.
We talked a while back, as we covered when Stripe grabbed OpenRouter, about how much leverage sits with whoever controls model access. This is that same story from the other end of the telescope. You are a distribution channel for these labs right until the day you become a threat, and then you're a liability they cut loose. So if you're building on a single model provider, ask yourself the ugly question: what happens to my product the day my supplier decides my new owner, or my new partner, or my new pivot makes me the enemy? Have a backup harness. Have a fallback model. Don't let one API be the load-bearing wall of your entire company.
Alright, from the labs cutting cords, let's go out to the physical world, because I've got two stories about deploying AI into the real, messy, dusty, human world, and they connect in a way I want to draw out.
First one's Caterpillar. Yeah, the yellow bulldozer people. TechCrunch's Kate Park caught up with their CTO, Jaime Mineart, at a conference, and the framing here is genuinely useful for founders because Caterpillar has been doing autonomy in the physical world for decades — automated haul trucks, drilling rigs, all of it, mostly in mining, where the conditions are so nasty and the labor's so short that robots just make sense.
Now they've got this thing called the Cat AI Assistant that lets a field technician standing next to a broken machine use voice commands to pull up repair procedures, troubleshoot, and figure out what parts he needs before he even starts. And it runs on Caterpillar's own proprietary data — Mineart says they've got about 1.6 million connected machines out in the world and more than 16 petabytes of structured data. That's the moat, folks. Not the model. The data nobody else has.
But here's the line from Mineart that I underlined twice. She said, quote, "The hard part about autonomy and about physical AI is incorporating that technology into the customer jobsite and into the workflows." And then this: deploying an autonomous machine, she says, is not the same as transforming a site to use AI. You gotta rethink how the people work alongside the thing.
And that's where they're putting real money — a hundred million dollars over five years to retrain their 118,000 employees for AI, autonomy, and robotics. Because as the machines get more autonomous, an operator who used to run one bulldozer might end up overseeing five of them from a command center. That's not a firing. That's a whole different job. And somebody's gotta teach it.
Now hold that thought — the human cost of automation — because there's a second mobility story that shows you the flip side, the part nobody puts in the press release.
Kirsten Korosec's TechCrunch Mobility newsletter surfaced reporting from her colleague Sean O'Kane, and I want to frame this carefully — this is a report drawing on data from 2024 and 2025, so it's not a today event, it's a recent look back. O'Kane dug into filings submitted to OSHA and found that the human test drivers who ride in Waymo and Zoox robotaxis — the safety operators, the folks sitting there watching the car drive itself — sustained more than two dozen injuries. From what? From the cars braking hard. Sudden, violent stops. Whiplash. Some of these workers got sidelined for months.
Think about the irony there. The whole pitch for robotaxis is that they'll make the roads safer, reduce crashes, save lives. And there's some evidence they do. But the person whose job is to babysit the robot while it learns to drive gently is getting their neck snapped by the robot's jerky braking. And the piece makes a sharp point — this is only Waymo and Zoox because those two happen to fall under OSHA's reporting rules. Other AV outfits are probably injuring their test drivers too, but they may be exempt from having to report it, so we just don't see it.
So here's the thread between Caterpillar and the robotaxis. Caterpillar's approach says: automation reshapes the human job, and if you're serious, you fund the transition — the retraining, the new command-center role. The robotaxi story is what happens when you treat the human as a disposable buffer while the machine gets good — a body in the seat absorbing the mistakes. Both are true right now. If you're a founder building anything autonomous, the question isn't just "does the AI work." It's "what happens to the person standing next to it." Caterpillar's writing that check. Somebody in the robotaxi world apparently wasn't. Watch which kind of company you're building.
Alright, let's stay in Elon's neighborhood but pivot to power, because there's a fascinating one here. And I want to flag up top — TechCrunch's Connie Loizos wrote this, and the underlying announcement from Musk came on a Saturday, so this is very recent, but the deeper backstory on the pollution fights goes back years, so I'll keep the timing honest as I go.
Musk confirmed what a secret foundry SpaceX has been building down in Bastrop, Texas is actually for. It's for casting turbine blades and vanes. Now stay with me, because this is one of those bottleneck stories that tells you where the real constraint in the AI boom lives. Everybody knows about the GPU shortage — Nvidia's newest chips still have months of lead time. But there's a second wall the whole industry's slamming into, and it's the power grid itself. The International Energy Agency figures data center electricity use roughly doubles by 2030, and the big turbine maker GE Vernova says it's basically sold out of production capacity through 2030. Sold out. Five years.
So why sold out? Loizos's piece explains it, and it's wild. The blades inside the hottest part of a gas turbine run at three thousand to thirty-six hundred degrees Fahrenheit — which is about eight hundred degrees hotter than the melting point of the metal they're made of. The only reason they don't melt is these microscopic internal cooling channels and special coatings, and each blade has to be cast as a single, unbroken crystal, grown slowly in a vacuum furnace, with no seams that could crack under stress. Only four companies on earth have mastered doing that at industrial scale, and every one of them is tapped out.
So Musk's move is: I'll just do the casting myself. He claims doing it in-house at SpaceX could accelerate gas turbines coming online by up to eighteen months. And if he pulls it off — big if — he owns a manufacturing capability that every other AI infrastructure builder on the planet has to rent from a tiny oligopoly. That's a genuine, hard-to-copy edge. You can't fundraise your way past a vacuum furnace and metallurgy.
