MASTERING CHATGPT
There can’t be an APC reader out there who hasn’t spent at least ten minutes noodling with ChatGPT – or one of its many derivatives, such as the Bing Copilot.
While there’s plenty of AI power and much to explore with the free version of ChatGPT, the real power is unlocked when you subscribe to ChatGPT Plus. Not only does that (largely) ensure you can get access to the chatbot during the regular periods of peak demand where the free version is switched off, it also opens up extra features such as plugins and advanced data analysis. Let’s look into those two features in greater detail.
Plugins
Plugins allow you to combine the intelligence of ChatGPT with third-party sites or services, such as Wikipedia, WolframAlpha or OpenTable. For example, you could ask the AI to create a summary of a long article you’ve found on Wikipedia or use it to research facts. And it goes further. Ask ChatGPT to “give me the release dates of the past five versions of Windows from Wikipedia” and it will even dig out the correct answers from the online encyclopedia, without you having to wade through different articles yourself.
With the Expedia plugin, on the other hand, you can ask the AI to do things such as create a three-day trip to Oslo for you, and it will come back with hotel and flight recommendations that suit your budget. It’s not perfect: it sometimes gets trip dates wrong when you click through to book flights, for instance, and it only deals with the US site, but it’s good for generating ideas and rough budgets. Irritatingly, you can only enable three plugins at a time and the Plugin Store is poorly presented and erratic. Often plugins don’t install at the first time of asking, for example. It’s a work in progress, but one that holds tremendous promise if it can be made to work more smoothly.
Advanced data analysis
This is arguably the best feature of ChatGPT Plus, although you may still need to dive into Settings | Beta features and switch it on first. Advanced data analysis lets you upload spreadsheets and other files and get the AI to do the hard graft that you might normally do in Excel.
For example, we fed ChatGPT Plus raw survey data from the PC Pro awards in an Excel workbook. Each tab represented a different awards category, and the AI automatically figured this out, but checked this with us to ensure it was making the correct assumption.
Once it had worked out the format, we could ask it to produce graphs showing each company’s performance in different categories or to tell us how company A fared against company B in different categories. We could also use it to calculate different weightings for each category, for example giving “very satisfied” responses a +2 weighting, but mere “satisfied” a +1, to see what effect that had on the overall results.
All of this could be done in Excel, but only if you have the skills with pivot tables or can apply the relevant formulae. Advanced data analysis lets you do all this with plain
English commands, and the results are much more reliable than they are with general chats. There are no hallucinations (at least none that we’ve seen), and it will tell you if it can’t perform a particular request or needs more guidance on what you’re looking for.
ChatGPT prompt engineering tips
Although AI services such as ChatGPT are growing ever more intelligent and are remarkably capable at determining intent, there are techniques you can apply to guarantee faster, more accurate results. This so-called prompt engineering is often sneered at in tech circles, but knowing how to get the best results from AI tools is a skill – one that can pay handsomely, too, if you check the job ads.
We’ve only got space to cover the raw basics here, but even these tips should help improve your hit rate with services such as ChatGPT.
There are different types of prompt, the most basic of which is known as zero-shot prompt. This is basically where you give the AI a task without any clues as to the type of output you’re seeking. For example:
Tell me the sentiment of the following social media post: the latest issue of APC was one of the best in years
When we entered that into ChatGPT, it correctly identified it as positive. Of course you don’t need an AI assistant to tell you that, but now imagine that you’re pasting 3,000 of these social media sentiments into a spreadsheet and want to tell how many are positive – that’s a job AI will perform far more
quickly than any human.To help improve the accuracy of results, you could give the AI clearer instructions of the output you’re looking for and provide examples. This is called a few-shot prompt. So, to elaborate on the example above, we might enter the following few-shot prompt:
Tell me the sentiment of the following social media post and categorise it as either positive, neutral or negative. Here are some examples: The latest issue of APC was one of the best in years.
Answer: positive Chris Szewczyk couldn’t write a bus ticket.
Answer: negative The Labs was good, but I didn’t get much from features.
Answer: neutral
Here we not only define what output we’re seeking (positive, neutral or negative), but we give the AI something to go on with supplied examples. These are, of course, very basic examples that any AI model should be able to determine on its own, but you can apply the same technique to far more complex, nuanced scenarios where the answer might not be as clear-cut to help improve your results.
In the example above, ChatGPT rightly identified the third example as neutral – it contained good and bad feedback. So now you could let it loose on a spreadsheet of many thousands of lines of feedback via advanced data analysis and let it work out the sentiment expressed in each, giving you quantitative results from qualitative data that would normally require human grunt work (in other words, how many of these social media comments on the latest issue of APC are positive?)
Chain of thought prompting is another technique that can help avoid errors. Here you’re asking the AI to think through its answers step-by-step, to show its working, if you like. For example, you might ask:
I’ve been offered two loans, each for $100,000, repayable over three years.
The loan from ANZ has a 4.5% APR and an arrangement fee of $500.
The loan from Westpac has a 4.9% APR and an arrangement fee of $600.
Which has the cheapest overall cost? Think step-by-step and show your working.
When we first entered this into ChatGPT Plus, it gave us a long set of workings, but that revealed its answer was based on annual payments, and not the monthly payment terms that most loans are based on. To confirm, we asked:
Are you assuming payments are made monthly?
It then went off and calculated the correct figures, answering that the ANZ loan would be around $800 cheaper over the three-year term (it provided more detailed answers than this).
There are a few key lessons, then:
1 Be as precise as possible with your initial prompt (for example, using “APR” rather than just “interest rate”).
2 Get the AI to show its working, so that you’ve got a better chance of spotting mistakes, misassumptions or hallucinations (stuff it’s made up).
3 Challenge the AI on its methodology, so that you can be as sure as possible that the answers you’re getting are correct.