Artificial Intelligence No 96: A three step generic strategy for GPT-3/LLM development

Artificial Intelligence No 96: A three step generic strategy for GPT-3/LLM development

Introduction

Welcome to Artificial Intelligence No 96:?

We are nearing 70K subscribers :)

In this edition, we discuss a three step generic strategy for LLM development for the use case of reading books with chatGPT.?

This is a special edition for me

Due to my unique circumstances, books have played a major role in my life. I have shared before - how to read a book a week in the age of facebook and twitter

It’s a good post for this month because february is #LowCodeFebruary and it aligns to our new course Low code data scientist: Low-code AI Apps including LLMs and ChatGPT (online). Finally, I also think its a way to think of chatGPT for education.?

Background

I was inspired by a post (reference below) which enables you to take a top-down approach for speed reading using chatGPT to rapidly absorb the main ideas of the book without having to spend hours ploughing through the entire text. First, we ask ChatGPT to list the book chapters. Next, we delve into the most interesting chapters asking for key points and summaries. Finally, we delve even more into specific points.?

I like this prompt engineering strategy (which I call drill down prompt engineering strategy)

Another prompt engineering strategy is from Saxifrage: prompt engineering - which could be seen as an? example driven

1. Examples: Provide examples of how to answer your prompt

2. Direction: Give guidance on what kind of answer you want?

3. Params: Change what answers you get by adjusting settings

4. Format: Make it clear how you want to receive the answer

5. Chaining: Link multiple AI calls together to complete the task

A generic three step strategy for LLM development

So, a three step generic strategy for LLM development could be?

1) Start with domain knowledge

2) Build it up using prompt engineering (here there could be multiple strategies as we see above)?

3) Then engage with data engineering and other strategies such as LLM model chains (ex using strategies like langchain)?

I also like this quote which helps you overcome a lot of media misconceptions about LLMs

“The purpose that it is serving is not to inform you about things you don’t know. It’s really a tool for you to be able to do what you do better,” said Margaret Mitchell source WSJ

So, the three step process starts with what you know (which is the correct way to use LLMs and GPT-3)

Impact for educators: Applying this idea to reading books

How do you apply this idea to reading books?

Considering the three step process

It starts with domain knowledge and then building up a prompt strategy based on domain knowledge.?

To acquire knowledge on reading books (the problem we are addressing), I very much recommend Mortimer Adler’s how to read a book First published in 1940, it is a classic which emphasises reading for comprehension as opposed to reading to collect facts.?

The book covers themes like

  • Understanding the overall structure: table of contents, index, and key paragraphs of major chapters.
  • Keywords, Most Important Sentences
  • Identifying Propositions in the book
  • What is the overall message or theme of the book?
  • How does the author's argument unfold?
  • What are the main principles and supporting evidence?
  • Is the author's argument valid?
  • What are the implications?
  • If you agree with the author’s argument, how will you act on it?
  • The Limits of the author’s perspective
  • Comparing with other authors / similar books

All of these could go on to design prompts

You can see other #LowCodeFebruary posts also

Especially also see these two links

OpenAI for educators by Christoffer Noring and Introduction to Azure OpenAI service

If you want to study this subject with us, please see our new course Low code data scientist: Low-code AI Apps including LLMs and ChatGPT (online).

References

Revolutionary method of book speed reading with chatgpt?

Saxifrage prompt engineering

Image source: Mortimer Adler’s how to read a book

Michael Zeldich

President at Artificial Labour Leasing, Inc

1 年

As I already explained, only one possibility exists to have an artificial teacher/lecturer, starting the development of artificial subjective systems. That kind of artificial system will build its own subjective experience and could muster the profession of teacher/lecturer as any other. Programmable systems of any kind cannot do that because they cannot be a person. Students building their own subjective experiences in that "interesting" way will have difficulties using it to solve creative tasks.

回复
Warren Powell

Professor Emeritus, Princeton University/ Co-Founder, Optimal Dynamics/ Executive-in-Residence Rutgers Business School

1 年

This is a classic example of how AI can be used to help people (not replace them!).

Sebastian Britz look fwd to discussing in Berlin also!

mohamed karim

Network Coordinator

1 年

Congratulation

回复
Dr Marsia Bealby

Innovative Leader ?? Management, Academic Leadership, Teaching in HE, e-Learning ?? PhD in Archaeology. 4 Master's degrees in: Archaeology, Museum Studies, Business, Public Relations. Ongoing MEd as a pathway to FHEA.

1 年

Thanks for sharing, Ajit.

要查看或添加评论,请登录

社区洞察

其他会员也浏览了