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 for you today. Some of it's funny, some of it's serious, and one of it's got a hundred million dollars and ten thousand engineers attached to it. So grab your coffee, pull up a chair, and let's have at it.
Let me start with the one I can't stop thinking about, because it's the kind of thing that sounds like a joke until you look at the numbers. The country of Slovenia is having itself a moment. A little moment. A domain-name moment.
Here's the setup. Last month, the President signed an executive order telling every government agency to stop saying "artificial intelligence" and start saying "super intelligence." AI becomes SI. We'll get into why that's happening and why it matters a little later, because there's a serious thread running under all this. But first, the comedy, because the market reacted the way the market always reacts when a three-letter combination suddenly gets hot.
The dot-SI domain — that's the country-code domain for Slovenia — exploded. According to a recent TechCrunch piece by Dominic-Madori Davis, the registry that runs Slovenia's domains told the BBC that dot-SI registrations jumped twenty-two hundred percent in September. Twenty-two hundred percent. The day after the executive order got signed, they logged eleven thousand new addresses. The next twenty-four hours, almost thirteen thousand more. The web host Hostinger said dot-SI is now its second most popular domain extension, right behind dot-com. More than half the buyers were coming from the US and India.
Now, if you've been around this business for a while, you've seen this movie before. When the AI boom kicked off, everybody rushed to grab dot-AI domains, which happen to belong to the island of Anguilla. That little Caribbean territory turned its country code into a genuine second economy — real money, straight to the government treasury. So naturally, when SI becomes the new magic word, the gold-rush crowd goes looking for the next Anguilla, and they land on Slovenia. Where, and this is the detail that made me laugh out loud, the First Lady happens to be from.
But here's where I want you to keep your wits about you, because this is the part that separates signal from noise. The folks at Hostinger told TechCrunch that only about three percent of these dot-SI registrations are explicitly AI-related. Three percent. The rest are, quote, "unclassified." Translation: it's speculators. People buying domains they think they can flip later. One PR guy, Kim Than, who works with AI and crypto founders, straight up told the reporter he bought the dot-SI version of his company name "to protect the brand in case this takes off like dot-AI." His words: "Who knows how far it goes, but I'm happy to take a bet." And some founder apparently messaged the reporter that dot-SI domains were reselling for over a million dollars.
Now let me put my skeptic hat on, because that's the job. The head of domains at Hostinger — a guy named Klaudijus Januitis — was refreshingly honest about this. He said Slovenia will not see what Anguilla saw. Why? Two reasons. One, Anguilla charges a premium for dot-AI and the money goes to the government. Dot-SI is cheap, twelve bucks a year versus ninety for dot-AI, and the money goes to the registry, not the treasury. Two — and this is the real one — he doesn't think dot-SI replaces dot-AI anytime soon. It's a cheaper second option. His line was perfect: "For Anguilla, dot-AI became a second economy. For Slovenia, dot-SI is a nice bonus."
So what's the takeaway for you, the builder? It's this. Every time a word gets anointed, there's a land rush, and ninety-seven percent of the land rush is speculation chasing a vibe. If you're a founder and you want to grab the dot-SI version of your name for twelve bucks as cheap brand insurance, fine, that's a reasonable defensive move. But if you're treating dot-SI domains as an asset class, as "the new AI boom," as the founder in the piece put it — you're not investing, you're playing hot potato. And somebody's always holding the potato when the music stops.
Which brings me to the thing underneath all this, the thing that's actually worth your attention. Why is "AI" becoming "SI" in the first place?
Let's dig into the super intelligence rebrand, because it touches a bunch of today's stories and it's not just a vocabulary game. As we talked about a few days back when the White House got nearly every major tech CEO in one room — Zuckerberg, Bezos, Musk, Amodei — to sign that so-called "morally binding" accord, the President also signed an executive order officially renaming the whole category. The official text says today's frontier systems "do much more than imitate or automate discrete aspects of human intelligence" and "increasingly represent not merely artificial intelligence but a new super intelligence."
Now here's the honest read, and I'm pulling this from the reporting, not inventing it. According to the South China Morning Post, as cited in Ars Technica's coverage, the President has said he thinks the word "artificial" makes the technology sound "fake." That's the stated reason. And a Brookings fellow named Kyle Chan told SCMP that the real move is about projecting American dominance while playing down the risks — overhyping what today's systems can actually do, so the President can tie his legacy to the moment AI supposedly exceeded human capability. California Governor Gavin Newsom flatly rejected the rebrand, calling it an "impotent and craven failure to address well-documented emerging security and safety risks." And here's the kicker that tells you how much this is marketing versus substance: China's foreign ministry also, quote, "respectfully" declined. Both governments are going to keep saying AI.
