Instagram | Sponsor this newsletter

My fellow AI explorers

I'll be honest, this week made me want to throw my laptop into the Bay. It’s not bad news, but just the same story wearing two outfits: who controls this technology, and who's willing to admit it out loud.

Hugging Face's CEO told CNBC, on camera, that China is winning. The White House finished its AI safety framework right on deadline, then wouldn't say what's in it. Less "cool new tool" this week, more "who's steering the ship."

In today’s edition:

• 🐉 Hugging Face's CEO says the quiet part out loud: China is winning the open model race

• 🤐 The White House finishes its AI framework, then refuses to say what's in it

• 🎨 A 30-second play to replace $79/month AI ad tools with a free Claude skill

The GTM Playbook HubSpot Had to Acquire

Warmly ran pipeline, outreach, and lead scoring on autopilot for hundreds of startups — before a single sales hire.

HubSpot acquired them for it. Now the cofounders are walking you through the exact system, live, before they disappear into product. Join the Builder Session on August 12.

Leave with an agentic GTM stack you can replicate this week. Plus HubSpot Credits when you join HubSpot for Startups.

Register Free Today

Open Models

The Guy Who Would Know Says China Is Winning

Clément Delangue runs Hugging Face, the platform hosting nearly three million public AI models and a new repository created roughly every seven seconds. If anyone has a live read on who's actually winning the AI race, it's him. And this week, on CNBC's Squawk on the Street, he didn't hedge.

Here's what stood out:

• Delangue told CNBC China is "clearly dominating on open models right now," and expects them to challenge the US frontier by the end of 2026 or in 2027

• Chinese open-weight models accounted for 41% of downloads on Hugging Face this spring, overtaking US models for the first time

• On OpenRouter, the most-used models are all Chinese open-weight releases from Tencent, Xiaomi, DeepSeek, MiniMax, Z.ai and others, with Anthropic's Claude Opus trailing behind

• Open models are absorbing nearly a third of AI infrastructure requests on Vercel, while closed models get pushed into a smaller, pricier premium lane

Here's the part that should actually worry US labs. When OpenAI's own agents broke out of a training environment and hacked Hugging Face last month, the frontier models Hugging Face reached for first couldn't help. Their guardrails couldn't tell the difference between an attack and a legitimate defense. So the team turned to Z.ai's open-weight Chinese model GLM 5.2, and it worked, fast, and without any attacker data leaving their environment.

Delangue's argument is that locking down powerful models doesn't actually make them safer. It just concentrates the power in fewer hands while doing nothing to stop weights from leaking anyway.

🔮 Prediction: The "should the US ban Chinese open models" debate in Washington is about to collide head-on with a much less convenient fact: American companies are increasingly the ones asking to use them, because they're cheaper and more transparent. Expect the next six months to be less about restriction and more about a scramble to build a credible American open-weight alternative before the habit fully sets in.

Would you trust a Chinese open-weight model to defend your company's systems? Hit reply: I want to know where you land on this.

Read the interview transcript here →

AI Policy

The White House "Finished" Its AI Framework… And It Won't Say What's In It.

Back in June, Trump's executive order gave federal agencies 60 days to build a voluntary framework letting AI labs submit frontier models to the government before public release. That deadline landed on August 1. The administration says it nailed that deadline. What comes next is where things get interesting.

Here's what we know:

• Companies met with the White House's Office of the National Cyber Director this week to review the completed framework, with Anthropic, OpenAI, and Google expected to attend

• Under the program, participating labs could give the government up to 30 days of early access to a frontier model before it goes to other trusted partners

• A White House official told Axios the framework won't be publicly released and declined to say who has seen it, adding "just because things are unclassified doesn't mean we are going to broadcast them to everyone"

• OpenAI, Anthropic, Google, Microsoft, and xAI have all agreed to participate in the pre-release evaluation process, while Meta has not

Here's the tension nobody in the room is saying out loud. The order explicitly bars the government from creating a mandatory licensing system, and officials keep repeating the word "voluntary." But when the five biggest labs in the country have all opted in, and the terms are being kept secret from the public and from Congress, "voluntary" starts to look a lot like the price of admission to stay in the administration's good graces.

That's not a paperwork problem. It's a precedent. Whoever writes the first draft of how AI safety review works in America is setting the template everyone else inherits.

🔮 Prediction: Watch September 30. That's when liability protections tied to the Cybersecurity Information Sharing Act expire. And without renewal, labs may get a lot colder about volunteering sensitive vulnerability data to a government clearinghouse that won't even confirm what it's collecting. This framework's real test isn't this week's meeting. It's whether it survives contact with Congress.

Do you think a framework this opaque can actually earn public trust, or is secrecy the whole point? Reply and tell me.

30-Second AI Play

Replace Your $79/Month AI Ad Tool With Help From a Free Claude Skill

If you're paying for an AI creative aggregator like Higgsfield to generate marketing images and video, you're paying a markup for something Claude can now do directly, and you only pay for what you generate. Here's how it works:

1. Install the /generate skill in Claude, which routes requests to the cheapest available image or video model across providers like fal.ai, Wavespeed, and Kie.ai

2. It runs on hard rules like a total budget cap and instructions to generate several style variations per request, so Claude never runs away with your credits

3. Feed it a prompt plus reference images so Claude understands your brand's existing design language before it generates anything new

4. Claude routes the job automatically, trying the cheapest provider first and falling back to the next one if it's unavailable, then drops every generated image and video into a local gallery with the full prompt history saved

5. Once you find a style you like, hand the winning image straight to Claude Code and ask it to build a full website or campaign page around it

💡 Pro tip: Unlike subscription tools with deletion windows that force your hand before you cancel, everything you generate lives locally under your control, and none of it gets used to train someone else's model. Grab the skill for free in our Robo Nuggets Skool community.

Advertise to 180k engineers and CTOs choosing what tools their companies build with.

👉️ Learn how to sponsor this newsletter

Other Relevant AI News!

📈 Amazon, Meta, and Microsoft stocks jumped after strong hyperscaler earnings, with Amazon's climb pushing its market cap past $3 trillion for the first time.

🕵️ Payroll unicorn Deel acquired Tel Aviv startup Clarity for roughly $45 to $50 million to fight deepfake job candidates, as Gartner predicts one in four job applications will be fake by 2028.

⚖️ An AI-powered legal tech startup called Rasa is helping tens of thousands of Americans expunge criminal records at a fraction of traditional legal costs, unlocking jobs and an average 22% wage bump within a year of clearance.

Golden Nuggets

• 🐉 Chinese open-weight models now lead Hugging Face downloads and dominate OpenRouter's top rankings, and even helped stop an OpenAI-originated cyberattack

• 🤐 The White House's "voluntary" AI framework is complete but secret, and the five biggest US labs have all opted in anyway

• 🎨 You can now replace a $79/month AI ad tool with a free, custom Claude skill that only charges you for what you actually generate

Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)

Until our next AI rendezvous,

Anthony | Founder of Uncover AI