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 AI and tech morsels cooling on the windowsill today. Grab your coffee, pull up a chair, and let's get into it, because there's a lot on this menu and I want to make sure you leave full but not stuffed with junk.
Let me tell you where we're headed today so nobody gets lost. We're going to spend real time on a court ruling that says the government can't just slap a "national security risk" label on a company because it doesn't like being criticized. That's our main course, and if you build anything that touches the federal government, you're going to want to hear it. After that, we've got Andreessen Horowitz putting a billion-one behind steel and copper instead of code, an Anthropic paper where the AI is doing its own alignment homework at four bucks an hour, the surveillance company Flock finally hitting a wall, and a few smaller bites to round out the plate. Let's have at it.
So let's start with the one that I think matters most for anybody in this audience who's building a company, because it's about the rules of the road when your customer is Uncle Sam. And it connects to something we've been circling for a couple weeks now.
A federal judge in California ruled Thursday evening that the Trump administration's designation of Anthropic as a supply-chain risk was illegal. This is TechCrunch reporting from Rebecca Bellan. Now let me back up and set the table here, because you need the whole picture. Earlier this year, Defense Secretary Pete Hegseth and President Trump labeled Anthropic — the company that makes Claude — a supply-chain risk. And that's not some slap on the wrist. That label ordered every federal agency, not just the defense folks, but everybody, to stop working with the company. You get branded a supply-chain risk, and the entire federal government is supposed to slam the door on you.
Now why'd this whole fight start? Here's the part that's actually about you, the builder. The dispute came out of Anthropic drawing hard lines on certain safety guardrails. Specifically, the company wouldn't let the Pentagon use its models for fully autonomous weapons and for mass surveillance of American citizens. Anthropic said no. The Pentagon turned around and said, hey, we bought these models, we paid for them, and now you're trying to control how we use our own property. That was the standoff.
And U.S. District Judge Rita Lin looked at all of it and did not mince words. She called the government's action "unlawful retaliation" in violation of the First Amendment. She said it was "arbitrary and capricious." And she said Anthropic was denied due process under the Fifth Amendment. That's a triple. That's not a judge finding a technicality. That's a judge saying the whole thing was rotten.
But here's the part I want you to really sit with, because this is where the judge basically caught the government contradicting itself. Lin pointed out that the label made no sense next to everything else the government was doing. On one hand, they're calling Anthropic a threat to national security. On the other hand, Hegseth had floated applying the Defense Production Act to the company — and as the judge noted, that would mean the company was essential to national security, not a threat to it. You can't have it both ways, fellas. Is it a threat or is it essential? Pick one. And on top of that, the Department of Defense was still out there pursuing a contract with the company, and the government was collaborating with Anthropic's newer model on cybersecurity work. So the deeds didn't match the words.
The judge wrote, and I'll quote her here because it's worth hearing exactly: "The empty invocation of national security is not a blank check to punish and retaliate against government critics." And she added that the government's own "words and deeds confirm that the challenged actions were based on a desire to make a public example out of Anthropic for its 'arrogance' in criticizing the government."
Now let me be straight with you about the age of this. The underlying event — the labeling, the lawsuit — that goes back to March of this year when Anthropic filed two complaints, one in California, one in Washington D.C. This ruling is the fresh piece. And it's only the California suit. The D.C. case is still going. So this is a first win, not a final win.
Here's why I'm making this the center of the plate instead of, say, a fund announcement. If you're a founder, you probably think about the government as a customer or as a regulator. What this case is really about is whether a company can set terms on how its product gets used, and whether the government can retaliate when you do. Anthropic drew an ethical line — no autonomous weapons, no mass surveillance of Americans — and the government tried to financially strangle them across every agency for it. A judge just said you can't do that. That's a real precedent for anybody whose values might someday put them crossways with a powerful customer. You are allowed to say no, and the invocation of "national security" doesn't automatically override your rights.