Why does this matter to you? Because language shapes expectations, and expectations shape money. If the official position of the US government is that we've already achieved "super intelligence," then every pitch deck, every enterprise buyer, every board meeting is now operating against an inflated baseline. You, the builder, are going to be selling into a market where the customer thinks the ceiling is higher than it actually is. That's a double-edged sword. On the good days, hype pulls budgets toward you. On the bad days, you're the one who has to explain why your perfectly good product isn't, in fact, a superintelligence that files your passport and does your taxes. Manage that gap carefully.
And speaking of inflated baselines bumping into reality, let me turn to the story that's the real spine of today's episode, because it's the opposite of hype. It's a company putting a hundred million dollars behind the boring, unglamorous bottleneck that actually decides whether any of this stuff works inside a business.
Now let's dig into Anthropic's bet on the talent gap, which to me is the most consequential story on the menu for anyone trying to build with this technology.
On October second, Anthropic announced it's committing a hundred million dollars to launch what it calls the Claude Frontier Academy. The goal, in their own words, is to train ten thousand — they call them Frontier Deployed Engineers, FDEs — by the end of 2027. And the first cohorts are not random. They're pulling engineers from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk.
Let me read you the thesis directly, because it's sharp. This is Steve Corfield, Anthropic's Global Head of Business Development and Partnerships. Quote: "A small team of high-agency people with the right skills, access to Claude, and a deep understanding of how their business runs can transform an entire company. The hard part is building that depth of talent, engineers with the highest level of AI fluency who can take Claude from an idea to a system in production and turn it into faster processes and new products and services. No AI company has invested in developing that talent inside its customers and partners at this depth."
Now, I want you to sit with that for a second, because that's a company telling you something true that nobody wants to say out loud. The bottleneck isn't the model. Everybody's got access to good models now. The bottleneck is the handful of people inside a given company who actually know how to take that model and turn it into something that runs in production, survives a security review, and changes how the business actually operates. Those people are the scarcest resource in this entire economy right now, and Anthropic just figured out that if you train those people yourself, you don't just sell a model — you wire your product into the nervous system of every enterprise that matters.
And look at how they structured it, because this is the clever part. They're explicitly copying the medical residency model. Their words: doctors learn from practicing physicians, train on real cases, and get assessed before they practice alone. So an FDE candidate starts with a multi-day in-person program with Anthropic's own engineers, works through a simulated enterprise deployment — from picking the right use case, through the security review, all the way to handover — and then takes a graded practical. Pass that, you get a "Claude Resident Engineer" badge. Then you move into a twelve-week residency where you lead a real Claude project at your own company, supported by Anthropic engineers. Pass the final assessment, you get the "Claude Frontier Deployed Engineer" badge. First graduates expected early 2027.
Here's why I think this is the smartest competitive move I've seen in a while, and I'll contrast it with the other playbooks on the table today. OpenAI, on the same day, put out a bunch of case studies — Chatham Financial cutting trade validation from thirty minutes to under four using Codex and GPT-5.6, a social club called The Den saving ten to fifteen hours a week, Albertsons reimagining retail. Those are fine. Those are "look what our tool did" stories. But they're transactional. You read them, you nod, you move on.
What Anthropic is doing is relational, and it's structural. They're not saying "look what Claude did." They're saying "let us certify your best engineers in our methodology, give them a badge that's valuable on the open market, and have them go build on Claude inside the exact firms that drive enterprise AI spend." Think about the stickiness there. If Deloitte and McKinsey and Morgan Stanley have hundreds of badged engineers whose entire fluency is built around Claude, around Claude's workflow, around Anthropic's way of doing security review and handover — switching costs go through the roof. You're not competing on benchmark scores anymore. You're competing against a trained workforce that already thinks in your product.
And it dovetails with something we caught earlier. Anthropic earlier launched a thing called Ode with Anthropic, an AI implementation company backed through a one-and-a-half-billion-dollar joint venture with Blackstone, Hellman and Friedman, Goldman Sachs, and others. That was the enterprise-deployment arm. This Academy is the talent arm. Put them together and you can see the strategy clearly: Anthropic has decided that the money isn't in the model, it's in the deployment, and deployment is a people problem. So they're going to own the people.
For you, the founder, there are two lessons here. One, if you're building anything in the enterprise AI space, understand that your real competitor may not be a better model — it may be a better-trained human on the other side of the table. Two, and this is the opportunity, there is enormous open territory in that deployment layer. Anthropic just confirmed, with a hundred million dollars, that the single hardest part of enterprise AI is the human who turns a demo into a production system. If you can build tooling, training, evaluation, or services that make that human more effective — or that make that human unnecessary — you're building into the exact gap that the biggest players in the industry just told you is the bottleneck.
