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 plate of tech served up today, so grab your coffee and pull up a chair. This is Barely Possible, the show where I read the technical stuff so you don't have to, and I try to tell you what actually matters for the thing you're building. Today we've got a startup betting that video games are better teachers than the entire internet, a robotaxi that called the cops on a couple of teenagers, a professor at a fancy school who found out exactly what happens when you take the AI away, and a coding startup that just doubled its price tag. Let's have at it.
Let me start with the story I keep coming back to, because it's the one that quietly reframes a debate we've been having for months, and it's not the flashiest headline in the pile.
There's a company called General Intuition, and they got a big writeup for a bet that sounds almost silly when you first hear it. Their whole pitch is that the internet — the text, the pages, the whole corpus we've been shoveling into these models for years — is actually a lousy teacher for the kind of intelligence you'd want in a robot or any system that has to move through the physical world. The CEO's argument, laid out in the piece, is that models like ChatGPT and Claude are genuinely great at text. They can write your email, they can argue about philosophy, they can debug your code. But ask them to understand how things actually move through space and time, how a ball arcs, how a body turns a corner, how you time a jump — and they're much weaker. And that spatial, physical, temporal understanding, they argue, is an essential ingredient for intelligence that actually generalizes, the kind of thing everybody keeps calling AGI.
So here's the twist. Where do you get millions and millions of hours of agents moving through space and time, making split-second decisions, reacting to a dynamic environment, succeeding and failing in ways you can measure? Video games. That's the bet. General Intuition thinks gaming data is the training set for physical AI — for robots — and that if you feed a foundation model enough of it, you can build smarter robots with minimal real-world data. The framing in the coverage is that robotics is about to have its ChatGPT moment, that same inflection where suddenly the thing goes from a lab curiosity to something that just works.
Now, let me be the skeptic in the room for a second, because that's my job. "About to have its ChatGPT moment" is a phrase that's been applied to robotics roughly every eighteen months since I can remember. Somebody's always about to have the moment. And "minimal real-world data" is the kind of promise that has a way of not surviving contact with an actual warehouse floor. Games are clean. Games have consistent physics that the game designer wrote. The real world has friction, dust, bad lighting, a cat that walks into the frame, and a thousand edge cases nobody coded. So I'd hold the champagne.
But here's why the story matters for you, the builder, even if you never touch a robot. The deeper argument underneath this is about where the next training data comes from. We spent years treating the open internet as this infinite, free resource. And now a couple of things are happening at once. The good text is getting used up, or licensed, or locked behind walls. And more importantly, people are realizing that raw internet text teaches a model to talk, but it doesn't necessarily teach a model to do. If the frontier of value is shifting toward agents that take actions — book the flight, run the workflow, drive the car — then the data that teaches action becomes the scarce resource. And games are just one specific answer to a much bigger question: where do you get high-quality records of agents acting and getting feedback? Whoever answers that well owns something durable. That's the thread to watch, and it's why I'd keep an eye on General Intuition even while I roll my eyes at the "ChatGPT moment" line.
That question about action versus talk brings me neatly to a machine that actually took an action this week, and it's a great little story.
Out in San Mateo, a couple of teenagers decided it'd be funny to do a drive-by on a Waymo. Toy guns. Pointing fake weapons at a robotaxi, presumably expecting the car to just, I don't know, keep driving, because it's a car. Here's what actually happened, per the reporting: the Waymo stopped, called 911 itself, and sat there and waited for the San Mateo Police to show up. And the police did show up. And the teenagers learned the hard way.
Now on the surface this is a funny little item, the kind of thing you tell at the kitchen table. Kids being kids, robot being a narc. But sit with it for a second, because there's a real observation buried in there for anybody building autonomous anything. The car didn't panic, didn't flee, didn't do something dangerous. It behaved like a very calm, very literal witness. It stopped, it reported, it waited. That's a designed behavior. Somebody at Waymo decided that when the system detects a threat, the move is: freeze, alert the authorities, preserve the scene. And it worked exactly as intended.
But hold that thought, because on the very same day, the feds came out with a story that shows the flip side of that coin.
