Conversation

Fun paper that shows that gzip-based kNN sentence classification can outperform BERT.

https://aclanthology.org/2023.findings-acl.426/

Understanding that every predictor is a compressor and vice versa was the most important insight I learnt since 2018.

Also: Concat/compress with a SotA compressor is a great way to measure similarity between two pieces of data - an old insight, as Steph Wehner did gzip based malware clustering in the early 2000s...