OpenAI becomes a service

OpenAI made headlines last year with the release of GPT-3, a sophisticated piece of AI engineering. It’s not the snappiest name for a product, but unpacking the acronym gives some useful clues about what it does.

What is it?

GPT-3 stands for Generative Pre-Trained Transformer 3. It generates content and is built on a pre-trained deep learning model. In this case, the model was trained against large swathes of internet content, and the content it generates is text.

As the model has been trained against such a large amount of text using a significant number of different parameters (over 175 billion!), it really is rather good at generating text in a wide range of forms with relatively little initial context.

What’s exciting about it?

It’s capable of translating content between languages, writing stories, writing code, generating news articles, and more. The output is astonishingly coherent, and – depending on the inputs – disarmingly whimsical in a very human-like way. See AI-generated pigeon breeds from one of my favourite machine learning blogs for an example.

You won’t believe quite how human-like the text it generates can be until you try it for yourself – which has historically been a bit tricky. Running the full trained model yourself would likely cost six figures due to the amount of computing power required.

OpenAI recently removed the waitlist to access their Beta program, meaning developers can get their hands on a hosted version of the technology. Which is excellent news. However, another organisation in this space have also announced their own flavour of GPT-3 as a service.

Enter Azure OpenAI Service

Microsoft recently revealed that they will be offering access to GPT-3 through their new Azure OpenAI Service.

This is good news – developers have a couple of options now, but also Azure brings with it a much more enterprise-friendly set of services for security, compliance, regional availability, and scalability.

Microsoft also has a license to allow it to embed GPT-3 capabilities into its products – expect to see the fruits of that in the near future. 

In practice

An early example of a practical application of the technology is Github Copilot, which is capable of producing code to help developers solve problems and even explaining the behaviour of existing code in plain English. GPT-3 is capable of generating code from scratch in a variety of programming languages given simple descriptions as a prompt.

From a marketing perspective, this technology could be used to help generate content variants for personalisation purposes, to create synopses / summaries of longer articles, or even to write coherent articles from scratch using simple prompts. There are all kinds of other potential applications - which are more easily realised now thanks to the availability of GPT-3 via APIs.

It’s invitation only for now – so sign up and await the email. In the meantime, the OpenAI Beta is public now, and I would definitely recommend signing up and throwing some prompts at the model and seeing what it comes back with. It really is quite remarkable.