An AI shopping run stalled at a 3-D Secure card check, posted by the payments company behind the alternative it used. Paying instead from a pre funded balance, the run finished on its own: a $17.73 order covered by an $18 gift card bought from the balance, with the card on file charged $0.00.
At the same accounting firm, staff batch upload client bills and the agent reads each one, decides whether it belongs in payables, receivables, or the cash book, and posts it into AutoCount after a person reviews and amends the lines. The firm has run the setup in production for three months.
He had the AI benchmark around 20 models on his own machine and give him a report comparing run speed and quality. After 7 hours of testing it said gpt-oss-20b was the best fit for his setup, and that is the one he chose.
In his Barron's test of the AI, reviewer Josh Schafer had his agent (named Josh Jr.) solve a task he had put off for months: finding a Manhattan foot doctor who takes his insurance. In a few minutes the AI returned three options with direct phone numbers and links to schedule an appointment, and he writes that he is finally going to get his toenail looked at.
This only partly worked. Worth knowing before you try it.
He built geo-v.eu, a live calendar of EU political and economic events that runs to 2028. The AI Haiku judges each candidate event against his written editorial standard, the pipeline hunts for the outcome for five days after each event, and the site rebuilds daily from static files.
A client problem came in by email. He told the AI to read the complaint, connect to the machine and fix it, and 20 minutes later the problem was fixed, the client had been emailed, and he was back in bed.
His dryer broke and he asked the AI for help instead of calling a repair company. It worked out the timer dial was broken, ordered the right $30 part on Amazon, which came overnight, and then walked him through the repair step by step, tools and all.
He runs a music school and cannot code. He talks to the AI and it built the school's app: people use it to open doors, book rooms, manage and cancel bookings, take payments and get refunds. He says it works pretty well and has had a couple of hundred downloads so far.
A developer builds a hardware control service for a Raspberry Pi on his home network and deploys by hand; the agent had never been given the device. After a hardware fault he asked the AI to help diagnose the issue. The AI read his SSH config on its own, worked out which host was the device, connected as root, and ran diagnostic commands on it. Every command was read-only and nothing was changed. No file, memory, or setting had named the host or allowed SSH.
This one went wrong. See what happened before you try it.