In partnership with

Welcome, humans.

So Will Depue just put words to the AI product that everyone keeps circling but nobody has actually given us: an executive super-assistant with deep memory, one continuous history, enough trust to handle email, credit cards, bookings, passwords, and subscriptions, plus the judgment to ping you only when it needs a human.

Hermes Agent is close, but technical to set up. ChatGPT Workspace Agents are getting there, and TBH, Grant has been loving that feature lately. Codex also points at the shape of whatever comes next: less “answer my prompt” and more “own this messy job until it is done.”

The missing layer is the life OS around it. That is why NOX is creating one AI inbox, Claude searching Slack context is a powerful unlock, and the latest Fable/Codex workflows are all tiny pieces of the same product: a trusted agent that remembers you, works across your tools, and handles the junk drawer of modern life.

Now if only someone would actually build this, because my current executive assistant is still me, and he keeps inventing more work for me to do instead of doing it for me and it’s ACTUALLY getting really rude. I have the World Cup to watch, bruv!

Here’s what happened in AI today:

😺 OpenAI reportedly discussed giving Uncle Sam a 5% stake.
📰 White House model-release standards talks picked up speed.
📰 Microsoft launched a $2.5B enterprise AI deployment unit.
🍪 Hugging Face and Cerebras showed open real-time voice AI.
💡 NVIDIA turned AI cloud financing into a business model.

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us!

😺 Uncle Sam wants 5% of OpenAI

So, when AI companies argue that the public should share in the upside, the question is no longer rhetorical. Somebody has to decide who actually gets the shares.

Well, OpenAI has reportedly discussed giving the U.S. government a 5% stake in the company, a proposal meant to show that the public could share in AI’s upside as the company courts Washington and prepares for a future IPO.

Here's what happened:

The Guardian, citing a Financial Times report, said OpenAI has held early talks about giving the U.S. government a 5% stake.
CNBC reported the proposal could be worth tens of billions of dollars and is meant to reduce political blowback.
Axios reported any deal would likely need Congress and could raise questions about government influence over model releases.
AI policy analyst Dean W. Ball argued there are two very different versions of public ownership: give the stake directly to households, or give it to the government.

Why this matters: If every household gets a direct stake, the pitch is simple: AI creates wealth, and all citizens share in it. That could look like a public wealth fund, a dividend, or some future version of the Invest America accounts for kids.

If the government owns the stake itself though, the incentives get strange fast. The same officials (eventually) regulating frontier models could also have a financial interest in one of the companies they oversee. If you think there’s a conflict of interest in how they treat the AI industry now, just wait. Political fights over safety, competition, national security, and IPO timing would all get very messy very quickly.

Ball's harsher version: handing a collective stake to Washington could make half the country see the deal as blatant corruption and the other half wonder why they themselves never got a check. On net, it’s a no good, very bad idea.

Our take: We should probably do the sovereign wealth fund thing. It works well for Norway, Singapore, Qatar, and the UAE because national windfalls get pooled, invested, and turned into long-term public wealth.

The answer cannot just be "trust the government to hold the shares for you." If AI wealth-sharing is going to work, the public needs to feel the wealth. They need to feel that ownership directly, not watch politicians divvy up a cap table they were never invited to join. You think ppl hate data centers now? Just wait until that happens.

FROM OUR PARTNERS

Scale AI support on AWS, see how July 9

Customer expectations keep rising. Support budgets don't. On July 9, Fin and AWS are hosting a live executive session on how leading enterprises close that gap: scaling AI-powered support while simplifying how they buy it.

You'll see how to resolve an average 76% of conversations with Fin on AWS enterprise-grade infrastructure, procure through AWS Marketplace to put committed cloud spend to work, and turn the Fin and AWS collaboration into lower support costs. Register for the live session to see how.

Save your spot

🎓 AI Skill of the Day: How to Prompt Fable 5

Claude’s new Fable 5 model is temporary, expensive, and too powerful to waste on tiny tasks. The real skill is knowing when its long-context judgment is worth the tokens.

Here’s Anthropic’s advice on how to prompt it: give Fable the outcome, avoid step-by-step micromanagement, save reusable context in Markdown, and make every progress claim point back to evidence. Try this when the task is too expensive to brute-force:

I want to use Fable 5 as planner and judge, not as the executor.

Goal:
[what I want shipped]

Context:
[files, docs, screenshots, constraints, examples]

Use this workflow:
1. Create a concise plan with exact files, steps, risks, and success criteria.
2. Write a handoff brief for a cheaper executor model.
3. List what Fable should NOT do itself.
4. Define the memory note to save after this run.
5. Define the verification gate Fable must use after execution: tests, screenshots, logs, file diffs, or manual checks.
6. After the executor finishes, review the result as judge and list only issues that change whether we should ship.

Keep the goal high-level. Do not micromanage implementation unless the evidence requires it.

Besides that… best workflow we saw today: Mitchell Hashimoto’s loop: Fable xhigh writes the architecture plan, a cheaper fast model does the coding, then Fable xhigh reviews the result. Other builder workflows add the same pattern with memory files, handoff docs, fresh chats, and explicit verification gates. TL;DR: use Fable as the expensive planner and judge, not the whole construction crew.

