Summary
Sue Khim, co-founder and CEO of Brilliant, sits down with Peter Yang to discuss how her company has spent over a decade thinking about AI and education — and why their approach deliberately limits what the LLM controls. The conversation opens with a frank assessment of the US education system: test scores are falling while grades are rising due to grade inflation, with some university STEM programs now sourcing remedial algebra materials for incoming students.
The core product insight Khim shares is Brilliant’s specific division of labor between humans and AI. Every pedagogical sequence — the logic of which concept follows which, and at what difficulty — is designed by master teachers. The LLM then implements those human-designed lesson structures in real time, adapting to individual learners without being trusted to design the learning experience itself. Khim is explicit that companies letting LLMs swallow their entire product end-to-end will disappoint users, because models are highly uneven across sub-tasks: excellent at implementation and code generation, poor at architectural judgment and designing engaging interactive experiences.
She frames this as a “middle-to-end” rather than “end-to-end” philosophy — constraining AI to the tasks where it demonstrably excels. The conversation also addresses AI cheating, the counterintuitive case for withholding explanations to force productive struggle, and why Brilliant avoids LLM-generated explanations despite their surface plausibility. A grounded reference for product teams thinking about where to draw the human-AI boundary in high-stakes applications.
📺 Source: Peter Yang · Published September 06, 2026
🏷️ Format: Interview
