Feature suggestion for better user interaction

so as some of you might know but to others it might be a surprise as i was to me today when i found out, GPT doesnt actually learn from interacting with the user, as a privacy concern, which explained a lot to me actually, as i usually have a ruleset for my chats, that GPT constantly ignores, and while i have been spending about 20% of my GPT 4 limited tokens correcting it in great detail, it clearly has not been learning. so the suggestion is: allow personalized learning for paying users, or hell, sell it as an extra feature, take my privacy and give me ability to teach it a few simple things, as the constant youve reached your limit for gpt 4, try again at this time is so annoying to get, especially when youve just spend 3 tokens teaching it to do it correctly and not get it wrong…
its not that it forgets the rules, asked about 400 response chat about the rules, and it knew them perfectly, its just that it cant learn that 1 thing, which is to follow those rules at all times. and if it could learn that, i would not mind paying even 50% more for that feature on top of the plus package.

if you cant spare resources for it on your end, have it as a program on users end, blasting the CPU and GPU at 100% load when its learning, a few hours of learning and it would be miles better already, a small price to pay on my end for stalling my grayness a few years, considering how many of my nerves die thanks to the annoyance of it not being able to learn a basic thing.
also if privacy is your concern, have the users who want it agree to 10 different forms to sacrifice their privacy for that feature, that way if you care for imaginary thing called privacy in this world, you can keep thinking you have it, and those that dont care for privacy can have it taken away by it, not that they had privacy to begin with.

You should understand more how language models and their chatbot memory work before wishing for magical things.

What is context length?

  • Context length refers to the maximum number of tokens the model can receive as input to generate a response. Each token represents a word or piece of a word.

For example, GPT-3.5 has a maximum context length of 4096 tokens, while GPT-4 currently available goes up to 8192. This means when you provide a prompt, GPT-3.5 can only consider the most recent 4096 tokens to generate its response, unless there’s a database engine that selects particular turns of old input based on a semantic search.

Why does context length matter?

  • Context length limits how much previous conversation the model can take into account. A longer context length allows the model to “remember” more of the conversational history.

  • A short context length forces the model to generate responses based only on the most recent input, without broader context. This can make conversations disjointed.

  • A longer context length results in more coherent conversations that reference earlier parts of the dialog. But it also requires more compute resources.

Does context length give the model long-term memory?

  • No, context length just determines how much access the model has to previous inputs. The model does not actually “remember” or learn from those inputs.

  • Each time you provide a new prompt, the model generates a response from scratch based only on the context length window. It does not have true long-term memory of past conversations. Chatbot software provides a limited number of earlier conversation turns to give the illusion of memory.

Why doesn’t the model learn from conversations?

  • These models are pretrained on a fixed corpus of text data. They do not continue learning or updating their training after deployment.

  • User interactions do not get incorporated into the model’s training. So conversations do not expand the model’s knowledge over time.

This lack of learning prevents the model itself from building long-term memory, consistency, or personality during conversations. Each response is independent.

So in summary, while context length affects responsiveness, the fixed context window and the pretrained nature of transformer AI cannot truly provide conversational memory or learning for large pretrained models like GPT. The model resets with each new prompt, and any “memory” it has of you is just a conversation database from which limited amount can be retrieved.

makes sense, however when i tell it to follow a rule, like not use a particular word, it follows it for a bit, and then starts getting it wrong, when corrected about it, it works for a few tokens(basically the thing i send is a token here, not the length of it), and then it keeps getting it more and more wrong over time, however if i ask what the rule is, it knows it. so its not that i cant remember it, its that it cant learn to follow it. why a small learning thing would be nice to have, as it would moderate the models responses, to see if they comply with what its learned or not, like dual channel in a way, it would moderate the models generated response, and if needed make it change it so it fits the rules so to say. and only if both channels are in agreement can i see the response generated. and the 2nd channel would be unique to every individual who uses gpt for an extended period of time. it woudnt be such an issue if the constant corrections and such would not cost me a token with the 50 token limit… even editing and regenerating now costs a token, not to mention sometimes something goes wrong 5 times, and every error also consumes one of my 50 tokens per 3 hours.

basically its not a memory issue, its that its not properly trained to follow user set rules.