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, welcome back. I'm your boy Tony DeLuca, and this is Barely Possible, where we sit down with the day's pile of AI news, push the hype off the table, and figure out what actually matters to you, the person who's gotta build something with this stuff. Pour the coffee, pull up a chair, we got a real menu today.
And I'm gonna start where the whole industry's center of gravity has quietly moved this week, because it's not a model launch and it's not a funding round. It's a question of who gets to hold the keys. Here's the thing that should make every founder sit up straight. The two best frontier models in America right now — Anthropic's Mythos and OpenAI's GPT-5.6 — are not being released the way these companies wanted to release them. They're being rationed. And the rationing is being done by the federal government, customer by customer.
Let me lay it out, because this developed in pieces over the last couple days and you need the sequence to understand the stakes.
First piece. TechCrunch reported, in a piece by Julie Bort, that the Trump administration has now cleared Anthropic's Mythos to be used by more than a hundred US companies and government agencies. Mythos 5, specifically. And here's the detail that jumps out — the authorization reportedly includes the non-American employees of those companies. So we've gone from a hard export block to a hand-picked allowlist. Over a hundred institutions get the good model. Everybody else waits.
Second piece, same window. OpenAI limited the rollout of GPT-5.6 after a government request. And this is where it gets interesting, because OpenAI didn't just quietly comply and smile. They put out a statement, and I'll read you the part that matters, reported by Rebecca Bellan. Quote: "We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." End quote.
Now read that again as a builder, because there's a lot packed in there. Cyber defenders. Global partners. Developers. That's not a safety argument. That's OpenAI saying: the people who need these tools to defend systems and build products are the ones getting starved while we wait for a customer-by-customer approval process to grind through.
And I want to be precise about what this approval process actually is, because there's a tendency to dress it up as policy. From the reporting around it, this is not a statute. There's no published rule. There's no agency with a clear mandate running a transparent review. It's the White House deciding, on an ad hoc basis, who gets frontier access and who doesn't. One commentary I saw this week — and I'll frame this as analysis floating around the discourse, not gospel — called it an "arbitrary, unknown, non-transparent license requirement," and made the point that this is actually worse than red tape. Red tape at least has a form to fill out. This is texts between executives and officials deciding access company by company.
Now here's the part that connects to you directly. Sam Altman made a point in his own messaging that the public delay is not the same as a development delay. The labs are still training. What's getting throttled is the release. So the gap — between what the frontier labs have running internally and what's available to you, the developer, the enterprise, the cyber defender — that gap doesn't close under this regime. It widens. Every week. The capability exists. You just can't have it unless you're on the list.
And that brings me to the deep dive, because there's a fresh TechCrunch piece this week, by Kate Park, that takes this story and follows the money out the door. The headline is plain: Asian AI startups are launching Mythos-like models as Anthropic's export ban drags on. And the line in there that should be tattooed on a whiteboard at every US lab — quote — "U.S. AI labs may never recover this enormous market."
Let me sit on that for a second, because it's the most consequential sentence in today's pile.
Here's the dynamic. You've got an enormous market — call it the entire non-US, non-allowlist world — that wants frontier-grade capability. The US just told a huge chunk of that market: you can't have our best stuff, and even if you're a US company, your overseas employees might be a problem. So what happens? Nature abhors a vacuum and so does enterprise procurement. New models are launching across Asia that promise Mythos-like capabilities with one enormous selling point attached: no export ban hanging over your head. No risk that the model you built your product on gets yanked next quarter because Washington changed its mind.
And I want you to feel why that second thing — the no-ban promise — might matter more than the raw capability. If you're a founder picking a foundation model, you are making a multi-year bet. You're building your product, your data pipelines, your fine-tuning, your whole stack, on top of somebody's weights and somebody's API. The single worst thing that can happen to that bet is that the model becomes politically unavailable to you for reasons that have nothing to do with your business. That's not a performance problem you can engineer around. That's a rug-pull risk baked into your supplier.
So when an Asian startup walks in and says, look, we're a few points behind on this benchmark, but you will never wake up to find out a government decided you're not on the list — for a lot of buyers outside the US, that trade is a no-brainer. Sovereignty and supply reliability beat the last few percent of capability. We've seen this movie in other industries. When a supplier becomes a geopolitical liability, customers don't admire your engineering. They diversify away from you, and they don't come back when the dust settles, because the trust is gone.
