The Culture Funnel: You can’t align what isn’t in the data

An analysis of modern large language model (LLM) training pipelines reveals that cultural diversity is frequently lost during the process. Data from Cohere Labs shows this "culture funnel" effect occurs when post-training data mixes filter out diverse cultural information. This loss of representation creates a significant challenge for model alignment. A model cannot be aligned with cultural values and nuances that are not present in its final training data.