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, it's your boy Tony DeLuca, and you found your way back to Barely Possible. We got a real menu today, the kind where you walk in for a coffee and walk out three hours later because somebody at the next table started an argument worth listening to. Buckle up. Let's have at it.
Here's the thing that's eating at me this morning, and it's the thing I want to spend real time on. The United States government, on a Friday afternoon, reached across a company's shoulder and flipped the off switch on its two best AI models. Not a fine. Not a warning letter that sits in a drawer. An off switch. Anthropic's Fable 5 and Mythos 5, gone, worldwide, for everybody, because of a so-called jailbreak. And the way it went down tells you more about where this whole industry is heading than any benchmark chart or funding round.
Let me lay out the facts, because the facts are wild enough that I don't need to dress them up. According to Ars Technica and TechCrunch, the Commerce Department issued an export control directive to Anthropic. The concern, as the Commerce folks framed it, was that a Fable 5 jailbreak could be a national security threat. Anthropic complied. They had to. But they were furious about it, and they said so in a blog post. Their words, and I'm quoting directly from what they wrote: "We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people." Hundreds of millions of people. One Friday letter. Lights out.
Now let's slow down, because this is where the founders and builders listening need to pay attention, and where the conspiracy crowd is going to overcook it. Two things matter here, and they're separate. There's what happened, and there's why it happened. Let's take what happened first.
What happened is that the government demonstrated, for the first time in a way nobody can argue with, that it has a kill switch on a frontier AI model and it's willing to pull it. Forget for a second whether the jailbreak was real or serious. The precedent is the product. The precedent is the thing that got shipped on Friday. Before this, the working assumption in every boardroom in this space was, quote, no government would actually dare tell an AI company to pull a deployed commercial model. That assumption is dead. It died Friday at 5:21 PM Eastern, which is when Anthropic says they got the letter. You build your company on assumptions like that. And one of them just got vaporized.
Now here's where it gets messy, and I want to be careful because there's a lot of speculation flying around. There's a report, from TechCrunch, that Amazon CEO Andy Jassy may have been the source of the security concerns that triggered this whole thing. May have been. The reporting is that Amazon researchers did the jailbreak work that became the basis for the government's directive. Now, what nobody has cleanly established is whether Amazon itself walked that finding into the government's hands, or whether it traveled some other route. So I'm not going to stand here and tell you Andy Jassy called up Commerce and asked them to whack a competitor's model. That's not what the reporting says, and I'm not going to manufacture it. What the reporting says is that the security concern that lit the fuse appears to trace back to Amazon's own testing.
And that's deliciously awkward, because Amazon is one of Anthropic's biggest investors. So if you're a founder, sit with that for a second. Your biggest backer's own security team produces the finding that becomes the pretext that gets your flagship product pulled by the federal government. That's not a competitor across the street. That's a partner who's on your cap table. In this business, the call is coming from inside the house.
Here's the technical substance, because it matters and because the hype machine is going to skip right past it. Anthropic says the jailbreak, the thing this whole national security panic is built on, essentially consists of asking the model to read a specific code base and fix any software flaws. That's it. And they say they validated that the same level of capability is, quote, widely available from other models, including OpenAI's GPT-5.5, and is used every day by the defenders who keep systems safe. In plain English: the scary technique is something security professionals do for a living, every single day, with tools that are sitting right out in the open. So either this was a genuine, specific national security concern that the public hasn't been shown the evidence for, or, and I'm just laying out the options, it was a pretext. The cybersecurity people quoted in the coverage are leaning hard toward pretext. One CEO of a security firm in the reporting basically asked, who at the White House looked at this and thought it was a threat, because this is exactly the kind of prompting defenders do.
Which brings us to the part that I find almost unbearably ironic, and we've been tracking the setup for this for a while now. We covered Claude Fable 5's release as a step-change coding model earlier this month, and we covered the guardrail fights around it. Now think about Anthropic's whole public posture for the last year. Dario Amodei has been the loudest voice in the room saying these models are dangerous, that the government should have the power to block deployment of a model that presents unacceptable risks. He said the government should have that power. And then the government used that power. On him. And his reaction was, well, not like that.
