Can anyone use GPT-3 now that OpenAI has removed its API waitlist?

Yesterday, 18 November 2021, OpenAI removed the waitlist for its API, the service that lets programmers send text to GPT-3 and get text back. Until now, you had to join a waitlist and be approved before you could even try it. Starting yesterday, developers in supported countries can create an account and begin experimenting immediately.

OpenAI says progress on safety tools is what made the change possible. Along with opening the doors, it published clearer rules about what the API may and may not be used to generate.

From waitlist to open sign-up

OpenAI launched the API in June 2020 as a private beta, a test phase open only to invited users. An API, or application programming interface, is a way for one program to use another over the internet. Here, your app sends a prompt, a piece of text, and the API returns a completion, text that continues the pattern you set. You can "program" it simply by showing a few examples of what you want.

At launch, OpenAI explained why it chose a waitlist and an API rather than releasing the model itself. The models are very large and expensive to run, the business helps fund its research, and controlling access makes it easier to cut off harmful uses. Early users included Algolia, Quizlet and Reddit. Every user was vetted individually.

The platform has grown since then. OpenAI says tens of thousands of developers already use it. VentureBeat reports that, as of March, OpenAI counted GPT-3 in more than 300 apps, generating about 4.5 billion words a day. VentureBeat describes yesterday's change as a move from private beta to a more open beta.

What changed to make it possible

OpenAI points to several improvements made over the past year:

• Instruct Series models. Versions of GPT-3 tuned to follow plain instructions better. VentureBeat gives an example of the kind of request they handle well: a short explanation of the moon landing pitched at a young child.
• Question-answering endpoints. An endpoint is a specific service within an API. These let developers supply their own documents, such as product manuals or a knowledge base, so answers can be based on trusted sources and be more truthful.
• A free content filter. A tool that flags generated text that may be sensitive or harmful, so apps can block it.
• Review and monitoring. OpenAI still reviews applications before they launch publicly, and watches for misuse.
• Tools for getting started. An improved Playground, a web page for trying prompts without writing code; an example library with dozens of ready-made prompts; and Codex, a model that turns plain-English requests into code.

The new content rules

OpenAI's updated content guidelines list categories that users may not knowingly generate, or let others generate:

OpenAI says adult content is banned for now because its systems cannot yet reliably separate harmful uses from acceptable ones, and that it hopes to allow some acceptable uses later as its filters improve. It is also testing more targeted filters for specific categories with some customers.

VentureBeat adds context: OpenAI acknowledges that these improvements do not solve the deeper problem of toxic or biased output. GPT-3 learned from a vast amount of web text, and research shows it can amplify the biases in that text. Researchers at the Middlebury Institute and Georgetown University have shown it can be used to write extremist or misleading material.

A worked example: your first hour with the API

Here is a sensible path for a newcomer, using the tools OpenAI describes:

1. Sign up on OpenAI's website, if you are in a supported country.
2. Read the usage and content guidelines before you build anything, so you know what needs review and what is off-limits.
3. Open the Playground and pick a prompt from the example library, such as a summarizer or a question answerer. Run it and look at what comes back.
4. Change one thing at a time: the wording, the examples or the length, and compare results.
5. Write your own prompt with a clear instruction and one example, like this:

Rewrite each sentence in plain English for a 12-year-old reader.

Sentence: The municipality will implement a new waste-collection schedule.
Plain English: The town is changing the days it picks up garbage.

Sentence: Residents are advised to retain their receipts for warranty purposes.
Plain English:

1. Check every output for accuracy before trusting it, and turn on the content filter if you plan to show results to others.

What should you build first with the OpenAI API?

Start with something small, useful to you, and easy to check. Good first projects turn text you already have into a different form: summarizing your own notes, rewriting jargon into plain language, drafting replies you will edit, or turning a description into a short piece of code with Codex. In each case you can judge quickly whether the answer is right, which is the most important habit to build with any language model.

Avoid starting with tasks where mistakes are costly or hard to spot, such as medical, legal or financial advice. GPT-3 can sound confident while being wrong, and its training data carries biases. Projects that face the public will also need to pass OpenAI's review before launch, so it pays to read the guidelines early and design with the content filter in mind.

Finally, keep an eye on how the platform evolves. OpenAI says it expects to keep streamlining the process for developers and to allow more use cases as its safeguards improve. Skills you build now with prompts and testing will carry over as the tools change.