Read this comic as text
Slide 1: The rough-sheet student
Teacher: Remember board exams? The student who wrote the answer immediately, versus the one who used the rough sheet first?
Student: Yeah, the rough-sheet one usually scored better, even with the same knowledge.
Teacher: Exactly. Same brain, same syllabus. The difference was thinking before writing the final answer.
Student 2: Wait, is this about exams or ChatGPT? I thought these models get smarter every day from our chats.
Slide 2: Frozen after training
Teacher: Common myth, actually false. The model is frozen after training. It doesn’t learn from your chats; it only gets smarter when a new version is released.
Student: Then why does it feel sharper lately?
Teacher: The big recent jump wasn’t a bigger brain. Models learned to use the rough sheet. They’re called reasoning models.
Student 2: Meaning what, exactly?
Teacher: Instead of blurting the first answer, it writes rough working: breaks it into steps, tries an approach, checks itself, then answers.
Slide 3: Still next-word prediction
Student: But isn’t it still just predicting the next word, like always?
Teacher: Yes, underneath it’s still next-word prediction. It just predicts its own rough work first, and that becomes its plan.
Student 2: So it didn’t get a new brain. It just learned to slow down and think before answering, like the rough-sheet student.
Slide 4: Same brain, more time to think
Teacher: Exactly. Same brain, more time to think, and that’s where a lot of the “smarter” comes from.
Takeaway: Models don’t learn from your chats. Reasoning models just write rough work before the final answer.
The concept in plain words
A reasoning model is an AI model that writes out rough working before it gives you the final answer.
Think of board exams. Two students know the same syllabus, but the one who works it out on the rough sheet first usually scores better. Reasoning models do the same: instead of blurting the first answer, they break the question into steps, try an approach, check themselves, and only then answer.
This also clears up a common myth. A model doesn’t get smarter from your chats: it’s frozen after training and improves only when a new version is released. A lot of the recent jump came from this rough-sheet habit, not a bigger brain.
Underneath, it’s still next-word prediction. The model simply predicts its own rough work first, and that becomes its plan. Same brain, more time to think.