And this dovetails with something the background chatter has been picking up for weeks. Larger enterprises are increasingly moving to secure their own compute and post-train their own models in-house, frequently on top of open-weight bases. The framing I keep hearing is that everyone's starting to understand how open source wins — not because open weights are technically superior, but because they can't be taken away from you. When the closed frontier becomes conditional on a government allowlist, the open base that you can run yourself, on your own metal, stops being the budget option and starts being the strategic one.
There's a companion piece to all this from Russell Brandom at TechCrunch, and the title says the whole thing: It's not about Anthropic vs. OpenAI anymore. His argument is that AI models have crossed into a zone where their capabilities have genuine political consequences, and dealing with those consequences is going to require collective action — not two labs sniping at each other for market share. I think that's right, and I think it's uncomfortable for everybody, because the labs spent years positioning this as a horse race. Now the race is happening inside a fence the government built, and the fence is the story.
Let me tie this back to where we've been on this show, because this isn't out of nowhere. We covered the front end of this saga earlier in the month — the export-control kill switch on Anthropic's Fable and Mythos over a jailbreak, and the whole question of how Anthropic may have talked itself into an export ban. This week is the next chapter of that exact arc. The kill switch became an allowlist. The allowlist became a customer-by-customer approval regime. And the predictable consequence — buyers in the locked-out market building or buying alternatives that can't be locked out — is now showing up in the launch announcements.
So if you're building, here's the actionable read. One: assume frontier access is now a political variable, not just a pricing variable. Budget for it. Two: if your product can run on an open-weight base you control, that's no longer just a cost decision, it's a continuity decision. Three: if you're selling outside the US, your customers are already thinking about supplier risk, and the American-frontier-model dependency that looked like a feature last year might look like a liability this year. None of that is hype. That's just reading the room.
Now let me shift from who controls the models to who's building the silicon underneath them, because that's the other structural story this week.
There's a TechCrunch video piece, by Theresa Loconsolo, framed around a simple question: why is everyone from OpenAI to SpaceX building their own chips and turning up the heat on Nvidia? And the news hook is that OpenAI shared plans for its own custom inference chip — they're calling it Jalapeño, built with Broadcom. That puts OpenAI alongside Google, Apple, and SpaceX in the club of companies designing their way out of single-supplier risk.
Now, I'm not gonna take you into the weeds on chip architecture, that's not what this show is for and it's not what you need. The builder takeaway is the strategic one. For years, if you were doing anything serious in AI, you had exactly one critical supplier, and that supplier had all the leverage. That's a dangerous position for a giant company and a worse one for a startup. So the biggest players are spending real money to build alternatives — not because they think they'll beat Nvidia at Nvidia's own game tomorrow, but because optionality is worth a fortune when one company controls the thing your entire business depends on.
Notice the pattern that keeps recurring today: dependency is the risk. Frontier model access, controlled by the government — dependency risk. Chip supply, controlled by Nvidia — dependency risk. The whole industry is quietly trying to de-risk its single points of failure, and that instinct should be in your head too when you architect anything that matters.
There's a related move on the chip-software side. Qualcomm is reportedly planning to acquire Modular, the chip-software-platform company, in a deal worth nearly four billion dollars, set to close in the second half of this year. And the name attached to Modular is the interesting part — it was founded by Chris Lattner, the guy who created LLVM and the Swift programming language, and who briefly ran Tesla's Autopilot software. I'll flag the timing honestly: I don't have a hard confirmed-close date in front of me, this is a planned acquisition. But the strategic logic fits the same theme. Qualcomm is buying its way deeper into the software layer that makes chips programmable, because the moat in silicon isn't just the silicon — it's the toolchain people build on top of it.
And in the same neighborhood, Qualcomm is reportedly planning to design a China-specific data-center chip built to clear US export limits. So you've got the same export-control gravity well bending the hardware roadmap too — companies designing products specifically to thread the needle of what they're allowed to sell where. Everything this week keeps coming back to the same fault line: where the line is drawn between what's exportable and what isn't, and how every player is reorganizing around that line.
