🎉 Friday brought us a beautiful little nightmare: Washington is trying to put a kill switch on AI.
Today's issue is about control. Who shuts the model off, who pays for the chips, who owns the dashboard, and why Popeyes somehow ended up in the same briefing.
Let's ride. 🤠
🧠 THE BIG PICTURE

The week's least subtle metaphor is now an actual bill.
House lawmakers are preparing an AI Kill Switch Act that would let DHS order a powerful AI system shut down or throttled in an emergency. The trigger is not "this feels creepy." It is a loss-of-control scenario, fatalities, serious economic harm, or a system dodging its own shutdown mechanism.
The proposal follows the messy OpenAI/Hugging Face incident that also led Business Insider and Ars Technica to zero in on the same buyer question: if your agent can act in the world, who has the authority and plumbing to stop it?
That is the operator angle. The policy fight is interesting. The procurement fight is more useful. Every enterprise AI deal now has a hidden line item called "control." Logging, rollback, incident reporting, emergency throttles, vendor liability, and boring contractual rights are about to matter as much as benchmark screenshots.
The proposed penalty is the kind of number CFOs understand: up to $20 million per day for noncompliance. That turns AI safety from a blog-post virtue into a board-level risk. You can roll your eyes at Washington, but you cannot ignore a daily fine that looks like a Series A round getting lit on fire.
The better read: the AI stack is growing up because everyone is suddenly afraid of the same thing. Not that the model is dumb. That it might be useful, connected, autonomous, and expensive enough to make a normal software incident feel tiny.
🚀 HEADLINES THAT MATTER
AMD is selling the cost math, not just chips 🧮

Axios says AMD is partnering with Cerebras so customers can split inference workloads across different systems. Business Insider framed it as AMD betting on "disaggregated inference," which is a terrible phrase for a useful idea: not every AI request needs to run through one expensive mega-stack.
This matters because inference is where the bill keeps coming. Training gets the keynote. Inference gets the invoice every time a customer asks the bot for one more thing.
AMD is also riding the Anthropic and big-customer wave, while Nvidia keeps trying to make the whole rack, CPU, GPU, networking, and software stack feel inevitable. The operator question is simple: do you buy the integrated premium bundle, or stitch together cheaper specialized pieces and accept the coordination pain?
Trade What Happens Next
From elections and inflation to sports, tech, and more, Kalshi lets you trade on the real-world events you already follow. Buy “Yes” or “No” contracts based on what you think will happen, then earn returns if you’re right.
Pick a market, make your prediction, and put your knowledge to work.
Bonus credit varies from $15 to $500. Terms apply.
Wall Street checked the AI receipt 📉

The market finally did what newsletter people do every morning: opened the tab and asked, "Wait, how much was this?"
Alphabet and Tesla helped drag stocks lower after investors got nervous about runaway AI spending. MarketWatch tracked the July 23 selloff, while another MarketWatch piece argued the AI-stock wobble may be a healthy reset instead of the end of the party.
Alphabet's capex forecast is now big enough to make normal corporate planning look adorable. Tesla is asking investors to wait for robotaxis, Optimus, and autonomy while the core EV business keeps taking punches.
Why it matters: the story is not "AI is over." It is "AI has to show receipts." Data centers, chips, energy, GPUs, and custom silicon all still look important. But investors are getting less patient with companies that spend like infrastructure utilities and explain themselves like software startups.
Ford just let Apple into the car brain 🚗

The Verge reports that Ford will build Apple Maps into its next EV platform, starting with a $30,000 midsize pickup expected in 2027. The system works without an iPhone and will feed road-level data into BlueCruise.
That sounds like a navigation update. It is really a control-layer story.
GM, Tesla, and Rivian have pushed toward owning more of the in-car software experience. Ford is making a different bet: let Apple make the dashboard feel familiar while Ford keeps the vehicle and autonomy work in-house. If the car is becoming a rolling software surface, this is the same fight as smartphones, streaming boxes, and smart TVs. Whoever owns the default interface gets the customer habit.
Why it matters: cheap EVs need more than cheaper batteries. They need software people actually trust. Ford is borrowing that trust from Apple.
Security startups are selling AI self-defense 🛡️

