How LLMs work, part 4: the transformer, block by block

The transformer is the architecture behind ChatGPT, Claude, Gemini, Llama and nearly every other language model. This is part 4 of our series on how LLMs work. You will walk through the 2017 design block by block (encoder, decoder, embeddings, positions, residual connections, normalization and masks) and then see the three variants that matter today: decoder-only models, mixture of experts and multimodal models.