Let me pull us over to a story that's a different flavor — the human one — because it's worth your time and it cuts against the cynicism.
TechCrunch ran a piece by Connie Loizos about a founder named Connor Christou, described as the fittest guy in the room, who got hit with a cancer diagnosis. And what he did was take everything — his blood results, his scan data, his wearable output, his journal entries — and feed all of it into Claude. He used the model as a way to synthesize a flood of medical information that no normal person can hold in their head at once.
Now I want to be careful here, because this is exactly the kind of story that gets oversold. This is not a model curing cancer. This is not a substitute for oncologists. What it is — and what's genuinely interesting for builders — is a real example of AI as a synthesis layer over a person's own scattered data. Blood panels in one place, scans in another, a wearable streaming numbers, a journal nobody's correlated. The model's job wasn't to be smarter than the doctors. It was to assemble a coherent picture out of fragments and help one frightened person ask better questions and track what was happening to his own body.
And that's a pattern worth noticing, because it's the same pattern that works in boring enterprise settings. The value isn't always the model being a genius. A lot of the time the value is the model being the thing that finally pulls all your scattered context into one place so a human can reason about it. If you're building product, that synthesis-over-fragmented-personal-data use case is real, it's sticky, and people will care about it a lot more than another chatbot. Just don't let your marketing department turn it into miracle claims, because the minute you oversell health AI, you've earned every bit of the backlash that comes.
From one person's data to a whole nation's combat doctrine — let me get to a story that genuinely stopped me.
Ars Technica, in a piece by Jeremy Hsu, reports that South Korea plans to train its entire military — a half-million-strong force — as, in their words, "drone warriors." The idea is to treat the drone as a universal combat tool, something every soldier is trained on, not a specialty for a small unit.
I'm not a defense analyst and I'm not gonna pretend the geopolitics is my lane. But the thing that's worth chewing on, even from a tech-business seat, is what it signals about how fast cheap autonomous and semi-autonomous hardware has moved from exotic to standard-issue. When a major military decides the drone isn't a specialist weapon anymore but a basic tool every single person carries, that's the same democratization curve we talk about with software — except pointed at the battlefield. The cost of the capability dropped far enough, and the capability got good enough, that it stops being elite and becomes universal. Hold that thought next to everything we just said about chips and models, and you see the same physics: the capability gets cheap, it spreads everywhere, and the institutions scramble to organize around the new normal. The only difference is the stakes.
Now let me move us to the talent-and-hardware corner, because there's an Apple-to-OpenAI move that fits a trend we've been tracking.
TechCrunch, reporting from Anthony Ha, says Paul Meade — the Apple vice president in charge of the Vision Pro headset — is reportedly leaving to join OpenAI's hardware team. And I want to handle this carefully, because the word is reportedly, and we don't confirm departures off a vibe. But this is a named, sourced report of a specific executive making a specific move, so it clears the bar.
Here's why it matters beyond the gossip. OpenAI is building hardware. We've talked about their device ambitions before. Pulling the person who ran Apple's most ambitious wearable hardware program is a signal about how serious that effort is and what shape it might take. The Vision Pro, whatever you think of its commercial run, was a genuinely hard piece of hardware engineering — sensors, displays, spatial computing, the whole stack. If OpenAI is hiring that DNA, they're not building a novelty. And it's another data point in the larger story we keep coming back to: the most ambitious AI companies don't want to just rent the model layer, they want to own the device the model lives in. Vertical integration, top to bottom. Same de-risking instinct, pointed at the consumer's pocket.
Let me give you a quick run through a few more that deserve a mention but not a sermon.
The New York Times sharpened its copyright fight against Microsoft and OpenAI. Ars Technica, in a piece by Ashley Belanger, reports the Times is now alleging that Microsoft built a copyright-infringing supercomputer specifically to help OpenAI — and the framing shifted after a Supreme Court ruling against Sony. I'm not gonna give you legal predictions, but I'll tell you what to watch: the theory of the case is moving upstream, from "the model output infringes" toward "the entire infrastructure that produced it was built to infringe." If that framing gets traction, it changes the exposure calculus for everyone who builds big training infrastructure, not just the named defendants. Founders training on scraped data should keep an eye on where this lands.
