Brilliant launches Koji, an AI tutor that teaches through interaction rather than explanation

Brilliant's new AI tutor, Koji, is a small green animated character that watches how learners interact with on-screen problems and responds by pointing, sketching, and annotating directly on the screen alongside them. Sue Khim, Brilliant's co-founder and CEO, describes it as a graphical tutor designed to feel like someone sitting next to you rather than a chatbot fielding questions.

“We just launched Brilliant's new AI tutor. His name is Koji. He's a little green guy. He helps you learn how to think and problem solve in math and coding... this tutor can see how you're interacting with all the concepts and the problems you're solving and point and sketch and annotate on the screen with you. We call it a graphical tutor because it's designed to feel like someone sitting next to you looking over your shoulder. A big marker of success of the tutor is whether or not he can make himself unnecessary.”

The core design bet

The distinction Khim draws is between explanation and understanding. Most AI education tools are optimized for the moment of explanation, she argues. Koji is designed for the moment of understanding, which means getting the learner to do the work rather than reading a wall of generated text. Research on human tutoring supports the approach: the more interactive the tutoring conversation, the more time the student spends on the material, and that time-on-task is the primary driver of tutoring's effectiveness.

Lowering embarrassment is part of that. Students who would never raise their hand in class or push back on a human tutor for a third explanation are more willing to sit in confusion with an AI tutor until something clicks.

What chatbots can't do

Khim's argument against raw chatbot use for learning is structural. When a student asks a question in a general-purpose chat interface, gets an answer, and leaves, the model receives no signal on whether the learner actually understood anything. There's no user model, no learning graph, no dense stream of real-time reps that would allow genuine personalization.

Building that infrastructure, she says, is not something a foundation model company is going to prioritize anytime soon. The UI has to be purpose-built for teaching. And some elements have to be fully deterministic, meaning the system must guarantee correctness rather than approximate it.

Success means becoming unnecessary

Khim frames the tutor's goal somewhat against typical consumer product logic. A good session ends when Koji has made himself unnecessary, when the learner is asking the questions Koji used to ask them. She draws the comparison to OkCupid, noting that a board member built that company knowing churn was a sign of success. Brilliant measures churn, but differently from most consumer apps.

Brilliant's product is live at brilliant.org and available on iOS and Android.