"A human brain doesn't work on probabilities it works on what you learn based on your five senses." – completely irrelevant to trainable algorithms.
"Models goes through training before they can learn to speak - no different to a toddler learning to speak words." – yes, models need to be trained, but completely different to a toddler learning to speak!
"They are not probability models. They are data models" – The GPT model uses data, and therefore there is a data model, but without the code (the algorithm) the data is useless. In AI when professionals refer to a "model" they are referring to the algorithm and the data as a whole. Hence GPT: Generative Pre-trained Transform.
A GPT model – a deep neural network – is used by ChatGPT to produce probability-based predictions, hence it is a probability model.
With "parameter" and "weight", you've picked up a couple of terms that refer to the GPT model's data, but they don't define the model as a whole nor do they explain how the model is used to get results. Another key term is "token", which refers to a meaningful fragment of the input data. In GPT the tokens can consist of multiple words ("will be", "has a", "neural network"), or parts of words ("brown"-ish, etc). Each token has a set of parameters associated with it. When applied, the model will produce "predictions".
Simple models will generally make a single prediction based on the inputs, and you often have little idea how likely the prediction is to be true. More complex models can produce multiple predictions with a probability of correctness. A program like ChatGPT can select from amongst these candidates. When you use the web interface to "talk" to ChatGPT it will normally select just one response. Occasionally you will get to chooose between two alternatives. But if you are using the developer's API you can ask for multiple responses to every prompt.
The API reference[2] describes a couple of parameters you can use to control this selection using probabilities. For example, you can use the "temperature" setting to control how tightly ChatGPT will control the probability of the response selection. At a temperature of 1, ChatGPT will choose only the most probable predictions. At lower temperatures it will select from lower probabilities, giving you more creative (i.e. less likely) answers. The other probability parameter is called "top p". With a "top p" of 0.1 you can tell ChatGPT to select predictions from only the top 10% most probable responses.
Very briefly:
A GPT is trained using these tokens and associated parameters as data
Training is achieved by adjusting the weightings applied within the neural network[1]
Once trained the weightings don't change
ChatGPT analyses your prompt to extract the tokens and associated parameters
ChatGPT also tries to determine the subject, focus, style, etc you are interested in
The parameters and (probably many) other values are submitted to the neural network
ChatGPT uses the resulting values from the neural network to identify the potential response tokens (the "predictions")
It filters these potential responses to measure relevance to the topic, focus, style, etc determined above
It then chooses from among the remaining candidates based on other criteria including probability
There's a lot more to it than this, of course. For example, the NN will only give ChatGPT a starting point for the response. It will use an iterative process using the NN to continue generating response tokens, while applying various rules to maintain coherence, style and relevance, while avoiding toxicity and repetition.
(20min delay)