And I'll tie this back to something. On Thursday we talked about the wave of companies — OpenAI, Anthropic, Google, over a hundred of them — signing a joint call for action to defend against rogue AI. And I noted then that these labs are trying to position themselves as responsible actors while the government's relationship with them stays messy. Well, here's the mess in a courtroom. The lesson isn't that Anthropic is a saint. The lesson is that the guardrails a company chooses to enforce are becoming a live legal and commercial battleground, and the courts are starting to referee it. Watch the D.C. case. If Anthropic wins that one too, you've got a pattern, and a pattern is what turns one ruling into a real shield for the next company that tells the Pentagon no.
Alright, let me shift from the courtroom to the checkbook, because there's a venture story today that tells you something about where the smart money thinks the bottleneck is.
Andreessen Horowitz — a16z — launched a new fund. They're calling it the "Machine Age" fund, and they raised one-point-one billion dollars for it. Sean O'Kane wrote this up over at TechCrunch. And here's what makes it interesting, and it's not the dollar amount, because a billion-plus is almost a Tuesday for these folks now. It's what they're pointing it at. The whole stated goal is to "open the throttle and accelerate the physical buildout of AI."
Physical. That's the word. a16z made its name and its billions on software — the beauty of software being that once you write it, copying it costs basically nothing. That's the whole magic of the software business. And now they're deliberately, in their own words, bucking that focus to go put money into hardware. Chips, memory, data centers, robots. In the post they wrote, and I'm reading it because it's a laundry list that tells you exactly what's keeping people up at night: "We need faster, more efficient systems. We need cheaper and higher-bandwidth memory across the memory hierarchy. We need faster and more scalable interconnects between nodes and systems. We need power efficient edge devices." And then the cooling, the materials, the electrical, the real estate.
Now here's my read for you as a builder. When the most famous software-first venture shop on earth stands up and says, we need to go pour a billion dollars into cooling systems and electrical buildout and memory chips, that's a tell. That's them saying the constraint on AI right now isn't ideas, it isn't models, it's the physical world's ability to keep up. It's power. It's real estate. It's whether you can actually get enough high-bandwidth memory, which — and this connects to a smaller story I'll get to — is in genuine shortage right now.
And I want to be honest and a little skeptical here, the way I try to be with all this stuff. a16z also called AI "the strongest tool ever developed for solving problems and bestowing abundance" and called its advancement "a social and national imperative." That's the kind of language that gets my eyebrow up. When a venture firm starts talking about "national imperatives," part of what they're doing is dressing up an investment thesis in flag-and-country clothes, which is very useful when you'd like the government to subsidize the buildout you're betting on. So take the mission talk with a grain of salt. But the underlying signal — money moving from bits to atoms — that's real, and if you're building anything in the infrastructure layer, cooling, power, memory, interconnects, there's now a very large, very well-connected buyer looking for you.
And speaking of the money chasing the physical buildout, here's a quick companion bite that shows you the machinery underneath. Lambda — that's the AI cloud company that buys chips and rents them out — just raised a billion dollars in private, short-dated debt to buy more Nvidia chips, which it's going to lease to Microsoft. Rebecca Bellan again at TechCrunch, citing Bloomberg. Now the older piece of this — the earlier credit facility — goes back to May, so that part's not new. What caught my eye is the scale of the whole game. According to data Bloomberg compiled, banks and tech companies have raised over four hundred billion dollars in AI-related debt globally in 2026 so far. Four hundred billion. In debt. This year.
Now here's the thing that a careful person notices. That's not equity. That's not people betting their own patient money. That's borrowed money, short-dated, with the bet being that you can deploy those chips fast enough and generate revenue fast enough to pay the loan back before it comes due. When you hear a16z pouring a billion into the physical layer and Lambda borrowing a billion against future chip rentals, you're hearing the same song from two directions — everyone's convinced the buildout can't come fast enough. I'm not going to tell you it's a bubble, because I don't have the crystal ball. But I will tell you that when the buildout gets financed on debt instead of equity, the margin for error gets thin. If demand hiccups, debt doesn't care about your feelings. Debt wants to get paid. File that one away and watch it.
Now let's move from the money to the machines improving themselves, because Anthropic dropped a research paper that's worth a couple minutes even for you non-researchers, and I'll keep it to the what and the why.
