make it actually remember the conversation and not bullshit

it’s pretty bad at understanding when i say look at this drop box link and summarize it’ll spew random bs sometimes

Structured responses would be a game-changer so that the data can be easily processed.

Having this ability to define an objective that when achieved the conversation ends.

A numeric value that is assigned to each response that measures how accurate or confident the model believes its response is

I'm noticing that the model does not always listen to the "system" prompt. For example even when I tell it to stay on a certain topic the user is able to derail the conversation to something completely random.

Token counting. It would be nice to take the guesswork out of when something will hit the limit. Also, if I knew (exactly) how many tokens my prompt is then I could fine-tune max_tokens so the output fits into whatever I want it to fit into. Exposing the tokenizer via an API?

keep the conversation thread in gpt-turbo (so we don't need to send the whole history and system instructions each time)

and better rate limits please

Conversation or chat ID