What is the temperature of ChatGPT
Well, ChatGPT is HOT right now, but that isn't what I mean.
You have probably seen cases of ChatGPT saying random things, giving incorrect information, or a case of asking the same question and getting slightly different answers.
It turns out if you just let an LLM like ChatGPT use the highest probability to predict the next word it uses, it ends up being more accurate, more grammatically correct, really dry, and boring...you know...like a robot. 🤔
So to make them more exciting and capable of more creative ideas and responses, they are now powered by a six-year-old. Oh...nope, my bad. Actually, a parameter called "temperature" was introduced. The parameter allows for adjusting how random you would like your responses to be.
A generalized description of how this works is that temperature artificially raises the probability of a random set of low-probability predictions. If one of these predictions that have been artificially raised via temperature has a probability that exceeds all others, then that is chosen as the next sequence in the response.
Why did I tell you this? Well, it will relate to an upcoming post and is the start of a series of posts I will do on simple Prompt Engineering.
Note: I asked Stable Diffusion to create a "boring robot" image for me, and this image is the result. I guess the boring part is the toilet paper rolls for wheels? The one that Dall-E created was a robot reading a book which I felt was a personal attack. I am going to go back to reading a book now.
Original Post on LinkedIn:
John Rice