Washington now picks who gets AI

The Frontier Got a Velvet Rope

For the first time, the most powerful AI models in the country aren't something you can just go buy. Whether you get the best model now depends on whether the government says you're cleared to be near it. And that one change is quietly handing the edge to the models nobody's bothering to gate: the open-weight ones, the models any company can download and run on its own machines for free.

It happened twice in three weeks. The White House gave Anthropic 90 minutes to pull Claude Fable 5 offline on national-security grounds, then personally signed off on the roughly 20 partners allowed near OpenAI's new flagship, GPT-5.6 Sol, and kept it from everyone else. A new executive order now routes the most capable models through up to 30 days of government review before release. The best AI in the country now ships on the government's say-so.

AI spent two years as a product race, smartest model, cheapest tokens, fastest ship date, and this month it turned into an arms race. Once a government decides a system can map its enemies' weaknesses, it stops treating it like software and starts handling it like enriched uranium or a fighter jet: classified, licensed, allies-only. Intelligence just joined the export-control list, and which model you're cleared to run depends on your passport as much as your credit card.

But you can only run an arms race on something you can actually hoard. Uranium, you can hoard. Open-weight AI is just math anyone can download, and the US bolted the door on its closed frontier in the very weeks Chinese open models hit good-enough parity on real work and started spreading for free. So the move meant to starve rivals of the best AI mostly hands the rival's own models the world's default instead. Every team that can't get Fable 5 or GPT-5.6 reaches for DeepSeek or GLM. Our read: wall off your own supply and the substitute on the shelf belongs to your competition, and for the company actually buying, available-and-good-enough wins.

The uncomfortable part is what all of this does for China. If you're Beijing, the prize was never a benchmark crown; it's becoming the default layer the rest of the world builds on, and you get there by giving your models away while Washington gates its own. You don't have to be better than American AI when you're the one that's actually in reach. Whether China planned this from the start or backed into it, US policy just made the bet pay off, and from here it's hard to tell which. Brilliant if it was on purpose, and more dangerous if it wasn't.

Lock your best weapon in the vault long enough, and you may look up to find the other side already won by handing theirs out for free.

We'll keep covering this one closely. My honest read: we're at the very start of a political fight that reorders who controls the most important technology of the decade. I hope I'm wrong.

Build the Plugin That Runs Your Morning

A plugin is a kit you set up once that gives Claude real skills and a memory of how you work. Instead of re-explaining yourself every morning, you hand it a task and it runs the whole thing in your voice, so the work you'd normally grind through by hand happens in the background while you do what actually needs you. Here's one to build today: the Daily Driver, a plugin that triages your inbox, scans your Slack, and hands you a prioritized to-do list plus first-draft replies, all from one command at 9 a.m. The person who built it isn't an engineer and never touched a terminal.

Under the hood, a plugin is just a folder of text files, and if you can write a Google Doc you can build one. Three plain-English pieces do the work: A skill is a saved instruction set, your SOP for one job: how you triage email, how you write a post, how you run a weekly report. A chain strings skills together so a single command fires all of them in order. A plugin is the folder that holds the skills plus the memory that makes the output sound like you: who you are, your voice, your team, your projects.

Ask Claude a one-off question and you get a smart answer you'll re-ask tomorrow. Give it your context once, in files it reads on every run, and it stops being a search box and starts handling the parts of your job you keep redoing by hand. You don't have to be the one who builds it, either; one person on your team can stand up the first version in an afternoon, and it quietly absorbs the busywork that eats everyone's first hour of the day. For anyone in sales, marketing, content, or ops, it's the best hour you'll spend this week.

1. Make the folder: A plugin is just a folder on your computer, no setup required. Everything else lives inside it.
2. Write your memory files: Three short docs in plain English. About Me (role, team, current projects), Brand Voice (how you write, with a few real examples), and Preferences (how you like work done). Claude pulls these into every run on its own, so you write your context once instead of re-pasting it into each prompt. This is the step that makes the output yours instead of generic.
3. Build one skill at a time: Start with email triage: tell it how you sort, what counts as urgent, what to ignore, and how to draft a reply. Then add the next one, an X writer fed your best posts, then Slack triage, then a daily productivity check-in.
4. Chain them into one command: Bundle the skills into a single "morning brief" so one trigger runs email, Slack, and your to-do list back to back.
5. Schedule it, then run it anywhere: Put the brief on a timer so it fires at 9 a.m. before you sit down (Cowork has a built-in scheduler; Claude Code runs it on a cron job, which is just a timer for code). The same folder works in Claude's desktop app, Cowork, and the Claude Code developer tool, so it runs wherever you already do.

Also worth your time

Spotify runs agents across 20 million lines of code

Spotify's chief architect Niklas Gustavsson dropped the number everyone's chasing: after they gave agents a way to check their own work, success rates jumped from 20-30% to around 80%. Good proof that what separates a flashy demo from real production is whether the agent can verify itself.

Meta can type words straight from your brain, no implant

Meta's new Brain2Qwerty v2 decodes full sentences from brain activity in real time without surgery, building on a v1 that just published in Nature. Still slow and stuck in the lab, but real-time, non-invasive brain-to-text went from science fiction to a measured benchmark this week.

OpenAI built its own chip

OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom inference processor, designed from scratch to run models cheaper than Nvidia GPUs and taken from design to tape-out in about nine months. When the company spending the most on Nvidia starts making its own silicon, that tells you where the real cost pressure is. (via TechCrunch)

Cheaper AI starts with understanding the work

Box CEO Aaron Levie made a sharp point on cutting AI costs: the tricks only work once you actually understand, in concrete terms, what you're asking the model to do, and the real savings come from a smart routing layer that sends each task to whichever model can handle it cheapest.

A free toolkit for seeing how AI agents are built

A developer shipped "tau", an open, teaching-focused project that cracks open the machinery behind the agent tools you use every day and walks you through building your own.

A prediction that agent "loops" will vanish

Cursor's David Pan figures that a year from now nobody will talk about agent loops (the back-and-forth where an AI tries something, checks its own work, and retries until the job's done), not because they flopped, but because they'll be baked into the tools and invisible. A useful gut-check before you over-engineer something that's about to become a default.

See you next Tuesday. Sven