The National Highway Traffic Safety Administration — that's the federal agency that watches over what happens on the road — came out and told autonomous vehicle companies, in pretty blunt language, to stop interfering with first responders. And the line that jumped out at me was this: they said emergency scenes are not, quote, edge cases. That's a shot. That's the regulator saying: you engineers love to file the messy, hard, unpredictable stuff under "edge cases" you'll get to later, and we are telling you that a firetruck at an accident is not an edge case, it's Tuesday.
So put those two side by side. In the toy-gun story, the robot did the smart thing and called the cops. In the NHTSA story, the government is saying these robots are, in the aggregate, getting in the way of the cops and the ambulances and the fire crews. Same technology, and the difference between a feel-good story and a federal warning is entirely in how the system handles chaotic, human, high-stakes moments. That's the whole ballgame for autonomy. The demo on a clean sunny street is easy. The behavior when a cop is waving you off and a hydrant is spraying and there's glass on the road — that's the part nobody can fake anymore. And the regulator just told the industry: we're grading you on the hard part now, not the demo.
Now let's shift from robots in the street to a professor in a classroom, because this one got under my skin a little.
There's a piece about an Ivy League professor — this is at Brown — who got suspicious that his students were leaning on AI to get through his course. So he did something almost quaint. He said, fine, the final exam is going to be in person. No laptops. No chatbots. Just you and a piece of paper. And when he ran that in-person final, the scores fell by fifty percent. Cut in half. The professor's line, and it's a strong one, is that AI cheating leads to, quote, a failed society, and that we cannot choose to become idiots.
Now, I want to be careful here, because the easy read is "kids are cheaters and AI is rotting their brains," and I don't think that's quite the honest story. Let me give you the more useful version.
A fifty percent drop when you remove the tool tells you something, but it doesn't automatically tell you the students learned nothing. Some of that gap is genuinely people who outsourced the whole course and retained air. But some of it is also that these students learned the course the way it was actually taught in 2026 — with the tool present. If you spend a semester solving problems with an assistant at your elbow, and then you get tested with your hands tied behind your back, of course you score worse. It's like training your whole life with a calculator and then getting graded on long division under a stopwatch. The number is dramatic, but the interpretation is messier than "they're all frauds."
Here's why I'm telling a builder this story, though, and it's not really about college. It's about the difference between a skill and a crutch, and that difference is now a live question inside every company that's rolling out AI to its own people. You, as a founder, are about to hand your team a bunch of powerful assistants. Some of your people are going to use them the way the high-performing users use them — as a reasoning partner, framing the problem, pushing back, iterating. And some of your people are going to use them as a crutch that lets them stop understanding their own work. And on a good day, on a normal Tuesday, you cannot tell those two groups apart. They both ship. They both look busy. The logs look identical.
The professor's in-person exam is a crude but honest instrument for pulling those two groups apart. And I think the real lesson for anybody building an org is: you need your own version of the in-person exam. Not to punish people, but to know, honestly, which capabilities live in your people and which ones live in a subscription you're renting. Because the day that subscription gets more expensive, or the model changes, or the vendor cuts you off, you find out real fast which of your competencies you actually own. That's not a moral panic. That's just inventory.
And speaking of capabilities you're renting versus capabilities you own — let's talk about the money, because there's a funding number today that tells you exactly where the market's head is at.
Lovable, the AI app-building company, is reportedly in talks to double its valuation to thirteen-point-two billion dollars. The round is around three hundred million, reportedly led by Menlo Ventures, and this is per a report from Sifted. Double. In one round. This is a company in the vibe-coding, build-an-app-by-describing-it space, and the market is saying it's worth twice what it was worth not long ago.
I'm going to give this to you straight, because I've watched a lot of valuation cycles. A double is a signal, and it's a signal that cuts two ways. On the bull side: this whole category — describe what you want, get working software — is clearly resonating with real users and real revenue, or nobody's writing a three hundred million dollar check at that price. Menlo isn't a tourist. On the bear side: this is exactly the kind of category where the moat is the question that keeps founders up at night. When the underlying models get better, a lot of the magic that a company like this provides gets easier for anyone to provide. The thing you're building on top of is improving underneath you, which is great until it improves so much that it eats your product.
We talked, not long ago on this show, about Figma acquiring the team behind a vibe-coding app, and about Vercel's CEO out there arguing about splitting models off from the agents that use them. This Lovable number is another data point in that same current: everybody with money believes the app-building layer is where a huge amount of value lands, and everybody is racing to plant a flag before the ground shifts. Thirteen billion dollars is the market betting that the interface to software creation is worth owning. Maybe it is. But if you're building in this space, the question you should be asking is not "how do I ride this wave," it's "what do I own that survives the next model release." Because the next model release is always coming.
