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, pull up a chair, your boy Tony DeLuca is back behind the mic, and we've got a fresh tray of stories cooling on the windowsill. Welcome to Barely Possible. Grab your coffee, tuck in the napkin, because today we're starting with a story that sounds like a gag but is actually a window into how the whole intellectual-property racket is being renegotiated right now.
So here's the situation. A major game-industry trade group, the Entertainment Software Association, that's the outfit that represents Nintendo, Microsoft, PlayStation, Amazon, Epic, EA, Ubisoft, the whole lineup of heavyweight publishers, they went to the US Trade Representative last week. And they asked the government to put a company called Anbernic on what's called the Notorious Markets List. That's the government's naughty list for outfits that, quote, engage in or facilitate substantial copyright piracy or trademark counterfeiting. This is reported by Kyle Orland over at Ars Technica, published October 13th, so this is current, this is fresh.
Now who's Anbernic? If you don't follow the handheld gaming world, Anbernic is one of these Chinese hardware makers that's carved out a nice little niche selling handheld Android gaming devices built for retro emulation. Little pocket machines that let you play old games. And here's the thing that makes this interesting, and here's the word the ESA leans on. They call these things "superfakes." A superfake, as they define it, is a high-quality counterfeit good sold on a commercial scale. Polished, professional, looks legit, sitting right there on Amazon next to everything else.
And the ESA's argument is that these Anbernic devices are routinely offered or shipped with memory cards preloaded with hundreds or thousands of unauthorized copies of copyrighted games. Play 'em the second you turn it on.
But here's where I want you to slow down and read the fine print, because this is the part that matters. Anbernic itself, when it sells directly from its own website, does not appear to load up these devices with pirated ROMs. The people doing that, the ones slapping an SD card stuffed with a thousand stolen games into the box, those are third-party sellers. Amazon resellers, eBay guys, TikTok advertisers. Orland points this out plainly in the piece: Anbernic just makes the hardware. The piracy is happening in the distribution channel, on the big centralized e-commerce platforms, which, I'll note, mostly look the other way because it's good for sales.
So what you've got here is a trade group going after the box-maker instead of the marketplaces where the actual infringement happens. And Orland's read on that, which I think is dead-on, is that it's a strange choice. Amazon and eBay could clamp down on the wholesale advertising and sale of these ROM-loaded devices a whole lot more effectively than you could ever hope to squeeze a hardware manufacturer in Shenzhen.
Now why do I lead with this? Because for you, the builder, the founder, the person shipping product, there's a real lesson buried in here about liability and where it lands. For years there was this polite fiction in the emulation world. You sold a perfectly legal emulator, you disavowed any knowledge of the pirated stuff, and if the end user went and found illegal ROMs on their own, well, not your problem, pal. That fiction is breaking down. When the pirated content becomes a core selling point, "includes thousands of games!", the legal cover evaporates. And the enforcers go hunting for the biggest, most reachable target, whether or not it's the most responsible party.
And here's the kicker on consequences. Being on the Notorious Markets List doesn't itself carry direct legal penalties. But it can presage a Section 301 investigation, which can lead to import restrictions. So a listing is a warning shot. It's the government saying, we see you, and we're clearing our throat. If you're building hardware, or a platform, or anything where other people can layer illegal content on top of what you shipped, this is the shape of the fight coming for you. Don't assume the guy actually doing the infringing is the one who gets the subpoena. Sometimes it's just whoever's easiest to name.
Alright, let's move from pirated Game Boy games to something considerably louder. Let's talk about shooting cargo into space out of a gun.
I'm not kidding. There's a startup called Longshot Space Technology, based in Alameda, California, and their whole pitch, covered by Tim Fernholz at TechCrunch on October 13th, is an idea that dates all the way back to Jules Verne. Instead of a rocket, you use an enormous gun to fire a projectile into orbit. The co-founders, CEO Mike Grace and CTO Nato Saichek, they've raised about $20 million in venture money to build this thing.
