Not all Transformers are identical

“Transformer” is a family, not one exact blueprint. The layout changes with the job.

• Encoder — strong at understanding / representing an input (classic for classification-style work).
• Decoder — generates sequences autoregressively, one piece at a time.
• Encoder-decoder — understand an input, then generate another sequence (classic translation).
• Decoder-only — many modern chat LLMs: repeatedly predict the next token from available context.

As an application developer, the practical takeaway is simple: the assistants you call through chat APIs are usually decoder-only models specialized for next-token generation after instruction-style training.