Underwhelming to see researchers continuing to develop bespoke architectures for "interpretability" rather than adapting existing, widely-understood models. https://www.reddit.com/user/y3i12