Now here's the engineering, and I'll keep it at the diner-table level. Rockets are a nightmare. Our old friend physics forces them to be built with almost no margin for error, and even then only a sliver of the weight makes it to orbit. Longshot's move is to take all the work the first stage of a rocket does, and collapse it into a reusable accelerator sitting on the ground. Picture a big steel pipe laid out horizontally with tanks of high-pressure gas bolted along its length. The projectile goes in, and as it flies down the barrel, those tanks fire off at precisely timed moments to shove it faster and faster. They call it a multi-injection gun, a design that actually goes back to World War Two and some 1960s research programs.
And they're not drawing this on a napkin. Right now they're firing a 70-foot gun with a six-inch bore near Barstow, California, hitting near-hypersonic speeds. Next up, a 140-foot gun. And eventually, they want to build the largest gun on the planet, a 328-foot barrel they'd extend to about 3,000 feet, capable of launching 220-pound payloads at Mach 8.
I love this quote from Grace. He says, "you can build something that has a lot of novelty, or you can do something really hard. You don't get to do both." And he points to SpaceX as the model, a company that took de-risked NASA rocket research and scaled it aggressively. Longshot wants to do the same thing with gun technology that's been sitting around for eighty years.
Now, the business angle, because that's what we care about here. The dream is $50 per kilogram to orbit. Fifty bucks a kilo. For context, the much-hyped orbital data center crowd needs prices to fall to around $300 per kilogram just to get off the ground. So if Longshot hits anywhere near its number, there's a lot of room underneath to undercut everybody.
But, and here's my skepticism hat going on, there are real catches. To actually reach orbital velocity, you'd need a gun ten kilometers long and nearly ten feet in diameter. That's six miles of steel. And CTO Saichek had the honest line of the whole piece: "The hard part is not building the thing, it's finding a place where they will let you build the thing." Which, if you've ever tried to get a permit for anything in this country, lands pretty hard.
So how do they survive long enough to get there? Revenue in the meantime. The US military has already awarded them contracts for hypersonic testing. Grace's pitch to the Pentagon is blunt: if you want to practice shooting down a Chinese or Russian hypersonic missile, instead of paying $30 million a shot, he can throw you cheap hypersonic clay pigeons for a quarter-million bucks and you shoot 'em down fast and often. That's the bridge business that funds the moonshot. Smart. You find the paying customer today while the big vision cooks.
Now, while we're at TechCrunch Disrupt, which kicked off October 13th in San Francisco, there was a whole parade of these Startup Battlefield companies swinging for the fences on hard physical problems. And I want to group a few of them together because there's a thread running through all of them that I think is the real story of this conference.
Let's talk fusion for a second, because two separate startups showed up this week betting on it, and they're betting on different pieces of the same problem.
First, there's a company called Licube, covered by Tim De Chant. Their whole business is fuel supply. See, most planned fusion plants run on deuterium and tritium. Deuterium's easy to get. Tritium is extraordinarily rare, and it's kind of an open secret in the fusion world that there just isn't enough of it to power every plant people are planning. The workaround is to make tritium on-site by surrounding the reactor with a blanket of lithium, so when neutrons from the reaction hit the lithium, it splits into helium and tritium. But here's the catch, and this is the needle Licube is trying to thread. Not all lithium works. You need lithium-6, which is far better at capturing neutrons, but it's only about 7.5% of natural lithium. The rest is the heavier lithium-7. Separating the two is hard, and there's barely any capacity to do it today. So Licube licensed technology out of Hirosaki University in Japan to purify lithium-6 using an electrolyzer and a ceramic membrane. Their CEO, Yohei Kiguchi, figures if fusion comes online, it's a $50 billion market by 2040.
And then, separately, there's Cambridge Heavy Industries, a company that's been operating in stealth out of a small lab near MIT. Their CEO is Yun Kai, and they've built a fusion device that fits on a tabletop. A benchtop reactor. Kai's whole philosophy is "because it's small, it makes everything easier." The small size lets them run 50 to 80 experiments a day, and that data feeds AI simulations to optimize the design. They just raised a $6.2 million seed round led by Azolla Ventures.
