How to learn with AI, one project at a time

Choose something worth making, understand what it takes, and put your learning into practice.

There are things I want to learn because I want to do something differently.

Explain an idea clearly when it matters. Make something I care about. Understand enough to ask a better question or make a more considered choice.

That is what interests me about learning with AI. You can bring it something you want to do, work through a part you don’t understand, and put the explanation to use while the question is still fresh.

The practical question is where to begin.

I’m exploring an approach that starts with a project you care about and a small part you can attempt. Get useful help, try it yourself, and examine what happened.

Choose something that matters. Start small. Learn from your attempt.

Start with something you care about making

“Learn to communicate better” is a large ambition.

“I want to contribute an idea in a meeting and explain why it is worth considering” gives the learning somewhere to go. Your first task might be to practise a short explanation and answer one reasonable question.

That first part has a purpose beyond completing an exercise. It prepares something you want to be able to do in the conversation.

Your project might instead be a family video, a simple tool, or understanding a topic well enough to explain it to a friend. Choose something you want to change, begin or explore, then find a manageable part to work on.

Dan Koe’s writing on learning with AI emphasizes starting a project and letting the work expose what you need to learn. Ali Abdaal illustrates a similar approach through building a home security system: the thing he wants to make gives unfamiliar concepts a purpose.

For the meeting, being able to explain your proposal may matter more than making slides. The project helps you choose the gap that serves your purpose now.

Separate two aims:

• What I want to do: contribute a proposal in the meeting.

• What I want to learn: explain its reason with an example and answer a reasonable question.

A polished paragraph and the ability to use its reasoning in conversation need different checks.

Of course, curiosity needs room to wander. A book can be worth reading without becoming a project. And some skills need foundational study before you can use them responsibly. This approach is useful when you want to turn an interest into something you can practise.

Give AI a specific job

Once you have a project, try a small part with what you know. An unfinished paragraph or rough sketch can reveal a question that a broad request for a learning plan would miss.

If you cannot begin, a relevant example can give you something to examine. You don’t have to name every skill you lack before learning from a small part of someone’s work.

Another idea I’m taking from Dan Koe is to choose an example of the result you want, examine how it works, and use that understanding while making your own attempt.

Suppose your meeting proposal is to protect a period of focused work while keeping time-sensitive contact possible. Here is a small, invented example of the distinction you need to explain:

“Before working on something that needs your attention, silence alerts that can wait. Keep a way for someone to reach you if a response cannot wait.”

You could ask AI:

“Point to one choice in this explanation that I could learn from. Show me the words you’re using as evidence, explain how they help, and help me try that choice in my own example.”

There is a choice you can see: the explanation separates alerts that can wait from contact that cannot. That gives the listener a decision to make. We can discuss that effect without pretending to know what an author intended.

Now try removing three words: “that can wait.”

“Silence alerts” is shorter. But imagine you’re waiting for a time-sensitive call. The shorter version no longer tells you which alerts the advice applies to. It could lead you to silence the very contact you need.

That is a useful question to bring to AI: which detail can I remove, and which detail keeps the explanation accurate?

For your attempt, explain the proposal in your own words: what you suggest, why it could help, and how someone could reach you when a response cannot wait. Use a routine reminder as the example.

Then answer the question: “What happens if something needs a response now?”

Can your explanation distinguish the two cases? If it cannot, restore the condition and try again. You can practise this alone, then take the explanation into the actual conversation.

Keep the original beside you while you examine AI’s interpretation. A polished explanation or a popular example can still contain a weak claim. Learn from a choice you can point to, and check whether its reason fits your project.

When you have made an attempt, give it that attempt:

“Here’s my explanation. Point to the first place the reasoning becomes hard to follow, tell me why, and help me improve that part.”

One extra question is useful when an answer rests on something uncertain: what do we know, what are we assuming, and what is still missing?

Suppose the AI recommends simplifying your talk because “the audience won’t know the subject.” Did you tell it that? Or did it fill in the gap?

Naming the assumption gives you a choice. You might keep it provisionally, correct it, or find out before proceeding.

Give understanding something to do

A clear explanation is a good beginning. I still want a way to find out whether I can use it.

Nick Saraev’s discussion of learning with AI includes a step I find particularly useful: after getting help, return to the task and try it again independently.

Close the explanation and try the relevant part yourself. For the meeting, say your proposal without reading, then answer the time-sensitive-contact question.

If you return to AI for feedback, give it the words you actually used. Ask it to preserve what works, explain one important gap and help you try again.

Compare the two attempts against the same check: can a listener tell when an alert can wait and how a response that cannot wait would reach you?

A revision that makes that distinction clearer is a specific improvement in this explanation. You still need the real conversation to find out how people respond.

There are two checks here. Does your explanation match the source? And is the source’s claim well supported? Understanding an answer doesn’t mean you have to agree with it.

AI can help compare passages and identify a disagreement. Open the original material before relying on that comparison. Check whether a cited source actually supports the claim, and whether supposedly independent accounts trace back to the same place.

When the consequences matter, bring the work to someone qualified to assess it. An encouraging AI response cannot certify your understanding.

Even a successful attempt has limits. It tells you something about this task, under these conditions. A later attempt or a different situation may expose another gap.

This newsletter is a useful place to try

This newsletter is one project I could use. I want to explain the approach well enough that a reader can begin, and understand when to adapt it myself.

My attempt would be an explanation with one example whose details I can defend. My check would be to close the draft, explain why those details matter, trace a claim to its source, and try a situation where the advice changes.

That is proposed practice, not a completed experiment or a result I’m claiming.

Put it to work on your own project

You can begin with two sentences:

I want to make or do…

By the end, I want to be able to explain or do…

Choose a small part to attempt. Get help with the gap you find. Then try again, with a way to check the result.

You do not need to map the whole subject before beginning. You need enough understanding to make the next attempt useful, and enough honesty to notice what remains unclear.

The companion for this issue, Learn Through a Project, helps you start with something that matters to you. Bring your project and, if you know, the part you want to learn. It will explain one relevant part and give you a small attempt you can make now, with a way to check it.

Take that into your own practice, or return with your actual attempt for one useful correction and another try. The point is to use what you learn in something you care about doing.

Try Learn Through a Project

What would you like to make, and what would you like to understand by the time it is finished?