We are well and truly in the honeymoon phase of our relationship with Generative AI, especially large language models (LLMs). The hallmarks are telling: feverish excitement, lots of chatter and (of course) an excess of unbridled possibility. Amidst all the noise, it’s still challenging to move from high-level conversation to user-level application.
Thankfully, most of us don’t need to do so just yet. Until recently, I haven’t had any real use for ChaptGPT and other chatbot tools. Admittedly, that hasn’t deterred me from spending copious amounts of time being immersed in them – I’ve dabbled my fair share.
Recently, however, I needed to automate a manual spreadsheet task. More specifically, I needed to copy and paste data from different spreadsheets. While it’s true that I could do this manually, I found the repetition to be tedious and draining. That being so, I set out to build version 2.0 of the spreadsheet, with automation to do the heavy lifting.
It was the perfect opportunity to test the waters. Here’s how it went.
A Crash Course In Prompt-Engineering
Using ChatGPT as my tool of choice, I prompted it for a script to “copy and paste data from one sheet to another.” The first few iterations revealed just how vague this instruction was. Some basic questions emerged.
Firstly, the script needed to filter out empty cells before pasting it. This way, I wouldn’t have to delete blank cells whenever the script ran. Also, the content had to go in a specific column. I needed to figure out how to prompt ChatGPT for this in the automation.
These are just two examples of the details I needed to specify in my prompting. I resorted to making a list of steps for the automation to follow. Steps that come naturally to me and are easy to overlook. I eventually got the intended outcome, albeit after an embarrassing number of iterations. I even added other automation features to the spreadsheet. That said, the arduous process of using ChatGPT was quite sobering. It forced me to reflect on the downside of using such a tool.
Not Exactly Flawless
There’s a hard limit to what we can expect from large language models like ChatGPT. While they are remarkably capable, they often fall short in equally remarkable ways.
In her TED talk titled Why AI is incredibly smart and shockingly stupid, computer scientist Yejin Choi demonstrates how models like ChatGPT and similar LLM tools that have pulled off incredible technical feats, could fail spectacularly in the realm of common sense.
She further explains that language captures only a glimpse of how the world works. Since LLMs are trained on textual and language data, this insight is crucial to understanding why the models have such sizable blind spots. Many unspoken rules of life are not part of our language – such as human norms and values. Hence, training models on even larger data sets will not necessarily result in the human-like reasoning we know and appreciate as common sense.
This concept explains why the situational intelligence needed to translate my basic prompts into detailed instructions was lacking in ChatGPT. I had to refine my prompts, adding more detail with each iteration, to get the desired results.
That said, we still have an incredible opportunity in LLMs.
Unpacking The Value
Beyond the experiment I shared, there are other ways to extract value from ChatGPT and similar tools. In a recent conversation with an academic, for example, I learned that LLMs are hugely effective at summarising lengthy journal articles, saving researchers countless hours.
When working with LLMs, we should provide effective prompts with sufficient instruction and context for the best results. Compare this to Google searches, where a single phrase (or sentence) is enough. LLMs need a bit more input. The adage applies: garbage in, garbage out. If you want ChatGPT to write an email, for example, you’ll need to tell it what to write and your preferred writing tone at a minimum.
What if you’re using the tool to size up a problem or topic? In these situations, the best prompts are clear but open-ended. You want to ask wh-type questions: what, when, who, where and why? I asked ChatGPT to “explain why income inequality is a difficult problem.” The 473-word response featured a brief introduction, 11 well-explained points, and a helpful outro – a great starting point for anyone researching the subject.
In hindsight, I could have applied this idea to my spreadsheet challenge by starting with a better prompt like:
List the steps needed to copy data from one spreadsheet into another.
The results would have saved me quite some time.
AI And You
As exciting as the times may be, they’re just as chaotic. There are uncertainties about how the cards may fall. It is OK to feel anxious about job security and your place in this brave new world of LLMs and Generative AI more broadly.
Regardless of how these concerns may land, it helps to appreciate that the real value lies in what we do next. Channelling that anxiety into curiosity and initiative will help you position yourself to leverage these technologies in profound ways.
I’m on a journey to learn as much as I can about Gen AI and tech more broadly. As I do so, I hope to see the opportunity it presents more clearly and share it with those around me.
Your relationship with technology doesn’t need to look like mine. What’s important, however, is to recognise the tidal wave of tech in this moment, and get yourself on the right side of it.