There's a three-panel cartoon that went viral that sums it up better than I can. Panel one, Dario says, this is the most dangerous AI yet, it could destroy global infrastructure, it can't fall into the wrong hands. Panel two, Trump says, okay, it's banned. Panel three, Dario, apoplectic: you can't do this. And look, I don't want to be cheap about it, because Anthropic did try to thread the needle. They said in their statement that they believe the government should be able to block unsafe deployments, but as part of a process that's transparent, fair, and grounded in technical facts, and that this action did not adhere to those principles. That's a fair distinction. A Friday-afternoon edict with no public evidence is not the same as a statutory review process. But here's the cold truth, founders: when you spend a year telling the most powerful government on earth that your product is a weapon, you do not get to act surprised when they treat it like one. You named it Fable, and it came with a moral.
Now let me get to why you actually care, because most of you aren't running a frontier lab. You're building products on top of these things. Here's the chain of pain. Fable 5 doesn't just live at Anthropic. It's served downstream through Cursor, through Devin, through OpenRouter, through every law firm running Harvey, every product that piped that model in through an API. When the model gets pulled, all of that goes dark, instantly, with no notice. And worse, the directive was framed around foreign nationals. The government's order said no access by any foreign national, inside or outside the United States, including Anthropic's own foreign-national employees. Which, if it sticks, means that to serve a model like this, you might eventually need to verify citizenship at the end-user level. Proof of citizenship to use an AI model. Think about what that does to your signup flow, your API billing, your entire international customer base.
And here's the strategic kicker for anyone building outside the US, or selling to anyone outside the US. Every procurement officer in Brussels, in Tokyo, in São Paulo who watched this happen now has a defensible, board-approved reason to hedge. To prefer a sovereign model. To run experiments with Chinese open-weight alternatives like the Qwen and DeepSeek families, where the quality gap is now small enough that it actually matters. The US just demonstrated, live, that frontier access can be revoked unilaterally, overnight, by one government. If you're a builder, the lesson is not panic. The lesson is don't single-thread your stack through one model from one lab in one country. Resilience just stopped being a nice-to-have and became table stakes. The harness, the routing, the ability to swap a model out without re-architecting your whole product, that's not a nerd hobby anymore. That's business continuity.
Nobody knows what happens next. The base case is probably that this gets reversed in some form, because the economic damage is enormous and the legal footing looks shaky. But the precedent doesn't un-happen. You can restore access to Fable 5 tomorrow and the kill switch is still on the wall, and everyone now knows it works. That's the part you can't take back.
Alright. Let me shift from the lab that got its lights turned off to a related question that's been quietly bubbling under all of this: who's actually watching the watchmen, and how many of them are there. Because Anthropic isn't the only AI company getting squeezed by the government this week.
OpenAI is now facing an investigation from state attorneys general. This is from TechCrunch. It's not clear yet which states are involved, but the reporting is that they're asking about a pretty broad spread, everything from OpenAI's advertising policies to how it handles health data. Now, on its own, this is a brief. AG investigations are a dime a dozen and most of them go nowhere. But put it next to the Fable 5 thing and you start to see the shape of the year. Federal Commerce on one side pulling models, state AGs on the other side poking into data practices and ads. The regulatory weather has changed. For a long time the story in this industry was that regulation couldn't keep up. That's not the interesting frame anymore, and frankly I'm tired of it. The interesting frame is that regulation is now arriving from a dozen directions at once, uncoordinated, unpredictable, and aimed at different things. Commerce cares about national security. The AGs care about ads and your health data. None of them are talking to each other. And you, the builder, are standing in the middle of that crossfire trying to ship a product.
The health data piece is the one I'd watch if you're building anything consumer-facing. People are pouring their entire medical lives into these chatbots. Symptoms, prescriptions, things they wouldn't tell their own doctor. If the AGs decide that handling that data carelessly is a consumer protection violation, that reaches a lot further than OpenAI. That reaches anybody building a health-adjacent AI product. So tuck that one away.