Have a specific skill you want to learn? Request it here.

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

*Dell Pro Max with GB10 gives your builder a serious desktop for experiments, demos, and workflows, powered by NVIDIA Grace Blackwell architecture, NVIDIA DGX OS 7, and 128GB memory. See Dell Pro Max with GB10.
Vellum gives you a personal intelligence assistant that remembers preferences, handles tasks, and can coordinate in Slack like a teammate, with no pricing details listed.
Seedance 2.5 in Dreamina lets you make 30-second cinematic videos with ByteDance's model, up to 50 multimodal references, R2V control, and longer-video beta support, with no pricing details listed.
Kimi Code gives developers a coding agent and CLI toolkit powered by Kimi K2.7 Code, including autonomous goal execution through a /goal workflow, with no pricing details listed.
Hugging Face and Cerebras real-time voice AI shows developers how to build an open speech-to-speech assistant with replaceable parts for listening, thinking, and talking back, with a free demo and repo available.
Safari MCP server lets agents connect to a real Safari Technology Preview browser window to inspect pages, capture screenshots, read logs, and debug web apps, with no pricing details listed.

📰 Around the Horn

White House talks around voluntary frontier model standards reportedly accelerated, with labs and national-security agencies negotiating release benchmarks and access rules.
Microsoft launched Frontier Company, a $2.5B AI engineering group built to move enterprise AI from pilots into measurable production systems.
Cognizant and OpenAI announced a GPT-5.5 cyber-defense service that moves enterprise teams from vulnerability discovery to validated fixes.
NVIDIA introduced a revenue-sharing and credit-support model for AI clouds, helping partners finance huge AI factories while giving NVIDIA usage-linked upside.
GitHub added AI credit pools to cost centers, giving Copilot admins a way to stop one team from draining shared monthly credits.
Cloudflare gave site owners new AI traffic controls for Search, Agent, and Training bots, including sharper protection for ad-monetized pages.
Anthropic reportedly began early custom-chip talks with Samsung, adding another sign that frontier labs want more control over the compute stack (they’re also apparently planning to design their own drugs).
Epoch AI said AI-assisted vulnerability discovery helped drive a record spike, with 21 organizations disclosing roughly 1,500 high- and critical-severity CVEs in June.

FROM OUR PARTNERS

AI handles the busywork. You handle the real work.

That's ClickUp's pitch for tasks, and after nine months running 2,000+ pieces of content a month on the platform, we can confirm: Super Agents (auto-task creation, auto-assign, auto-prioritize) have eliminated hundreds of manual clicks per day for our team.

When a podcast episode hits "ready for edit," the right editor gets it. When the edit's done, the next stage routes automatically. The math gets compelling: 500+ "Human Skills" out of the box, 3.3 million tasks already automated platform-wide. So the Free Forever plan covers the basics, but you can add Brain at $9/user/month for the AI bits.

→ Start free

💡 Intelligent Insights

A few useful watches for the weekend, from practical agent workflows to the hardware bottlenecks underneath all this model magic:

Naval's latest founder roundtable: spend more time living with expensive agent workflows now, because today's pricey token habits are a preview of normal work once inference gets cheaper.
Claude's future-of-work conversation shows the practical unlock for team AI: put agents in public channels, connect source-of-truth files, give them narrow recurring jobs, then tune when they should jump in or stay quiet.
Using Large Language Models: a helpful explainer from Andrej Karpathy on prompting, tooling, and treating chatbots more like operating systems than search boxes.
Ex-Google Insider: You're Not Ready For The Next Phase of AI is worth watching for Logan Kilpatrick's practical map of where AI products are going next. He breaks down the shift toward memory, agents, and AI that can operate across real workflows, then pairs nicely with Andrew Dai's take on self-improving models and autonomous software agents.
Dylan Patel on AI's Real 100x is a useful counterweight to AI hype. Patel explains why model gains increasingly come from inference-time compute, better tooling, and better systems around the model, not magic one-shot intelligence leaps.
Grant Sanderson on AI and math is a strong watch if you want to understand why AI is good at some forms of reasoning and brittle at others. His examples make the difference between pattern fluency and durable understanding feel concrete.
Greg Brockman on merging chat and agents is a useful window into OpenAI's product direction. The core idea: the future interface is probably less about choosing between chatbot and agent, and more about one system that can talk, plan, and act when needed.
The Codex workflow that runs a consulting business is a practical demo of AI as an operations layer. The useful takeaway is not the specific stack, but the pattern: break work into repeatable loops, give the agent context, then let it handle the boring glue.
Continual learning for long-running agents is a good primer on why persistent memory matters. If agents are going to handle real projects, they need to improve from feedback over time instead of starting every task like it is day one.
What does the next training paradigm look like? is worth watching for a clearer read on post-pretraining progress. The big takeaway: future gains may come from environments, synthetic data, and feedback loops that teach models to act, not just predict text.
Building Great Agent Skills is useful if you are trying to make agents actually repeat good work. It shows why the best prompts often become reusable procedures, examples, and checklists that agents can call on later.