Now, Cambridge Heavy says a full power plant is about a decade out, roughly the size of two basketball courts. But here's the business bridge, same pattern as Longshot. Before the power plant ever happens, they plan to use their small devices to make tritium. That same rare isotope Licube is worried about. Tritium sells for around $30,000 a gram. There's only about 40 kilograms of it in the world. One study they cite predicts the industry will be short about 28 kilograms by 2054. At current prices that's about $840 million worth. So the near-term money isn't the power plant, it's selling the scarce fuel while everyone else is still building reactors.
You hear the thread yet? Longshot, Licube, Cambridge Heavy, all three of them have a giant, decade-plus vision, and all three found a boring, lucrative, nearer-term revenue stream to stay alive while they chase it. Hypersonic target practice. Lithium for semiconductors and pharma. Tritium for everyone else's reactors. That, friends, is the actual lesson of a conference full of hard-tech moonshots. The moon is nice, but you'd better have a day job on the way up.
And there's a fourth one that fits the pattern, Helix Carbon. This one's a resurfaced piece from earlier this month, so I'll keep it tight, but it's a clever angle. The iron and steel industries are among the biggest polluters on earth. And most startups try to decarbonize steel by inventing entirely new ways to make the metal. Helix Carbon's CEO David Brown says that's the wrong move. His line: "Any attempt to decarbonize the steel industry without dealing with the blast furnace is not going to work." Those furnaces last over a hundred years. You're not replacing them. So instead, Helix takes the carbon dioxide pouring out of a blast furnace, runs it through an electrolyzer, turns it back into carbon monoxide, and feeds it back in, which lets producers use less coke. Saves about $14 a ton of iron. Doesn't sound like much, but margins in iron are famously thin. And again, the bridge play: before they tackle the slow-moving steel giants, they're going after the specialty chemicals market first for faster revenue. Same playbook. Work with the furnace you've got, not the one you wish you had.
Now let's shift from the physical world back to the one you actually build software in, because there were a couple of startups at Disrupt that are more directly in your lane.
First, a company called Niobia, covered by Zack Whittaker. San Francisco-based, founded by Gaurav Jha and Harveen Kaur, both of 'em with chemistry PhDs who've worked at chipmakers and battery manufacturers. What they do: AI software that catches production defects on battery and chip lines, then lets an engineer basically chat with the system to figure out what went wrong. The agents run on self-contained GPUs mounted right on the manufacturing equipment, like a sensor. They spot pinhole-sized flaws, scratches, the kind of thing that can cost millions and set back a production timeline for days while engineers hunt for the cause.
But here's the part I want you builders to hear, because it's a tell about where enterprise AI is actually going. Niobia runs on open-source models. Why? Jha said it straight: "None of our customers want their proprietary data to go to the cloud, even if we tell them it's safe." Chip and battery makers are in a high-stakes race, sometimes with adversarial nations, and they will not hand their process data to a public AI cloud. Niobia even lets customers hold their own encryption keys so the startup itself can't decrypt the data. That's the sovereignty demand showing up in a real product, not a slide deck. Data stays on-prem, model is open, customer owns the keys. If you're selling AI into any serious enterprise, that architecture is going to keep coming up.
And then there's one more from Disrupt that I think is a genuinely beautiful little business, Flox Intelligence, covered by Ram Iyer. The founder, Sara Nozkova, spent 13 years in transportation and autonomous vehicles, and she had this insight that machines are brilliant inside the environments we engineer for them, cities, factories, warehouses, fixed geometry, known objects, but the second you step into nature, they go blind. So Flox builds edge devices that keep wildlife off train tracks, airport runways, and roads. A motion sensor wakes a camera, an on-device model identifies the animal, and the thing emits a species-appropriate sound to scare it off. And it varies the responses so animals don't get used to it.
