Artificial Intelligence and dealing with relationships
No, we are not talking about when your daughter announces "I've painted my little brother a prettier colour".
We are talking about when Artificial Intelligence (AI) like ChatGPT can fail to understand simple relationships between objects. In contrast, humans are very good at understanding simple relationships between objects, and traditional computer systems are brilliant at processing even massively complex relationships.
AI systems can provide a huge boost to your business productivity, but sometimes they might be the wrong tool to use. We illustrate that here with an example of a simple problem that AI fails to solve efficiently and reliably.
Relationships in traditional computer systems
Relationships are a fundamental part of most computer programs. For example we could define some objects in a high level computer language and have a full computer program as simple as this:
Father has either zero or many Daughters
Object: Mother
Mother has either zero or many Daughters
Object: Daughter
Daughter has one Father
Daughter has one Mother
Object: Apple
Father has zero or many Apples
Daughter has zero or many Apples
Class Person contains: Mother, Father, Daughter
Class Fruit contains: Apple
That's it! We are done and now ready to get our trivial computer program to solve some problem for us. Here is the problem we are going to ask it to solve:
I had 15 apples. I then gave 6 apples to my daughter who then sold 6 apples to her father. How many apples do I have now?
When we attempt to program this problem we will quickly find that there is no single answer. Most likely our program will fail. We will have to go back and provide the missing information. Before we do that, let's go to ChatGPT4 and give it the same problem to solve:
A ChatGPT prompt
The answer from ChatGPT is incorrect. Is that worse than the program failing?
Providing missing data, or a better prompt
As we noted above, our traditional computer program would simply fail to run. While this may seem irritating, it's clearly a better response in this case. In contrast ChatGPT made a best guess, and it's best guess is obvious nonsense. If we are building a computer system processing customer bills or controlling an airplane, we don't want the system guessing like this.
Let us supply the missing information so our computer systems can solve the problem we presented.
The obvious bit of information to supply would be "I am the father", but let's make it slightly less obvious. Let's instead inform our computer systems that "I am not the mother". When we provide our traditional computer program with this information it will take a millisecond to deduce that I am therefore the father and I will get the 6 apples back form my daughter. I will then have 15 apples.
In contrast, ChatGPT responds like this:
At this point we might be tempted to admire ChatGPT's misplaced confidence that it has the correct answer. However, in real-life situations it may be dangerous; we don't want our AI machine to convince our users that it is correct when it is in fact wrong.
Some may raise the point that our prompt was needlessly complicated. However, if we are using AI to digest data from various sources in your organisation, this is typical of the sort of data input we will have to manage.
We could continue prompting ChapGPT in this manner and it would not only never provide the correct answer, it will contradict itself and forget information we have provided. We won't do that now, instead we will provide a clear prompt that allows ChapGPT to resolve it's mistakes:
ChatGPT now finally understands a relationship that one would assume was obvious. Certainly our traditional software would have understood it in a tiny fraction of a second.
Conclusions
LLM AI like ChatGPT is an exciting technology that can address many problems that have been difficult to program with traditional computer systems. At Business Landing we use AI systems to improve our productivity. However, it will be a long time, if ever, before AI replaces all traditional software systems.
For many business software applications AI will be a waste of resources, and might even be a danger to your business. It's important to understand when AI is appropriate. For example, Business Landing have developed various AI products, including a next-generation AI search bot for websites. This AI search bot takes the best of AI and combines it with enterprise quality software. Such winning combinations are likely to be the future of AI.
If you want to know what software systems are best for your business, contact us or leave a question below.
PS:
If you want another example of ChatGPT 4 getting confused, then you might want to take a look at this one. We noticed that ChatGPT 4 could solve a famous problem when given in it's traditional form, but was hopeless when we slightly adjusted it. This clearly shows that the AI LLM model is simply recognizing a pattern of words without truly understanding the logic.
A similar recognition failure occurs here. In this case the failure appears to be that the AI recognizes the text and assumes it is the more complex version of this problem. The more complex version being the version that is usually encountered.