The AI Tutor That Makes Kids Actually Think | E2298

What It Covers

Sue Khim, CEO of Brilliant.org, presents the platform's new AI tutor named Koji, built on seven years of interactive lesson infrastructure. The episode covers Brilliant's evolution from student loan comparison tool to STEM learning platform, the Socratic tutoring methodology, consumer-vs-B2B pricing decisions, and why AI that makes users think is resonating with parents.

Key Questions Answered

• •Socratic AI tutoring architecture: Brilliant's Koji tutor constrains LLM roles to what they do well while keeping mathematical correctness inside deterministic lesson infrastructure built since 2019. Every interactive canvas has an API that the model reads and writes to, enabling graphical tutoring with full screen visibility. This separation of concerns — LLM for conversation, deterministic systems for accuracy — is the core technical decision founders building AI education tools should replicate.
• •Pricing benchmark shift: Brilliant prices against tutors at $30/month, not against casual apps at $60–100/year. The average in-person tutor costs $80/hour in mid-tier cities, scaling to $150–300/session in major metros, totaling $10,000/year for high-dosage tutoring. Brilliant targets a 95% cost reduction versus that benchmark. Founders should identify the expensive analog their product replaces, not the cheapest digital competitor, when setting price anchors.
• •Consumer over B2B for feedback velocity: Selling direct to learners rather than through school districts keeps Brilliant one step from the actual user. The team reads every App Store review with multiple humans and receives detailed parent emails including photos and videos of product issues. This feedback volume directly shapes the product roadmap. B2B education sales insulate companies from real user reactions, allowing poor products to persist because students are forced to use them regardless.
• •Verifiable reward signals for tutoring AI: Frontier models like GPT-o1 have plateaued in tutoring effectiveness because they lack verifiable reward signals for reinforcement learning specific to pedagogy. Brilliant's tutoring sessions generate real learning loops where the reward signal is measurable student comprehension. Founders building AI in high-stakes domains should identify domain-specific verifiable rewards rather than relying on general model improvements, which plateau on specialized tasks.
• •Scaffolding removal as mastery signal: Koji's lesson design progressively removes visual scaffolding — tile labels, area model diagrams — as students demonstrate understanding, ending each lesson in a test-like environment with no tutor present. This graduated independence model mirrors how effective human tutors operate. Product designers building learning or coaching tools should build explicit "exit ramp" mechanics where AI assistance decreases as user competence increases, rather than maintaining constant support.
• •Brilliant's pivot origin — first principles funding conversation: Brilliant emerged when investor Chamath Palihapitiya challenged the All Tuition team to abandon student loan rate comparison and instead ask what would make the biggest possible difference. He argued that helping people find cheaper rates preserves the broken system rather than eliminating it. Founders should pressure-test whether their product optimizes around a problem or actually removes it, particularly when incentive structures reward keeping the problem complicated.