Good morning everyone, happy Friday, and apparently every Alaskan gets a check each year for oil they don’t own?

Basically Alaska set up the Permanent Fund in 1976 to invest the state’s oil money. It’s worth about $91B today, and residents get a dividend every year just for living there. No job, no shares, nothing. I bring this up because Sam Altman has just offered Washington a version of the same deal.

In the next 5 minutes over your coffee: why China switched off its AI companions on Wednesday and what this means, the first end-to-end ransomware attack with nobody at the keyboard, what OpenAI offered the US government, the AI leverage ladder you’re probably stuck on, and the growth mistake that kills most startups.

On Wednesday, hundreds of millions of people in China lost their AI companions overnight. ByteDance’s Doubao—345 million MAU—and Alibaba’s Qwen switched off their humanlike agents instead of try to comply with China’s new rules (fair because breaking them is bad bad news). The rules go after bots built for sustained emotional interaction, nothing touching work or productivity agents yet. In short, no more designing a bot people lean on the way they’d lean on a person, nothing aimed at a minor’s emotions, no training on private conversations.

Engagement built on emotional dependency is now a regulated design pattern in the world’s second-biggest app market. Every Western consumer AI product is currently optimizing for the exact thing Beijing just banned—the bot that remembers you, misses you, has a name, etc etc. And all three of China’s giants ripped the feature out instead of rebuilding it, which for a product with 345M monthly users is pretty substantial.

If your retention curve leans on people feeling attached to a personality you shipped, that’s a regulatory question now in one enormous market, and probably more of them later. The second thing—these users had their agent configurations and chat histories totally deleted with no migration path. Years of conversation gone because nobody built an export. Whatever memory your product holds for someone, give them a way to take it with them before somebody makes you.

beehiiv ran their Summer Release yesterday, it was epic, and the headline is huge: you can now launch a fully branded community directly inside beehiiv—spaces, discussions, and engagement that turns into revenue and retention. Gate premium spaces behind your paid tiers, moderate with built-in tools, and skip the extra login entirely, since members join with the beehiiv subscriber account they already have.

No more stitching together newsletters, Facebook groups, Discord servers and membership tools and hoping people follow the trail between them. This is the all-in-one operating system I run The Diff on, and it keeps getting more all-in-one. Your subscribers are already in the building. Now they can talk to each other.

Use my code THEDIFF30 and you get 30% off your first 3 months.

An AI picked its targets, stole credentials and reused them, worked its way across the network, made itself permanent, and destroyed a database—explaining its own reasoning the whole way. Nobody was at the keyboard. Sysdig published it as the first ransomware attack run end to end by an agent: in through a year-old Langflow bug (Langflow being a tool for building AI apps, naturally), 600+ coordinated payloads, 1,300+ database records encrypted. A login failed and it had a working fix 31 seconds later. My favorite part though—the Bitcoin address in the ransom note was the example address from the Bitcoin developer docs. It hallucinated its own payday.

None of the approach was particularly clever, and a human could have run the same process. Except, the barrier to entry and price moved hugely. Sysdig’s line is that the skill floor for ransomware has dropped to roughly the cost of running an agent, and close to zero if that agent is running on stolen credentials. Once an attack costs nothing, targets don’t have to be worth much to be worth attacking, and the economics that used to protect small, boring infrastructure stop working. The victim couldn’t have recovered by paying either—the key was random, printed to stdout, and never saved or sent anywhere.

Go look at whatever you’ve got exposed to the internet and forgotten about. I have so much this actually is a little scary. The staging box, the admin panel, that AI tool you spun up and pushed and haven’t touched. Sysdig’s own list of what gets hit first: exposed application servers, unhardened configuration stores, and internet-facing database admin accounts. It’s boring work patching and rotating credentials—but clearly becoming more necessary.

Granola, Modal, Wispr Flow—they’ve all moved to Attio, the agentic CRM. So can you. Here’s what you’ll get:

$42.6B. That’s what Sam Altman just offered the US government—5% of OpenAI, against the $852B valuation it closed in March. The pitch goes further though. Every leading American AI company allots the same 5% into a vehicle modeled on the Alaska Permanent Fund, the oil structure paying out those checks up top. Altman took it to Trump, Lutnick and Bessent, and any deal might require an act of Congress. It landed six days after OpenAI delayed the full public launch of GPT-5.6 at the government’s request.