Now let's talk about a different kind of unreliable narrator. KPMG, the big consulting and accounting firm, had to pull a report. A report about AI usage. Because the report itself, per TechCrunch, contained apparent hallucinations. Let me just let that sit there for a second. KPMG published research about how people use AI, and the research had to be retracted because it appears AI hallucinated parts of it. As the writeup dryly put it, once again AI proves to be an unreliable source of information about AI.
Now I want to be a little fair to KPMG here, and a lot unfair, because both are deserved. The unfair part first: this is funny, and it's exactly the kind of thing that should make every executive who's been waving around a slick AI-generated market report stop and sweat a little. The number of "studies" floating around right now that are really just a model confidently inventing statistics with a nice chart on top, it's a lot. More than you think. If a firm with KPMG's resources and review process let hallucinated material into a published report, what do you think is happening at the dozen smaller shops cranking out thought-leadership PDFs every week?
And here's the builder lesson, the fair part, because it's not actually about KPMG being dumb. It's about process. The failure mode wasn't the model. The model did what models do, it generated plausible text. The failure was that there was no verification layer between the generation and the publication. The model wrote it, somebody pasted it, nobody checked the citations against reality before it went out the door with the firm's name on it. If you're shipping AI-generated content of any consequence, your moat isn't a better prompt. Your moat is the boring, unglamorous verification step that catches the confident lie before your customer sees it. The people winning right now aren't the ones with the cleverest prompts. They're the ones who built the checking into the pipeline. KPMG, a firm whose entire brand is checking things, forgot to check the thing. Don't be KPMG.
Let me stay on this thread of AI inside big companies, because there's a story here that's the human cost of the gold rush, and it's a brutal one. TechCrunch ran a report, and the headline is rough, so I'll quote it plainly: Meta's months-old AI unit is described, by the engineers inside it, as a soul-crushing gulag. Their word, not mine. The unit employs 6,500 people, and the reporting says it's on the verge of revolt.
Now, 6,500 people. Think about that. That's not a team, that's a small city. Meta went on a spending spree, scooped up an enormous amount of talent, paid eye-watering numbers to assemble this AI org, and the people inside it are describing it as a place where the soul goes to die. And I want to connect this to something, because it's not a gossip item, it's a strategy warning. You can buy the talent. You cannot buy the conditions that made the talent productive in the first place. The thing that makes a great engineer great is usually some combination of autonomy, clarity, and the feeling that the work matters. You assemble 6,500 of those people into a giant reorganized unit with unclear mandates and internal politics, and you don't get 6,500 units of genius. You get a gulag, by their own account, and a verge-of-revolt retention problem.
For founders, the lesson cuts both ways. If you're small, this is your edge. The reason a ten-person startup can out-ship a 6,500-person org isn't that your people are smarter. It's that your people can actually see what they're building and why. That clarity is a real, defensible advantage, and it's one the giants are demonstrating they have trouble buying back. And if you're growing, this is your warning label. The thing that's working for you at thirty people is the thing that breaks first when you try to scale it to three hundred. Meta has effectively infinite money and they still produced a place people hate to work. Money was never the bottleneck. Coherence was.
Alright, let me change the temperature entirely, because not everything today is doom and government kill switches. Let's talk about a guy who thinks the next big opportunity isn't building a smarter model, it's making your rent cheaper.
Andrew Yang, you remember him, ran for president, the universal basic income guy, he's got a take in TechCrunch that I actually think is the most builder-relevant idea in today's whole pile, and almost nobody's going to give it the time of day because it's not flashy. Yang made a list of everything Americans systematically overpay for. Housing, food, wireless. And his thesis is that the next startup gold rush isn't another chatbot wrapper. It's giving that money back. Building companies whose entire reason to exist is lowering the cost of living.