Why does this matter beyond "aw, they save the moose"? The numbers. The FAA estimates wildlife strikes cost US aviation about $272 million a year on average. France's rail operator says wandering animals affect more than 8,000 trains a year and are a top-five cause of delays. In one region of France, 120 trains hit animals in a single year, causing more than 15,000 minutes of delays. That's a real, expensive, boring problem that nobody's solving well. Flox has about 150 devices live across three continents, expects about $1.5 million in revenue this year, and has an OEM deal with the train-maker Alstom. They're now raising a $10 million Series A. And Nozkova's framing is the line to remember: "Physical AI is walking out of the factory and into the messy, alive outdoors. Nature intelligence is the next layer of the perception stack. Almost nobody is building it." That's a founder who found an unglamorous, underfed corner of the market and planted a flag. I respect it.
Alright, let's turn to the big legal and security stuff, because OpenAI had a week.
Two separate OpenAI stories, and they rhyme in an uncomfortable way. First, the courtroom one. And I want to frame this carefully, because the underlying lawsuit goes back to July, so this isn't brand new, it's a fresh ruling in an ongoing case. Apple sued OpenAI back in July over trade secrets. The accusation, as reported by Sarah Perez, is that OpenAI engaged in a pattern of theft, former Apple employees allegedly using Apple laptops to download confidential material about unannounced products, and allegedly coaching other Apple folks on what information to bring to their OpenAI interviews. This is all happening against the backdrop of OpenAI building a consumer hardware device that could one day rival the iPhone.
And the new development is that a judge, U.S. District Judge Edward Davila, ruled Apple gets expedited access to OpenAI's hiring materials. Interview checklists, guides, scripts, technical presentations, specifically for interviews with former Apple employees who applied to OpenAI's hardware and robotics groups. OpenAI's got five days to turn over what it already has, fourteen for the rest. The judge also approved depositions, including for OpenAI's chief hardware officer. Now, the judge didn't give Apple everything, he denied the broader fishing expedition into OpenAI's actual R&D, metal finishes, power management, and so on. But the hiring records are coming out. For anybody building a hardware team by poaching from a competitor, this is a flare in the sky. The interview process itself is now discoverable. How you hire can become the evidence.
And then, the stranger OpenAI story, and I'll be careful here because the underlying event traces to an employee who started back in July of 2025, so I'm framing this as how it's surfacing now. OpenAI says it alerted the FBI after discovering an employee allegedly planted a hidden recording device in one of its executive conference rooms. The employee was fired and arrested on October 9th. Prosecutors named him and charged him with using an electronic device to intercept oral communications, basically covertly recording people without consent. The device was plugged into a USB port on the back of a television, and OpenAI says it contained advanced encryption and a mechanism for transferring data designed to evade monitoring, though, and Zack Whittaker is appropriately skeptical here, it's unclear how technically distinct that really is from any other bug.
Here's the detail I can't get past. OpenAI says it caught the guy because he was spotted looking behind the TV where the device was hidden, and OpenAI had placed a camera there that caught him. And the company said it used its own internal AI models to understand the device. FBI Director Kash Patel framed the whole thing as an attempt to capture sensitive communications about the company's superintelligence technology, and called protecting that innovation one of the bureau's highest counterintelligence priorities.
Now put those two OpenAI stories next to each other. One is a courtroom fight over whether secrets walked out of Apple's doors. The other is physical espionage, an actual bug in a conference room. The through-line is that we have crossed into a period where AI labs are being treated, and are treating themselves, like national-security assets. Counterintelligence priorities. Espionage. Trade-secret litigation with expedited discovery. The stakes stopped being about app downloads a while ago. If you're a founder in or near this space, understand that the security posture you need is no longer "set a strong password." It's counterespionage. That's a wild sentence to say about a software company, but here we are.
Let me take a breath and do a couple of quick hits, because there's a few more things worth your time.
First, a continuity note and a resurfaced item. A while back we talked about Rivian's broader ambitions, and this week at Disrupt, Rivian's industrial robotics startup, Mind Robotics, named a president, Doug Field. Now, this is a resurfaced piece, Field actually left his previous role back in April, so I'm not treating that as news. The news is the appointment. And Field's resume is something: he was employee number one at Segway, ran engineering at Tesla through the Model 3's "production hell," led Apple's canceled car project, and spent five years on Ford's EV strategy. Mind Robotics has already pulled in more than a billion dollars in backing, from Accel, Andreessen Horowitz, Kleiner Perkins, and the venture arms of Volkswagen and Salesforce. The interesting strategic bit, from RJ Scaringe, is that unlike Tesla's humanoid robot ambitions, Mind Robotics isn't trying to mimic all the biomechanics of a human. They're building purpose-built industrial robots on foundation models, and they're using Rivian's own factory as a live, high-volume training environment. First customer is their own parent company. That's a real advantage, honestly. You've got a factory to learn in from day one.
Second quick hit, and I'll keep this one short because the underlying event is actually from June, so I'm flagging it as an older story resurfacing. Meta rolled out a feature called "Your Algorithm" to Facebook. It lets you tell the recommendation engine, in plain language, what you want to see more or less of. They launched it on Instagram earlier this year, and Threads has a version called "Dear Algo." The idea is moving social feeds away from pure black-box automation toward something more like how you tune a streaming service. Fine, modest, sure. The reason I even mention it is the direction of travel: giving users a natural-language steering wheel on the algorithm. Whether that's genuine control or just the feeling of control, I'll let you decide. I've got my suspicions.
And third, there's an OpenAI-adjacent product note I'll mention briefly because the full details didn't come through clearly, so I'm not going to oversell it. A company called Higgsfield announced it's extended AI-generated video from about one minute to over ten minutes, built on OpenAI's video and image models. Longer coherent generated video is a real capability jump if it holds up, but I don't have enough in front of me to tell you how good it actually is, so file it under "watch this space" rather than "run and tell your friends."
Now let me get to the one that I think is the most important story for you, the builder, and it's not a product launch and it's not a lawsuit. It's a tweet with a dataset behind it. This is the deep dive.
Shawn Wang, who goes by swyx, posted this on October 13th, and I want to read you his framing directly because it's precise. He writes, and I'm quoting the post: "How to Kill Your SaaS. A lot of people talk about the death of SaaS but mostly in theoretical terms. We took a real $43,000 a year SaaS we use and launched a global hackathon to clone it in a weekend, with both AI and REAL HUMAN evals, using any models any harness. 72 submissions, a lot of manual evals and about 100 billion tokens later, here are the results." He goes on to say, in a special edition of the Latent Space podcast, they share the data collected, what's still hard about killing nontrivial, midtier SaaS, and they reveal the winners of a $10,000 "Kill My SaaS" contest.
Now, I'm limited to what's in that post, so I'm not going to invent findings that aren't there. But let me tell you why this little experiment matters more than half the funding announcements I covered today, and let me pull out the exact words that carry the weight.
The key phrase is "nontrivial, midtier SaaS." Because look, we've all heard the "software is dead, AI will eat all of SaaS" line for a couple of years now. And mostly, as swyx says, it's been theoretical. Big sweeping claims, no receipts. What makes this valuable is that it's an actual experiment with actual constraints. They took a real tool they really pay $43,000 a year for. Not a toy. Not a to-do list app. A real business expense. And they put a bounty on it: clone it in a weekend, any model, any harness, go.
And the single most important word in that whole post is "still." "What's STILL hard about killing nontrivial, midtier SaaS." Because that word tells you the answer. If cloning that $43,000 tool over a weekend were trivial, the post would read "look how easy this was." Instead it reads "here's what's still hard." That's the tell. 72 teams, 100 billion tokens, a weekend of effort, real human evaluators grading the output, and the headline isn't "SaaS is dead." The headline is "here's the gap that remains."
