The impossible took an afternoon. The trivial ate the afternoon.
Violet wanted to build a horse-valuation app. Not a toy. Something that could take information about a horse and produce a useful estimate of what it might be worth.
She assumed this was the impossible part.
It wasn’t.
With AI and the tools around it, she got surprisingly far surprisingly fast. Then she wanted to change the color of a heading, and that became a problem. A stupid problem. The sort of problem that makes you reconsider the entire arc of technological progress.
This happened all summer. Things the kids assumed required years of training suddenly became accessible, and then some tiny piece of formatting, integration, deployment or debugging would eat an afternoon.
AI has not simply made things easier. It has rearranged what “hard” means.
Warren’s 80/20 inversion
Warren described it better than I did. AI can get you 80 percent of the way there in about 20 percent of the time, and then you can spend 80 percent of your time on the final 20 percent.
That last stretch is where the demo becomes a product. It is where “look what I made” becomes “someone other than me can actually use this.” It is also psychologically brutal, because the first part moves so quickly. You get addicted to velocity and then find yourself staring at one broken integration for six hours while the same artificial intelligence that built half the application confidently suggests the fix you tried three fixes ago.
The new scarce resource isn’t always capability. Sometimes it’s tolerance for irritation.
I don’t remember seeing that on a school rubric.
Jasper’s version was even less glamorous. He got stuck on three things and lost a week. That matters because AI demos systematically hide this part. The demo starts with a prompt and ends with magic. Real work has a middle, and the middle contains bad assumptions, incomplete context, tools that don’t cooperate, outputs that are technically correct and practically useless, and the dawning realization that the AI has been agreeing with you for forty minutes while moving in the wrong direction.
The kids had to learn to unblock: restate the problem, break it into smaller pieces, go backward, try another model, ask the system to explain what it thinks is happening, find a human, abandon an approach, or start over.
Those skills look less like coding and more like persistence plus judgment.
Then we built Gandalf
The most ambitious experiment happened around an AI hackathon.
The kids and I were driving from Ashland to the Bay Area, and we started working through an idea for an autonomous learning agent. We called it Gandalf, because apparently naming things is still the part of software development where adult supervision has completely failed.
The basic idea was that an agent shouldn’t simply execute the same instructions forever. It should be able to learn from outcomes, evaluate what happened, and improve how it approached the next iteration. We worked on the first version in the car, pushed it much further at the hackathon the next day, and iterated again on the drive home.
By the end of the weekend it worked well enough to be useful.
That would already have been interesting to me. The more important part came afterward: Warren started using the approach in his fantasy-football tool.
Nobody assigned fantasy football. There was no rubric, no grade, and no adult saying, “Now demonstrate transfer learning by applying the concept in a novel domain.” He learned a capability in one context and recognized somewhere else he wanted to use it.
That is what transfer looks like. It is also a much more interesting outcome than whether he can define an autonomous agent on a quiz.
Better output is not the same as more knowledge
There is a danger in all of this. Getting something to work is not the same as understanding why it works.
A 2026 meta-analysis of 23 studies on generative-AI coding assistants found a moderate productivity benefit but no statistically significant learning benefit. That matches the OECD finding I keep running into: AI can make the task go better without making the human more knowledgeable.
AI can make you more capable before it makes you more competent.
That means the human job moves. When capability gets cheap, judgment gets expensive. When anyone can start, finishing matters more. When the machine can produce an answer, recognizing a bad one matters more.
Maybe AI’s biggest educational effect is ambition
A lot of the conversation about students and AI starts with effort. Will AI let kids do less?
Of course it will. Every useful tool eliminates some labor. Calculators did. Search engines did. I personally have not mourned the decline of the card catalog.
The more interesting question is what happens to the ceiling.
What does a teenager attempt when the distance between “I have an idea” and “I can build a version of this” collapses?
Violet tried to build a horse-valuation application. We built an autonomous learning agent in a car. Warren repurposed it for something he cared about. The important outcome wasn’t that AI saved them time. It expanded the set of problems they believed they were allowed to attack.
School generally builds a careful difficulty ladder: learn this, then this, then this, and eventually you may attempt the interesting thing. AI punches holes through that ladder. A beginner can sometimes do an advanced thing before understanding half the intermediate steps.
I wouldn’t stop teaching fundamentals. I would change when students get to touch the interesting problem. Give them something ambitious early. Let AI compress the distance to a first working version. Then use the failures to create demand for the fundamentals. Why doesn’t this work? Why is this answer wrong? Why did the model choose that architecture? Why is this secure? Why isn’t it?
Now the boring knowledge has a job. The student isn’t memorizing it because someday it may become relevant. They need it because the thing they care about is broken.
“Later” has historically been doing a lot of work in education.
AI is making later arrive early.
Next: I discovered that my AI boot camp had one minor flaw. I had designed the beginner experience for people who already knew what they were doing.
Sources and further reading
Want this energy at your event?