In the previous tutorial, we learned that LLMs like ChatGPT are "taught" through machine learning, the technological process involved in the input. But what about the output? How does ChatGPT formulate its responses?

The answer lies in a technology called generative AI.

What's generative AI?

While machine learning and deep learning focus on understanding and interpreting data, generative AI takes it a step further by creating new data that mimics the original in novel ways. It's akin to a chef who’s able to think up new dishes based on what they learned from their training and studying recipes.

Generative AI is used for diverse creative and practical applications, such as generating realistic images, videos, and text for chatbots and virtual assistants, creating new video game environments, and more.

So what happens behind the scenes during a ChatGPT interaction?

When you type a query into ChatGPT and it responds, there are several processes working in tandem to make everything work:

1. You type your question, e.g., "What was Isaac Newton's role in the development of classical mechanics?"
2. ChatGPT uses a process known as natural language processing (NLP) to comprehend the meaning, context, and intent behind the question.
3. ChatGPT then uses knowledge retrieval to locate the most relevant information from its training data to create a response.
4. Once ChatGPT has compiled the information it needs, it uses a process called "language generation" or "text generation" (this is where generative AI comes into play) to generate a coherent and meaningful response.

Is it always accurate?

LLMs have been through extensive training, meaning they’re impressively correct most of the time. But they’re not 100% reliable.

At times, AI can produce an answer that’s misleading or factually incorrect. This is known as a hallucination, and can be caused by:

• Being asked an ambiguous question, which causes the AI to make incorrect assumptions.
• Limitations in the AI's knowledge, particularly if the required information is missing from its training data.
• The AI's inherent design constraints, like the inability to access current information or comprehend context beyond its training cut-off date.

How does ChatGPT compare to other AI assistants?

While ChatGPT shares many capabilities with other AI assistants like Claude, there are often distinctions:

• Model quality: As newer, more advanced models are released at different times, the best-performing model can vary. When GPT-4 was introduced, it was the most capable model, but Claude's subsequent release of Opus 3 seems to have taken the lead. This is correct as of today and may change anytime!
• Features: ChatGPT boasts a rich feature set, including image analysis, code generation, and more. While Claude offers similar functionalities, ChatGPT currently provides a more comprehensive suite of features. However, these differences are often minimal, as new features are continually added to all assistants.
• ChatGPT is guided by its own set of principles and guidelines, which may differ from those governing other AI models, resulting in varying outputs and behaviours.

Now you understand ChatGPT's inner workings, let's explore in the next tutorial how to use its potential in everyday tasks.