Ask not what AI can do for me ...

THE NEW REVOLUTION

Ask not what AI can do for me ...

When large language models like ChatGPT were opened to the public to use and play with, artificial intelligence became part of our lives on all levels. But what can we actually do with it? How can it help us in our everyday activities? ALI VAN WYK spoke to Prof. Duncan Coulter, an AI specialist at the University of Johannesburg.

ANGELA TUCK
ANGELA TUCK

WE’VE seen computers do astonishing things over the last 60 years, from playing a big role in putting a human on the moon to most of us carrying around a small computer in our pockets. Called a cellphone, it is more powerful than a mainframe Nasa computer from the 1960s, and connected to a global digital network, just for our convenience. It’s been a revolution.

And now, according to no less a computer boffin than Bill Gates, we are in the middle of another revolution – artificial intelligence, and particularly large language models (LLMs). I asked Prof. Duncan Coulter about the relationships we are about to develop with machines.

We’ve seen what computers can do. To normal people, it seems things like CGI [computer-generated imagery] in the movies perform miracles. How is AI different from what computers could do in the past?

It is not like we made some amazing breakthrough in understanding the deep nature of intelligence, or anything like that. It is really a question of scale. We just got better at extracting meaningful patterns from data that is largely unstructured, and then, very importantly, predicting what future data would plausibly look like, based on those patterns. So, that covers a wide variety of things, like texts written by people, music, and videos, it is really just patterns.

Prof. Duncan Coulter
Prof. Duncan Coulter

You can think of it as a particular collection of writing. We can now generate new texts that feel like [they come] from that writing in terms of the style and idiosyncrasies of the authors, that can all be plausibly replicated. Therefore, even if it did enhance the extraction and generation, what it produces is not fundamentally new. The problem that really took people by surprise, is that a lot of what we do in our daily lives is not actually that difficult or requires deep intelligence. Much of what we do is just following patterns, repeating patterns and going through the motions. So, a lot of that is now plausibly automatable. Deep insight, and things like that, would still require human input, but most of what we do every day isn’t actually that intelligent.

We are suddenly surrounded by examples of what can be done with AI, but somehow it is still difficult to apply it to our own situation in our private and work lives. In a small business, for instance, how can we find out what AI applications are available to automate tasks? Is there an audit that can be done to match what is available with what a business needs? I used to own a business buying, renovating and selling vintage furniture. We had one admin person, two upholsterers, two woodworkers and a driver. Can AI applications improve a business like that?

I don’t have any expertise in robotics, but I can say that anything requiring interaction with the physical world seems to lag a bit behind. So, regarding all of the mechanical work or reupholstering, while there has definitely been a lot of progress, I anticipate that human skill will still be more cost-effective in the medium term. While there are now robots that can do very delicate tasks, they are certainly not anywhere near mass production and their usefulness at scale has certainly not been tested. They also require expensive rare elements for batteries. So, scaling them up is certainly not a foregone conclusion.

For the admin and backend side there's definitely scope for co-ordination and customer management, things like relationship management, handling queries and orders and so on. So, many of the initial points of contact can be automated and things can be summarised.

Let me ask a more specific question. Many small online businesses use Shopify as the online platform for their shop and a cloud computing program such as Sage for their bookkeeping. When you use these programs, you quickly learn that you have to copy a lot of information manually between these two programs. For instance, if a customer buys a dining set on Shopify, and they ask for an invoice or receipt, you must copy all their information manually from Shopify into Sage. It would have been much more efficient and less costly if the new customer’s information was automatically transferred to Sage and a customer profile created there so that an invoice could quickly be churned out. I looked for little programs or as they are called, plug-ins, online, for this purpose, but could only find expensive ones for large businesses. Here’s my question: Could I as a layperson work with a system like ChatGPT to create a little program for me to do this job? Are we there yet?

I would say it is technically possible but probably unwise to do it without some expert supervision for now because the tools aren’t exactly mature yet. The companies are basically throwing whatever they can against the wall to see what sticks and no clear standards for those improbability outcomes have emerged. It will be a bit of a Wild West for a while with a lot of competing options. So, I would not do it without someone with experience.

