You never ask about what you've never heard of

Every conversation with ChatGPT or Claude starts with an empty box where you type a question. This works well when you have a specific question and already know enough about the topic to judge the answer.

Say you've baked bread for a while and your last loaf came out dense. You ask why. Because you already know your dough and your oven, you can check the answer against what you know. You can tell whether the explanation fits what happened, ask a better follow-up question if it doesn't, and try the suggested change on your next loaf.

With a subject you've meant to learn for a while, you have much less to check an answer against, because you don't yet know what the subject contains. Take Byzantine history. You might have heard of the Byzantine Empire and know that Constantinople fell in 1453, but you don't have a specific question to ask. So you start with the one thing you've heard of, and you ask why Constantinople fell.

ChatGPT or Claude answers that question. After that, what you learn next depends on the next question you type, and that question can only come from what you already know. A paper from 1979 used this idea as its title: "To ask a question, one must know enough to know what is not known."¹ After the answer about 1453, most of what you know about Byzantine history comes from that one answer. So your next question comes from that answer too.

Because each new question comes from the last answer, you learn the parts you happen to come across, and you miss the parts you don't know exist. Questions that start from 1453 lead to the siege, the city walls and the Ottomans who captured the city. You can keep asking questions like these for a long time and never learn how the Byzantine Empire was governed for more than a thousand years. The answers about the siege don't mention government, and you didn't know to ask about it, so you never ask.

You probably won't notice that anything is missing. Each answer fully answers the question you asked, so it feels complete. It doesn't list the other parts of the subject, so nothing on the screen shows you what you haven't asked about.

Byzantine history takes more than one evening to learn, so you come back to it another day. A chat answer only answers the question you asked. Nothing in the answer keeps track of how far you've got in the subject or tells you what to learn next. So each time you come back, you start again with an empty box and have to work out what to ask.

One solution is to ask ChatGPT or Claude for a syllabus, which is a list of the topics to learn, in order. You can save the syllabus in a project, which keeps your chats about the subject together, and set a reminder to come back to the subject. This helps, because a syllabus lists parts of the subject you wouldn't have known to ask about. But the syllabus is then a plan you have to manage yourself. Each time you come back, you have to find where you stopped, decide whether you understood the last topic well enough to move on, and ask the assistant to teach you the next topic.

ChatGPT and Claude also each have a mode made for learning, which sounds like it might manage this plan for you. These are ChatGPT's study mode and Claude's learning mode. Both modes change the way an answer is given. Instead of a direct answer, you get questions that guide you towards it, and ChatGPT's study mode also gives hints and quizzes.² But both modes still start from the question you type. Deciding what to learn next, and remembering to come back, is still your job.

For a subject you've meant to learn for a while, I think the alternative is a course that manages the plan for you. Here, a course means a series of lessons on the whole subject, in an app or on a website, that you work through over as many sessions as you need. The whole course is planned before you ask your first question. So the course can start with the basics and put each idea after the ideas it depends on. Between sessions, the course stores the plan and keeps track of where you are in it. With a syllabus, you had to do that yourself.

Before you start, you can read the course outline: the list of everything the course will cover, in order. So you can already see the parts of the subject you wouldn't have known to ask about. A course on Byzantine history would include how the empire was governed, at the right point in the order. You would read the lessons on government when you got to them, whether or not you had thought to ask about government.

Sophia is an app that builds private courses around what you want to learn. You only need to type the subject, such as Byzantine history. If you want the course to use a paper, an article or a YouTube video, you can add it. Sophia plans the course with language models, the kind of AI behind ChatGPT and Claude. The course is split into sections of lessons, in order, starting with the basics. Each lesson is a short card that explains one idea. You can look at the course outline before the lessons are written.

You can still ask questions in Sophia. Each course has a chat, and you can open it over the card you're reading to ask your question. The answer takes that lesson and the course outline into account. To answer your question, Sophia can also search the course's other lessons, the sources you added and the web. A question about one lesson is specific, and the lesson gives you the background to judge the answer. That is the kind of question a chat box works well for.

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

Footnotes

1. Naomi Miyake and Donald A. Norman, "To ask a question, one must know enough to know what is not known", Journal of Verbal Learning and Verbal Behavior 18(3), 1979. People learned a computer text editor from a manual: those with no training asked more questions about an easy version of the manual, and those with a little training asked more about a hard one. ↩

2. OpenAI, "Introducing study mode", 29 July 2025, which lists goal setting and progress tracking across conversations among features still being explored. Anthropic, "Introducing Claude for Education", 2 April 2025, which describes learning mode as guiding students' reasoning rather than providing answers. Both as of October 2026. ↩