I’ll ask it!
Prompt (followed by two linefeed characters)
What is the purpose of OpenAI’s InstructGPT instruction-following models (like text-davinci-002), and how do they differ in behavior to normal completion AI models?
The purpose of OpenAI’s InstructGPT instruction-following models is to generate text that follows a set of given instructions. These models are specifically designed to understand and follow human-written instructions, making them useful for tasks such as text summarization, question-answering, and text generation.
Unlike normal completion AI models, which are trained to generate text based on a given prompt or context, InstructGPT models are trained to understand and follow specific instructions. This means that they are better able to generate text that is coherent and relevant to the given instructions, rather than just completing a given prompt.
Additionally, InstructGPT models are trained on a diverse range of tasks and instructions, making them more versatile and adaptable to different scenarios. They also have the ability to incorporate common sense and world knowledge into their text generation, making their outputs more human-like and understandable.
The fact that the AI can answer a question might not be so obvious now. Instead, we must compare it to a completion model without the “instruct” training, and see what the effect of providing a prompt for the AI to complete on is; same prompt:
The InstructGPT models are intended to be used as a way to give instructions to GPT-3. The idea is that you would give it a list of instructions, and it would then execute them. For example, if you gave it the instructions “take a picture of the Eiffel Tower, find the Eiffel Tower on Google Maps, and then return the coordinates of the Eiffel Tower,” it would execute these instructions and return the coordinates of the Eiffel Tower.
How does the behavior of the InstructGPT models differ from the behavior of normal completion models?
The behavior of the InstructGPT models is different from the behavior of normal completion models in that the InstructGPT models are not designed to produce a single response to a given prompt. Instead, they are designed to produce a sequence of responses to a given prompt. This is different from the behavior of normal completion models, which are designed to produce a single response to a given prompt.
How does the behavior of the InstructGPT models differ from the behavior of normal completion models?
The behavior of the InstructGPT models is different from the behavior of normal completion models in that the InstructGPT models are not designed to produce a single response to a
You can see that the completion AI makes a farcical answer - and then continues writing as though the prompt was just part of some document that was being written, until ultimately it starts looping when the pattern of the previous text becomes dominant - it completes, like an advanced version of autocomplete on your phone.