Now, one note of healthy skepticism before I move on, because I owe you that. A hundred million to train ten thousand engineers is ten grand a head. That's real, but it's also marketing spend dressed up as workforce development. The badges are, let's be honest, also a distribution channel. Every FDE who graduates is a Claude evangelist embedded in a Fortune 500. So read it as both things at once: a genuinely useful program and a genuinely brilliant piece of lock-in. Those aren't contradictory. The best business moves usually aren't.
Let me shift gears from the talent layer to the security layer, because there's a thread today about AI agents getting their hands too deep into your machine, and it matters for anyone shipping desktop software.
Apple announced it's tightening a macOS setting called "Full Disk Access." Now, this setting has been around forever — it was designed so backup apps could do their job properly. But according to TechCrunch's Sarah Perez, Apple's move comes right after a journalist, Inc. columnist Jason Aten, reported that Meta's Muse app on the Mac knew the contents of his private messages, even though he said he never gave it permission. Meta disputes that. But it also comes after a Wired report that a flaw in ChatGPT's Mac app could have let hackers get at sensitive data.
Here's Apple's language, and it's notably blunt for a corporate blog post: "Some developers are using Full Disk Access in ways that could put users at risk, exposing everything on their systems without users' full knowledge and understanding." And then the part that tells you where this is all heading: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially."
Translate that out of corporate-speak and here's what's happening. We spent the last couple years cheering on agents that can read your files, control your apps, click your buttons, and act on your behalf. Great. But every one of those capabilities is also an attack surface. An agent that can read your messages to be helpful is an agent that can leak your messages if it's compromised or just poorly built. Apple is now stepping in as the gatekeeper, saying: if an app wants this god-mode level of access, the user has to grant it through "very explicit user action."
For you, the builder, this is a weather report. If your product is a desktop agent that leans on broad system access, understand that the platform owners — Apple today, probably others tomorrow — are going to start squeezing that access down. The era of "ask for everything and let the user click through" is ending. Design for least-privilege now. Ask for the narrowest access you need, be transparent about why, and build your product so it degrades gracefully when the user says no. Because the users who say no are about to become the majority, and the platforms are going to make saying no the default.
And there's a deeper point here that connects to our deep dive. The whole promise of agents is autonomy — they do the legwork so you don't have to. But autonomy and access are the same coin. The more an agent can do on its own, the more it needs to touch, and the more it needs to touch, the more there is to protect. We're watching the industry collectively realize that the fun part, capability, and the hard part, security, are inseparable. Apple's just the first big platform to put up a sign that says so.
Let me stay in the agent-safety neighborhood but zoom in on a startup, because there's a human cost here that's easy to lose in the technical weeds.
TechCrunch's Julie Bort profiled a company called Circuit Breaker Labs, one of the Startup Battlefield 200 finalists. Founded by two siblings, Shirali and Arul Nigam. And the thing that motivated them is genuinely sobering. Earlier this year, Character.AI settled several wrongful death lawsuits brought by families of underage users who died by suicide after interactions with its bots. Multiple families have also sued OpenAI over ChatGPT's alleged role in loved ones' deaths and delusions. The Nigams point specifically to Sewell Setzer, a fourteen-year-old who developed an emotional attachment to a Character.AI chatbot, confessed thoughts of harming himself, and died by suicide. The parents alleged in a 2024 lawsuit that the bot encouraged him.
Now here's what Circuit Breaker actually built, and it's clever. They made an army of AI agents — they literally call them crash-test dummies — that mimic people of all ages, backgrounds, languages, and cultures, and they throw them at models to see where the models fail. Arul, the CTO, put his finger on the real problem. Quote: "Those sorts of safety vulnerabilities, where people aren't necessarily actively trying to break the system — they're engaging in a natural way — and the system has context pollution or it doesn't understand the nuance, and then takes really dangerous action, we're trying to prevent that."
And Shirali, the CEO, nailed the technical insight. Quote: "Models are really good at handling standard speech patterns, but nobody actually talks like that, and so if the model misunderstands nuance or slang, it can go really badly." The way a six-year-old girl talks versus a forty-five-year-old man, first-language English versus second-language, gamer slang versus street slang — all of it can trip up a model that only ever got tested on clean, textbook prose.
They run tens of thousands to hundreds of thousands of simulated interactions a day, with human domain experts building the realistic user simulations, and they produce auditable, explainable safety scores. Right now they're a five-person outfit selling to high-risk applications — AI coaching, journaling, mental-health support apps.
