How did ChatGPT learn to follow instructions?
ChatGPT, and the underlying language model GPT-3 from OpenAI have the incredible capability of following instructions . . despite never having been explicitly designed to.
ChatGPT is the new chatbot that’s taking the world by storm due to its unprecedented ability to intelligently dialogue with humans.
In fact, GPT-3 is only specifically trained to do one thing: predict the next word. GPT-3 is an example of a ‘large language model’ (LLM) that traditionally was used to aid spell checkers.
It’s arguably not hyperbole to suggest that ChatGPT is an important inflection in human history given it’s the first useable general purpose embodiment of human thought into software.
And instruction following is a big part of that.
To some of us, ChatGPT is virtually AGI, artificial general intelligence. Perhaps only a few tweaks are required to achieve a simple AGI-like first effort.
To others it can’t even do basic human reasoning. But they are wrong. These researchers are overly pedantic and are being blinded from GPT’s clearly measurable capabilities by their own paradigm bias. It’s a not-invented-here thing.
But how did GPT-3 learn to follow instructions?
