I love this. As a former ML engineer, we saw this in forecasting streamflow. The best forecasting models were not the decades-old physical models, the purely 'statistical' models (AKA throw a bunch of data at an ML model), but what we called 'theory-guided ML', which stood on the shoulders of decades of research in hydrology and computing. https://www.upstream.tech/posts/not-all-ai-has-the-same-iq
Totally agree with these algorithms being closer “collective” or inferred intelligence. I was thinking about that when a recent Nature paper claimed that AI-generated poetry could be better rated on metrics of beauty and rhythm than poets like Emily Dickinson and Chaucer. The model was trained on their work to start with! And then was essentially just using more accessible language.
I love this. As a former ML engineer, we saw this in forecasting streamflow. The best forecasting models were not the decades-old physical models, the purely 'statistical' models (AKA throw a bunch of data at an ML model), but what we called 'theory-guided ML', which stood on the shoulders of decades of research in hydrology and computing. https://www.upstream.tech/posts/not-all-ai-has-the-same-iq
Totally agree with these algorithms being closer “collective” or inferred intelligence. I was thinking about that when a recent Nature paper claimed that AI-generated poetry could be better rated on metrics of beauty and rhythm than poets like Emily Dickinson and Chaucer. The model was trained on their work to start with! And then was essentially just using more accessible language.