There is, however, a family of automation technologies which are confusingly called robotic process automation (there are no real robotic elements), but they can take values from a user, where they basically simulate a user. They move the mouse around, copy and paste things and act like a synthetic employee. Those have started to incorporate some of the more advanced techniques that have been developed in the last five years. So, they would be able to do the things you describe, for sure, but I wouldn’t just implement it myself. It will be the bigger companies that roll that stuff out since they can afford expensive consultants to help them. It would be a bit of a risk for a smaller company to do it.

In our personal lives, are there any applications developed through AI that can help to automate things?

Probably. I am not so big on home automation. I know that there are many frameworks for controlling various devices and making them publicly accessible on the internet of things. One thing worth noting is that our academy has two focuses, AI and security, and a lot of these technologies for home and business automation through the internet of things, are very insecure. There is a big risk of things going wrong with hackers and so on, and also just mundane things like companies going bankrupt and all of your stuff not working anymore. Or the ways you did things stopped working but they don’t offer any backwards compatibility. So, you’ve got to install new things.

Can you train yourself to speak to AI in a better way to get better results?

There are some techniques, but that’s one aspect where the progress has been substantial and the language interfaces to AI systems have gotten quite good, to the point where … you know, some people are pretty bad at speaking to other people and I’d say the average LLM is better at speaking to people than the average person. There is a family of techniques called prompt engineering – little tweaks that you put into the way you brief questions to get better results. One of the obvious ones is, if you’re asking it to solve a problem for you that requires a little bit of extensive reasoning, you can just add in the phrase,Think this through, step-by-stepand it will improve the quality of your answers.

But a lot of these techniques are built into the newer versions of systems anyway. The one I mentioned became what we now calltrain of thought”, and the reason for this is, that before you start interacting with the system, it is in fact prompted by the company itself to do what is called a system prompt. The companies are continuously tweaking and refining their systems to improve performance. So, I don’t know if it is worth your while to break your back trying to learn how to speak to an AI system. They are more likely to learn how to speak to us than we to them.

In general, what are the best applications that came out of the AI revolution?

The low-hanging fruit is customer retention and customer management. So, that’s why there are all the chatbots, the front-line queries to answer common questions and so on, that seems to work reasonably well. In a sense, some of these LLMs store a lot of knowledge in addition to just being able to produce intelligible sentences. So, they can operate like Wikipedia or a sort of way to get yourself broadly familiar with a topic. But if you want to do a real deep dive into a topic, you’ll have to move beyond that and engage with the original sources.

A big problem with the LLMs, and it is something we cannot control, is that they don’t know how right or wrong they are. Effectively, they are just projecting the next action they’re going to take or the next word they’re going to produce. So, they don’t really have a measure of how good that is. All they know is what is likely next according to the patterns they’ve internalised. They are not intentionally lying to you.

For younger people heading out into careers, which jobs are most likely to become redundant because of AI?

The problem is a wide variety of white-collar work can be affected and it is particularly at the junior and lower mid-levels. There is a big, looming problem. Let us take programmers, for example. The code generation systems can produce code that is as good as the code that a junior programmer produces, for sure. But the problem is they cannot yet reliably produce code that is as good as what an expert senior or even mid-senior coder can produce .

But by cutting out your junior level of programmers, how do we ever get people to your mid- and senior levels? A similar problem exists in law and other kinds of things where the groundwork is how you cut your teeth and learn and get to the next level. So, those pathways are so easy to automate away. We tend to make provisions to benefit us as a society in the short term but [that] might catch us in the long term. Therefore the companies might automate their entry-level work without thinking about their long-term survival.

So, it will be decades before AI takes over complex coding?

For a long time, coding was almost a default career option for people who knew they wanted to do a white-collar job. They knew that coding paid well and could therefore be oversubscribed. There are people who really deeply engage with the subject material and internalise the more computer science side of things but they aren’t the majority working in IT.

So, I think there is definitely going to be a reduction in the number of jobs in IT as a whole. You’ll still get the experts who maintain systems, or those who do the things that these AI’s aren’t good at, but those sort of experts and programmers aren’t the majority of programmers. The bulk will be affected. I anticipate a culling and that will put wages that computer workers will be able to demand under pressure.

♦ VWB


BE PART OF THE CONVERSATION: Go to the bottom of this page to share your opinion. We look forward to hearing from you.


Speech Bubbles

To comment on this article, register (it's fast and free) or log in.

First read Vrye Weekblad's Comment Policy before commenting.