A product manager asks why one pretrained foundation model can support summarization, drafting, and question answering instead of requiring a separate model trained from scratch for every task. Which explanation is most accurate?

1. Pre-training eliminates all need to evaluate the model on the company specific tasks.
2. The model automatically updates its weights whenever a user includes a document in a prompt.
3. Broad pre-training learns reusable representations that can be applied or adapted to multiple downstream tasks.
4. The model contains a verified database entry for every answer it will ever provide.