Thank you for this contribution! I was familiar with the idea of ML models capturing a compressed snapshot of the data, but that work on exploring its limits in DTs looks very interesting.