BERT vs GPT vs T5: Architecture Choice

Question: Why would you choose BERT (Encoder), GPT (Decoder), or T5 (Encoder-Decoder) for a given project? Key points: BERT is encoder-only, trained with masked-language-modeling for bidirectional context - well suited to understanding/classification tasks (e.g. sentiment, NER, embeddings). GPT is decoder-only, trained autoregressively - well suited to open-ended generation and creative writing. T5 is encoder-decoder, framing every task as text-to-text - flexible for translation, summarization, and other sequence-to-sequence tasks. Architecture choice should follow from whether the task is primarily about understanding input, generating output, or transforming one sequence into another.