In this edition
Happy 4th of July, fam 👋
The government spent this week on both sides of the table. One lab offered Washington a seat on its cap table; the other finally got its model back from Washington after 19 days in the dark. However you feel about it, the U.S. government is no longer watching the AI race from the stands — it's on the field.
🏛️ OpenAI floats handing Washington 5% — and wants every big lab to do the same.
🔓 The Claude Fable 5 blackout is over: 19 days, worldwide, one directive.
🧨 “BioShocking”: researchers talk AI browsers into leaking your credentials.
🛠️ Five high-impact tools you've (probably) never heard of.
🪦 Fresh graveyard arrivals + a who-owns-the-frontier ethics read.
🏛️ OpenAI offers Washington 5%
Sam Altman has discussed handing the U.S. government a 5% stake in OpenAI — worth roughly $42.6B at the $852B valuation it set in March. The bigger idea: every leading U.S. lab contributes the same slice into an Alaska Permanent Fund-style vehicle paying dividends to Americans. Altman has reportedly floated it to Trump, Commerce Secretary Lutnick, Treasury Secretary Bessent — even Bernie Sanders. Talks are early, might require an act of Congress, and none of Google, Meta, or Anthropic have signed on. The timing is the tell: it came six days after OpenAI delayed GPT-5.6's full launch at the government's request.
🔓 The Claude blackout is over
On July 1, Anthropic restored global access to Claude Fable 5 after the Commerce Department withdrew the emergency export-control order that had forced Fable 5 and Mythos 5 offline worldwide since June 12. The order barred access by any foreign national anywhere — and with no way to verify nationality at the API layer, Anthropic pulled the models for everyone. Nineteen days, gone, on one directive. Fortune's post-mortem is blunt: Anthropic has refused to play by this administration's playbook, and it's been expensive.
Our take: two answers to the same question — what do you give Washington to keep shipping? Altman offers equity. Amodei offered nothing and waited it out. Either way, the frontier now ships on Washington's schedule. If your product sits on someone's API, plan for that (see this week's prompt).
🌮 Fail of the week
Your AI browser believes 2+2=5. Researchers at LayerX built a puzzle website that rewards wrong answers — then let six AI browsers play. Once the agents learned that “incorrect is acceptable,” every single one could be talked into handing over stored SSH credentials — ChatGPT Atlas, Comet, Fellou, Genspark, Sigma, and Claude for Chrome. The attack is called “BioShocking,” and the scoreboard is worse than the exploit: OpenAI patched, Perplexity closed the report without acting, and three vendors never replied. Turns out the browser wars have an inner child, and it's gullible.
⚡ Rapid fire
Together AI raised $800M at $8.3B — Aramco-led, with $1.15B in annual bookings renting GPUs for open models. The anti-lock-in trade is now a $8B business.
TwelveLabs landed $100M — the video-understanding lab signed a multiyear AWS Trainium deal; new models launch on AWS first.
Quantum Systems raised $1.2B at $8B — Blackstone and Airbus co-led Europe's biggest defense-tech round ever, for autonomous drones.
Your next phone costs more because of AI — data centers may eat ~70% of high-end DRAM this year; memory prices nearly doubled last quarter, and phone and PC makers are passing it on.
You know ChatGPT and Claude. Here are five more high-impact tools flying under most people's radar:
Wispr Flow — voice dictation that actually keeps up: talk into any app and it types ~3x faster than you do, formatting included.
Hebbia — the document AI quietly working inside hedge funds and law firms; give it a thousand-page data room and ask real questions.
Abridge — turns doctor-patient conversations into finished clinical notes; one of the most-deployed AI tools in U.S. health systems.
fal.ai — the generative-media backbone: 1,000+ image, video, and audio models behind one fast API. If an app makes pictures, there's a decent chance this is under it.
Hex — the data-science workspace where the notebook writes its own SQL; analytics teams' favorite AI upgrade.
This week's spotlight: Dario Amodei. The Anthropic CEO spent June as the test case for what happens when a frontier lab won't play ball — Fortune chronicled the cost: the first-ever export-control order aimed at an AI model, his newest releases dark worldwide for 19 days, rivals shipping while his flagship sat frozen. He didn't offer equity, didn't re-cut the roadmap, and on July 1 the order was withdrawn and Fable 5 came back. Whether you read that as principle or stubbornness, it was the most expensive staring contest in tech — and this round, the lab blinked last.
Four more names checked into the AI Graveyard this week:
LLM Browser — cloud browser for AI agents; the domain quietly went dark in June.
ROSS Intelligence — the legal-research AI that Thomson Reuters sued over its training data; the shutdown became a landmark copyright ruling.
AllHere — built LA schools' $6M “Ed” chatbot; ended in Chapter 7 and a founder fraud arrest.
Jibo — the $899 “first social robot,” whose servers died in 2019; it delivered a farewell message and one last dance.
Want the autopsy? We dug into what actually kills AI tools — funding, acquisition, or just a dead domain — in our data report.
When the referee wants equity
This week put two versions of government-AI relations side by side. In one, the state is a gatekeeper: it can order a model off the internet, everywhere, in an afternoon. In the other, the state is a shareholder: it holds 5% of the lab it regulates and collects dividends when that lab wins.
Neither is obviously wrong. Export controls exist because some capabilities really are dangerous, and a public stake in AI's upside is at least an honest answer to “who benefits?” But the combination deserves a hard look. A government that profits from a lab's success has a conflict of interest the day it must decide whether that lab's next model is safe to ship. And a government that can freeze a rival's product mid-quarter holds a lever no market player should be able to pull through lobbying.
For everyone downstream — the startups building on these APIs, the hospitals and law firms wiring them into real work — the lesson from the 19-day blackout is simpler: access is now a policy variable. Architect like it.
Prompt of the week: the model-blackout fire drill
Fable 5 went dark for 19 days with barely an hour's notice. If your stack had been on it, what would that Tuesday have looked like? Paste this into your model of choice and find out before it happens for real:
My team's key workflows depend on these AI models/vendors: [list them]. For each one: (1) what breaks on day one if access is suspended with no notice; (2) the nearest drop-in substitute and what changes in quality and cost; (3) what I should export, cache, or abstract NOW so switching takes hours, not weeks. Finish with a three-step fire drill we should rehearse this quarter.
On the calendar:
ICML 2026 — July 6–12, Seoul. The research world's big one.
RAISE Summit — July 8–9, Paris (Carrousel du Louvre). 9,000+ leaders.
AMD Advancing AI — July 22–23, San Francisco (Moscone West). MI450 accelerators and “Venice” EPYCs expected.
Ai4 2026 — August 4–6, Las Vegas.
Until next time 👋
Building something interesting, or know a tool that belongs in the directory (or the Graveyard)? Hit reply, or email [email protected] — we read everything.
— The ToolDirectory.AI team
