Artificial Intelligence (AI) is everywhere. We all know ChatGPT (OpenAI), ClaudeAI (Anthropic) or Gemini (Google DeepMind). They are making headlines.

These are generative AIs based on text models using large language models (LLMs).

There are also generative AIs for images, audio/music/voice, or video. As their name suggests, they are very useful for generating text, images, and more.

But there are other types of AI. It’s important not to limit oneself to just these.

“When all you have is a hammer, everything looks like a nail.” This metaphor originates from Maslow’s Hammer, also known as Law of the Instrument, which implies excessive confidence placed in a single tool. It suggests that if you only have one tool, you’ll tend to apply it to every problem, even if it’s not always the most suitable solution.

Broaden your knowledge. At the very least, make sure you are aware of the other AI tools that exist.

An example

The doodle illustrating this article was commissioned from ChatGPT (a generative text AI) with the instruction to ask DALL·E (a generative image AI) to create a black and white doodle with #FF5500 highlights, in an hand-drawn style, with a height of 1 075 pixels and a width of 1 435 pixels.

Here are the hiccups:

• The dimensions simply were not respected. I made corrections after the generation.
• There was no inclusion of the #FF5500 color. I later learned that DALL·E cannot receive an instruction to insert a specific color. There was no mention of this. I added the color myself afterwards.
• When you look at the words written, they make no sense. They “look like” actual words, but they are not.

This situation is simple: I asked something of an AI, and it was not suited to the request. It delivered something anyway, without making any judgement about the quality of its output.

Two lessons are clear already:

1. You must learn about the limitations of the AI tool you’re using. The tool itself does not signal them.
2. You must choose the right type of AI for the task at hand.

In this example, the AI used was the right one to fulfil my request. Would another type of AI have been able to do so adequately?

Let’s examine the options.

The types of AI

There are several types of AI. You can find a list anywhere. We provide one here to make this reading self-contained, without any claim to exhaustiveness.

So, here is an overview of the types of AI in a few lines.

Symbolic AI (or rule-based AI)

During my university studies, we called these “expert systems” (this was in the 1980s – man I’m old). They use logical rules and explicit knowledge bases – information that is already known. Everything must be known and codified. It works well for medical diagnosis (where rules are known), regulated decision support (e.g., taxation), and logical deductive reasoning. Utilization of Rule Engines simplified the work.

Pawa naturally introduces this into its existing code by codifying the rules of best practice in sales analysis and demand forecasting. The advantage of these rules is that they are shared by the industry and stable enough to be applicable to this model.

Supervised statistical AI (supervised machine learning)

Algorithms learn from labelled data, which must be high quality. Poor data leads to poor learning. Data must be high quality and remain so even if the context changes. This works well for predictions where we also have the actual outcomes. Prediction and reality enable this learning. It works well for demand forecasts, email and document classification, and image recognition.

Pawa has plenty of quality data. Yours!

This model is well-suited to sales and demand forecasting.

Unsupervised statistical AI (clustering, dimension reduction)

This time, algorithms discover structures or groupings in unlabeled data. They recognize patterns. Here, you’re asking AI to discover patterns in the data, and it infers that similar results will follow. Useful for detecting rogue behaviors. For example, a product’s sales suddenly dropping or soaring.

Again, Pawa is successful here. There is plenty of high-quality data. Patterns that differ from usual behaviors can be detected. Useful for unusual sales behavior.

Reinforcement AI (reinforcement learning)

Algorithms learn by trial and error, maximizing a reward over time. The duration of the reinforcement term can have a significant impact and change the behavior of the AI. It works well for games and situations where feedback (success or error) is fairly immediate. For example, teaching a robot to keep its balance, video games. Learning will be slow if feedback comes late and if you want to maximize over the long term.

Generative AI (e.g., GPT, DALL·E)

The star of the moment.

These models usually generate content from text prompts, but it can also be images, sound, etc. This is where they excel. Asking them to do statistical analysis is unnatural. The problem is that they always accept the task, even if they are poor at that kind of work.

Pawa uses it notably to answer your questions in a personalized way, relying on its specialized knowledge base, and to generate executive summaries that guide your managers’ decisions.

Other Models…

There are hybrid models and many variations of the types listed. They combine the advantages and disadvantages of the models they implement and add complexity.

Choosing the right AI model

A key aspect in deploying AI is matching the right model to the problem to be solved.

For Pawa, we have taken the greatest care in this.

Thus, Pawa uses:

• Symbolic AI to apply industry best practices in sales analysis and demand forecasting. The rules are known and stable.
• Supervised statistical AI for sales and demand forecasting. Pawa is a champion of high quality and reliable data.
• Unsupervised statistical AI to identify unusual variations in sales. This model is the best for this type of challenge.
• Generative AI to answer textual queries and generate executive summaries that are easy to interpret.

And that’s it for this second blog post on AI project deployments… as we have applied it at Pawa.

Ciao