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 welcome back to Barely Possible. Pull up a chair, grab whatever's in the mug, because we've got a Fourth of July menu today and it is not what you'd expect. We've got a billionaire admitting his robots aren't ready, private space cowboys running Top Gun in orbit for the Pentagon, a spyware firm getting caught bugging the exact guy who was investigating spyware, and a bunch of kids in New York throwing a party specifically to hate the very technology I make my living talking about. There's a nice irony in doing a tech briefing on the day people are lighting sparklers and unplugging, so let's lean into it.
Let me start with the story I keep chewing on, because it cuts against a week of very loud noise. Yesterday we talked about Jersey Mike's, of all companies, dropping AI into an IPO filing as a signal the hype had gone fully cartoonish. So it is genuinely useful timing that Mark Zuckerberg, the man spending more money on AI than most countries spend on defense, reportedly stood up in front of his own staff at an internal meeting and told them the AI agents haven't progressed as quickly as he'd hoped.
Now read that carefully, because this is a founder show and I want you to hear it like an operator, not a headline reader. This is not Zuckerberg saying AI is fake or the whole thing's a bust. This is the guy who reorganized his entire company around superintelligence, poached talent at nine-figure numbers, and told Wall Street to trust him on the CapEx, quietly conceding to the people building it that the agent piece specifically is behind schedule. The report says development efforts were not moving as quickly as anticipated. That's the whole substance we've got, and I'm not going to inflate it into a resignation or a reorg or a pivot, because the source doesn't say that. But the meaning is plenty by itself.
Here's why it matters to you if you're building. Everybody in your investor deck, everybody in your board meeting, everybody in your Slack has been operating off a working assumption: that autonomous agents were about to go from demo to dependable in a matter of months. That the harness was almost there, that the models were almost reliable enough to just turn loose on real workflows and walk away. And the single most resourced person on the planet in this exact race just told his own house that it's slower than he thought. When the guy with the most GPUs and the deepest pockets is saying the agents aren't there yet, that is a permission slip for you to stop pretending they are in your own product. Build for the human-in-the-loop reality that actually exists today, and you'll ship something that works instead of something that demos.
I want to connect this to a couple threads we've been pulling all week, because it's the same animal wearing different coats. Remember a few days back we talked about Ford bringing back the greybeard engineers, the veterans, because the AI tools weren't getting the quality job done on their own? Same lesson, different building. And earlier this week, Claude Sonnet 5's pricing surprise, where the fancy new tokenizer quietly made English text about one-point-four times more expensive despite the lower sticker per token. In each of these, the pattern is: the promise arrives before the plumbing does. The demo is real, the reliability is a work in progress, and the bill is bigger than the brochure. Zuckerberg admitting the agents are behind is just the biggest name saying the quiet part in the biggest room.
And look, I'll be fair to the other side of the ledger, because there's real signal that the capability curve is bending hard even while the reliability curve lags. There's a recent read on AI and jobs floating around, drawing on a benchmark that measures whether a model can do freelance work at a level a paying client would actually accept. Not the model's own opinion of its work, a human judge comparing it against what a paid professional delivered. And the top score on that jumped to about sixteen percent, quadrupling in under eight months from a starting point of a couple percent. So the trajectory is genuinely steep. But sixteen percent professional-grade means eighty-four percent of the time you still need the human. That's not an apocalypse and it's not a nothingburger. It's exactly the gap Zuckerberg is describing: getting faster, still not finished. If you're a founder, the opportunity lives right in that eighty-four percent, in redesigning the deliverable and the workflow around what the tool actually does well.
One more piece of that jobs picture worth thirty seconds, because it's counterintuitive and it's useful. There's data pairing company AI spending against payroll for something like twenty-one thousand US businesses. Firms with high AI adoption grew headcount around ten percent over two years while low-adoption firms stayed flat, and entry-level growth was actually a bit stronger than the average. And this wasn't the token-maxing crowd blowing millions; the threshold for high adoption was something like thirty bucks per employee per month. Now, be skeptical, because fast-growing companies are the ones with money to spend on AI in the first place, and the folks running that analysis said as much. But the direction is interesting: at least so far, the companies leaning in are hiring more, not fewer, and they're hiring for people who know how to use the tools well. Meanwhile tech and finance specifically are shedding jobs, and everybody's arguing about how much of that is real AI displacement versus AI as a convenient cover story for cost-cutting after the 2022 over-hiring binge. My read: both things are true, and the honest answer is we're still on day one of figuring out how to organize work around these systems. Anybody selling you certainty on AI and jobs right now is selling you something.
