What is Attention in LLMs? Why are large language models so powerful

3-Second Blueprint (English)

1. Static word embeddings assign one fixed vector per token, leaving homonyms like 'Apple' ambiguous between fruit and brand.
2. Self-attention dynamically shifts a token's vector representation based on the other words appearing in the same sentence.
3. Contextual tokens like 'orange' pull 'apple' toward the fruit subspace, enabling LLMs to disambiguate meaning in context.