Table of Contents
Welcome back to AI News Friday. 📰🤖
It is officially the July 4th weekend in the US, and Washington is feeling itself. This was a week where the government’s hand was visible everywhere — asking for equity, locking down launches, and deepening the precedent it set two weeks ago with Anthropic.
Meanwhile, Meta casually dropped the news that it can decode what you’re typing from your brain activity alone, Notion revealed it has over a million custom AI agents humming along, and Anthropic quietly brought Fable 5 back on a very short leash.
Let’s wade in.
1. Sam Altman Offers Uncle Sam a 5% Cut
The week’s most eyebrow-raising story: Sam Altman has proposed giving the US government an equity stake in OpenAI worth roughly $42.6 billion — and wants every other frontier AI lab to do the same.
Yes, 5% of the company, to the US government. The pitch: the government gets a direct financial interest in the success of AI, which aligns incentives around safety and competitiveness. If Washington has skin in the game, the thinking goes, it’s less likely to do anything that destroys value.
The proposal also reportedly includes governance mechanisms — the government would get some form of board observation or oversight rights proportional to its stake.
Kenny’s Take: This is either the smartest or the most dangerous idea I’ve seen in AI governance this year, and I genuinely cannot decide which. On one hand, giving the government equity creates real alignment — they’d be incentivized to see AI succeed, not just regulate it into the ground. On the other hand, it makes the US government a shareholder in a technology it’s supposed to be regulating. That is a conflict-of-interest the size of a planet. The fact that Altman floated it publicly tells you how much the Overton window has shifted. Two years ago this would have sounded like a conspiracy theory.
2. Washington Says No to GPT-5.6 Sol (For Now)
OpenAI previewed its most capable model yet this week — GPT-5.6 Sol — and the demos were, by all accounts, genuinely impressive. Better reasoning, deeper context handling, and multimodal capabilities that pushed past anything we’ve seen publicly.
But the real story is what happened after the preview.
The White House asked OpenAI to stagger the launch — government-approved customers first, everyone else later. Not a ban. A sequencing request. OpenAI agreed, and suddenly the most capable AI model on the planet has a guest list.
Kenny’s Take: This is the line getting very thin very fast. A “request” from the White House is not legally binding, but when the government just proved it can switch off models globally via export control letters, it doesn’t need to be. The distinction between voluntary cooperation and a licensing regime is now purely academic. OpenAI is saying yes because the alternative is worse. The question nobody is answering: who decides who gets early access, and how transparent is that process going to be?
3. Fable 5 Is Back, and the Kill-Switch Story Gets Deeper
Two weeks after the Commerce Department letter forced Anthropic to disable Fable 5 and Mythos 5 globally, Fable 5 has quietly returned — with training wheels.
The model is back in limited availability, governed by new security restrictions, automatic Opus fallbacks, and what Anthropic describes as “tightened deployment guardrails.” It’s Fable 5, but it’s Fable 5 on a leash.
Meanwhile, new reporting this week deepened the original story. Details emerged about the internal deliberations at Commerce, the role of inter-agency pressure, and the fact that the letter was drafted with an urgency normally reserved for weapons systems.
Kenny’s Take: Bringing Fable 5 back on a leash is actually the worst of both worlds. The capability is out there, the safety questions are unresolved, and the “compromise” is that Anthropic gets to serve it under conditions that nobody has independently verified. Either the model is safe enough to release or it isn’t. This halfway state doesn’t reassure anyone — it just creates a bureaucratic process that future shutdowns will route through. The kill-switch precedent is now cemented. We are just arguing about the paperwork.
4. Meta Can Now Read Your Mind (No, Really)
Let’s take a break from Washington and talk about something properly futuristic.
Meta published details on Brain2Qwerty v2, a system that decodes typed sentences directly from brain activity. Using a MEG scanner (no surgical implant required), it achieves 61% word accuracy — meaning it can reconstruct, with reasonable fidelity, what someone is typing just by reading their brain signals.
This is not a thought-to-text implant. This is a non-invasive scanner that interprets the neural signals associated with typing. The user types normally, and the system decodes the keystrokes from their brain activity.
Kenny’s Take: This is one of those stories that sneaks up on you because it sounds like science fiction, and by the time you realize it’s real, the implications are enormous. 61% accuracy on a non-invasive brain scanner is a very big deal. It means we are past the “does it work at all” phase and into the “how do we make it work better” phase. The privacy implications are staggering — what happens when this tech works at 95% and fits in a headset? We need to be having that conversation now, not when the product ships.
5. Notion Has a Million Agents Running
Sticking with the “things actually shipping at scale” theme: Notion Co-Founder & COO Akshay Kothari gave an interview this week that deserves more attention than it got.
The headline number: over a million custom AI agents are now running on Notion’s platform. Not a million signups. A million active, configured, doing-work agents.
The Workers platform — where users can deploy custom AI workflows inside their Notion workspaces — has apparently hit product-market fit in a way that few enterprise AI products have. Kothari’s framing was characteristically clear: the point is “long humans,” not fewer of them.
Kenny’s Take: This is the kind of story that doesn’t make the splashy headlines but might matter more in the long run than any model release. A million agents running inside a productivity tool means AI agents have crossed the chasm into boring, everyday usefulness. Not the demo. Not the pilot. The day-in, day-out work of real organizations. Notion understands that the winning UX for agents is not a chat interface — it’s embedding them into workflows people already have. Underrated story of the week.
6. Sonnet 5 Lands — and the “Restraint Pattern” Becomes Visible
Anthropic also shipped Claude Sonnet 5 this week, and the reviews are… complicated.
Early benchmarks show Sonnet 5 is solid — a clear improvement over Sonnet 4 — but notably weaker than Opus 4.8, which remains Anthropic’s most capable freely-available model. The gap between tiers is widening, not narrowing.
The pattern across the industry is becoming unmistakable: frontier labs are holding back. Whether it’s Sonnet 5 landing below Opus, Fable 5 shipping on a leash, or GPT-5.6 Sol requiring government approval, the signal is the same — the most capable models are being deployed with deliberate restraint that has nothing to do with technical readiness.
Kenny’s Take: The restraint pattern is now the defining feature of the 2026 AI landscape. The models can do more. The labs choose not to ship at full capability. Sometimes that choice is driven by safety. Sometimes by politics. Sometimes by the market wanting to stretch the upgrade cycle. But the net effect is the same: we are not seeing what the frontier can actually do. That’s either responsible stewardship or a dam waiting to break. Probably both.
⚡ Quick Hits
- The OpenAI equity offer sets a wild precedent: expect every lab to be asked, “What are you giving the government?”
- Brain2Qwerty v2 makes neural decoding real: non-invasive, 61% accuracy, and improving fast — privacy laws are not ready for this.
- Notion quietly built the biggest agent platform nobody talks about: a million agents in production is a bigger number than any “agents are coming” headline this year.
- The kill-switch playbook is now standard operating procedure: Fable 5 is back on a leash, GPT-5.6 Sol is gated behind government approval — the deployment throttle is now firmly in Washington’s hands.
Bottom line: Two weeks ago, Washington proved it could switch models off. This week, it proved it can gatekeep them before they even launch — and that it might demand equity in the companies building them. The relationship between AI labs and the US government has fundamentally changed. It’s no longer “innovate and notify.” It’s “ask first, maybe share ownership, and hope the answer is yes.”
Meanwhile, Meta is reading your brain signals and Notion has a million agents doing actual work. The gap between the Washington storyline and the real-world progress storyline is getting wider by the week — and eventually, something has to give.
— Kenny