With everyone is focussing on LLMs, I'd like to remind people that "classical" methods like xgboost are as relevant as ever even for text based use cases (e.g. sentiment classification, document classification) because they use a fraction of the resources that an LLM does.