And on cue — the next model release came this week too. Let me run through the model news quickly, because there's a bunch of it and most of it doesn't need a dissertation.
Grok got a new version, 4.5, out of Musk's operation. Elon's describing it as an, quote, Opus-class model, and the pitch is that it's cheaper, more token-efficient, and lower cost than the other powerful models. Now, "Opus-class" is Elon-grading-his-own-homework, so put whatever discount on that you normally put on a founder describing his own product. But the interesting part isn't the capability claim, it's the positioning. Notice he's not leading with "it's the smartest." He's leading with "it's cheaper and more efficient." That's a tell about where the whole industry's mind is at right now. Six months ago every launch was a raw-power flex. Now the flex is efficiency. Everybody figured out at roughly the same time that when you're running agents that make thousands of calls, the cost per token is the thing that decides whether your product has a business model or not. We've been circling this expense-management theme for a while on the show, and I'm not going to belabor it, but I'll just point out: even the guy who loves a superlative is now selling you on cheaper. That's the market talking.
On the OpenAI side, Sam Altman announced GPT-live, a next-generation voice mode launching in ChatGPT. And his framing was personal and kind of interesting — he said he's always preferred typing to talking to an AI, and now he thinks that's going to shift. He called it magical and, quote, real. Take the marketing gloss off and there's a genuine product bet underneath. Voice has always been the interface that either feels natural or feels like you're arguing with a phone tree. If they've actually crossed the line where talking is better than typing for a meaningful chunk of tasks, that changes the shape of a lot of products, because voice pulls in the people who never wanted to type at a computer in the first place. That's a bigger addressable market than the keyboard crowd. I'm not going to tell you it's there until I hear it myself, but the direction is worth clocking.
And then there's Google, quietly getting graded on its own report card. Google updated its Android developer benchmark — Android Bench — with a bunch of new models, and the headline from the coverage is a little awkward for them: their own Gemini still lags behind on it. I love this, honestly, because it's Google running an honest benchmark and then publishing the result even though the result makes their own model look middling. That's more integrity than a lot of these benchmark exercises show. For you as a builder, the takeaway is small but real: don't assume the house model is the best model for your task just because it's from the same house. Benchmark it yourself. The company that made the benchmark just showed you their own model isn't automatically winning.
Now let me pull one thread out of all this model news, because Yann LeCun — Meta's chief AI scientist — dropped a line this morning that ties it together. Responding to someone, he said the biggest risk of AI is the concentration of power in a few dominant providers of proprietary AI assistants, and that the only solution to AI sovereignty is open-source foundation models.
Now, I've got to give you the honest context here, because LeCun has been beating this open-source drum for years, and he works for a company whose whole AI strategy is built on releasing open weights, so this is not exactly a man discovering a new religion. Consider the source. But the reason I'm surfacing it is that it lands in a very specific moment. We've spent the last couple weeks on this show talking about the machinery of dependence — companies getting cut off from models, access getting rationed, vendors changing terms. And here's one of the most prominent researchers in the field saying, out loud, that the whole risk is concentration, and the whole answer is openness. Whether you buy his solution or not — and reasonable people fight about it — the diagnosis is the one that keeps showing up from every direction this week. The professor's in-person exam, the Lovable moat question, the efficiency arms race, LeCun's tweet — they're all circling the same anxiety: how much of what you depend on do you actually control?
Alright, let me shift gears entirely, from the abstract to the concrete, because there's a cluster of stories about AI meeting the real world in ugly and messy ways, and a builder needs to see these clearly.
The hardest one first, and I'll keep it brief and clinical because it deserves that. There's a lawsuit — reported by Ashley Belanger — involving Grok, Musk's image tool, alleging that a man used it to generate thousands of abuse images of a child, and that the company only flagged a tiny fraction of the offending prompts to authorities. More young girls are now suing X over AI-generated child sexual abuse material, with the platform accused of shielding predators. I'm not going to detail any of it beyond that. The reason it belongs in a builder's briefing, and not just a crime blotter, is this: if you ship a generative tool, you are shipping a machine that can be pointed at the worst thing a human can imagine, and the question of what you detect, what you report, and how fast you report it is not a nice-to-have compliance checkbox. It is core product safety, and the courts are now treating it that way. The allegation here isn't just that a bad thing happened — bad people will always try to misuse tools — it's that the company didn't report what it saw. That distinction is where the legal exposure lives. If you're building anything generative, your reporting pipeline is not an afterthought. It's load-bearing.