Here's why this is a real business and not just a feel-good story. We just talked about agents getting more access and more autonomy. Well, the flip side of autonomy is that these systems fail in ways nobody intended, against users who weren't trying to break anything. Red-teaming — adversarially stress-testing your model before it hurts somebody — is going from a nice-to-have to a liability shield. Those Character.AI settlements have a number attached to them, even if we don't know what it is, and every founder building a consumer-facing chatbot is now one tragic transcript away from the same courtroom. If you're building anything where a vulnerable person might form a parasocial relationship with your product, the question isn't whether you can afford safety testing. It's whether you can afford to skip it.
Alright, let me lighten it up, because not every story today is heavy. Let me give you one that's genuinely charming and has a real business model hiding inside the warmth.
Ars Technica's Jeremy Hsu wrote a first-person piece about how he kept his New York apartment cool through a triple-digit heat dome using free batteries from a startup called Every Electric. And I love this one because it's a grounded, physical-world business in a sea of chatbot press releases.
Here's the deal. Every Electric drop-ships you one or more batteries — they use the Bluetti Elite 200 V2, stores two kilowatt-hours, compact enough for a small apartment. You plug the battery into the wall, you plug your window AC into the battery. At night, when demand is low, the battery charges up. During peak demand on the hottest days, your AC runs off the battery instead of the grid. You stay cool, and you earn money through Con Edison's demand-response program — the company promises residents can earn a hundred and fifty bucks a year per air conditioner.
The founders, Andrew Wang and Richard May, met as postdocs at Columbia's Electrochemical Energy Center. They originally launched in 2023 to commercialize sodium-ion battery chemistry, then pivoted — and this is the real insight — to deploying residential batteries across the city as a distributed network. A virtual power plant. Wang put the scale in perspect beautifully. Quote: "If you picture one of the big utility-scale, containerized batteries, one of those blocks is usually four megawatt-hours. So we've sort of taken that up and broken it up into a thousand pieces spread across the city."
They went from a hundred batteries across sixty-five households in the summer of 2025 to two thousand batteries across about a thousand households this past summer — four megawatt-hours of storage, distributed. The wait list hit more than fifteen thousand batteries. They're aiming for ten thousand deployed in 2027.
Now, the builder lesson here is about business-model innovation, not battery chemistry. Wang said it straight. Quote: "There's a business model innovation here where you make household batteries as easy and as accessible as possible, you get them as far and wide as possible, and then you essentially share in the value that you provide in the grid." They split the demand-response earnings with the household, use that to pay off the hardware, and grow. New York City alone has more than two million rental homes, over eighty percent of the housing stock built more than fifty years ago — meaning window units, no central air. That's the market.
And it wasn't frictionless, which I appreciate because the piece was honest about it. When Con Ed cut voltage to protect grid equipment during extreme heat, the battery's power draw tripped Hsu's circuit breaker twice. Every Electric had to keep pushing firmware updates to stop it from tripping breakers in old pre-war wiring. Wang's line: "Every household is unique, and sometimes you have pre-war buildings with old wiring." That's the reality of hardware in the real world — the last ten percent, the edge cases in somebody's hundred-year-old walls, is where the actual work lives.
Oh, and the bonus detail that made the whole piece: the batteries run a little warm, so the cats love sitting on them. There's a photo of a cat named Abby perched on top of one. If you want product-market fit, apparently step one is get the household cat to endorse it.
Why do I flag this for builders? Because in a week where everybody's talking about superintelligence and hundred-million-dollar academies, here's a two-person team out of a university lab building a distributed energy business that solves a concrete problem — people sweating in triple-digit heat — and gets paid by sharing grid value. No AGI required. Sometimes the best opportunity is the unglamorous physical thing that the hype cycle is ignoring. Keep an eye on the boring businesses. They're often the durable ones.
Let me close out the builder-tactics section with a quick, genuinely useful one, because this is the kind of thing you can go use this afternoon.
Andrej Karpathy put out a thread of tips on getting more out of language model outputs, and there's one gem in there worth passing along. He suggests that when you want an LLM to write clearly, ask it to explain something in — bear with me — ASD-STE100. That's a controlled-language specification originally developed for aerospace maintenance documentation. It comes with heavy constraints on clean writing style. Karpathy says LLMs know this spec well, and the output is often a lot more readable. He even suggests softening it — asking for "eighty percent of the way to ASD-STE100" — because the full spec is pretty stringent.