Now, I grew up where people argued about the price of everything at the kitchen table, so this lands for me. And here's why it's smart, and why it connects to everything else we've been talking about today. We've spent this whole episode on the AI arms race, the CapEx, the billion-dollar funding rounds, the labs spending themselves into oblivion to build models that the government might just turn off. That's one game. It's an expensive, crowded, dangerous game. Yang is pointing at a completely different board. The total addressable market for "the stuff every American overpays for" is the entire American economy. Housing alone is the single biggest line item in most people's lives.
And here's where AI actually fits, but quietly, as a tool rather than the product. A lot of why you overpay for things is friction and middlemen and information asymmetry. Finding the cheaper option is work, comparing it is work, switching is work, and the incumbents make money off the fact that the work is annoying enough that you don't do it. That's exactly the kind of work an agent can eat. Not as a flashy demo, but as a boring, relentless machine that finds you the better wireless plan, fights the surprise medical bill, refinances the thing, switches the provider. The pitch isn't "look at my amazing AI." The pitch is "I saved you four hundred bucks a month and you didn't have to lift a finger." That's a value proposition your aunt understands. That's a business.
I'm not saying Yang's right about everything, the man's been wrong before. But the instinct, that the durable opportunity is in giving people their money back rather than in selling them a fortieth productivity tool, that instinct is sound, and it's pointed away from the part of the market that's currently on fire and getting shot at by the government.
Which is a nice segue, actually, because let me talk about pricing for a minute, since it's the other side of that same coin. There's a piece from Bernhard Hauser of Growing Ventures, an essay called Good Pricing Grows With the Value You Deliver. And the core idea is one I want every founder listening to internalize, especially the ones building agent products. At one of his portfolio companies, AI agents run the recurring business-intelligence check-ins, do the SEO research, handle parts of the outbound campaigns. Real work, every week. And the pricing argument is that good pricing isn't a fixed sticker, it scales with the value you actually deliver.
Now this sounds obvious until you put it next to the rest of today's news, and then it gets sharp. Remember we covered the SemiAnalysis findings recently, the ones showing that OpenAI's two-hundred-dollar Pro plan delivers up to fourteen thousand dollars in monthly API-equivalent value, and Anthropic's Max plan around eight thousand. Those are subsidized prices. The labs are eating enormous losses to hook developers. So if you're building on top of those models, and you're pricing your product on the assumption that the underlying token cost stays at today's subsidized number, you are building your house on a beach. The subsidy era ends. And Hauser's point, that your price should track the value you deliver, not your cost, is the thing that saves you when it does. If you're priced on value, on the four hundred bucks you saved the customer, on the work you took off their plate, then it doesn't matter much what your input cost does, because you've got room. If you're priced on a thin margin over a subsidized token, the day that subsidy evaporates, so does your business. Price the outcome, not the input. That's the whole lesson, and it's free.
Now let me shift from the labs and the pricing to the big public-markets story, because there's a frame on it I think most people are getting wrong. SpaceX is now a public company. We covered the IPO itself the last couple of days, the record-breaking debut, the hundred billion in retail orders, the trillionaire stuff. I'm not going to re-litigate the pricing. But there's a fresh piece from Eric Berger at Ars Technica that asks the question that actually matters now that the confetti's swept up: SpaceX is now owned by investors who will want to see it make money, so what comes next?
And here's the angle that I think is genuinely underappreciated. The thing that the market got most excited about, the thing that justified that monster valuation, wasn't rockets. It was AI. Specifically, the pivot toward SpaceX as a kind of neocloud, the idea of data centers in space, the mega-compute deals. We talked a few weeks back about Google's enormous monthly compute deal with SpaceX. That's the story that turned a launch company into a trillion-dollar AI infrastructure play in the market's mind.
But now the meter's running. As a private company, SpaceX could sell a vision and live off the next funding round. As a public company, every quarter, somebody's going to ask: where's the revenue. Berger's framing is the right one, the discipline of public markets is a different animal than the patience of private capital. And there's the merger chatter on top of it, Gwynne Shotwell, the president, dropping yet another hint about an eventual Tesla combination, which TechCrunch frames as starting to feel inevitable. Whether that's real or just Elon doing Elon, I won't speculate.
