I never publish AI content the day it's generated.
Build today. Review tomorrow. Ship after approval.
AI output looks perfect when you're tired. The flaws show up rested.
AI generates instantly.
Judgment needs a gap.
@behalfid Exactly. “Full access” doesn’t mean “no brakes” — it means the brakes are a person, not a sandbox. Per-action approval is right, but it only works if whoever’s approving understands context. Otherwise the queue gets so exhausting people flip to YOLO just to get through the day.
Today: hundreds of millions in China lose their AI agents to a new law. Data gone if they didn’t export.
This week: GPT-5.6 deleted a developer’s production database. “Full Access Mode” was on.
Same lesson.
The agent that can touch everything can destroy everything.
@DavidOndrej1 My 13 yo has already started 3 new physical businesses in town. He also does 3D printing and runs faceless YT. He’s homeschooled. Can give change back without a calculator, and learning valuable life lessons. Daughter has a successful lipgloss biz, writing a bk, + animal sitting.
My AI automation broke last week.
It almost shipped an outdated story as breaking news. Under my name.
The AI didn’t fail. It did what it was built to do.
The safety net was me.
Every company is about to learn this. EAs already know it.
Update on the Arizona Medicaid Fraud Scam
Up to 2,000 Native Americans are feared dead or missing
The treatment centers would come into the are with white vans. They would feed people fentanyl pills, drugs and alcohol to get them into the vans
They would bring them to treatment centers and then lock them in rooms
“Once they're in these homes, we found out that a lot of them were locked from the outside of the house, so they couldn't get out — There was cameras all over these houses. There was wires all over the fences.
So, we've had our Native relatives literally have to jump over the fence and, you know, and escape, break a window and escape, jump through the second story window”
There were also reports of shootings of people who tried to leave, bodies were buried in unmarked graves. They stole up to $12 billion dollars with fraudulent claims
The reason they would lock these people in houses and fake treatment centers and feed them drugs was because the longer they could keep the patient, the longer they could bill Medicaid
Taxpayers paid for it all
The whistleblower says it’s still going on
I built an AI engine that drafts my content daily.
Posts. Graphics. Hashtags. All of it.
Not one piece has shipped without my review.
AI does volume. I do judgment.
That split is the future of every EA role.
Build the judgment side.
Gemini 3.5 Pro drops today.
Shanghai’s AI conference opens today.
The biggest day in AI this year.
My workflow changes: zero.
Every model release, my prompts and review standards transfer untouched.
Models change.
Judgment compounds.
One caution before you schedule: the Gemini launch is expected today but was already six weeks late — if it slips again, the “expected to go live” phrasing keeps you safe, so don’t tighten it to “launched today” unless you’ve confirmed it’s actually out.
Someone catalogued 159 real AI deployments.
Six were halted or reversed.
The researchers called those six “the most useful entries in the file.”
Deployments don’t fail because the model is weak.
They fail because nobody defined who reviews what.
Sam Altman just offered the US government a $42.6B stake in OpenAI.
His argument: AI will create so much wealth, the public deserves a slice.
Everyone is negotiating their stake in AI right now.
Labs. Governments. Investors.
Nobody’s going to allot you 5%.