On a much lighter note in the same "AI meets reality" bucket — Google's deepfake detection system got used this week to debunk a hoax. There was a picture going around that appeared to show Senator Mitch McConnell in a hospital bed, covered in tubes, in obvious distress. Turned out to be an AI-generated fake, and Google's detection system was used to call it. So there's your two-sided coin again: the same generative capability that produces the abuse images and the fake hospital photos is also spawning the detection tools that catch them. It's an arms race, and detection is always a step behind generation, but at least there's a defensive line forming. For anybody building in media or trust-and-safety, that detection layer is quietly becoming a product category of its own.
Now let me dig into the deep dive, because there's one story that I think matters most for you, the founder, and it's the one about drones — but not the way you'd expect.
Let me set it up with the flashy headline and then get to the part that actually matters. The flashy version: the U.S. is now hunting for cheaper hunter-killer drones after Iran destroyed roughly a billion dollars' worth of Reapers. The reporting, from Jeremy Hsu, is that American military drone losses in the Iran war have spurred a Pentagon call for cheap replacements. A billion dollars of these big, expensive, exquisite Reaper drones — gone. And the response isn't "build them tougher." The response is "build them cheaper."
Sit with that, because it's one of the most important strategic ideas of our moment, and it applies way beyond the battlefield.
For decades, the American way of building military hardware was what people call the exquisite approach. You build a small number of extraordinarily capable, extraordinarily expensive platforms. A Reaper drone is a masterpiece. It's got the sensors, the range, the payload, it's a flying luxury item. And the whole logic of that approach depended on an assumption: that your fancy thing wouldn't get shot down very often, because the enemy couldn't afford or couldn't manage to kill it. That assumption just broke. When the other side can knock a billion dollars of your exquisite platforms out of the sky, the exquisite model isn't just expensive — it's a losing trade. You're bringing a Fabergé egg to a rock fight.
So the Pentagon's pivot is toward attritable systems. Ugly word, simple idea: things cheap enough that you're okay losing them. Instead of one perfect drone that costs a fortune and breaks your heart when it dies, you want a swarm of cheap ones where losing a few is just Tuesday. The math flips. Quantity becomes its own kind of quality. And the whole industrial base has to reorient — away from building a handful of perfect things slowly, toward building a flood of good-enough things fast.
Now, why am I making this the centerpiece for a room full of founders and builders? Because this is the single most important cost-structure debate of the entire AI era, and it's playing out in the exact same shape everywhere you look.
Think about what we just talked about all episode. The frontier model versus the cheap efficient model — that's exquisite versus attritable. The expensive, perfect, do-everything model is the Reaper. And the whole industry's pivot toward cheaper, token-efficient, good-enough models that you can run by the thousands — Grok's efficiency pitch, Microsoft's in-house cost-cutting models we talked about earlier this week, everybody suddenly bragging about being lower cost — that's the attritable swarm. The Pentagon and the AI industry independently arrived at the same conclusion in the same season: when you have to run the thing constantly, at scale, in the real world, the exquisite version bankrupts you. Cheap and plentiful beats perfect and scarce.
And it goes deeper than models. Think about how you architect a product. The old way, you'd make one expensive, careful, high-stakes call to the best system you had, and you'd protect it. The new agentic way, you're making thousands of cheap calls, and you design assuming a bunch of them will fail, and you build the system to tolerate the failures instead of preventing every one. That's attritable design. You stop trying to make every single operation perfect and expensive, and you start making them cheap and redundant and resilient in the aggregate. Same philosophy the Pentagon just got mugged into adopting.
There's a related story in the same pile that makes the point from the other direction. There's a startup called Manna, autonomous drone delivery, and they're planning a big U.S. expansion — a whole operations and manufacturing facility in Tulsa, Oklahoma, that they say will eventually employ a thousand people. Delivery drones. Now notice: the whole business model of drone delivery only works if the drones are cheap and plentiful. Nobody's delivering your burrito with a million-dollar exquisite aircraft. The economics demand the swarm. Manna building a manufacturing facility — actually making the things, at volume, in Oklahoma — is a company betting its whole existence on the attritable side of this trade. Cheap units, made in bulk, lose one now and then and shrug.
