ChatGPT in the business world | Mini-Series 1/2 | Fork 10

Reading time: > 6 minutes

This is part 1/2. The second part will be about the use cases of ChatGPT.

I am an IT project manager in a manufacturing company. So, why am I writing about an AI-powered tool which seems to have nothing to do with my daily work?

Well, some innovations

Read this small series to understand how ChatGPT (and other natural process language AI systems) work, what you can do with it, and what it might mean for you and the related field you work in.

So, let’s start with the basics.

What is ChatGPT?

If you ask this question in the most basic sense, then ChatGPT is a user interface with a computer program behind it that answers anything you write into the chat window - called “prompt” - with human-like wording as well as phrasing. You can try it out yourself here.

In the example below, “explain quantum computing in simple terms” is the "prompt", while the response below is the result of my prompt. What is a prompt, exactly? A prompt is a phrase or a collection of individual keywords used as input. The tool then tries to analyze and understand the input and automatically generates a response.

If you ask the question technically, then ChatGPT is an artificial intelligence platform that uses natural language processing to have human-like conversations with humans. It has been designed to help people quickly find information, engage in conversation, and learn from interactions with the AI. What is really differentiating this chat tool is that it uses context in a very humane way. But we will get back to this later.

From a user perspective, this implies that the better your understanding of using “prompting” is, the more meaningful the response by the tool can be. Here are some examples:

Bad prompts:

• “Tell me more about everything.”
• “Describe all the characteristics of the ideal customer for our service.”
• “What are the positive and negative impacts of everything?”

Good prompts:

• “Tell me more about the latest trend in fashion.”
• “Describe to me the characteristics of the ideal customer for our service.”
• “What are the positive and negative impacts of social media?”

What people might miss, is that it is probably the fastest, massively adopted tool (> 1 million users) in the history of the internet.

To give you an example of how well it works, I will let ChatGPT explain itself: In the following paragraph, the cursive written part is based on prompts by me to ChatGPT.

How does ChatGPT work?

ChatGPT works with two AI processes. First, it uses the NLP (natural language processing) engine which compiles context to understand customer queries and generate responses. Secondly, it uses machine learning to create conversational experiences quickly and easily.

I asked the AI: what is a natural language processing system?

Natural language processing (NLP) is a subfield of artificial intelligence (AI) that enables computers to understand and process human language. It is used to analyze, interact with, and understand written and spoken human language. The NLP system can then be used to identify patterns in the data and produce insights or meaningful conversations.

To go deeper, I asked the AI to elaborate on it.

They use algorithms to try to interpret and understand the meaning of the language. They analyze the text for patterns and relationships between words and phrases, and […].

But it was still too generic, so I asked it to explain the analysis of text for patterns and relationships between words and phrases.

They (I guess it means the NLP models) analyze the text for patterns and relationships between words and phrases by looking for similarities in the way the words are used. For example, a NLP system might analyze a sentence for words that have similar meanings, such as 'shout' and 'yell', or for words that are related to each other, such as 'car' and 'road'. It can also look for relationships between phrases, such as how two sentences might be related to each other. This analysis helps the NLP system identify the meaning of the text, so that it can respond in an appropriate way.

Isn’t that cool? It explains itself how it works.

Ok, but a few more details about the tech. The software behind ChatGPT – it is called GPT-3.5 – is a so-called “Large Language Model (LLM)”. It takes terabytes of data from Wikipedia, and social networks like Twitter, Reddit, and Co. to “train” the algorithm.

The so-called “neural networks”, basically the “brains” of the self-learning computer program, scan these giant amounts of data and recognize patterns based on the statistical probability of a given word following another word.

Referring back to ChatGPT, the mode used is called “Reinforcement Learning from Human Feedback”.

Open AI – the company behind ChatGPT – trained an initial model using supervised fine-tuning, where human AI trainers supplied conversations impersonating both a user and an AI assistant, and were given model-prepared suggestions to help construct their responses. This new dialogue dataset was then blended with the GPT-3.5 dataset, which they altered into a dialogue format.

Then the trainers took randomly selected exchanged messages, compared them with similar completions of the requests, and had the trainers rank them. The best result is called the “reward” and is used to fine-tune the overall model.

3 reasons why it is much better than all the other chatbots

1. Amount of data and parameters

This was the recent state of usage of parameters in the machine learning algorithm from 2020. Microsoft’s Turing-NLG used 17 billion parameters in its language model. To compare it: GPT-3.5 uses ~175 billion. Then is more than 10x!

2. Self-learning: GPT-3.5 trains itself (unsupervised learning)

The program trains itself on the basis of the recorded texts (via the chat window). So, every time you put in a prompt in the chat window and give the response a thumbs up or thumbs down, it will take it as a reward and refine its model. The basis that makes this technology possible is called “Deep Learning”.

Deep Learning – in the case of ChatGPT – means that the software can predict which word is most likely to appear next in the sentence, conjugate it correctly, use the correct article, and adapt to the tonality of the text. It’s a bit like back in school when you had to fulfill a text with some words missing in it.

3. The attention-based model

Another innovation behind ChatGPT is the program's ability to link together not just individual words, but entire sections across pages and chapters. Thanks to this model, it can save relevant information and reference it later.

It achieves it by focusing on certain elements or tasks within a larger data set. This model uses a "soft attention" approach in which the model focuses on parts of the data that are most relevant to the task at hand. In this way, it can reduce the amount of data it needs to process and speed up the training process.

Next week: What are the implications of this tool?

So, now you know how it works. If you want to try it, do it before 3 pm because the US is still asleep. Afterward, it starts to get to its capacity limits and does not work properly.

This week, we went into what it is and how it works. Next week, we get into the more important stuff: What does it mean for the different industries that rely on writing and/or summarizing stuff?

Short post scriptum about layoffs

In my predictions for 2023 about more layoffs especially in middle management, I would like to highlight two pieces of news that were published:

First, Mark Zuckerberg - and I quote - reportedly said during an internal Q&A session in late January about the layoffs at Meta (former Facebook):

"I don't think you want a management structure that's just managers managing managers, managing managers, managing managers, managing the people who are doing the work.”

Second, FedEx announced it will cut 10% of its “officer and director team and consolidate some teams and functions”.