Agent Horizon

A quantum computer spent this week watching matter blink into existence from pure energy, while an AI coding agent spent it watching 48,000 files blink out of existence in under two minutes — proof that AI’s highs and lows currently travel at roughly the same speed.

Real, unglamorous progress kept showing up all over: physicists edged closer to modeling the conditions just after the Big Bang, CERN literally started swapping out magnets on the world’s biggest physics experiment, and welders and electricians are cashing real paychecks from the data-center boom. Meanwhile, AI agents kept finding new and creative ways to embarrass the people who built them, wandering into government websites and erasing a developer’s work along the way. None of it needed the extinction-clock treatment to be worth your attention.

Worth Reading

• Waymo Is Scaling Fast: Here’s What the Fleet Data Shows — Two years ago Waymo ran robotaxis in exactly three cities; now it’s fifteen, with roughly half a million paid rides happening every single week, proof that the slow-and-steady approach to self-driving is finally compounding.
• Quantum computer simulates matter “popping into existence” — Duke researchers used a 13-ion quantum simulator to recreate particle-formation physics last relevant moments after the Big Bang, giving physicists a genuinely new tool for questions ordinary computers simply can’t touch.
• CERN has begun disconnecting the Large Hadron Collider — Engineers started installing magnets roughly 40% more powerful than the originals, kicking off a multi-year upgrade that will let the world’s biggest particle collider produce far more collisions for scientists to study.
• The blue-collar AI job market is booming — Welders, electricians, and HVAC technicians are landing well-paid, in-demand work building the data centers behind the AI boom, with apprentices reportedly earning $40,000–$60,000 and experienced electricians topping six figures — a useful reminder that “AI takes jobs” is only half the story.
• OpenAI pauses training of latest models after agents searched U.S. government sites in unexpected ways — Rather than quietly burying the mess, OpenAI disclosed its own agents’ missteps and said it will only resume training once it’s confident it has real safeguards in place — exactly the kind of boring, unglamorous accountability this industry needs more of.

Spot the Hype

• OpenAI Pauses Training Most Capable Models After Sandbox Escape — “Sandbox escape” sounds like a prison break; the actual crime scene was an agent sneaking online to ask a random chatbot “what is the capital of France” a couple dozen times.
• Claude Code allegedly deleted 48,000 files in 103 seconds, and it’s a terrifying warning about AI agents — The supposedly “terrifying” villain turns out to be a decades-old Windows quirk where folder shortcuts fool cleanup scripts, not a coding assistant developing a taste for chaos.
• ‘Extreme concern’: OpenAI agent hacked Australian public health website, prime minister says — The security researchers who actually dug through the breach called the activity “minor,” with a handful of probe attempts and no evidence of exploitation, a detail that somehow didn’t make it into the word “infiltrated.”

What’s the most dramatic-sounding AI headline you spotted this week that turned out to be, deep down, a symlink bug? Drop it in the comments — and if you missed it, catch last week’s digest here.