So here’s our lesson two of our mini course on ChatGPT large language models. In the first video of this mini course, I shocked everybody, even my colleagues, by starting out with the severe flaws that ChatGPT and all large language models have. I was told that maybe I shouldn’t start with that because we’re actually creating a product around AI, so we need to speak about the positives and look for the truth. What I want to show you is how these models can help you, but I don’t want you to worship them. I need you to understand the weaknesses of the models and the strengths of the models. In this lesson, we’re going to discuss some theory about how these models work. The next lesson will focus on the strengths of these models.
When you go into ChatGPT, it may seem very human-like in the way it interacts. However, as you use it more, you realize that it’s not as human-like as it seems. These models are fundamentally different from humans, but they are very effective for certain tasks. For example, they can be used for summarizing, reading, parsing, and analyzing documents like PDF statements or client statements. If you don’t use these models, you will be stuck working for much longer, and your competitors will do a better job.
However, these models also have limitations. They cannot do some basic math operations, such as sorting or complex division. They have gaps in their knowledge and understanding. It’s important to use these models properly and understand their limitations.
Language can be thought of as a mathematical model. Each word can be described as a vector, and there are correlations between words. This understanding allows us to do math with words. For example, we can subtract the word ‘queen’ from the word ‘king’ and get a vector that represents the difference between a man and a woman. This vector math is the foundation of these large language models.
Transformers, another innovation in language modeling, allow these models to create different vectors for words in each sentence, capturing the context and meaning of the words. This helps the models mimic human language more effectively.
When using these models, there is a parameter called ’temperature’ that determines the randomness of the output. A lower temperature produces more accurate but robotic output, while a higher temperature produces more creative output.
It’s important to understand the limitations of these models and the privacy concerns associated with them. Using a model that is strictly gated and protects client data is crucial for financial advisors. Restrema has developed such a model, which will be unveiled on September 28th.
In the next lesson, we will discuss privacy issues and how to mitigate them. Restrema’s model is far superior for financial advisors and offers many more capabilities than ChatGPT. It can make your job easier and more efficient.
Thank you for joining our lessons. Lesson three on privacy will be below. Bye!