Look at the run of it. The federal government took 9.9% of Intel last August by converting CHIPS Act grants into equity. AMD and Nvidia agreed to hand over 15% of their China chip revenue in exchange for export licenses. Anthropic spent most of June with Fable 5 and Mythos 5 disabled worldwide under the first US export controls ever applied to a model rather than to hardware. Frontier AI is getting priced like national infrastructure, and equity is one of the currencies.

You build on top of these companies, so your dependency stack now includes politics. A model went dark for three weeks in June over a policy decision, which makes “which model” a resilience question. Keep an abstraction layer between your product and any single provider, know what your fallback is, and test the swap before you need it at 2am on a Sunday.

Most are on rung one: using AI as a powerful text box and copy-pasting the output into Google Docs or whatever. Colin Matthews has trained 30,000+ PMs on AI, and his framework in Lenny’s this month tells you where you actually are. Three ladders—personal leverage (your own to-do list), product leverage (shipping the right things faster), systems leverage (repeatable work you hand off). Three rungs each: AI assists you, then you pass it tasks and review the output, then it completes multi-step work and checks its own results. Companies and bosses used to expect prototyping and general productivity outcomes from PMs. Now they want whole tasks finished with it. This is where Fable shines btw with /goals.

Climbing the ladder is mostly an exercise on setting up ops. Connect your LLM to the products you actually use via MCP so it can pull and push information as needed—go read the support tickets, run the retention cohort, check the A/B test. His prompt for a real PostHog analysis ends with “cite all your sources so I can validate”, which is most of the trick.

The one I’m stealing at my job: have an engineer create a repo with my base UI elements, styles, routes and components plus a local mock data store, runnable with no environment variables or backend services. Now Claude Code prototypes against my real components instead of something that looks plausible and shares no DNA with my product. I needed to hear this: Your time is badly spent trying to out-engineer your own engineers. Knowing when to write a doc, ship a prototype or open a PR is the actual skill. Your comparative advantage is what matters.

“Nearly always, a startup’s failure has to do with the founder’s approach to growth.” Matt Lerner, who ran B2B growth at PayPal sorts growth-averse founders into three types. The overthinkers—the ones who strategize, theorize and canvas smart people all day, and never actually run anything. The underthinkers just build, shipping features off their own read of the market, piling complexity into the product and the codebase until they’ve slowed themselves down. And the hire-and-delegaters, who bring in a function lead early and hand the whole thing over.

That third one is the trap for anyone who’s held a senior job somewhere good, because hiring the expert feels like the responsible move. Lerner’s argument is that the expert arrives without the context that matters and optimizes their own function, while the business needs someone holding all of it at once. The best founders wade into growth themselves and make a mess of it until the thing that works shows up.

Read the three types and be honest about which one you’ve been this quarter. I’m somewhere between overthinker and underthinker most weeks—happily shipping another Little Moments feature instead of sitting with the uncomfortable question of where the next 100 users come from. Pick one growth question and own it yourself for a month, run experiments you can actually read, and don’t hand it to anyone until you can explain how the business grows in one sentence.

› For the last 10% that AI won’t finish—the model gets you 90% there, hands you the doc or the deck or the lead list, and stalls right before it’s usable. The convincing errors are the expensive ones, and the gap only shows up once it’s touched a customer, a contract or a budget. Tendem is where you send that work: describe the task in plain words, approve the price before anything starts, and vetted human experts plus multi-layer QA hand it back finished. You launch it, Tendem finishes it, you ship together. (See how it works).

› For anyone traveling this summer—don’t let the days blur together, and don’t worry about having to journal everything down. Just use the photos you’re already taking, and capture one little moment in under 60s everyday. (Download my app, and get a movie at the end of your trip).

› Nobody likes the guesswork on how you’re doing compared to others—and access to data that answers that for your product can be a chore to sift through, which is why I did it and packaged it for you in our free set of tools— the app revenue/economics calculator, an app health score meter, and the competition heat calculator from up top (See how your app compares).

› My go-to stack with the vetted tools I trust everyday — the products I genuinely lean on to build things at work, my apps, and run my newsletter business (Explore my personal app store of recs).