Creating a synthetic training dojo for edtech
I've been teaching 100,000 fake students for 2 weeks.
— Austin Way (@AustinA_Way) March 14, 2026
and used them to build the best AP prep system in the world.
I took Qwen 3 8B models and gave them simulated human memory.
Now every night thousands of simulated students start with zero knowledge of the social sciences.… pic.twitter.com/7GHIQ7n6l3
An interesting use case for AI agents to improve curriculum and content: using a bunch of simulated “dumb” student agents who cannot rely on any latent knowledge about the subject (e.g., social sciences), and instead only use information provided by the teacher model before taking a real test. The teacher then evaluates the results, identifies areas of weakness, re-strategizes the content, and iterates.
Average scores improved from 3 (45th percentile on the test) to 4.43 (80th percentile) in 2 weeks (don’t have the actual numbers to validate this, but feels like a very large effect size - 2 sigma??) . Combined with large population models, this approach could open up an interesting way to test products and services, and improve curriculum or training before releasing it into the wild - a kind of training dojo for edtech products.
Outside of education, I’m imagining running marketing campaigns on population-scale personas to identify ideal PMF or ICPs or testing how people would react to policies. Very exciting!