Exponential Learning with AI
How I used AI to train my son to outperform his peers in selective school exams.
My 14 year old son sat his selective school exams this weekend. These exams are for admittance to the elite public schools in our part of Australia. They are very difficult exams set well above the level of average intelligence, used to filter out top performers from the entire pool of students within the state.
My son is a bright student and wanted to take part and do well. I have deep misgivings about our current educational systems, but that is the game that he is playing, and I want him to win. So I decided to train him.
This is a very brief story of how I used a totally new “recursive” learning methods utilizing AI (that we have developed at Polynize) to spark exponential learning and performance.
The Co-Learning Loop
The fundamental innovation of this learning method is to put humans and AI into a co-learning loop as shown below.
The loop begins with a challenge (in this story, the exam question). I get my son and the AI (ChatGPT) to both create answers for exactly the same challenge. I then put the answers side-by-side and analyze them according to a number of skill metrics. The analysis is immediately shared back to my son, and we discuss areas in which he failed to outperform the AI.
Once the analysis had been understood, I would ask my son to iterate by answering the same challenge again. Then I would run the loop again.
I ran the loop about 3 times for different exam questions in creative and persuasive writing, verbal reasoning and reading comprehension. I found very rapid improvement in each of those areas within a very short time.
Exponential Learning
For this experiment, I purchased a large number of practice exams from a private supplier which were indicative of the type of question that would be encountered in the live exam. These were high-level and difficult challenges within each of the exam area.
For quantitively measuring performance, I focussed on the sections which had close-form multiple choice answers (verbal reasoning and reading comprehension). We used about half the practice exams for learning, and then completed and graded the other half under test conditions.
Over about 2 weeks, my son completed about 15 full practice exams, with co-learning in between. I found that his performance within the loop increased very fast even within the same exam question (his answers after 3 iterations of a loop were much, much better) and overall test scores increased exponentially.
Early on, he was getting test scores of around 50%. By the end, he was consistently getting test scores of 85%-90%+ on every exam section [exams are set so that most students get about 50–60% correct / complete].
Conclusions
This is empirical work, I am not an academic or an educationalist — I am a researcher, technologist and entrepreneur building platforms that create real-world outcomes.
In this case, not just the data but my observations of this process show me that the interaction between the human mind and the machine “mind” is accelerating learning, making it faster and more autonomous.
We ran very fast iterations, perhaps 5x to 10x faster than linear, conventional learning. This meant I could fit learning into smaller time periods that could be done at convenient times.
The benchmarking against AI was motivating — there is an inherent sense of competition and incentive (and my son understands the deeper logic — if he cannot outperform AI, he will not have future success).
Finally, he loved being able to see his performance through fine-grained data that made his learning process transparent. Just as running / biking is more fun with Strava, this made learning more engaging too.
I’m now automating this learning process, generalizing it across all areas of skills and work and opening it up for access to the world through my work at Polynize.
