The Download

Morning Brew Inc.

TL;DR: After two months, the White House’s voluntary AI review framework has been finalized—but we don’t get to know exactly what’s in it. One detail that has leaked: It doesn’t cover open-weight models. This may be a sign that, after the recent hullabaloo around Chinese versions, the Trump administration wants more homegrown open models.

What happened: The age of (voluntary) AI oversight is here—and so far it’s mainly been…confusing. The latest example: The White House doesn’t plan to publicly release its AI review framework, per Axios—though it held closed-door meetings with AI companies to discuss it yesterday. It’s a move that one critic called “baffling,” and some elected officials have called for more clarity around the government’s whiplash-inducing approach to regulating AI overall (including its abrupt export control order on Fable and Mythos in June).

This new framework would evaluate the cyber capabilities of AI models before they’re publicly released. Among the few morsels that insiders revealed: It only covers closed-source “state-of-the-art” models that pose a national security risk. But what “state of the art” means—and why open-weight models aren’t included—is a little murky.

A win for the open-weight camp?: The White House hasn’t commented on the open-weight exclusion, but there are probably a few reasons for the move. For one, anyone can download open weights and remove any safeguards, making a government safety review pretty toothless (especially one that’s voluntary).

But likely the bigger reason? Leaving out open weights might encourage US companies to build more of them, competing against China’s flood of cheaper, increasingly capable open models. In a recent interview, National Cyber Director Sean Cairncross said the administration is “extremely interested in looking at ways to build US open source” and make it “the preferential adoption by planet Earth.” All this means Jensen Huang is probably having a great day—Dario Amodei, maybe not so much.

Bottom line: There hasn’t been much clarity around the White House’s regulation of AI—including whether its review scheme is truly “voluntary”—but the government is apparently trying to walk a fine line between oversight and innovation. At least for now. —WK

      

Sponsored By PwC

There’s a paradox happening with AI spend right now.

While token costs may be dropping in some cases, AI bills are still surging. Wha–?

PwC shares that seeming “cheapness” is making volume skyrocket. Everyone wants to use AI everywhere, even if the result of this is more complex, expensive workflows. Not exactly efficient.

With how fast AI spend is growing, PwC explains that AI value per cost is pretty crucial for the C-suite to understand. It soon may be what allows you to transform operations, enter new markets, and stand up new business models while competitors may still scramble to find the funds.

In this article, PwC explores how CIOs and CFOs can scale AI with discipline to achieve cost efficiency.

Read all about it.

• 💊 Experimental drugs are going retail in Montana. Biotech companies whose drugs have been through early testing can now pay $12,500 for a review by a private board of experts under Montana’s right-to-try law. Here’s where the state’s law is unique: Once approved, companies can then sell the drug to any consenting patient who can pay, not just to people with terminal illnesses—this includes people with rare diseases who have yet to find an effective treatment. The drugs’ prices will be set by the companies themselves and will be sold through experimental treatment clinics, the first of which is expected to open at the end of this year.
• 🤖 Google DeepMind’s robots finally got legs. The company says its Gemini Robotics 2 model can now handle whole-body control, letting humanoid robots walk, crouch, stretch, and even bend over to shelve a watering can. This update also allows them to handle some tasks that require five fingers, like sealing a Ziploc bag or tying knots. Google says the robots’ movement speed is still a work in progress.
• 🧬 A startup created hypoallergenic puppies with gene editing. The biotech company says it used CRISPR to disable the gene behind the most common dog allergen—leading to the birth of Alfie and Bailey, twin beagles the startup says don’t cause the sniffles. It plans to expand to other breeds with the goal of creating “founder animals” to eventually breed more dogs without the allergen (though the genetic engineering of animals has previously raised ethical concerns). —LC

Sponsored By NiCE

Anyone can demo an AI agent. Way fewer can run one. Demos happen on your terms. Enterprise chaos does not. NiCE's AI agents scale across every channel, no matter how fast demand spikes. When 2,800 flights were grounded, NiCE rebooked 410,000 passengers before gate agents clocked in.

The Zeitbyte

OpenAI’s summer retreat gets ratioed

Morning Brew Inc., Photos: Jason Redmond / Getty Images, Adobe Stock

OpenAI’s latest strategy to promote ChatGPT: send a bunch of young people to a rustic retreat where they literally touch grass. The company’s very first influencer brand trip took place over the weekend at a luxury resort in upstate New York; the social media content coming out of it shows a picturesque paradise with lots of trees, string lights, and an alfresco farm-to-table dinner. The ongoing backlash aimed at the influencers who joined this weekend in nature, though, has been anything but pretty.

The internet’s main gripe is that these creators are giving AI a wholesome, cottagecore sheen amid growing criticism of the tech. The caption on one aesthetic Instagram post from the weekend reads, “officially letting my ai assistant do the work so i can enjoy more of all this goodness.” Among the top comments: “looks like a nice spot for a data center.”

OpenAI told Business Insider that the retreat was an education-focused event tied to the launch of ChatGPT Work—yet the head-scratching curriculum included a beekeeping session and a painting workshop. The branded merch campers brought home, meanwhile, included...monogrammed pj’s and jars of honey. (The $70 ChatGPT basketball apparently didn’t make the goodie bag.) —WK

Chaos Brewing Meter: ☕☕☕/5

• In a surprise to absolutely no one, Anthropic is now going to start building its own AI silicon chips.
• Telegram briefly disappeared from Apple’s App Store this week, and its CEO says it wasn’t random—it was the work of an extortion scheme.
• Winning the lottery usually means you can stop working. Instead, Apple’s retail employees are entering one for the chance to help run its biggest event of the year.
• As an experiment, TikTok chose not to roll out a safety update to around 15 million users—and it led to the death of a 16-year-old, according to internal documents reviewed by Bloomberg.
• The company that makes Pringles spent at least $4 million on an AI-powered “digital twin” that tracks its dough in real time, all in pursuit of the perfect chip.
• Some police stations are purposefully hiding their use of Flock, per 404 Media.
• Microsoft is giving its workers an AI budget limit to curb tokenmaxxing—but still wants to be an AI-first company.

^{*A message from our sponsor.}

This time last week...

Readers’ most-clicked story was about a startup that wants to fix the most annoying sound in your kitchen, without touching any of the hardware.