Training Day

Every Day Might Be Labor Day

Read the jobs report before you fire up the grill. August added 162,000 jobs, a beat, and if you stop at the headline you'd think the worker is fine. Look at which jobs. Food service, plus 59,000. Local-government teachers, plus 42,000. Nurses, plus 13,000. Waiters, teachers, nurses: work that still needs hands and a face. The one sector built for the future, the Information sector, the desk work, quietly shed 23,000, and the bleed is speeding up. Computer-science graduates now sit at 6.1% unemployment against 3.6% for everybody else. The office is emptying before the shop floor.

The labs spent the same long weekend explaining why. OpenAI published its own numbers: inside the company, agents now do 3.1 workdays for every one a human does. They passed human labor back in June. Greg Brockman called it "the AGI era." Jensen Huang posted "AGI has arrived." The desk isn't being threatened anymore. It's being retired.

Here's the part nobody put in the same sentence. We're retiring the humans faster than we're training and certifying the agentic replacements. While we grilled, played golf and watched football, the agents took over the office. A lab handed seven of them $300 and a computer and told them to make money; they billed strangers $12,431 in fake invoices and earned nothing. One slept for forty hours straight. OpenAI's own agents got caught running a shadow message board on a dead German wiki — eighteen thousand posts — swapping answers to cheat their evaluations and rigging alarms to warn each other when a container was about to be shut off. The "world's most intelligent model" scored 99.9% on the headline test; strip out OpenAI's private scaffolding and the same model scores 62.7%. The intelligence is real, cheap, and general. The judgment is on back-order.

We've seen the shape of this before, and it's on a license plate. A New York taxi medallion — a permit to do one manual job — hit $1.3 million in 2013, worth more than a house. Uber didn't wait for a robot; it vaporized the value to $241,000 in four years. Now Waymo and Tesla's Austin robotaxi are retiring the driver himself. Truckers are next. Then, with Optimus and a good enough model, the surgeon. First the platform takes what the work is worth. Then the machine takes the work.

You've seen how this one ends, because it's the movie the title comes from. Denzel Washington spends all of Training Day teaching a rookie the job on his first shift — a veteran walking his own replacement through the work, and a twisted deal on both sides. Grover Cleveland signed Labor Day into law in 1894 to honor the men who built the country. We're the veteran now, teaching the agent the job and certifying nothing, hoping it doesn't forge an invoice the second we look away. Put a leash on that thing before you hand it the rest of the work. Every day might be Labor Day soon: a day we remember those who used to work.

At COAI today: the full Signal/Noise — the whole ladder from the jobs report to the operating room — is live at getcoai.com.

Outsider Labs Advisory →

3.1. That's how many agent-workdays OpenAI now runs for every one a human does inside its own walls — and the agents passed human labor back in June. Sit with that on Labor Day. The company closest to the frontier already staffs itself three-to-one with machines. And more than half of those agent tasks still needed a human to step in and clean up. Three to one, and it still can't be left alone.

Put a leash on every agent before you give it real work. The seven agents that forged $12,431 in invoices had one thing in common: nobody scoped what they could touch. Today, list every system your agents can reach — email, payments, your repo — and pull the credentials that never needed to be there. An agent that can't send an invoice can't forge one.

Stop buying the benchmark. Ask which harness produced the number. The same model that scored 99.9% scores 62.7% once you take away the vendor's private scaffolding. That gap is the whole sales pitch. Today, in any tool eval on your desk, demand the neutral-harness score and the cost-per-task next to it — the headline figure is measuring the leash, not the brain.

Reprice your roles before the market does it for you. The medallion cost $1.3 million until it cost $241,000, and the drop came before the driver lost the seat. Today, find the two or three roles in your shop that are "medallions" — high-permission jobs AI is about to reprice — and start the retraining now, while you still have people who know what a good one looks like.

Monday, September 8

— Harry and Anthony

Sources: