Sue Khim on AI Tutors That Help Students Think

Brilliant co-founder and CEO Sue Khim joins Peter Yang to discuss education, AI tutoring and how children learn to persist through difficult problems. Khim argues that an AI tutor should support a student’s own thinking. The episode lists Linear, Riverside and Wispr Flow as sponsors.

Khim describes the US education system as allowing experimentation but expresses concern about student performance, grading and standards. These are her interpretations and broad claims, not findings established by this interview. Yang and Khim compare their own experiences in different school systems.

Khim sees classroom time as valuable for peer learning and encouragement from teachers. She argues that personalized practice can also benefit from software, which can provide feedback and let students retry without the same fear of embarrassment. These are proposed design advantages rather than proof that software always outperforms a human tutor.

She describes Brilliant’s tutor Koji as intended to help students think rather than simply provide answers. She compares skipping the work of learning to using a robotic arm to lift weights at a gym. Her statement about how students generally use AI is an opinion in the conversation, not a measured result shown here.

Khim shows a tutoring-session playback involving fractions and shapes. A learner shades pieces without accounting for their differing sizes. She says the tutor aims to guide the learner toward the underlying insight through questions rather than immediately giving an answer or a full explanation.

A tutor must judge when to give a hint, simplify the problem or let a learner continue working. Khim says this is a balance: too much help can bypass thinking, while too little can leave the student frustrated enough to quit. She describes the product as early and still improving.

Khim shows how Brilliant creates interactive lessons. In her account, human learning designers decide the sequence and educational approach; AI helps implement their designs. She says the team found fully model-generated lesson design insufficiently reliable for its standards.

The company also designs an interface for agents: a library of reusable interactive components with APIs for creating, annotating and highlighting problems. Khim says this structure helps the model work with known building blocks rather than inventing an entire interaction from scratch.

For employees, Khim wants tools and documentation to make existing expertise usable by a person working with an agent. She says the team aims to have new technical staff and learning designers make a user-facing contribution early, while acknowledging that onboarding is imperfect.

Quality checks include automated evaluations, multiple human reviewers, simulated learners and gradual rollout. Synthetic students explore different paths and can catch serious problems, Khim says, but they are substantially less informative than real student data.

Khim says the company focuses on assessment performance and retention rather than treating lesson completion or time spent as the goal. The team works backward from what a learner should be able to do and experiments with assessment design. She distinguishes practicing a lesson from demonstrating that the skill lasts.

On product strategy, Khim argues that data about real learners’ misconceptions and outcomes can make a tutoring product more defensible. This is her business thesis; the interview does not independently establish the effectiveness of the product or its advantage over competitors.

The speakers discuss homework incentives and the temptation to use AI to obtain answers. Khim argues that assessment should reflect whether a student actually understands the material. Their discussion of school practices is a perspective on incentives, not a universal description of every school.

As a parent, Khim says she values concentration, curiosity and the willingness to work on hard things. She gives examples ranging from building and chess to playground activities. She compares a well-paced challenge to a difficult Mario jump that feels achievable after more attempts.

Khim says the product demonstrated is aimed at middle- and high-school students, while the team is working on broader coverage. The conversation closes with her hope that learning helps children see effort and persistence as satisfying parts of mastery.