Rens ter Weijde, co-founder and CEO of KIMO – Interview series
KIMO is a Dutch start-up founded by two Harvard alumni: Krishna Deepak Nallamilli (India) and Rens ter Weijde (Netherlands). The team is focused on building the artificial intelligence needed to generate individual learning paths through digital learning content.
As e-learning has taken off, it has dropout rates of up to 95%. Why is the success rate so low?
When we launched KIMO, we surveyed a few hundred users to better understand the situation. First of all, most online providers offer MOOCs (online courses), but users perceive MOOCs as a significant time commitment. They often used a lot of “shorter” workarounds like reading articles, listening to podcasts, asking questions on Google, etc., which better fit their daily schedule. Thus, learning is more multi-channel in practice than these providers allow. Additionally, many users reported that they had not been guided through their online journey. The result is that they spend a lot of time researching, trying to decide what to study, etc. A third reason concerns the relevance of the actual content. Online documents are often static, pre-recorded and not entirely relevant to them. You could argue that the content isn’t personal / relevant enough for them – or at least not relevant enough to justify the time spent.
Many users claim to be bored and frequently cite lack of engagement as a problem. Why do you think users feel denied their rights to e-learning?
I think it is possible to improve learning platforms on at least two main dimensions. First, better intelligence is needed to guide users on better journeys and provide better content recommendations. You could say that this is the R&D required for the education sector, strongly linked to AI algorithms. The second element is the other side of the value chain: UI / UX and the end user experience. Most LMS systems are considered boring and obsolete by users. They are a far cry from the fancy, real-time, social, and personalized software that users expect today.
Could you share the story of the genesis of KIMO and what attracted you to solving the e-learning problem?
Sure! KIMO started when Krishna, my co-founder, and I met at Harvard Business School. We loved the environment, but at the same time realized that the good parts of the experience were not adaptable to people around the world. We decided on a whiskey in the lobby of a hotel that we could try to create a “digital career coach”. This car was the first version of KIMO.
Can you explain how AI is needed to generate individual learning paths through digital learning content?
In fact, KIMO relies on a multitude of AI models in the pipeline. Some models are inherently simple, others are more complex. The common thread is that most models rely on natural language as input data (NLP, e.g. transformer models). These templates lead to the content recommendations you receive, the grouping of content into specific topics, or the recognition of essential skills required for jobs. We also have more experimental “generative” AI models, like the model that answers questions about content inside the KIMO app. If this works well enough, this is one more step towards teacher automation that we envision.
Can you explain how an AI system can learn to understand jobs in detail (e.g. technical skills or soft skills)?
Simply put: we decided to ignore the existing databases (O * Net, ESCO) for this job because they weren’t granular enough and out of date. Instead, we built a system capable of recognizing approximately 40,000 skills in market jobs in near real time. You could say that our system “reads” all of these job profiles in order to predict the type of skills currently required for the jobs. These recognized skills are then grouped into general and technical skills.
Can you explain how personalized learning works on the platform such as how the system will know what type of content works best for each user like articles, videos, podcasts, articles etc.
The simple answer is that we match users and content through vector matching, which is common practice in recommendation models. The hardest part is deciding how these vectors are constructed, in other words, which elements are weighted. For now, the system is relatively straightforward and works with user learning preferences and popularity scores for online documents. The future will be more interesting, as we try to weigh in the user’s current state (eg his work) and the desired end state.
What are some of the machine learning methodologies currently used in the KIMO system?
We use many different models, depending on the task. But I can say that we have a deep love for the attention-utilizing NLP models, hence the transformative models.
Where do you see the future of online education in 5 years?
In short, I see online education moving from “boring, lonely and unique” to “very engaging, social and personalized”. Online training companies need to escape the idea that they exist in a traditional, slow-moving industry. Instead, they should realize that they are competing in the age of information curation, at the heart of many important trends today.
Is there anything else you would like to share about KIMO?
Yes. KIMO is still a baby, or “beta” as we call it. Download the app, give it a try and send us your feedback!
Thanks for a great interview, readers who want to know more should visit KIMO.