Ben Thompson writes the tech newsletter and podcast Stratechery. Last week he interviewed Joanna Stern — a longtime tech journalist, formerly at the Wall Street Journal, who now runs her own YouTube channel and newsletter. The headline topic was Apple’s new folding iPhone.
I skipped past that part. The second half of the conversation, where two people who live in tech every day talked about how they use AI, was more useful.
The pickle test
Stern needed a new way to test iPhone battery life. Her producer joked that they should just get a phone poked over and over until the battery died. That night, she built a website for it using Codex, OpenAI’s coding tool — one that counts every poke, times the test, and logs the results. Then she had the AI buy the domain, pokethepickle.org, and publish the site.
By her own account, she has no real coding background. The whole thing took a few minutes.
Why this works
Thompson’s read: getting an AI agent to do something already built for humans — like booking a flight on an airline’s website — is hard. The agent has to fight through infrastructure that wasn’t designed for it. But building something new, for your own specific problem, means there’s no existing structure in the way. That’s where these tools do their best work.
Thompson manages several instances of a TV app called Channels and used to keep them in sync with a management app he built himself. He said he’ll never touch that app again — it’s easier to tell an AI agent, “go change the settings on all of them,” and it’s done a minute later.
Stern uploaded her kid’s school PDF calendar to Claude and asked it to build a Google Calendar and share it to her family’s Skylight display. No manual entry, no learning Skylight’s own interface.
The switch that has to flip
Thompson named the real barrier, and it isn’t capability. It’s permission. There’s a difference between vaguely knowing you could build almost anything with these tools and actually sitting down to do it. “There’s just this switch that needs to flip,” he said. Once it does, something like a pickle-poking website stops looking impressive and starts looking normal.
Where agents still stumble
Stern also ran three AI shopping agents through a back-to-school shopping test. All three struggled with real websites — one couldn’t book her a hair appointment without her logging into the account first, and another picked the wrong store, then had to start over. Same pattern, other direction: agents still trip on tasks built around existing logins and existing store layouts. They’re better at solving a problem nobody’s solved yet than at replacing the way you already click through a familiar app.
What to do with this
Stop waiting for someone to build the app that fits your exact problem. If it’s a real one-off — a calendar you need synced to an odd display, a tracking sheet only you’d use, a small script that saves you ten minutes a week — that’s the job AI is good at right now. Describe the outcome you want and let it build the thing. Save your patience for the tasks that route through logins and shopping carts. Those still need you.
;-)
Ernie
P.S. In the AI Workshop, we build small, practical things like this together every week — so the first one doesn’t have to be the hardest.