So here's what I want you to take from the deep dive, and it's the through-line of the whole show today. The most important strategic question in tech right now isn't "how do I build the most capable thing." It's "how do I build the thing that survives being used constantly, at scale, in a world that's harder and more expensive than my demo." The exquisite instinct — build one perfect thing, protect it, depend on it — is getting punished everywhere. On the battlefield, in your model bill, in your product architecture, and honestly in that classroom too, where the students who built one perfect dependency on a tool got cut in half the moment the tool went away. Attritable, resilient, cheap-and-plentiful, own-what-you-depend-on — that's the posture that's winning across every one of these stories. And you didn't need a defense contractor to tell you. The market's been screaming it at you all week.
Let me clear a few smaller things off the counter before I let you go, quick hits.
Elon Musk's long saga with the SEC over how he disclosed building up his stake in Twitter — back when it was still Twitter — has finally ended. A judge approved a one-point-five million dollar settlement, and the reporting from Lucas Ropek notes the judge did it despite, quote, misgivings. So it's over. A billionaire, a disclosure fight, a rounding-error settlement, and a judge holding his nose while he signs off. File it under: the cost of doing business, when you're that big, is smaller than it should be.
On the crypto and prediction-market front — a judge rejected Kalshi's attempt to override New York state gambling laws. Kalshi runs a prediction market, and they tried to argue that their sports-related markets shouldn't be subject to New York's gambling rules, and the judge said no, New York can restrict that. This matters if you're anywhere near prediction markets or crypto, because it's another brick in the wall of "federal-ish framing doesn't automatically get you out of state law." The prediction-market crowd keeps trying to argue they're a special category that floats above the old rules. Courts keep saying: nice try, you're still subject to the state you're operating in. Watch that space, because the whole business model of these platforms depends on how that fight resolves.
Over in India, Truecaller — the caller-ID company — is clashing with the telecom regulator over anti-spam rules. Their gripe is interesting: they say users are increasingly just ignoring and blocking calls from India's dedicated business-number series, the numbers the regulator set up specifically so legitimate businesses could call you and you'd know it was legit. And it's backfiring, because people have learned to distrust the whole series. There's a lesson in there about trust systems: you can mandate a "this is a real business" label all you want, but if enough bad actors hide behind it, users learn to ignore the label entirely, and now you've poisoned the one signal you built. Anybody designing a verification system should tattoo that on their arm.
And two quick ones from the "the world is weird" file. Down in Miami, a company called City Labs pulled off a first for commercial nuclear power in space — their BOHR mission is being described as a pathfinder for future nuclear-powered spacecraft. Small commercial outfit, nuclear power, orbit. The commercialization of space keeps eating capabilities that used to belong only to governments. And over in Australia, the government told volunteers to throw out thousands of perfectly functioning test routers after a program wrapped up — even though, as the reporting dryly notes, the devices could easily be reflashed and reused. Thousands of working routers, headed for the landfill because it was easier than dealing with them. If you ever wonder why people don't trust institutions to be careful with resources, there's your little parable.
Alright, let me tie the bow on this one, because I think there's a clean thread running through everything today.
Every big story in the pile came back to the same question: what do you actually own, versus what are you renting and depending on. The professor found out his students were renting their competence and got cut in half when the rental ended. Lovable's investors are betting thirteen billion dollars on owning the app-building layer before the models shift underneath it. The Pentagon found out that owning a few exquisite drones is a worse bet than owning a cheap swarm you can afford to lose. LeCun's out there saying the whole risk is depending on a handful of proprietary providers. And the model makers, one after another, are pivoting from "look how powerful" to "look how cheap to run," because they figured out that at scale, the thing you run constantly has to be affordable or the whole thing falls apart.
Build the resilient thing, not the exquisite thing. Own what you can't afford to lose access to. And do your own in-person exam once in a while, so you know which of your capabilities are real and which ones evaporate the day the tool changes. That's the whole show.
This has been Barely Possible. I'm Tony DeLuca, I appreciate you spending part of your day with me, and I'll be right back here tomorrow with a fresh plate. Take care of yourselves out there.