For the first time, a language model has demonstrated an ability to “solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting.” Even more surprising still, it’s performance is close to human-level performance.
Given such a powerful, new tool, the first question raised is, “How can I harness it?” Today we’ll take a look at some of the ways GPT-4 can be used within a secure cloud development environment (Gitpod) to augment and improve the productivity of developers.
GPT-4 uses a natural language interface for its API calls. It also uses what’s called a system role for telling the language model the identity to assume when answering questions.
The system query is a place for you to tell the language model how you’d like your responses made. Here you can suggest things like “respond using only code”, or “respond as an experienced frontend developer”.
According to Andrej Karpathy in his new Microsoft Build talk, we need to ensure that we tell our language model to assume the identity of an expert, because it’s been trained on data from non-experts, and can’t really know that expert level answers are what we want unless we tell it.