How to teach AI using chatGPT to a firefighter from Iceland? (ie domain experts/non developers) - with design and development prompts

Background

Yesterday, at an event at OpenAI in NYC, I was asked this question

How exactly did we teach AI to a firefighter?from Iceland? ie a domain expert / non developer

In my university of oxford course AI course, we were inspired by Magnus Smarason and Dr. Amit Chougule DPM,MD,MRCPsych

Their testimonials are here

Magnus (a firefighter from Iceland) and

Dr Chougule - an NHS doctor

We continue to work with them. In fact, I am pleased to say that we have now brought both Magnus and Dr Amit into our course to share their experiences.

let me share more here.

First, I wanted to set the context

• I believe that this is a huge problem (need to reskill) - but assisted by AI - also a huge opportunity.
• I chose the least common denominator ie pure prompting using chatGPT
• This is a moving goal post. We continue to collaborate and learn
• Despite the proven potential of education to provide upward mobility , sadly, in many cases, education does not tie to outcomes.?However, now with AI and tools like chatGPT, there is a unique opportunity to democratise the learning of AI - especially for domain experts(non developers).
• Being neurodiverse myself, I am also working with other neurodiverse students - who like me, find the potential of AI tools like chatGPT truly empowering
• Many thanks to OpenAI for helping us to share our views and best practices - especially to Bailey Richardson and Mamie Rheingold

Process

Note I have shared design and coding prompts for a simple example below.

What is AI? Broadly, we can see it as machine learning and deep learning algorithms .. but more simply these are a mechanism to make complex decisions using data.

The first step is to understand and quantify the business decisions using a business metric.

The next step is to translate (and visualise) the business metric to an AI metric. (machine learning or deep learning model evaluation metric) . I have explained this using a simple chat here

We then frame the requirements using the metric. The metric becomes the north star.

Next, we create a conceptual knowledge graph using prompts from this requirement

from this knowledge graph, we can use prompts to create three things

1) user stories

2) synthetic data

3) acceptance tests (ATDD, BDD and TDD )

This completes the prompts for design. from this, you use prompts to build the model. I have shared a simple example below. for code and build

Notes

1)? This approach is based on Coding - it is not low code as in drag and drop. We explain how code works so that people can understand what is generated

2_? We use prompts using chatGPT - hence no other tool?

3_ The innovation is in the pedagogy. Our focus is the domain expert and not the developer.

4. We are rethinking each step of the pipeline using LLMs and for domain experts.

5) Test driven development is slow but with the help of AI its accelerated

6) We use a conceptual knowledge graph using prompts and prompt against this knowledge graph. This gives powerful results. We have used this strategy for dynamic learning paths also??Using conceptual knowledge graph instead of a real knowledge graph with large language models

7) VIsualization?is important?

8) The process is iterative and reflective but the idea is we get a first version of the design and system very early - maybe in an hour and then iterate??

9) We find that synthetic data (generated by chatGPT) is very useful

10) We are also exploring the use of cursor and OpenAI canvas for code generation

Prompts -Overall steps using a simple example?

?"I want to predict stock prices. what is the machine learning model I should use"

"I want to model this as a regression problem. how would I go about it"

"Which model evaluation metrics could I use for this problem as a regression"

"Now, I want to model stock prices using a regression model and using the Coefficient of Determination (R2 Score) model evaluation metric. Create a schema for a knowledge graph for such a problem"

"Excellent. From this knowledge graph, create user stories. Ensure that they are always tied to the Coefficient of Determination (R2 Score)"

"Excellent. From this knowledge graph, create acceptance tests (TDD, BDD, and ATDD). Ensure that they are always tied to the Coefficient of Determination (R2 Score)"

"Great, now create some synthetic data for the basic regression problem for stock price prediction which I can use in this example to generate code"

"please write code fand show code or a regression model using this synthetic data and the target variable "Closing price "using the "Coefficient of Determination" (R2 Score) model evaluation metric"

Acknowledgements

Many thanks to Anjali Jain Ay?e Mutlu

Image by Magnus Smarason - a beautiful day in Iceland - which has inspired me to visit in the spring next year!