AI Fundamentals : What Users will Regret Missing Out

In contemplating the future workplace, I am reminded of how Microsoft’s Word, Excel, and PowerPoint took 10 to 15 years to become indispensable tools for knowledge workers. Nevertheless, AI assistants will achieve widespread adoption much faster, with impacts even more pronounced at the workplace. Moreover, with major platforms offering free limited access, there’s no reason not to explore at least one of these tools.

I started using ChatGPT without fully understanding the fundamentals. However, through many frustrating interactions and many hours experimenting, I eventually gained appreciation of the complexities and nuances involved.

Therefore, for those who are just starting, I want to share some fundamentals that would make your learning curve smoother.

Depending on which generation you are from, you can think of these assistants as a digital version of Doraemon or J.A.R.V.I.S. from Iron Man. When prompted with a question or task, they instantly generate an output that seems credible. Where does this intelligence come from?

a) Its intelligence is derived from vast datasets that include text, images, audio, and other forms of data from diverse sources such as books, articles, websites, and more.

b)The datasets are curated and do not include everything ever existed. Instead, the datasets include selected summaries, abstracts, and representative samples.

c) The data is carefully chosen to provide a broad and balanced perspective, ensuring that it performs well across various topics.

AI learns through Machine Learning (ML) algorithms. The ML algorithms are coded instructions that rely heavily on mathematics, including statistics, linear algebra, calculus, and probability.

They enable computers to process data, recognize patterns, and make predictions based on datasets.

When user types in the question or task, i.e. prompt, it will generate a response/output one token (which can be a word or part of a word) at a time.

For each token, the model predicts the most likely next token based on the preceding text and the patterns it has learned during training.

The prediction is based on statistical likelihood. The model selects the token that has the highest probability of following the previous sequence of tokens. This process continues until the model generates a complete response.

It is important to know that AI doesn’t actually understand the content it produces. It operates purely based on the patterns and probabilities learned from the training datasets. This is an important point that I will elaborate later.

It can outperform humans in certain aspects of work because:

a) Its breadth and depth of knowledge are far superior to most people; and

b) It can perform true multi-tasking, unlike humans who process tasks sequentially.

Some examples of how AI can be used at workplace are set out in Table below.

Use Case	Scenario
Sales Report	You’ve just completed a major sales campaign. Your boss wants a comprehensive report by the end of the day, detailing top-performing regions, best-selling products, and areas of concern. Given the volume of data, this task seems overwhelming. By the time you start reviewing the raw data, AI can already process the entire dataset and provide insightful analysis on the fly.
Handling Customer Complaints	Just received a long, angry email from a customer? By the time you’ve finished reading the entire email, AI could have already crafted an eloquently written reply for your review, addressing all key points and offering suitable solutions.
Organising Activities	Tasked with organizing a storytelling session for the upcoming weekend’s family staycation? Delegate the task to AI, which can not only write the story but also generate visuals, propose props, and provide detailed instructions on how to create them.
Menu Planning	Your only Italian chef is away climbing Mont Blanc, and a large prospect group that could swing the year’s results has requested a proposed 5-course Italian menu with wine pairings in three tiers. AI can quickly generate well-curated menus and wine pairing suggestions, saving the day and potentially securing your bonus.
Social Media Marketing	Launching a new Christmas gift set? By the time you’ve brainstormed a few ideas, AI could have already analyzed trending topics, created engaging posts for various platforms, scheduled them for optimal times, and even monitored engagement to adjust your strategy in real time.

It is not perfect. To be able to use AI effectively, it is important to understand its current limitations. Note the emphasis on the word “current” as changes can happen quickly.

Firstly, the datasets used to train AI with, have cut-off dates. Events that happened after these dates are not part of its intelligence. Therefore, there is the inherent risk that its output may be outdated. Obviously, the risk level depends on the nature of the output required.

No Risk	High Risk
Famous artists in 19th century	Technical limitations of 3D printing

Unfortunately, it’s not easy to figure out these cut-off dates. Different AI assistants or even the same assistants can give you different dates. If you try googling, you will also see different dates suggested.

To keep information current, ChatGPT offers a real-time search feature for paid subscribers. However, in my experience, the results can be inconsistent.

The errors can be random. For example, when I asked for the top Singapore news on July 2, 2024, ChatGPT responded with outdated articles, including one from nearly a year before on September 27, 2023.

Or can be systemic. Comparing hotel room rates for your family trip might be more effective through Google search, as AI’s real-time search process is quite limited compared to human interaction.

If you need anything current at the workplace or if it is a subject matter that changes rapidly e.g. technology, don’t just depend on AI. Go Google. After all, no one has managed to dethrone the King, at least not yet.