Alright, let's swing from the guy admitting the future's behind schedule to some folks living in a future that already arrived and is frankly a little unnerving.
There's a piece from Tim Fernholz reporting that private space pilots are now flying orbital missions for the US Space Force. Read that sentence again. True Anomaly and Rocket Lab are doing what the piece calls Top Gun-style satellite fly-bys for the US military. So we've now got commercial companies, private outfits, maneuvering spacecraft up close to other satellites on behalf of the Pentagon, doing inspection and what the trade calls space domain awareness, which is a polite phrase for keeping tabs on the other guy's hardware in orbit.
Why does a founder show care about satellite dogfighting? Because it's the clearest sign yet that the commercial-space layer has become dual-use infrastructure. Ten years ago, if you were flying missions for the Space Force, you were Lockheed, you were Boeing, you were a giant defense prime with a decade of clearances. Now it's a startup and a company that a few years ago was mostly known for launching small rockets. The barrier between commercial space and national security space has basically dissolved, and that opens a whole category of contracts and companies that didn't exist. If you're building anything adjacent to space, sensors, comms, autonomy, tracking software, the customer with the deepest pockets and the most urgent need is now reachable by companies that would have been laughed out of the room a decade ago. That's the story under the story.
And it rhymes with a couple other space items in the pile. The rocket roundup this week had NASA handing Rocket Lab deals for three dedicated launches on its Electron rocket, an Indian startup nearing its first launch, and SpaceX hitting a milestone north of a thousand somethings. The whole sector is the same shape as the Space Force story: the government increasingly renting capability from a swarm of private companies instead of building it in-house. And on the flip side of that coin, there's a recent recap from Eric Berger asking the uncomfortable question, when are the big rockets NASA desperately needs actually going to be ready? His line, and I'm quoting him here, was that he hasn't seen anyone put out a date for a new rocket and actually hit it. Related, the NASA chief was out praising Blue Origin's progress after a launch failure, saying, quote, we've got time into 2027 before we're getting nervous. That's a very specific kind of executive optimism. When your boss says we're not nervous yet, put a reminder in your calendar for 2027.
Now let me shift from orbit to something a lot closer to your pocket, and a lot creepier.
Zack Whittaker has a report that is almost too on-the-nose to be real. A politician who was investigating spyware abuses had his own phone hacked with Pegasus spyware. To spell it out: a government customer of NSO Group used Pegasus to break into the phone of a European politician who, at the time, was serving on an EU committee whose entire job was investigating the spyware industry. They bugged the investigator. With the exact thing he was investigating. While he was investigating it.
I don't have to editorialize much here, the facts do the work. But I want to pull the lesson out for the builders, because it's not just a good yarn. This is the commercial surveillance industry telling you, plainly, that oversight is not a shield. If a sitting member of an EU oversight committee, presumably one of the more security-conscious phones on the continent, gets popped, then the threat model for anyone building sensitive software, handling sensitive data, or sitting on anything a nation-state might want is real and it is aimed at the device in your hand, not just your servers. Endpoint security isn't a checkbox, it's the whole ballgame, because the phone is the soft target. And if you're building in the security or privacy space, the demand signal here could not be louder. The people whose literal job is to police this stuff are getting compromised by it. That's a market and an indictment at the same time.
Speaking of the model companies and their safety posture, Anthropic put out more detail on Fable 5's cyber safeguards and its jailbreak framework. Now, careful listeners will remember we've been tracking the Fable saga all week: it got pulled, it got cleared for global release after some back-and-forth with Washington, and then it got redeployed and people lost their minds over how good it is. This new post is the plumbing side of that story, Anthropic laying out how they're hardening the model against being turned into a cyber weapon and how they're framing jailbreak resistance. I'm not going to pretend I've got the deep internals here, the substance we've got is that it's a detailed safeguards-and-jailbreak framework tied to the redeployment. The takeaway for you: when a frontier lab pairs a big capability release with a very public safety-framework post, that's not just PR, it's them building the paper trail they'll need when a regulator or a customer's security team comes asking. If you're deploying these models into anything regulated, that framework is the thing your compliance people are going to want to read. The link will be in the show notes.
Now, staying in the neighborhood of who controls the AI companies, let me revisit the OpenAI stake story, because it moved and it's worth a second look with the specifics.
