Learning can seem optional when AI builds things for you

If you build things with an AI assistant, such as an app, a website or a spreadsheet, you describe what you want and the assistant does much of the work. Each thing you ask for is a requirement, and the assistant builds all of them, including parts you wouldn't know how to build yourself.

I think of getting from where you start, with nothing built, to the thing you want to build as a road, with the finished thing as your destination. On the way, the road crosses terrain: the subjects you'd need to know about to build the thing. The gaps in what you know are the uneven parts of the road's surface. Some gaps are rough patches, where you can guess and usually guess correctly, such as what a button should say. Other gaps are crevasses, where you can't cross until you leave the road for a while and learn something, such as how a website takes card payments.

An AI assistant fills in the crevasses and smooths the rough patches for you. Sometimes it builds the thing in a completely different way from the one you had in mind, which is a new road to the same destination. Either way, the AI builds the road and takes you along it: you get the finished thing, and you never have to learn the subjects you'd have needed to cross the crevasses.

If the AI can build the road and take you to the destination, it's fair to ask why you should learn anything. I think it's still worth learning some of the terrain, for two reasons.

The AI doesn't know your plans

The first reason is that the AI builds what you ask for, and it doesn't know your real-world situation or your long-term plan unless you tell it. Your real-world situation covers many things, such as who will use what you build and which laws apply to it. Your long-term plan is what you expect the thing to become in a few years. What you know of the terrain affects what you think to tell the AI and what you think to ask for.

Say you build an app that only your family uses. It runs on a computer at home, and only devices on your home network can reach it. It doesn't need logins, and privacy laws such as the EU's GDPR don't cover data a person handles only for their household.¹ But you expect to open the app to other people someday. Then it will need accounts, and privacy laws will apply, with rules on how long the app keeps people's data and when it has to delete it. You don't want accounts yet, but you want the app to store its data in a way that makes accounts and those rules easy to add later.

To ask for exactly that, you'd need to know enough of the terrain, which here includes some law as well as how apps work: how an app keeps each user's data separate, and what privacy laws ask of an app that holds other people's data. If you don't know how apps keep each user's data separate or what privacy laws ask, you'll probably ask for a family app and nothing more. The AI will then build one that stores everything as if only one household will ever use it, and adding accounts years later can mean changing how all of the app's data is stored and every part of the app that reads it.

What you know limits what you ask for

What you know of the terrain also limits what you imagine asking for. When people imagine something new, they tend to give it the features of things they already know.² Say you're building a website for a club, and a chat for its members would be useful. If you think a live chat would make the website much more complicated, you leave it out. But some backends, such as Convex, are realtime databases. If you build the website on one of those from the start, adding a basic live chat takes only a few lines of code.

Or say you build a free tax calculator and want people to be able to use it on their phones. If you don't know about progressive web apps, which are websites that people can add to their phone's home screen and use like an app, you ask for an app, and the AI builds you a React Native app for iPhone and Android. Then every new version has to pass each app store's review, you have two store listings to keep up to date, and people usually get a fix only after they update the app. For a simple free calculator, a progressive web app would probably have been enough, without any of that.

The AI usually won't suggest any of these on its own: storing the data so accounts are easy to add later, a realtime backend such as Convex, or a progressive web app. It builds what you ask for, and it can't plan for something you never told it about. So unless you mention your plan or ask what else is possible, it builds what you described and no more.

You can ask the AI to question you about your plans before it starts, or to suggest what else is possible. Both help, but the AI still works only from what you tell it, and it can't see the plan in your head. You also have to know which parts of your plan affect how the thing is built before you think to mention them.

Knowing the terrain also helps when the AI takes an odd route to the destination: it builds the thing in a way you wouldn't expect, or in a way that is too complicated or too simple for the job. If you know the landmarks and forks, meaning the main ideas involved and the points where a different approach was possible, you can ask the AI why it didn't take another route.

Learning how things work is enjoyable

The second reason to learn some of the terrain has nothing to do with getting to the destination. When the AI fills a crevasse, you can still choose to leave the road and learn the subject you'd have needed to cross it. Much of the technology you use every day without thinking is made of solutions people worked out to fix real problems. OAuth is what lets you sign in to a website with your Google account without giving that website your Google password. TLS encrypts what your browser and a website send each other, so someone else on the same Wi-Fi network can't read what you send, such as a card number. HTTP/2 lets your browser fetch many pieces of a web page over a single connection at the same time. Finding out how things like these work is enjoyable in itself.

I think this kind of learning also gives you more ways to think about a problem, and sometimes an idea from one subject helps with an unrelated part of your life or work. An idea usually gets used in an unrelated area because you went looking for where else it applies. People often don't use a solution they already know on a new problem until someone tells them it might fit.³

If learning how things work doesn't appeal to you, the first reason applies without the second. The plan for what you're building is still yours, and so is the choice of what to ask for.

You don't need to learn everything first

The examples in this post range from privacy law to how browsers talk to websites, which could make it sound as if you have to learn everything the road crosses before you build anything. But in reality, you need to learn much less than everything the road crosses: enough of the terrain to make the decisions only you can make, such as what to ask for now and what to leave for later. You may find that learning that much is enjoyable as well.

Sophia is an app that builds private courses around what you want to learn. You can ask it for a course on a general subject your project depends on, such as how apps keep each user's data separate, or on something you're curious about for its own sake.

Sophia is coming to the App Store, iOS first. Join the waitlist for an email at launch.

Footnotes

1. Regulation (EU) 2016/679, the General Data Protection Regulation, Article 2(2)(c): the regulation does not apply to the processing of personal data "by a natural person in the course of a purely personal or household activity". ↩

2. Thomas B. Ward, "Structured Imagination: The Role of Category Structure in Exemplar Generation", Cognitive Psychology 27, 1994. Participants asked to imagine and draw animals from another planet mostly gave them the features of animals on Earth, such as matching left and right sides, eyes and legs. ↩

3. Mary L. Gick and Keith J. Holyoak, "Analogical Problem Solving", Cognitive Psychology 12, 1980. Undergraduates read a military story, then tried a medical puzzle that the story's method could solve: 11 of 12 used the method when told the story might help, and 3 of 15 did without that hint. ↩