GPT History Quiz

Answers and explanations appear as you take the quiz.

1. What does GPT stand for in the guide?
   • Generative Predictive Transformer
   • General-Purpose Transformer
   • Generative Pretrained Transformer
   • Generative Prompt Tuning
2. What did the 2018 GPT work combine?
   • Training a task-specific supervised model from scratch using labeled examples
   • Using an encoder-decoder model without unsupervised language-model pretraining
   • Unsupervised language-model pretraining with supervised task adaptation
   • Prompting a frozen model with examples but skipping task-specific fine-tuning
3. What did GPT-2 research examine in 2019?
   • Whether each task still required its own supervised parameter updates
   • Whether a next-token model could attempt tasks described in text without separate task-specific training
   • Whether text prompts could replace language-model pretraining entirely
   • Whether every task required a separate model head and labeled dataset
4. What characterized GPT-3’s few-shot evaluation in 2020?
   • A new set of weights was trained for every task prompt
   • It was evaluated without examples or instructions in the prompt
   • It relied on a separate fine-tuning dataset for each evaluated task
   • Examples were placed in the prompt context without a task-specific gradient update for that evaluation
5. How does few-shot prompting differ from fine-tuning?
   • Few-shot prompting supplies examples in context; fine-tuning updates model parameters
   • Few-shot prompting updates weights, while fine-tuning keeps pretrained weights fixed
   • Both methods require new gradient updates for each individual task
   • Fine-tuning only retrieves documents without changing model parameters
6. What does the guide say about its historical timeline?
   • It lists every current GPT product and specification
   • It predicts future release dates
   • It covers research milestones rather than an exhaustive current product list
   • It is a live product availability page
7. Why distinguish a base model from a chatbot product?
   • A product is the original research model unchanged, including all of its development context
   • A product may add instructions, retrieval, tools, and an interface around a model
   • Adding retrieval or instructions cannot change product behavior
   • Research evaluation results transfer unchanged to every product built from the model
8. How should a historical benchmark result be interpreted?
   • As a timeless measure of every current product
   • As proof that the model solves every task
   • As a guarantee of current performance
   • As evidence from a particular dataset, prompt, and evaluation setting