Classroom time and individual practice serve different purposes
Source 07:09In this conversation with Peter Yang, Brilliant co-founder Sue Khim describes how she wants AI to help students learn without doing the learning for them. The discussion moves from a tutoring example to the product machinery behind it and the outcomes the team tries to measure.
Khim values what people can do together in a classroom: peers learning from one another and a teacher whose enthusiasm motivates the group. She contrasts that with the challenge of giving every student individualized practice and feedback in a large class.
She argues that software can make some forms of personal instruction more accessible and adapt its graphical interactions as a student works. These are the product opportunities she sees, alongside the distinct value of a teacher and peers, rather than a claim that every educational function should be replaced by software.
Giving an explanation can still remove the student's work
Source 09:58Khim compares using AI to skip learning with bringing a robotic arm to lift weights at the gym. Her aim is to use the tutor's intelligence to protect the student's own thinking time.
In a playback of a real tutoring session, a learner works on fractions by shading parts of shapes. The misunderstanding is that different pieces do not represent equal amounts. Khim does not want the tutor simply to give the answer, or even immediately deliver the explanation an adult might find obvious.
She says an explanation can feel satisfying while letting the learner avoid finding the insight. The tutor instead tries to keep the student exploring why the answer is wrong. The session's voice interaction is not shared for privacy reasons, so the visible playback is not a complete record of everything the student and tutor said.
At 11:56, the original playback shows unequal pieces in the same shape. The private voice conversation is not reconstructed.
Productive difficulty needs pacing and judgment
Source 13:52Protecting effort does not mean leaving a student stuck indefinitely. Khim describes non-AI work on pacing, feedback and hints, together with the tutor's decisions about what to do next.
Is the error small, or does it reveal a larger misconception? Is it time for a hint, a simpler problem or a drawing? Should the tutor let the student try for another minute? The balance is between useful effort and frustration that makes the learner leave altogether.
She describes the product as early and the judgment as an active area of work. Easier answers could improve short-term engagement, but the team's stated priority is helping students stay with something difficult and think more effectively. That priority requires more than maximizing the number of completed interactions.
Humans design the learning sequence; AI implements it
Source 18:05Khim shows a learning designer's sketch for a coordinate-plane lesson. The human chooses the pedagogical story, the questions and their order; AI helps implement the design.
She says attempts to generate the whole experience from a model could look almost right while falling short of a master teacher's sequence. Students encounter the weakness as confusing teaching and frustration. A product can be undermined by the task its model performs poorly, even if other parts are impressive.
Her framing is that AI can be a strong solution for the middle of a workflow. The team still chooses where human planning and verification are needed. In this account, the craft of the lesson is a constraint on implementation, not something automatically supplied by fluent generated content.
At 18:30, the original demo illustrates a human-designed lesson implemented with AI.
The model needs an interface as carefully designed as the learner's
Source 22:32The visible interface should be clear and enjoyable, but the agent also needs tools suited to the product. Khim describes a library of composable primitives, each with an interface a model can use to create or annotate a problem, highlight something or ask a subquestion.
An agent can combine a designer's sketch with an engineer's graphing module and the learning team's guidelines. It does not have to invent an entire interactive activity and every tutoring intervention from scratch.
This structure supplies grounding and limits the space the model must search. Khim connects it with latency, accuracy and reliability: the model reasons about the student's input and the next action while operating within a prepared set of components and schemas.
Internal work also needs a usable agent interface
Source 26:19The same idea changes the team's own work. People responsible for a domain make its expertise, tools and documentation accessible to colleagues using agents. A new employee should be able to benefit from what the organization already knows.
Khim describes a shift from making one thing to improving the system that makes things. She wants someone joining the team to find good information and make useful changes without repeatedly reconstructing all that context.
The onboarding philosophy remains learning by doing. Technical hires and learning designers should get the toolchain working and follow a change through to a user-facing result. She presents day-one impact as an aspiration the team works toward, not a guarantee that every new hire achieves it perfectly.
Quality checks include human review and real students
Source 29:46The primitives have evaluations across several dimensions, including correctness, physical behavior, visual overlap and tap-target size. Khim also says multiple human reviewers judge what the team releases, drawing on learning designers' craft and taste.
Synthetic students explore difficult or unsuccessful paths so the team can inspect whether the tutor's responses make sense. They help catch serious problems before gradual rollout.
She is clear that this does not replace real student data. Watching actual sessions and seeing many learners struggle with a concept reveals far more about how to teach it. Synthetic paths are an additional check, not a demonstrated substitute for observing learning.
Measure what the student can still do
Source 32:32Asked how the team balances difficulty against completion, Khim says completion rate is not the outcome they optimize for. They work backward from what a learner should be able to do at the end of a chapter and test that ability.
If a student has already learned it, completing fewer small steps can be better. If the student is struggling, returning to an earlier prerequisite may be more useful than finishing the current course.
The question extends beyond today: can the learner do the thing a week or a month later? Time spent and quantity completed can increase without answering that question. Assessment needs to distinguish apparent progress from capability that remains.
Solved homework is not evidence of learning
Source 34:54Khim's view of product durability follows from that outcome. The useful data includes real misconceptions, mistakes, attention and which interventions help actual students learn.
An answer-machine interaction may reveal what was on the homework, while saying little about whether the student understands it afterward. That difference matters to the feedback the team wants for improving a tutor.
The conversation also raises incentives. If the visible goal is merely to submit a correct sheet for a grade, students have a reason to take the shortest route. Khim argues for aligning practice and assessment with the ability the task was meant to build. That is her educational argument, not a universal description of how every school grades homework.
Help a child experience hard work as worthwhile
Source 39:21In discussing her own young children, Khim focuses on attention, concentration and space to become absorbed in an interest. The activity need not be mathematics; it could be building, chess, monkey bars or magic tricks.
Peter Yang raises the moment when a child enjoys making a game but becomes discouraged by bugs. Khim compares a well-paced learning experience to a difficult jump in a game: hard enough to require repeated attempts, but still inviting another try.
The aim she describes is willingness to tackle unfamiliar problems and discover that effort can be satisfying. At the time of the interview, she says Brilliant is built for middle- and high-school learners, even though the hosts discuss younger children as parents.
Adapted from Peter Yang's interview with Sue Khim, “AI Makes Cheating Easy. Here’s How It Can Make Kids Smarter Instead.” Product claims and teaching choices are attributed to the conversation. Sponsor material and unverified claims about national test trends are omitted.