But — and Loizos does not let this slide — turbines in the ground mean pollution coming out. Down in Memphis, where Musk's operation has been running gas turbines to power its data centers since 2024, the NAACP has repeatedly accused the company of running them without the permits or pollution controls federal law requires. We're talking smog-forming compounds, formaldehyde, stuff linked to asthma and cancer, in neighborhoods that already eat a lot of industrial pollution. And it's not just Memphis — out in Virginia's Data Center Alley, a study using the EPA's own health model estimated one facility's eight turbines could cause somewhere between three and a half and six and a half extra premature deaths a year, and tens of millions in health damages.
So here's the builder's takeaway, and it's a hard one. The AI arms race just became a heavy-industry arms race. The moat is metallurgy now. And the externality — the cost that doesn't show up on the balance sheet — is landing on somebody's lungs a mile downwind of the data center. If you're building in this space, that bill is coming due, in lawsuits and regulation, whether you priced it in or not.
Now let's shift from the infrastructure story to something a little quieter but genuinely important for anybody in a lab coat or building for people in lab coats.
Anthropic put out a note — dated the 27th, and the underlying program actually kicked off back in June, so I'll frame it as the recent expansion of an older effort — expanding their support for scientists. And the headline number is real: they're opening ten thousand seats for scientists around the world to get Claude subscriptions, free or heavily discounted, for a year. Standard seats free, premium seats with five times the usage for fifteen bucks a month. And they say they plan to go well beyond that initial ten thousand.
On top of that, they're broadening their "AI for Science" program — free credits, up to fifty thousand dollars per project — and expanding it beyond just biology into more compute-heavy fields. They name-check progress on the Riemann zeta function and Claude's work on protein design as the kind of thing they want to fuel.
Now, why should a founder who isn't a scientist care about this? Two reasons. One, this is a land-grab. Anthropic is planting Claude in ten thousand research labs during the formative years of these scientists' careers, the same way the old software companies used to give their tools to universities for free — you get 'em hooked young, they build their whole workflow on you, and they never leave. It's distribution strategy dressed up as generosity, and it's smart.
Two, and this is the part I respect — buried in there is a real note about dual-use risk. They say for biology and chemistry, researchers are still capped to certain model classes, and their more capable models will keep blocking professional biology and drug-development queries because of the potential for misuse. They're building an actual gated access program with the U.S. government for legit life-sciences pros. So even while they're throwing the doors open, they're keeping the dangerous rooms locked. That's the tension every AI company building for high-stakes fields is gonna have to live in — open enough to be useful, closed enough not to hand somebody a recipe for something awful. Watch how they thread it, because you might have to thread the same needle someday.
Alright, let me do a quick lightning round of a couple smaller ones before I let you go, because there were some tidbits worth your time.
Retro, that friend-focused photo-sharing app built by a couple of former Instagram engineers, Nathan Sharp and Ryan Olson — Sarah Perez at TechCrunch flagged an SEC filing showing they pulled in north of twenty-one million in a Series A. Now, framing this honest: the filing surfaced recently, but the raise actually closed back in December, so this is an older deal just now catching the light. The interesting bit is the thesis. Sharp told TechCrunch, quote, "Something that has to be true and will be true is that people will still want to see more of their friends." The whole pitch is that as the feeds fill up with algorithmic slop and AI-generated junk, there's enduring value in a place that just shows you your actual friends. No advertising, subscription-based. Around seven million downloads, and in-app spending up over 460% in six months. For founders, that's a signal worth chewing on — the counter-positioning play. Everybody zigs toward AI content, and there might be real money in zagging toward the human stuff.
And a little contrast to that: Hollywood celebs are piling into microdrama apps — those minute-long soap-opera-on-your-phone shows. This one's an older trend piece, the underlying moves go back to the spring, so I'll frame it as background color. But the number's eye-popping: U.S. revenue for microdramas is expected to hit a billion and a half this year. James Franco, Issa Rae, Taye Diggs, even Kim Kardashian putting money in. Issa Rae's series with TikTok's app pulled 150 million views. So on one side you've got Retro betting people want less algorithmic content, and on the other you've got Hollywood betting people want maximum bite-sized dopamine. Both can be true. Different customers, different needs. That's a whole market-map lesson in two stories.
Last one, and it's a palate cleanser. The President dialed into a NASA news conference. Ars Technica's Eric Berger reported this — and I'll be straight, this is a weekend event resurfacing today, not brand-new. NASA Administrator Jared Isaacman literally held up his phone at a post-launch press conference for the Nancy Grace Roman Space Telescope, and Trump spoke through it to the assembled reporters, saying, quote, "I'm supplying you all that money." Berger's dry note is the good part — presidents call astronauts orbiting the Moon; they don't dial into routine post-launch pressers. And there's a delicious wrinkle Berger points out: the same administration's first budget tried to cut NASA science by fifty percent and even sought to cancel the very telescope Trump was now taking a victory lap on. So the founder's lesson, if you want one, is eternal: the guy who tried to kill your project will happily take credit for it the second it succeeds. Get it in writing, kids.
Alright, that's the plate for today. Go check your dead infrastructure, don't build your whole company on one API's good mood, and remember the person standing next to the robot is a person. I'm Tony DeLuca, and this has been Barely Possible. Be good to each other out there, and I'll catch you next time.