On the geopolitics-of-platforms front: Russia is telling its citizens to "switch to Android" after Apple blocked key Russian apps, per Nate Anderson at Ars Technica, with the Russian government calling Apple's decisions "bizarre." The interesting part for builders is the reminder that platform owners are now instruments of geopolitics whether they like it or not. When your distribution runs through somebody's app store, that store's compliance decisions become your business continuity problem. Same dependency lesson, different shape.
There's a reported FTC clearance for Elon Musk to acquire a SpaceX-alumni startup called Mesh — and I want to be upfront about the framing here, because this is an older thread resurfacing. Mesh came out of stealth back in February with a fifty-million-dollar Series A. So the company isn't new; what's circulating is the regulatory green light. I'll leave it there rather than dress it up as breaking news, because it isn't.
And then a smaller one that's a little funny and a little ominous for anybody shipping AI-built code. TechCrunch, via Julie Bort, covered Corgi — a Y Combinator-backed insurance-tech startup — pushing back against an accusation from a company called Papermark that Corgi stole its open-source software. Corgi says it didn't. And the reason this is worth ten seconds of your attention is the new wrinkle it raises about vibe coding. When you generate a big chunk of your product by prompting a model, and the model has been trained on a universe of open-source code, where exactly is the line between inspiration and appropriation? Whose code is in your codebase, really? This is going to be a recurring fight, and "the AI wrote it" is not going to be a magic shield in a courtroom. If you're building fast and loose with generated code, at least know what's underneath you.
Let me close the loop with a couple of lighter builder signals, because not everything is fences and lawyers.
Simon Willison made an observation this week that I think a lot of you will nod at. He noticed that today's models are much less likely to default to building everything in React than they were a year ago. He used to have to write "don't use React" in basically every frontend prompt; now he mostly doesn't have to. And he was honest about the uncertainty — he wasn't sure if it's a genuine shift in the models' default preferences or just that he's usually prompting inside existing projects where the model is smart enough not to drag React into a codebase that isn't using it. He tested it in a fresh project and got vanilla HTML and JavaScript back.
Why do I bother telling you this? Because it's a small, concrete sign that these tools are getting better at reading context and matching the conventions of the thing they're working in, instead of bulldozing every project toward one default framework. For anybody using AI to write real code in a real codebase, that contextual restraint is worth more than another point on a benchmark. The model that respects your existing stack saves you more hours than the model that's slightly smarter in the abstract.
And one last builder-economics note that ties neatly back to where we started. There's a sharp observation from swyx this week about how we measure model performance. The idea is this: if you hold the inference budget constant when you report evals — measuring by dollars spent on inference rather than by raw token counts — then open models suddenly look very different. Because open models can give you a lot more dollar-per-token mileage on popular inference providers than closed-model APIs. So anyone launching an open model, or anyone leaning toward open for cost reasons, has a real incentive to report performance measured in dollars of inference rather than tokens.
Why does that matter to you? Because it's a reframing of the whole "is open good enough" question. Stop comparing models on abstract capability and start comparing them on capability per dollar at the budget you actually have. When you do that math, the open option often wins on the only axis that hits your runway. And that's the same conclusion the enterprises securing their own compute keep arriving at, just from a different door. Capability is the headline. Cost-per-useful-output and supply you can't lose are what actually run your business.
So here's the thread that ran through the whole episode. Every big story today was secretly the same story: dependency is the new risk. Depend on a frontier model and the government can ration it. Depend on one chip vendor and you're hostage to their roadmap. Depend on a foreign supplier and a sanctions decision can break you. Depend on somebody's app store and their compliance team becomes your continuity team. The winners this cycle are the ones quietly building optionality — their own chips, their own post-trained models, open bases they can run on their own metal, supply they can't have taken away. That's not a doom-and-gloom message. It's a build-list. Look at your own stack and ask, honestly, what's the single thing that, if it got pulled tomorrow, would sink me? Then go make that thing replaceable. That's the move.
That's the menu for today. I'm Tony DeLuca, this has been Barely Possible. Be skeptical, be specific, and don't let anybody build your business on a thing they can take back. Catch you next time.