Our ChatFPT model interacts in a conversational manner. By using the dialogue format, ChatFPT is able to answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests. Unlike InstructGPT, ChatFPT follows instructions in a prompt and provides detailed responses.
In order to learn about its strengths and weaknesses, we are excited to introduce ChatFPT. ChatFPT is free to use during the research preview.
A few samples
In the following sample, ChatFPT asks clarifying questions to debug code.
Unless you know more about the code, it's hard to tell what's wrong. Is there any more information you can provide about what the code should be doing and why it isn't working as expected? What part of this code is this? Is it the entire thing?
User
It's only part of it - the error never surfaces. The channel seems to be malfunctioning
ChatFPT
It's hard to say without knowing what the code is supposed to do and what happens when it's executed. This code has a potential issue in that the resultWorkerErr channel is never closed, so the code could hang if the resultWorkerErr channel is never written to. It may happen if b.resultWorker doesn't return any errors or if it's canceled before it can return any.
You can fix this issue by closing the resultWorkerErr channel after writing to it.
If you want to send an error on the channel, for example, you can add the following code:
Close
Despite the fact that b.resultWorker never returns an error, the resultWorkerErr channel will always be closed. Again, without more information about the problem you're experiencing, it's difficult to tell whether this is what's happening to you. Would you mind providing more details about what happens when you run the code?
Limitations
It is not uncommon for ChatFPT to write plausible-sounding but incorrect or nonsensical answers. There are currently no sources of truth during RL training, so fixing this issue is challenging; (2) training the model to be cautious causes it to decline questions it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on the model's knowledge, rather than what the demonstrator knows.
If you tweak the input phrasing or attempt the same prompt more than once, ChatFPT will react differently. The model may claim not to know an answer given one formulation but may be able to answer correctly with a slight rephrasing.
Models are often excessively verbose and overuse certain phrases, such as stating that they are OpenAI-trained. Training data biases (trainers prefer long, comprehensive answers) and well-known overoptimization issues are to blame.12
When the user provides an ambiguous query, the model should ask clarifying questions. Currently, we assume that the user meant something different from what they intended.
Despite our best efforts, the model sometimes responds to harmful instructions or displays biased behavior. Using the Moderation API, we are warning or blocking certain types of unsafe content, but we expect some false negatives and positives at this time. As part of our ongoing work to improve this system, we'd like to hear from users.