On Friday, Anthropic published a paper — this is Russell Brandom writing it up at TechCrunch — titled "Automated Researchers Can Reliably Mitigate Alignment Failures." Now don't let the dry title fool you, because what's underneath is genuinely a little eyebrow-raising. What they did was build an AI system whose job is to improve another AI model's behavior. Specifically, they gave it ten benchmarks for particular misaligned behaviors — think of these as ten different ways a model can go wrong — and the automated system was able to improve performance on every single one of them without wrecking the model's overall performance.
Here's how it works, and I'll keep it plain. This thing, led by an Anthropic fellow named Chen Yueh-Han, basically imitates what a human researcher does. It reads the available literature, it proposes a method, it trains the model on that method for about thirty minutes, checks if it worked, keeps the methods that work, throws out the ones that don't, and does it again and again. It's the scientific method, but running fast and running itself.
Now here's the two lines from the paper that made everyone sit up. The first: "The best AAR method" — that's their Automated Alignment Researcher — "beats what experienced humans propose, on average within six hours." And they went further: "Human guided research directions do not lead to stronger performance." Oof. And then, in case you weren't already feeling it, they put a price tag on it. Quote: "An AAR costs roughly four dollars per hour in API inference against the one hundred fifty dollars per hour we pay our human researchers." So the machine is faster, arguably better on these specific tasks, and roughly thirty-seven times cheaper an hour.
Now I want to be the guy at the kitchen table who says: hold on. Because the paper itself, to its credit, points out the catch, and the catch is a big one. This whole thing only works insofar as the benchmarks actually reflect the real alignment goals. In plain English: the machine is very good at winning the test you gave it. But if the test doesn't fully capture what you actually care about, then you've just built a very cheap, very fast machine for gaming a test. And who writes and maintains those benchmarks? Humans. Who maintains the research literature the thing reads? Humans. So the humans didn't disappear — they moved up a level, from doing the work to defining what "good" means.
Why does this matter to you as a builder and not just to the lab coats? Because this is the "recursive self-improvement" idea people have been whispering about — AI improving AI — showing up in an actual, modest, peer-reviewable form. It's early. It's narrow. But the direction is clear: more of the grunt research work getting automated, with the human role concentrating on defining the target. If you're running a lean team, that's the pattern to internalize. The leverage isn't in doing more experiments yourself. It's in being crystal clear about what you're actually optimizing for, because increasingly a machine can run the experiments for four bucks an hour, and it'll ruthlessly optimize whatever you point it at — including the wrong thing.
Alright. Let me take you from the labs to the street corner, literally, because there's a surveillance story with real momentum today, and it's one where I have to be straight with you about the timing.
Flock — the company that puts up those solar-powered license plate reader cameras all over the country — is having a rough month. New data compiled by an anti-surveillance group called Secure Justice, reported by Cyrus Farivar at Ars Technica, shows cities are canceling their Flock contracts at an accelerating pace. Now I want to be careful here, because we've circled Flock before — back on the eighteenth we talked about Wisconsin towns pulling out. So let me give you what's genuinely new. The numbers. Secure Justice counted 214 cities and counties that have dropped Flock since 2021. And of those, ninety of them — ninety — dropped it in August of 2026 alone. That's a fourfold jump over the previous month. Their executive director, Brian Hofer, looked at the chart and said, quote, "Looking at that image, it's gone parabolic."
And the reasons are the ones you'd expect if you've been paying attention. Out-of-control surveillance, unwanted data sharing, high costs, and reports of cops abusing the tool. And it's gone bipartisan, which is the part that tells you it's real and not just one side's talking point. Senator Josh Hawley, a Republican, announced he's investigating Flock. And Governor Ron DeSantis, also a Republican, gave a speech calling the cameras "out of control" in Florida after two cops there got arrested for misusing them. DeSantis said, and I'm quoting, "When I see, like, a police officer using this to track, like, an ex-girlfriend, I'm like, you know what the hell's going on?"
