Teaching Machines To Teach Like Humans
by Marcus Birkenkrahe
I’ve used DataCamp courses for years in all of my data and many of my computer science courses at Lyon College. Coupling classroom instruction with online exercises worked really well: It’s like adding a bunch of highly motivated, passionate, knowledgeable industry-expert teachers as guest speakers and educators to my courses.
However, there was a snag: While the students loved DataCamp courses because of the interactive exercises and the flexible learning, I noticed that the skills often did not “stick” as deeply as I wanted. I attributed this to the fact that coding exercises are usually cloze (gap-filling) exercises: partial code snippets are given that the learner must complete.
This makes it easy to check correct answers (for DataCamp), but it trains some students to become cloze rather than coding experts: They get really good at filling in the blanks without always understanding what exactly they’re doing or why. Which is why, in class, I supplement DataCamp instruction with drill exercises where no code is given, just the problem. For many learners, a blank canvas is important to push them to demonstrate (and develop) true understanding. This is not fun-free, but it’s a different type of fun—deferred to solving the problem rather than coming up with the correct answer quickly.
In the bright, generative AI-lit future of education, this limitation is now being directly addressed. With the acquisition of Optima and the launch of DataCamp’s AI-native learning engine, exercises, examples, datasets, and even entire project briefs can be generated in real time and tailored to the learner’s current level, goals, and context. The platform can now move fluidly from guided, scaffolded practice to fully open-ended, blank-canvas challenges—and still provide immediate, precise feedback on completely free-form code.
This is a meaningful step forward. After years of supplementing DataCamp myself with open-ended drills, it’s encouraging to see the platform beginning to handle that progression natively, at scale, for every learner.
That said, I do have two lingering concerns:
-
Hallucinations and factual accuracy – Even with the NotebookLM-style strong RAG that DataCamp’s Chief AI Office Yusuf Saber mentioned in the November 17 webinar with CEO Jonathan Cornelissen - limiting external input and relying heavily on validated, high-quality retrieval data - some hallucination risk may remain in edge cases. I’m curious how the team plans to monitor and iteratively reduce hallucinations, especially in technical subjects where small inaccuracies can derail learning.
-
Complex production environments – I wonder how well the AI tutor will cope when courses move into real-world data-engineering platforms (multi-service setups involving Airflow orchestration, Spark clusters, cloud permissions, dbt in production, etc.). Reproducing realistic, stateful, multi-tool environments inside a browser-based sandbox is already challenging today; adding generative variability on top of that feels ambitious. I’d be interested in the roadmap for handling those more infrastructure-heavy topics without losing the hands-on depth that sets DataCamp apart.
Overall, the Optima integration feels like the right direction—turning a great interactive platform into something closer to a true personal tutor. I’m looking forward to learning more about how DataCamp plans to design content that fully leverages these new capabilities while keeping the risks in check.