Week 5 - (part 2) Prompting

Question 1

What is GPT’s problem?

Answer 1

1) Bottleneck: token limit is at 8,192
2) There’s a hard limit on the input and output length
3) Training cut off date

Question 2

How far can you push automation?

Answer 2

No code software - helpful because it helps lay people visualise and develop software, allows them to take advantage of the databases in the system

Question 3

Getting over the hurdle of automation

Answer 3

Decomposing the problem and providing good prompts

Question 4

What is a system prompt?

Answer 4

To establish a standardised format that optimises the framing for ChatGPT and thereby enhances its performance across a broad spectrum of inputs. It’s a generic prompt

Question 5

What is zero-shot prompting

Answer 5

Doesn’t require any examples and is often used to demonstrate the power and capability of the LLM. Also used when there are not enough tokens to give examples.

Question 6

What is prompting?

Answer 6

The “prompt” is the user generated input to the model and is a way for you to guide the model

Question 7

What is few-shot prompting?

Answer 7

Providing multiple demonstrations in the prompt to steer the model to perform better. It provides the input context.

Question 8

What is chain of thought prompting?

Answer 8

Guides ChatGPT to explain its answers in a step-by-step format of its reasoning process.

Question 9

What is prompt compression

Answer 9

Sending a text prompt to GPT to compress it and used the compressed prompt instruction. Then uncompress it after

Question 10

What are structured outputs?

Answer 10

Come up with a prompt given a piece of text for it to extract the key entities for it. Useful for customer service type chat bots.