“Research into language model internals reveals that larger models are more likely to use shared, abstract neural representations for concepts across different languages, whereas smaller models tend to use separate, language-specific neurons.”