What is Chat GPT
ChatGPT is a (LLM) large language model-based chatbot.
But how large?
Open AI used about 175 Billion parameters from over 45 Terabytes of data taken from the Internet.
It's All Numbers
When you write a question for Chat GPT, it turns those words in to a set of tokens.
And each token has a unique number.
These numbers are used in the neural network to predict the answer that goes along with the question.
Chat GPT is a Predictive Text Answer System
Training Tokens
Depending on where you read on the internet, Chat GPT, 3.5, was trained on 200 to 500 Billion tokens.
That is a lot of tokens.
A token can be a word, a part of a word, or even just a letter as in a name like GPT, which is three tokens, but three special tokens as they reference the very thing that is reading it.
A token can also be a comma or a questions mark or any number of non letter letters.
Depending on the context a word could be one of many different tokens.
Question Tokens
When you write a question for Chat GPT, it turns those words or parts of words into a set of tokens.
And each token has a unique number.
These numbers are used in the neural network to predict the answer that goes along with the question. With Chat GPT 3.5, the free one, https://chat.openai.com/, your question can be able 3,000 words long and this equates to 4096 tokens.
So that is still quite a big question.
An example
Write a limerick about the Philippines.
Lines 1,2 & 5 must rhyme. As must 3 &4.
In the islands of the Philippines fair,
Colorful sights fill the air.
With pristine beaches so white,
And warm smiles shining bright,
A paradise beyond compare!
The question, “Write a limerick about the Philippines,” is turned into a set of tokens.
The words limerick & Philippines are key to this question as are write and about.
Chat GPT does not deal with words, only numbers, and from those numbers, it predicts an answer and turns the numbers that it predicts back into their words, and we get our answer.
The first word starts the answer rolling, “In.” In a limerick about the Philippines, what would be the most predictable second word? Chat gives us back “the.” As it is predicting the third word it is not thinking “Hmm, what sort of rhyming do I want to do.” No it is only predicting the next word using the question and the answer so far.
Beware of anthropomorphism
When computers are only using tokens and numbers, but appear to be created, we quickly forget that words like:
Define for me learning.
How do we go about learning?
Why do we learn a thing?
Adapt – Discover – Experiment – Produce – Create
These words do not mean what you might think they mean.
Chat GPT is all about information and lots of it.
Chat GPT turns this information into numbers (tokens), building a matrix of how these tokens relate to each other.
From this matrix, Chat creates a second matrix from your question.
This second matrix creates a prediction for the first word that will begin to answer your question, then the second, and so on. Billions of computer hours of learning and fine-tuning have gone into Chat GPT, such that when it was released to the public it went viral.
Now we have Chat GPT 4 that you can pay for and Chat GPT 5 is on it way.
Chat GPT has no…..
knowledge
Chat is numbers and predictions.
It does not have knowledge.
It does not know what it is doing. The software of Chat runs step by step, it does not know what the next step in its software is going to tell it what to do.
Even saying it learns is inaccurate, as it is refining its capacity for better predictions as it received positive feedback from us saying what we like, and don’t like.
Its predictive matrix becomes better not more knowledgeable.
understanding
Chat GPT has no understanding.
It does not take information and turn it into knowledge like a human does.
It does not refine its knowledge into useful understanding, rejecting poor knowledge.
Any quality it has about the information it has is a function of that information, not some knowledge/understanding synthesis.
All information it has, its tokens are numbers, and the programming takes those numbers and though trial and error and billions and billions of computations, a matrix is created to best use those tokens in a feedback loop to create a predictive framework to enable answer to flow from questions
wisdom
Wisdom has a strong spiritual quantum to it.
We can grow in wisdom through trial and error.
Watching others and not doing what they do.
Taking moral advice and following it.
We get to grow trust worthly sources of Information that we turn into Knowledge, Knowledge into Understanding, and Understanding into Wisdom.
Chat GPT may give us easy access to asking hard questions and gaining insightful answers, but we have to learn to ask these hard questions well. Otherwise, garbage in will give us garbage out and we may not even know we now have garbage in our essay.