Does chatGPT regret what it says?

Last updated: Oct 2024 | Estimated read time: 4 min

TLDR

The way chatGPT was built (i.e: the transformer architecture) suggests that when it's starting to answer a prompt, it's unlikely to suddenly change direction along the way. This is a similar phenomena when someone asks you your opinion about X, you start answering something, but during your explanation you feel that what you say is partially wrong or it would have been better to say something else. However, you also feel that it would be a bit ridiculous to admit that what you just said wasn't the best answer you could give. The (partial) solution of this problem is to abuse of the temperature parameter, which defines the randomness in chatGPT's answers. This implies that, for a given prompt, it will answer different things and give us a better panel of the possible answers.

Foreword: how LLMs work?

For a more detailed explanation about language models, check this article.

When using chatGPT on the official OpenAI website, you can see that its answers are coming word by word (or token by token to be specific). This isn't only a UI/UX decision by OpenAI but also the way chatGPT works. When it receives a prompt, chatGPT uses a probability distribution to determine the most likely word that should follow. For example, what's the most likely word, based on your experience: "I'm so hungry, I really [word]"? Personally, I'd say "need" or "want". Once it's done, the process continues again and again. This process is repeated iteratively, using the previous sequence as a new prompt, predicting the next word and continuing until the most likely word marks the end of the sequence.

It's essential to note that chatGPT works without access to the Internet or an external database. It relies solely on its past exposure to data to predict the most appropriate word at a given moment in the text.

An example with programming

Don't worry, there are no programming prerequisites to understand this section.

You now understand the basic functioning of chatGPT. This way of "reasoning" implies that chatGPT may "lock" itself in a direction that is not optimal. In order to make it intuitive, we will talk about programming.

When coding, you regularly need to import packages. Packages are a group of functions, coded by other developers, that do stuff. The point behind using packages is that it makes you gain lots of time since you can use super useful functions without even knowing how they work. For example, if you want to create some charts with the Python programming language, you regularly use the Matplotlib package since it has a lot of tools, very easy to use, that create very customizable charts. You can also use the Seaborn package, similar and related to Matplotlib, but with some different features. One last detail about packages is that generally a code looks like this: you first "call" the package, and then you write your code.

To this, if we add the way chatGPT works, we can reveal something interesting. When chatGPT is writing code, it will first write the import package1 etc, and then write the code. And since it uses past words to predict the next word, it means that it will only (theoretically, but that's usually true) use functions from the imported packages.

This also means when it has finished importing the packages (let's say it just imported the matplotlib package), it will not use the seaborn package, even though it would have been a good idea for this issue. You can think of it intuitively if you ask someone the best scientific argument that proves global warming, and the person starts saying that it's pretty hot outside today. It was one of the possible answers, but probably not the best one. Once it started, for most humans, it's hard to say "I was wrong, this argument is not a good one". And chatGPT is even worse than humans because the way it works reinforces the fact that once you've started making mistakes, you're stuck with them.

The solution: increase the heat

It would be super cool if it was possible to have let's say 3 or 4 chatGPT, trained (or finetuned) on different data, to compare their answers. So far it's not possible (as far as I know), but we can use a "trick" or a "masked" feature from chatGPT: the temperature. The temperature of a language model is a setting that affects the randomness of its responses. Have you ever seen that chatGPT give you slightly different answers even when you give it the exact same prompt?

When using chatGPT on the official website we can't change this parameter: we don't even know its current value. But when using it with the OpenAI API (this just means using it in another service like google doc, some coding program or anything else), you can choose this parameter value. It ranges from 0 (answers are deterministic, no randomness and the model tends to always answer the same thing to the same prompt) to 2 (highly random model, answers may be very significantly different for the same prompt).

The technique here is to use this parameter to see different responses to the same prompt, so as to be able to compare them. When ChatGPT generates text, it does so by predicting the next word in the sequence based on the previous words. The next word is predicted using a probability distribution over the entire vocabulary of the model. Each word in the vocabulary has an associated probability that indicates how likely it is to be the next word in the sequence.

Adjusting the temperature will affect the importance of each word. If the temperature is high, the probabilities will be more uniform, meaning that the model will be more exploratory and make less certain choices. This can lead to more creative and diverse responses, but it could also produce less consistent or riskier results.