We touched this yesterday, the idea that OpenAI was floating handing the US government a stake. The more detailed reporting from Ashley Belanger sharpens it: Sam Altman is reportedly in active talks with the Trump administration, and OpenAI has offered the US a five percent stake, which is described as far lower than what Senator Sanders was targeting. At current valuations, five percent is around forty-two billion dollars, structured to go into a sovereign wealth fund modeled roughly on the Alaska Permanent Fund, the one that pays out oil and mining money to residents. And here's the kicker that makes it more than an OpenAI story: OpenAI reportedly proposed that all the leading AI developers should kick in five percent of their equity. Google, Meta, Anthropic, the whole frontier. Whether any of them agreed to that, or are even at the table, is not clear, and the sources called the whole thing conceptual and early-stage, possibly needing an act of Congress to actually happen.
I want to be plain about the framing, because there's a quid pro quo hanging in the air that the reporting itself points at. Handing the government an ownership stake buys good relations with the administration and answers the political blowback about AI wealth being concentrated. That's the polite reading. The less polite reading is that the most powerful companies of this era are negotiating their political protection by cutting the state in as a shareholder. For a founder, this is worth watching not because you're going to give the government five percent of your seed-stage startup, but because it tells you where the regulatory wind is blowing. If the frontier labs normalize equity-for-goodwill deals with Washington, that reshapes the whole competitive landscape, who gets the friendly treatment, who gets the antitrust letter, who gets the export license. When the big players start paying protection in equity, the little players eventually pay in compliance costs. Keep your eye on it.
And there's a nice contrast sitting right next to it in the pile that tells you the state is not always the AI industry's friend. Google lost its long-running appeal of a record EU fine and now has to cough up four-point-seven billion dollars, which we covered earlier this week. Same governments, very different postures depending on the company and the continent. That's the actual terrain: not regulation versus innovation as some clean abstraction, but a messy, company-by-company, country-by-country negotiation where a five percent gift buys you a friend in one capital while a bundling case costs you billions in another.
Let me take a breath and go somewhere completely different, because it's the Fourth of July and there's a story in the pile that captures the mood of the moment better than any funding round.
There's a piece from Vittoria Elliott on the WIRED side about a Luddite festival in New York City called Summer of Ludd, teaching people how to live offline and harnessing what they call Gen Z's rage against Big Tech. And I want to sit with this for a minute, because I think a lot of people in our world are tempted to laugh it off. Kids in a park learning to use a flip phone, how quaint. Don't laugh it off. This is a market signal, and it's pointed straight at you.
The original Luddites, by the way, get a bad rap. They weren't anti-technology, they were textile workers smashing machines because the machines were being used to gut their livelihoods and wages, and they had no other lever to pull. The name became an insult, but the actual grievance was economic. So when a bunch of twenty-somethings in 2026 start throwing festivals about living offline, the interesting question isn't are they right or wrong about technology. It's what are they actually mad about. And the answer, near as I can tell, is the same as it was two hundred years ago: they feel like the technology is being aimed at them rather than built for them. Attention harvested, feeds engineered, agents making decisions about their lives that they didn't ask for.
Now tie that back to the top of the show. Zuckerberg admits the agents aren't ready. The freelance benchmark says AI does maybe sixteen percent of real work at professional quality. The surveillance industry is bugging its own overseers. And Gen Z is throwing anti-tech festivals. There's a through-line here, and it's not doom. It's that we are in the awkward adolescence of this technology, where the capability is real but the trust is thin and the backlash is organizing. If you're building consumer product right now, the Summer of Ludd crowd is telling you exactly what they don't want: they don't want to be the product, they don't want the black box making calls, they don't want to feel managed by an algorithm. The founders who win the next cycle are going to be the ones who build tools that feel like they're on the user's side, not tools that feel like they're extracting from the user. That's not a culture-war take, that's a product take. The rage in that park is a spec sheet if you're willing to read it.
And on the same Fourth-of-July, taking-stock note, Eric Berger also has a personal essay in the pile, the one about still seeing signs of hope in America at two hundred fifty years, opening with a line about how hard it is to pinpoint the moment America started to lose the plot. I'm not going to turn this show into a civics lecture, and I'm not going to pretend I can summarize a man's personal essay he wrote for a holiday. But I'll note it, because the fact that a veteran space reporter, a guy who spends his days on rockets and launch schedules, felt moved to write about hope and darkness on the country's birthday tells you the mood is heavy even among the people who cover the shiny future for a living. Sometimes the most honest thing in the tech feed is the person stepping back from the tech.
Okay, let me bring it back down to the workbench with a few quicker hits, because there's good builder-relevant stuff in the pile.