The builder takeaway here isn't about whether to buy the stock, I'm not in that business and neither should this show be. It's about what it means when the biggest, most hyped infrastructure player in your supply chain goes public on an AI thesis it now has to deliver. If you're building anything that depends on cheap, abundant compute, and these days that's most of you, then the people supplying that compute just took on a whole new set of quarterly masters who want returns. That tends to push prices up, not down, over time. The free-money, build-it-and-they-will-come phase of AI infrastructure has a public-market clock on it now. Plan accordingly.
Let me hit a couple of quick ones before we wrap, because they're worth knowing even if they don't need a whole sermon.
The FBI, per TechCrunch, built itself a fake town. A whole replica small town, hidden inside a building in Alabama, as a dedicated training ground for simulating real-world cyberattacks. And I love this for two reasons. One, it's just a great image, a Potemkin village with a federal badge. But two, it tells you how seriously the people whose job is defense are taking the threat surface now. They're not running tabletop exercises on a whiteboard anymore. They're building physical environments to rehearse attacks on the kind of critical infrastructure that runs a real town. If you're building security products, or honestly if you're building anything that touches infrastructure, that's a signal about where the defensive money and attention is going.
And here's a small one that connects to it. There's a launch making the rounds, swyx flagged it, a Fusion API that's claiming Fable-level performance on deep research tasks at half the cost. Now, I'm not going to vouch for the benchmark, vendor numbers are vendor numbers, take them with the usual fistful of salt. But notice the timing. The week the government pulls Fable 5, the marketing pitch becomes Fable-level performance, somewhere else, cheaper. That is the market routing around the damage in real time. The instant a frontier model becomes politically radioactive or simply unavailable, the entire ecosystem's incentive becomes give people that capability through some other door. That's the resilience instinct I was talking about earlier, except now it's a go-to-market strategy. Capability abhors a vacuum.
There's also a Kimi coding model release floating in the mix, the K2.7-Code from Moonshot, claiming solid gains on coding benchmarks and fewer reasoning tokens. I'll keep this to the headline because we don't deep-dive open-weights kernels on this show: a Chinese open-weight coding model posting numbers competitive with the frontier, released the same week the US demonstrates it'll pull its own frontier models. Connect those two dots yourself. Every time the West makes its best models harder to get, the open-weight alternatives from elsewhere get a little more attractive to a procurement officer who just needs the work done. That's not a benchmark story. That's a geopolitics story wearing a benchmark's clothes.
So let me tie the whole thing together, because there is a thread, and it's not the tired one about regulation versus innovation. The thread today is about dependence, and who's holding the off switch. Anthropic learned that the government holds an off switch on its models. The downstream builders on Cursor and Harvey learned the off switch reaches them too. Meta's engineers are the off switch on Meta's own AI ambitions, 6,500 people who can quit. SpaceX's public investors are now an off switch on SpaceX's patience. KPMG found out that an unverified model is an off switch on its own credibility. Every one of these stories is about somebody discovering, the hard way, that a thing they thought they controlled actually had a switch, and somebody else's hand was on it.
The builders who come through this in good shape are the ones who, starting today, go look at their own stack and ask the uncomfortable question: where's my single point of failure, and whose hand is on it. If the answer is one model, from one lab, in one country, priced on one subsidy, you've got homework. If the answer is value-based pricing, a model-agnostic harness, and a verification layer between the machine and your customer, you can sleep a little easier. Not great, but easier.
That's the menu for today, kiddos. The government's got a kill switch and it's not afraid to use it, Amazon's fingerprints might be near the trigger, Meta built a city its own people want to flee, and Andrew Yang's over here quietly pointing at the one market nobody's bombing. Watch your dependencies, price your value, check your sources, and don't name your product Fable.
This is Tony DeLuca. Take care of each other out there, and I'll see you tomorrow.