Homework up, exams down

AI keeps lowering the barrier to learning something new. You can ask about anything and keep asking follow-up questions until it makes sense. If you're curious, or willing to pick up new skills, that's great news.

It's also making surface-level learning more common. You get the answer, it sounds right, and a week later it's gone.

A study of 26,811 secondary school students in China put numbers on this. Within six months of starting to use AI, students cut their homework time from 64 minutes to 45 and raised their homework scores 18%. Their closed-book exam scores fell 20%. Students chose when to start using AI, so the study can't prove AI was the cause, but a randomized trial in Turkey found a similar 17% drop in math exam scores.

Most of the drop in China came from the four in five AI users who finished homework faster than the fastest students without AI, with unusually high scores. The researchers read that as handing the homework off. AI users who kept spending normal time on homework did about as well on exams as students who didn't use AI.

As that kind of shortcut gets more common, I think what will be rare is people with the patience and discipline to actually learn things.

The best way I've found is to learn by doing. Start building the thing, and when you get stuck, use AI to get moving again or to explain what you're missing. Then go back to building. Scott Young makes the same case in Ultralearning, this week's book note.

One warning sign from the study: if you're trying to learn something and AI makes it much faster, you may have skipped part of the learning.

What's something you learned with AI that you could still do without it?
Hit reply and tell me. I read every response.

Have a great week!

Graham

📚 Book Notes

Ultralearning - Scott Young

Ultralearning is a how-to book, so it's a fair test of the Ferriss argument below.

You could ask a chatbot for Scott Young's nine principles and have them in seconds: metalearning, focus, directness, drill, retrieval, feedback, retention, intuition, and experimentation.

The list on its own won't teach you much, and the book explains why. Young's main advice is to learn by doing the real thing, which he calls directness, and to test yourself instead of reviewing. He points to research where free recall beat re-reading, even though students felt re-reading worked better.

Reading a summary is a lot like re-reading. The learning starts when you close it and try the thing yourself.

📖 Article

Has AI Already Killed How-To Nonfiction? - Tim Ferriss

Tim Ferriss published the US print sales for his five books, and the drop is steep. Sales fell 13% in 2024 and 46% in 2025, and 2026 is on pace for another 57% decline.

His explanation is that books like The 4-Hour Body work like lookup tables. You flip to the answer for losing fat or fixing your sleep. A chatbot that has read his books, and thousands of others, can now give you a version adjusted for your weight, schedule, and injuries in a few seconds.

He also makes the case for reading the whole book. Years ago, busy friends asked him for a few bullet points instead of the 608-page book. He says none of them followed the advice, while thousands of readers who went through the whole book lost more than 100 pounds.

From the Blog

📝 How to Learn Anything Faster: Build Before You Feel Ready
The loop I use now: build something real before you feel ready, then retrieve what you know, get feedback, and drill whatever keeps breaking.

📝 Just-in-Case vs. Just-in-Time Learning
AI makes just-in-time answers almost free. This older post looks at when it's still worth learning something before you need it.

🔗 Things I Found Interesting

LLM Wiki (Andrej Karpathy)
An idea file for having an AI agent build and maintain a personal wiki from what you read. One of his examples is reading a book and filing each chapter as you go, so your notes keep linking to everything else you've read.

How to Take Book Notes
My older answer to the same problem: highlight on an e-reader, copy everything into a document, clean it up, file it in Evernote, and re-read. It's far more manual than Karpathy's version, but the goal was the same: remember more of what I read.

How to turn your AI into a world-class designer (Anshu Chimala, Lenny's Newsletter)
Chimala led software engineering and design teams at Apple for 12 years. His point is that models drift toward the safest, most predictable choice, which is why so much AI design looks the same. He walks through the process he uses to push them past it.

📣 Quote

"I'd rather write books for 10,000 people than make short-form video clips for 10,000,000." - Tim Ferriss