The Superintelligence Podcast

Open-source vs. frontier AI: a real-world test

Jonathan Cornelissen & Yusuf Saber

Jonathan Cornelissen & Yusuf Saber

CEO and Chief AI Officer, DataCamp

DataCamp

“The bottleneck really is the infrastructure layer. It’s not the model quality.”

In this conversation

How do you choose an AI model for a product people rely on? DataCamp co-founder and CEO Jonathan Cornelissen and Chief AI Officer Yusuf Saber join Kim Isenberg and Peter Thum on The Superintelligence Podcast to discuss how DataCamp builds and runs its AI Tutor. Saber, who built the system at Optima before DataCamp acquired it, explains how the tutor breaks good teaching into individual behaviors and uses “system reminders” to keep the model on track where it tends to slip. He describes how a friction detector called Sentinel has surfaced most of the tutor’s test expectations, and why the tutor should use a model’s intelligence but not its knowledge, and say “I don’t know” when it lacks a reliable answer. Cornelissen explains why cost has become the main pressure and why clients with approved model lists, such as banks and governments, make easy model switching essential. The conversation also covers where latency really comes from in production.

Key takeaways

  • All tutoring with real learners still runs on frontier models. Open-weight models are so far only tested in a simulated environment that replays real lessons, Saber says.

  • On DataCamp’s quality evals, the open-weight Gemma 31B model beats the frontier models it uses today. What holds back a switch is hosting infrastructure with the needed speed, reliability and caching.

  • DataCamp had more than 10 million hours of learning last year and expects around $100 million in ARR this year. With an hour of AI tutoring costing several dollars, cost has become the priority, Cornelissen says.

  • The tutor draws on more than 100 teaching behaviors. Loading all of them at once would take over 120,000 tokens, so the system selects the relevant ones at each moment.

  • More than 300,000 learners have already spent significant time with the AI tutor, and DataCamp expects that to reach millions in the next six to twelve months.

The daily AI briefing people actually read.

What shipped, who’s winning, and what it means. Six days a week, in under ten minutes.

The newsletter

The people building the AI future. Read by 160,000+ subscribers.

Subscribe free