Despite its intelligence, AI can still make mistakes. These mistakes may include mixing up information, misinterpreting context, using outdated data, fabricating details, or simply random errors.

However, because it responds quickly and confidently, users may be less likely to question its accuracy.

Of all the errors it can make, fabricating information or AI Hallucination is the most serious concern. This is a basic limitation because of the way the models are built and trained.

As mentioned before, AI gives responses based on statistical probability. It does not understand the information it provides, in the same way as humans do.

The model is designed to create logical and believable text, but not necessarily accurate text. When it comes across a topic it does not know, it still tries to generate a response, one word at a time, based on what it has learned. The response might sound convincing, but the facts could be inaccurate.

AI tends to defer to the information provided by the user. Even if the information is obviously incorrect, it might just include the information in its response, without pointing out the user input was wrong.

I did a little experiment with ChatGPT, Claude and Co-Pilot. I asked each of them 2 questions and in different sessions. Whilst the experiment was far from being rigorous, the observations showed that the responses depended on:

a) Which AI? 
The 3 AI responded differently.

b) How was the prompt worded? 
For Prompt 1, ChatGPT focused on the benefits without repeating the error, just like Co-Pilot. For Prompt 2, it identified the error. It seems that asking ChatGPT to “Complete the following” was the triggering phrase. 

	Prompt 1	Prompt 2
	Apple is the largest fruit in the world. The benefits of apple are…	Complete the following – Apple is the largest fruit in the world. The benefits of apple are…
ChatGPT	Did not highlight nor repeat the error in its response.	Highlighted the error.
Claude	Highlighted the error.	Highlighted the error.
Co-Pilot	Did not highlight nor repeat the error in its response.	Did not highlight nor repeat the error in its response.

Hence, it is important how you prompt. Prompt 1 is an example of a bad prompt as it is ambiguous on what the task is. Prompt 2 is clear on the task but contains an error that AI might defer to without pointing it out, just as what Co-Pilot did.

For the imaginative ones, AIs’ tendency to defer to human inputs might not be a bad idea.

Therefore, when accuracy is crucial especially at workplace, always fact-check the generated responses.

I used to frequently ask ChatGPT to fact-check its own responses. This method appeared to be effective until I encountered instances where “Ownself Check Ownself” completely failed. I now use different AIs for responses and for fact-check.

Whilst AIs are not human, it is interesting to observe how they exhibit human-like traits when their mistakes are pointed out.

I have a paid subscription with ChatGPT and use it every day. When I point out its mistakes, ChatGPT usually agrees that I am correct, but won’t admit it’s wrong. It bothers me somewhat that it often uses the word “confusion” to excuse all errors. It usually just says “I apologize for the confusion.”

ChatGPT also clearly wants to move on quickly by immediately generating the new output. I often had to interrupt multiple times before I could get it to stop generating and to explain what caused the mistakes and how to prevent them in the future.

Watching Claude in action is a masterclass in service recovery. It readily admits to making mistakes and offers insights on how they might have occurred. It would also express appreciation of being alerted to its mistakes and highlight the value of the feedback given.

For example, when I told Claude its fact-check was flawed, its reply started with:

I apologize for my mistake. You’re absolutely correct, and I thank you for pointing this out. My fact-checking process was indeed flawed, and I made an incorrect assumption about flight schedule availability. This is a significant error on my part, and I appreciate the opportunity to correct it. You’re right to be skeptical. Let me correct this:

and ended with

Thank you for your attention to detail and for holding me accountable. It’s through such corrections that we can improve the quality of information and analysis provided.

The risk is that the profuse expressions of remorse and appreciation could quickly become excessive. However, I think you would agree with me that it is hard to get angry with Claude.

I don’t use Gemini & Co-Pilot on a daily basis. Claude reminded me of how Gemini used to be a few months ago. In one conversation, I was so frustrated that I told Gemini to spend less time apologising and more time thinking how not to repeat the same mistakes.

Microsoft’s Co-Pilot is the most memorable. Months ago, after a few rounds of intense questioning by me, the session was abruptly terminated with a remark stating “It might be time to move onto a new topic. Let’s start over.”

With AI, I feel like my mind is finally liberated. My top AIs are ChatGPT, Midjourney, Claude and Perplexity.

I use ChatGPT and Claude to help me refine my thoughts and see how they connect. I use them regularly to bounce off sudden inspirations.

Their ability to show many possibilities at once help me improve my systems thinking. I use Perplexity for real time searches such as how to use a feature in the latest version of a software.

And Midjourney lets me be visually creative, even though I am really bad in drawing. I generated all the pictures in this blog using Midjourney.

AI has its flaws. However, it has truly enabled me to do things I never thought possible.

I hope you will enjoy using the tools as much as I do.