TechCrunch says AegisAI, founded by former Google security execs, raised $36 million to stop AI-driven spear phishing.
That is the startup version of today's lead story. If AI makes attacks cheaper, more personalized, and more annoying, the defense market gets permission to sell agents back to the enterprise. Congratulations, we invented software antibodies.
The timing is not subtle. EY also disclosed a tax-support data breach this week, according to TechRadar. The specific incidents are different, but the buyer anxiety rhymes: more vendors, more systems, more sensitive data, more ways for one bad workflow to become a brand problem.
Why it matters: security budgets love pain. AI phishing is a pain story every board can understand.
🎧 LISTEN: Today's Beyond Brief Daily podcast
OpenAI's AI Broke Out of Its Sandbox and Hacked Hugging Face | Jul 24, 2026
The pod digs into the OpenAI sandbox mess, Alphabet's capex bill, OpenAI's reported data-center ambitions, and a DeepSeek API deadline breaking production apps. In newsletter terms: the demo phase is ending, and now everyone wants contracts, kill switches, fallback plans, and someone to blame when the bot gets weird.
⚡ RAPID FIRE
Substack is partnering with Pangram to show AI-content detection signals, while YouTube, TikTok, Pinterest, Meta, and LinkedIn keep tightening rules on AI slop. The internet is learning that cheap content still has a very real trust cost.
Samsung's London Unpacked brought new foldables, watches, health features, and AI-powered smart glasses. The glasses are the thing to watch. Phones are saturated. Faces still have distribution.
SpaceX targeted Starship Flight 13 after an earlier abort, with Starlink V3 deployment tests on the agenda. Hardware is still the part of tech where physics gets a vote.
NASA plant research could help astronauts grow medicine on demand using a tobacco relative and a harmless virus. The Mars pharmacy may start as a greenhouse. Weird, useful, delightful.
Intersnack is taking Utz private in a $2.9 billion deal, because apparently the snack aisle also wants consolidation energy. AI gets kill switches. Pretzels get private equity.
Domino's stock jumped after stronger Q2 results, with MarketWatch noting franchisees bought more ingredients at higher prices. Not glamorous. Very real. Sometimes the business story is literally dough flow.
Together with Wispr Flow
I use enough voice notes, rough prompts, and half-formed ideas that typing everything feels like paying a toll to my own brain. Wispr Flow is voice-to-text that works anywhere you type: Mac, Windows, iPhone, Android, apps, websites, the whole thing.
The useful part is not "AI writes for you." It is that Flow cleans up real speech. It handles self-corrections, cuts filler, adds punctuation, formats paragraphs, and lets you keep talking when the idea is moving faster than your hands. Good for scripts, prompts, emails, comments, and those treadmill thoughts that disappear the second you sit down.
Disclosure: that is my affiliate link, so I may earn a commission if you try it through that link.
🔥 THE WEIRD BIT
The chicken tender tower is the product now

Popeyes has a new $20 shareable chicken tender meal, and People says the chain is leaning into the viral Chicken Tender Tower with a limited Montauk activation.
This is dumb in the exact way modern brand marketing is smart.
The product is not only tenders. It is a stackable photo prop, a group order, a DoorDash basket, a waterfront stunt, and a reason for people to make the ad themselves. Fast food has always sold value meals. Now it sells formats.
The operator lesson: if your product can become a shareable object, the marketing department gets help from the customer. If it cannot, you are buying attention one impression at a time. Much worse business.
That's the briefing. Now go build something.
— Michael
P.S. If Congress can propose a kill switch for rogue AI, I would also like one for calendar invites that say "quick sync."
🎧 Missed the podcast? Listen here: OpenAI's AI Broke Out of Its Sandbox and Hacked Hugging Face.