Now, in fairness, Flock isn't sweating publicly. Their spokesperson said new city partnerships are still outpacing the cancellations by roughly ten to one, and they point to over five thousand communities using the product. And there's a separate fresh piece here worth a beat — Zack Whittaker at TechCrunch reported a new YouGov survey of twenty thousand people where forty-six percent oppose these cameras in their communities and only thirty-eight percent support them. That's a reversal from last year, when the majority was in favor. Public opinion flipped.
Here's the builder lesson buried in this, because it's not really about cameras. It's about network-effect products that get sold on the promise of "everybody's using it, so you have to too." That flywheel spins beautifully on the way up and it spins the same direction on the way down. Once cancellations start clustering and it's an average of three localities a day pulling out, the "everyone's on board" pitch inverts. Now it's "everyone's bailing, why are we still in?" If you're building anything that depends on being the default that nobody questions, understand that the default status is exactly what makes you a target the second the mood shifts. Flock spent years being the thing nobody asked about. Now it's the thing everybody's asking about, and that's a very different business.
Okay, let me clear the plate with a few smaller bites before we wrap.
First, a nice one for the builders who feel like they're grinding uphill. TechCrunch's Dominic-Madori Davis had a piece on how Sweden built one of Europe's hottest startup ecosystems. You know the names already — Spotify, Klarna — but the newer wave is legal AI startup Legora and the vibe-coding tool company Lovable. Sophia Bendz, a general partner at Cherry Ventures, went on the Equity podcast and gave the simplest explanation for why it's working, and I actually love it. She said being a founder is now cool there. When she worked at Spotify years back, the cool thing to do was become a banker or go into consulting. Now, she said, "a lot of people optimize for creating stuff." The ecosystem raised two-point-eight billion this year, on track for five billion, and American investors are flying in to hand out term sheets. The through-line: the first generation of founders is now mentoring and funding the next, and people who worked at Lovable and Legora are peeling off to start their own things. That's how a scene compounds. Culture first, capital follows.
Second bite, a privacy tool that's actually useful. Sarah Perez at TechCrunch reports Brave's browser added email aliases — a feature that lets you sign up for websites without handing over your real email address. Any mail to the alias forwards to your real inbox, and Brave says it deletes the message from its servers after forwarding and doesn't read the contents beyond spam filtering. Why do you care? Because your email is the identifier that ad-tech giants use to stitch together everything you buy and view across the web, and it's the thing that leaks to data brokers when some random retailer gets hacked. Five free aliases to start. Small feature, real hygiene. If you're building consumer product, notice that "don't make me hand over my real identity" is quietly becoming a selling point.
And a last one that connects back to our hardware theme. Apple TV raised its subscription prices again — the fourth time in four years, now fourteen-ninety-nine a month, up from twelve-ninety-nine. Now the price hike itself is older news, went out earlier this year. But here's the thread I want you to catch: Amanda Silberling's piece notes that tech companies are also raising hardware prices because of RAM and component shortages driven by — you guessed it — the massive demand for hardware to build AI data centers. So the same shortage that has a16z raising a billion dollars for memory and interconnects is quietly showing up in your streaming bill and your gadget prices. The AI buildout isn't happening in some distant server farm you'll never see. It's reaching into the cost of the RAM in your next laptop. Everything's connected, kiddos. Follow the memory chips and you'll understand half the economy right now.
Let me pull it together. The spine of today, if you want one thread to walk away with, is about power and where it actually lives. A judge told the government its power to brand a company a threat has limits when it's really just retaliation. a16z is betting that the real power in AI now lives in the physical world — the chips, the cooling, the electricity — not the code. Anthropic showed a machine that can out-work its own researchers on a narrow task for four dollars an hour, which shifts the power from doing the work to defining the target. And Flock is learning that being the powerful default is only an asset until the room turns on you. Power, and who really holds it, and how fast it can shift. That's the plate.
That's the menu for today, kiddos. If you build in the infrastructure layer, this was your day. If you build anything with values that might cross a big customer, keep an eye on that D.C. courtroom. This has been Barely Possible. I'm Tony DeLuca, thanks for pulling up a chair, and I'll see you back here next time — save room.