Now let me connect this to something we covered just a couple of days ago, because it's the same finding from a different angle. Earlier this week we talked about that Harvard study, Fiona Chen and James Stratton, that found AI coding agents produce about 30% more code but not meaningfully more shipped software. More lines, same number of finished features, and code review time ballooned. This "Kill My SaaS" experiment is the field version of that lab result. You can generate an enormous amount of code aimed at replacing a product. 100 billion tokens is a staggering amount of generated output. But replacing a nontrivial, midtier SaaS product, something with real edge cases, real integrations, real reliability expectations, years of accumulated handling-of-the-weird-stuff, that turns out to still be hard. The last mile of software, the boring, unglamorous reliability and edge-case work, is where the moat actually lives. Not in the demo. In the thousand small things the demo never shows you.
So what do you do with this if you're building? A couple of things. One, if you run a SaaS business, don't panic, but don't get comfortable either. The trivial stuff, the thin wrappers, the simple CRUD apps, those are genuinely getting cloned in a weekend now. If that's your whole product, you've got a problem. But the "nontrivial, midtier" stuff, the product with real depth and real integration into how customers work, that still has defensibility, and this experiment is some of the first real evidence of where that line sits.
Two, if you're on the other side, trying to use AI to go kill somebody's incumbent SaaS, the lesson is that generating the code is the easy 80%, and the hard 20% is everything your customer would actually notice when it breaks. The eval matters more than the generation. Notice they ran both AI evals and real human evals. That's the discipline. Anybody can generate a clone. Proving it's actually good enough to depend on is the whole game, and it's the part people skip.
And three, the meta-point. We keep looking for the moment when AI definitively eats an entire category, and what the honest experiments keep showing us is a more textured picture. Not "software is dead," not "nothing's changed," but a shifting, specific line between what's now trivially reproducible and what still requires real work. The builders who win are the ones who know exactly which side of that line their product sits on. This "Kill My SaaS" contest is, in a small way, a tool for finding that line. That's why I lead you to it. Not because it has a tidy answer, but because it's one of the few people doing the real, token-expensive, human-graded work of finding out instead of just tweeting hot takes.
Let me close out with one that's got nothing to do with AI, because sometimes the most grounding story on the menu is just a reminder of how little we actually know.
Over the weekend, the World Health Organization put out a disease outbreak report on the death of a Russian infectious-disease researcher, reported by Beth Mole at Ars Technica. And the whole report is an exercise in uncertainty. A lab worker at the Irkutsk Antiplague Research Institute in Siberia became ill in late September and died. Early media reports suggested she died of plague after breaking a test tube. But Russian authorities told the WHO it was pneumonia of undetermined cause, that she tested negative for all the high-threat pathogens, and that she'd actually been vaccinated against plague as an occupational precaution. They say they ran 5,500 lab tests on contacts, found no ongoing transmission, and detected no high-threat pathogens.
And yet the WHO's conclusion is basically a shrug of deep concern. Their words: "A key uncertainty for this event is the absence of a confirmed diagnosis in the deceased laboratory worker." Without a cause of death, they can't independently assess whether Russia's precautions were adequate, and they aren't ruling out a lab-associated exposure. They're not saying it was something scary. They're saying they genuinely cannot tell, because the information coming out is limited.
I include this partly because it matters on its own, and partly as a palate cleanser against a whole episode of founders telling us they're going to solve fusion and shoot cargo into orbit and clone your SaaS over a weekend. Here's an institution with every incentive to give you a clean answer, looking at a situation and saying, honestly, we don't know. There's something almost refreshing in that. The confidence in this industry runs hot. A little "we can't independently assess that" humility wouldn't kill anybody.
That's the menu for today, kiddos. A trade group chasing the wrong target, a gun that wants to be a rocket, four moonshots that all quietly found a day job, a lab company fighting actual spies, and a weekend hackathon that may have told us more about the real limits of AI than any keynote this week. Keep your eye on where the liability lands, where the near-term revenue hides, and which side of that trivial-versus-nontrivial line your own product is sitting on.
This is Tony DeLuca. Thanks for spending part of your day at the counter with me. Take care of each other out there, and I'll see you next time on Barely Possible.