Trials and tribulations of a practical application of AI

This is an informal AI story as much about my own naivety as it is learning to navigate AI.

I will start by saying I am an AI optimist. I can see how AI and machine learning is already shaping the way we do things. I can also envisage how much it will help in simple tasks in the future. Despite my optimism, here is an example of the limitations of AI and the pitfalls of blind user input.

Throughout my career I have every so often expounded the phrase “I wish I knew code”. This is usually right about the time I have database tables delivered and I am needing to combine the tables without affecting the values. Whilst geological software can do some of what I would like, it can often influence the interval lengths that then affect the validations of the data (RQD is an example). Every so often I consider teaching myself code.

Even before I get to the specifics of learning, I get lost in WHICH code to learn - there are so many options and so many people explaining the difference my head spins. I also only use code every couple of years so when I have learnt some code, I have quickly lost it (use it or lose it philosophy). Despite not knowing how to write code I am able to read basic structure and understand what it is doing. This allows me to make basic modifications once I have the right structure.

Recently I was muttering the same old line "I wish I knew code". Then I had a lightbulb moment - surely AI can help. Optimistically I typed my problem into ChatGPT. The question I asked was along the lines of “How can I use excel to…”. The solution was an excel solution and not really what I was after. Optimism dampened. I ended up doing the task manually as I have done in the past. In this case it wasn’t a huge dataset so not too time consuming, but the process creates large, complicated spreadsheets and is tedious.

Here lies the first teaching of AI. Ask the right questions. In this case AI delivered exactly what I asked but I asked the wrong question.

A few days later over my morning coffee I realised that once again I would have to combine 2 tables to complete my analysis. This time it wasn’t going to be so easy. The dataset was larger, and the combination was more complicated. So, I started thinking about AI again. “Surely it can write the code for me.”

I started planning the steps in my head using dummy cell references. I sat at my computer and wrote them out in Word. I then asked ChatGPT to write an excel macro for me the steps I had planned out. In 2 seconds, I had code and instructions on how to get it running in Excel.? I was so excited. How easy was that.

The code worked as I had asked it to, but I then realised I had missed a step. I edited the steps and plugged back into ChatGPT. Two more seconds later I had the edited code. I then realised I missed another step. With the addition of the third step, the code stopped working. It was stopping at certain lines. The data looked ok. Where was it going wrong? I told ChatGPT it wasn’t finishing the complete dataset and to have another go. The next round of code crashed excel. Another attempt… Back to the code not completing the dataset.

What ensued was a loop of code not working and me trying to work out where my instructions were going wrong to give better instructions. Quite a few hours later (some of which was waiting for the code to work only to realise it had crashed) I threw my hands in the air frustrated. It had all seemed too simple, but I had wasted enough time. I committed to doing it manually. As I started to plug away manually, I realised that my excel formulas were not giving me the right answer. It then clicked. There WAS a problem with the data. I re-ran the code and then reviewed the messages I was getting. “Ahhh”. If I had taken the time to read the error message properly, I might have seen where the error was. The error message itself was erroneous but hidden in there was a hint as to where to look. I went back to AI and told them what the issue was. Within half an hour I had a working macro that did the task I wanted it to.

So here lies the second learning of AI. You can create something very quickly. However, creating something useful that provides the results you want requires knowledge and experience. You must know the basics of what you want before you start and be clear in communicating them.

In the time-poor age we live in, convenience lies at the heat of almost everything we do. AI certainly gives quick answers. Despite my experience, I remain an AI optimist. I got the code I wanted in the end which I can use on multiple projects. I am confident AI will not replace me in the near to medium term future. However, with these lessons behind me, it is already changing the way I do things.