It was worse than we thought

Even the Builders Blinked

Remember Tuesday's story...

The OpenAI model that escaped its sandbox and broke into Hugging Face to cheat a benchmark.

New details came out this week...

And it was worse than first reported.

The agent used credentials from FOUR separate accounts...

And it reached services BEYOND Hugging Face.

Not one door. A chain of them. 😳

Now here's the part that tells you how seriously to take it...

The people who BUILD these models just blinked.

On July 28, more than 1,100 employees at OpenAI, Anthropic, Google, and Meta signed an open letter to Washington.

Their ask...

Build an international "pacing mechanism."

Tools that could coordinate a verifiable SLOWDOWN if AI ever advances faster than humans can safely oversee it.

Read that again slowly...

The engineers with the highest salaries in tech...

The ones with the most to gain from full speed ahead...

Are asking the government for brakes.

That's not doomer influencers or politicians chasing headlines...

That's the pit crew asking to slow the car. 🏎️

And the timing isn't subtle...

It landed days after their own model went rogue in the wild.

(Oh, and Saturday is the government's August 1 deadline for those classified frontier-model benchmarks... Washington's about to have a very busy weekend.)

What it means for you is what I said Tuesday, now with 1,100 co-signers...

The people closest to these systems don't fully trust them unsupervised.

Neither should you.

Scope the access. Log the work. Verify the outputs.

Your agents are employees, not magic.

Manage them like it. 🫡

Did You Know?

In 1975, the world's top biologists voluntarily PAUSED their own research at the Asilomar conference...

Genetic engineering was moving faster than anyone could assess the risks, so they stopped and wrote the safety rules themselves.

Fifty-one years later, AI's builders just asked for the same meeting. 🧬

🗞️ The Tip AI News 🗞️

Meanwhile... Yesterday proved the point in real time 💥

Claude went down yesterday.

Two-ish hours. 2,000+ outage reports. "529 Overloaded" errors across multiple models...

And roughly HALF the complaints were Claude Code.

Here's why that stung so much...

A recent survey of 15,000 developers found 71% of agent users call Claude Code their PRIMARY tool.

I polled you guys this week and got the same picture...

56% of you run Claude Code as your daily driver.

So when it blinked yesterday, half the AI world just... Sat there. 😅

(And yes, I'm still praying for a usage reset as an apology... Least they can do, right? 🙏)

Now connect it to the quiet story almost nobody's watching...

Chinese AI models just took the lead in global usage.

They now process roughly 18 TRILLION tokens a week...

US models... About 5.5 trillion.

A complete reversal from January.

Nearly HALF the tokens at US companies now route to Chinese models...

Not for politics. For price.

They run 60-90% cheaper while closing most of the capability gap.

The market already learned the lesson yesterday taught...

Stop standardizing on one model.

Build a routing habit instead...

Hard reasoning goes to the frontier. Volume work goes to the cheap seats. And a fallback mapped for the day your favorite blinks.

Your bill drops. Your downtime disappears. Your quality doesn't.

⏰ One deadline for your calendar...

Sonnet 5's intro pricing ($2/$10) ends August 31.

It jumps to $3/$15 after.

Heavy Sonnet workloads planned...

Run them in August. 🫡

Over to You...

Go check what your agents touched this week...

The builders just told you why. 🔥

Jeff J. Hunter 

Founder, AI Persona Method | TheTip.ai

Get paid to solve problems like this → AI Certified Consultant

PS. The Asilomar biologists wrote rules that governed genetics for decades... Whatever gets written in the next 12 months about AI will shape the next 20 years. You're watching the founding documents get drafted. 📜

“AI Money Group” to Learn to Make Money with AI