LinkedIn·Friday, 21 August 2026·5d ago

Explaining how the AI tutor actually works with Yusuf Saber in interview by Kim Isenberg and @peterthum from Superintelligence. Link to…

Jonathan Cornelissen

Co-founder & CEO at DataCamp

Explaining how the AI tutor actually works with Yusuf Saber in interview by Kim Isenberg and @peterthum from Superintelligence. Link to full interview in the comments. There were 10M+ hours of learning on DataCamp last year, and our goal is to shift all of it to the AI Tutor. This creates massive opportunities and challenges. Top 10 things we cover: 1. The importance of open weight models in an educational context: open vs frontier, not in theory but in the context of helping millions of learners. 2. How we're thinking about (tens of) millions of dollars in AI tutor token costs, and why cost has become a key engineering problem. 3. Why a 27B open model is beating frontier models on our teaching evals, and why we still can't put it in production. 4. The one bottleneck stopping us from switching to open weights tomorrow. 5. What it takes to teach an AI to teach: 100+ tutor behaviors if you used them all at once. 6. How the tutor decides, at every learner response, which teaching behaviors matter right now, and which to leave out. 7. How we can swap the underlying model for millions of learners in minutes, and why enterprise clients with approved-model lists make that a core feature, not a nice-to-have. 8. Evals: 60% of our quality bar was sourced by a system (we call it Sentinel) that watches for learner friction. 9. What 300,000+ learners on the AI Tutor taught us: higher conversion, higher satisfaction, and an engagement distribution that shifted in a way we didn't expect. 10. Why we're not building an AI tutor for just data/AI skills. We're building the best AI teacher, full stop, and data/AI skills are just where the gap is biggest today. We have several principal and staff AI engineering positions open. Get in touch if you're passionate about building the #1 AI tutor.