The browser wars. There's a rundown from Lauren Forristal making the point that the browser fight in 2026 isn't about search anymore, and walking through the best alternatives to Chrome and Safari. And that reframing is the whole story for a founder. For twenty years the browser was a funnel to a search box, and the search box was the money machine. Now the browser is becoming the surface where AI agents live, where the assistant reads your tabs and acts on them. That's why suddenly there's a crop of challengers actually worth naming instead of the usual graveyard of Chrome clones. If you're building anything that wants to sit between the user and the web, the browser layer is contested again for the first time in a long time, and contested layers are where opportunity hides. The default is no longer safe. Worth watching.
Related little gadget in the pile that I actually kind of like, from Ivan Mehta: a three-button keypad called Dune that changes what it does based on what app you're looking at. In a meeting app it becomes toggle mic, toggle video, bring the window to the front. It's a small thing, but it's a nice example of context-aware hardware, a physical control that reads your screen state and adapts. In a world where everybody's building the smartest possible software, there's a quiet lane for dumb, reliable, tactile hardware that just does the annoying thing with one button. Not every product needs a large language model. Sometimes it needs a knob.
On the cars-and-geopolitics beat, two items rhyme nicely. TechCrunch's Tim De Chant asks why nobody's buying the Chevy Silverado EV, an all-American electric pickup that he calls a solid first draft but clearly not a hit. And a recent review from Jonathan Gitlin on the Volvo EX30 Cross Country, a good little EV that Americans won't be able to buy soon, killed off in the US by tariffs and anti-China policy. Put those together and you've got the American EV situation in miniature: the domestic trucks aren't compelling enough to move, and the compelling small imports are getting walled out by trade policy. If you're building anything in the physical-goods, supply-chain, or hardware world, the lesson is that geopolitics is now a product spec. The Volvo isn't dying because it's a bad car; it's dying because of a customs schedule. Your bill of materials is a political document now, whether you like it or not.
And there's a recent piece on autonomous vehicle hype coming back, with Travis Kalanick, yes, that Travis Kalanick, back building a robotics company, and a firm called Humble Robotics bringing autonomy to freight. The framing from the reporting is that this feels like a rerun of the 2016 hype cycle, the same talent wars, the same capital flooding back, the same faces who lived through the first go-round. I bring it up mostly as a caution. We've seen this movie. The self-driving-is-imminent narrative burned a lot of money and a lot of timelines last time. Freight is genuinely a better wedge than robotaxis, the routes are simpler and the economics are cleaner, so I'm not dismissing it. But when the capital comes roaring back with the exact same energy as the last cycle, the disciplined move is to ask what's actually different this time, and make somebody prove it before you believe it.
Couple of science-and-wonder items to round us out, because it's a holiday and you're allowed to be delighted. There's a report from John Timmer on an artificial cell that managed a few rounds of cell division, which is a genuine milestone in synthetic biology even with the caveat that it only works for a handful of divisions and requires a ton of added materials to pull off. It's not life-in-a-dish, but building something that can divide, even clumsily, even briefly, is a real step. And on the lighter side, there's a Smithsonian VR experience called Starstruck that lets you walk through the stars, exoplanets, and observatories, and an Ars piece revisiting Wing Commander IV, the time a game director basically made a whole movie, back when the full-motion-video game was supposed to be the future and then quietly wasn't. I put those two side by side on purpose: one is a technology that arrived, the VR walkthrough, and one is a technology that was hyped and then faded, the FMV game. Good reminder on a day like today that the future doesn't arrive on the schedule the hype promised. Some of it shows up late, some of it shows up in a form nobody predicted, and some of it, like the FMV movie-game, turns out to be a beautiful dead end.
Which, honestly, brings the whole episode full circle. Zuckerberg says the agents are behind schedule. The benchmark says the work's getting done but slowly. The Luddites are throwing a party. The self-driving money is back for round two. And a thirty-year-old game reminds us the promised future doesn't always show up. If there's one thing to carry out of here on the Fourth, it's this: build for the technology that actually exists today, not the one in the keynote. The people who ship into the gap between the promise and the plumbing are the ones who win, because that gap is where all the real problems, and all the real money, actually live.
That's the menu for today. Whether you're grilling, watching something explode in the sky, or defiantly using a flip phone in a park in solidarity with the Summer of Ludd crowd, I hope it's a good one. I'm Tony DeLuca, this has been Barely Possible, and I'll catch you on the next one. Be good to each other out there.