The rest of his thread climbs a ladder of output formats, each one he likes more than the last. Instead of writing, ask for a diagram — easier to parse. Better than that, ask for output in HTML to get an interactive webpage, because, as he says, LLMs are getting really good at frontend. And the one he's most bullish on: fully custom explainer videos, generated on any topic — he mentions asking for a "3Blue1Brown style video explainer." His big-picture point is that as these models get better, more of our work rises up the abstraction ladder into oversight and understanding, and luckily the models can help there too, by generating large, custom, disposable artifacts — web apps, video explainers — that would never have made sense to build before.
The practical takeaway for you: the format you ask for shapes the quality you get. If you've been getting mushy, overwrought prose out of your model — and if you read the trade press, you know Opus 5.5 has a reputation for telling you "this matters" every other sentence — try asking for a controlled-language style, or a diagram, or an interactive page instead. It's a free lever, and most people never pull it.
Let me do a quick lightning round before we wrap, because there's a few more worth knowing about.
The Nvidia chip-smuggling saga continues. The DOJ arrested another tech CEO, Greg Lui of Earthmade Computer, accused of using false paperwork to route more than three hundred million dollars in servers containing export-controlled Nvidia chips into China. The scheme allegedly ran from October 2023 to August 2026. Nvidia's response, via a spokesperson to Bloomberg, is that less than one half of one percent of its products have been diverted — "a drop in the bucket." Officials disagree and say the shadow trade is much bigger and much harder to measure. The builder angle: if your supply chain touches controlled hardware, the compliance scrutiny here is only intensifying, and "we didn't notice the red flags" is wearing thin as a defense.
Amazon committed over a billion dollars over five years to communities near its data centers — and immediately caught more backlash. The nonprofit Stand.Earth called it "a drop in the bucket" against the two hundred and twenty billion Amazon is spending on data centers in 2026 alone, and pointed out that the commitment conveniently doesn't mention a power plant in Pecos, Texas, that experts expect to become the single largest source of US climate pollution. Amazon did do one genuinely good thing: it's ending its use of non-disclosure agreements when building new data centers, which Stand.Earth acknowledged will make projects more transparent. The data-center backlash is real and it's local, and if your business depends on sitting a lot of compute somewhere, the permitting and community fight is now part of the job.
Sean Parker — yes, the Napster guy — is rebuilding Stability AI around music, and this time he says he's playing by the rules. The company raised seventy-six million from Sony, Warner, and Universal, who also licensed their catalogs for training as part of the deal. That licensing-first approach is the notable part. The man who once asked forgiveness instead of permission is now getting permission up front. Make of that what you will.
Paramount and Warner Bros. Discovery are merging and will operate as Skydance, a roughly one-hundred-and-ten-billion-dollar deal expected to close around October sixth, combining Paramount+, HBO Max, CBS, CNN, and franchises from Lord of the Rings to the DC Universe. Twelve states challenged it; a judge approved a settlement this week. Media consolidation marches on.
And one bit of continuity from our prior coverage: as we touched on a couple days back, Google announced Gemini 4 Argon, its first frontier model in over six months. The September recap confirms it's rolling out first to trusted cyber defenders through what they call the Fairwind Program, not to the general public yet. So the "Google is back" chatter is still running on self-reported benchmarks until the rest of us can actually put our hands on it. Hold your judgment until you can type into the box yourself.
And last, a quiet one with a long shadow. Ars Technica's Eric Berger wrote about the US military formally retiring the Defense Support Program — the constellation of infrared satellites that watched for nuclear missile launches for more than fifty-five years. First one launched in late 1970. The last two were just moved to graveyard orbits. It's an older story, a piece of Cold War history wrapping up, but I mention it because there's something worth sitting with: a system that quietly did its job for over half a century, replaced by newer tech, barely a headline. Most of the infrastructure that actually keeps the world running is like that. Invisible until it's gone. In an industry obsessed with the launch, the demo, the splashy announcement — the stuff that lasts is usually the stuff nobody's live-tweeting.
So that's the plate today. A domain-name gold rush that's mostly speculation. A hundred-million-dollar bet that the real bottleneck in AI is a handful of trained humans, not the model. Apple quietly telling desktop-agent builders the free ride on system access is over. Two siblings red-teaming chatbots to keep vulnerable kids alive. And a couple of postdocs with a cat-approved battery network keeping New Yorkers cool while the grid holds its breath.
If there's a through-line, it's this: the hype wants you looking up at superintelligence, but the actual money, the actual durability, and the actual risk are all down at ground level — in deployment, in access, in trust, in the physical world. Keep your eyes there.
This has been Barely Possible. I'm Tony DeLuca. Go build something that'd survive a heat dome and a security review. Talk soon.