Entering a prompt into a large language model (LLM) triggers a complex process where the text is converted into numerical tokens and processed layer by layer. LLMs utilize the transformer architecture, which allows them to analyze entire sequences of text in parallel, improving efficiency compared to older models. Each LLM consists of many transformer layers, each incorporating self-attention mechanisms and feed-forward neural networks. The number of parameters in a model, such as 7 billion in one example, reflects its capacity to learn from data and generate responses. Tokenization is crucial for transforming text into a format LLMs can understand. The common method, Byte Pair Encoding (BPE), breaks down words into smaller subword tokens, ensuring that both common and rare words can be efficiently represented. Each token is associated with an integer ID, which the model uses in its computations. The performance and cost of using these models can vary greatly depending on the number of tokens generated, especially for non-English texts that often require more tokens due to their linguistic structure. Once the text is tokenized, the model generates embeddings—continuous vector representations of each token. These embeddings are processed through multiple layers of self-attention and feed-forward networks. Self-attention calculates how much each token should focus on others in the sequence, using learned matrices to derive attention scores. The model operates in two phases: the prefill phase, where it computes for all input tokens simultaneously, and the decode phase, where it generates output tokens one at a time, relying on previously computed values to save on memory. This two-phase operation highlights the model's design for both efficiency and responsive text generation.