At Rise of AI 2019 in Berlin, KP Thai, senior learning and data scientist at Squirrel AI Learning, argues that AI can redefine education rather than just make it more efficient. Using the SAMR model, he distinguishes substitution and augmentation, which merely enhance existing teaching, from modification and redefinition, where AI changes the paradigm itself by moving students from passive to active learning and shifting teachers from the 'sage on the stage' to a coaching role supported by an AI 'assistant coach'. He describes Squirrel AI's after-school, business-to-consumer model in China, built on granular knowledge components, a probabilistic model of what each student knows, and a virtual personal assistant that uses natural language dialogue to diagnose misconceptions behind wrong answers. He reports that external randomized controlled trials across more than 18 studies show students using Squirrel AI outperform peers in traditional instruction, including some taught at favorable teacher-to-student ratios, learning the same content faster. Founded in 2014, the company operates over 1,700 learning centers in 200 Chinese cities and is expanding internationally through research partnerships.
Key takeaways
- KP Thai uses the SAMR model to argue AI belongs at the 'redefinition' level, changing the education paradigm rather than just making existing teaching more efficient.
- Squirrel AI reframes teachers as 'head coaches' supported by AI 'assistant coaches' rather than replacing them.
- The system diagnoses individual knowledge gaps using granular 'knowledge components' mapped as a prerequisite network, since mastery time varies by student and by topic, not by a fixed learner speed.
- A virtual personal assistant uses natural-language dialogue after wrong answers to diagnose the underlying misconception, not just that the answer was wrong.
- Across more than 18 external randomized controlled trials, students using Squirrel AI outperformed peers in traditional instruction, including some taught at favorable teacher-to-student ratios.
Detailed notes
From substitution to redefinition: the SAMR model
Thai uses the SAMR model to frame four levels at which technology can affect education. At the bottom, substitution merely replaces a teacher's activity, such as MOOCs like Coursera or edX offering the same lecture via video. Augmentation adds efficiency, such as automatic grading giving instant feedback. Modification changes the task itself, as in the flipped classroom, where students watch lectures as homework and do hands-on problem-solving in class with the teacher's guidance. Only the top level, redefinition, is where he places AI: it does not just change tasks but changes the paradigm, moving education from passive learning, where all students receive the same content at the same pace, to active learning, where students generate and search for answers themselves.
From sage on the stage to guide on the side
Thai argues group-based instruction wastes time because every student learns differently, at different speeds, with different prior knowledge and goals. He proposes reframing the teacher's role using a sports-team analogy: the human teacher becomes a head coach who strategizes what students need to learn, while AI-based tutors act as assistant coaches providing one-on-one attention. Freed from delivering content, teachers can focus on social-emotional support and collaborative group work rather than memorization drills. He stresses this does not remove teachers but changes their function, combining a human head coach with an AI assistant, plus continuous, contextual, actionable analytics rather than data collected only via a final exam.
What implementing AI in education requires
Thai lists four requirements: AI expertise, from elite researchers to engineers who can turn models into products; mature AI models and algorithms able to solve real-world problems; sufficient quantity, quality and ease of experimentation with data; and adequate computing infrastructure. He says a business-to-business model selling directly to schools does not work well, so Squirrel AI sells business-to-consumer through after-school learning centers in China. He warns against a one-size-fits-all algorithm: a model tuned for math does not transfer to reading, so implementation must be built from the specific content students need to learn, using comprehensive, contextual, historical data about each learner rather than single-course, one-dimensional data.
How the adaptive system works
Squirrel AI breaks subject content into very granular 'knowledge components' mapped as a prerequisite network, so the system can diagnose exactly what a given student does and does not know, for example a student needing to master only another 6 percent of a topic. Time required to master a component varies by student and by component, not by a fixed notion of a 'fast' or 'slow' learner, which Thai says justifies personalized content sequencing. A virtual personal assistant uses natural language processing to start a dialogue when a student answers incorrectly, asking them to verbalize their reasoning so the system can identify underlying misconceptions rather than only knowing an answer was wrong. The company also experiments with facial-expression capture, transdermal imaging and EEG headsets to gauge stress, attention and focus, though Thai says usefulness of these signals is still being researched.
Q&A highlights
Asked how AI can adapt to a student quickly enough, Thai says the system does not retrain in real time but iterates on a roughly weekly cycle, collecting performance data on how students interact with content and feeding it back into model updates. A neurotechnology company representative working on stroke rehabilitation notes parallels with their own EEG-based work and suggests possible collaboration. Asked how the system avoids giving students a one-sided experience that omits things like personal anecdotes a human teacher could offer, Thai says content still includes videos and varied question types, and that human teachers remain involved to supplement, sometimes replace, and inform the AI.
Quotes
Group based instruction is a waste of time, so instead what we should be doing is use AI based technologies to continuously adapt one on one.
KP Thai
Questions answered in this talk
How does Squirrel AI personalize learning for each student?
According to KP Thai at Rise of AI 2019, Squirrel AI breaks content into granular knowledge components mapped as a prerequisite network, diagnoses exactly what each student knows or does not know, and recommends personalized content and sequencing rather than group-based instruction.
Does AI-adaptive learning replace teachers?
According to KP Thai at Rise of AI 2019, no: AI acts as an 'assistant coach' handling one-on-one content delivery, while the human teacher becomes a 'head coach' focused on social-emotional support and collaborative work.
Is there evidence that AI-adaptive learning improves outcomes?
According to KP Thai at Rise of AI 2019, over 18 external randomized controlled trials found students using Squirrel AI's adaptive learning outperformed peers receiving traditional instruction, learning the same content faster and with better engagement.