Nov 18, 2021 - Technology

OpenAI's GPT-3 gets a little bit more open

Animated gif of a computer screen that reads "A.I" with the "I" as a blinking cursor
Illustration: Sarah Grillo/Axios

The artificial intelligence research company OpenAI will eliminate the waiting list for access to the API of its natural language processing program (NLP) GPT-3.

Why it matters: The move will accelerate access to the world's best-known reading and writing AI model, and is a sign that OpenAI believes the program is safe enough — and can be monitored sufficiently — to be disseminated more widely.

Driving the news: Developers from supported countries will be able to sign up to access GPT-3's API and begin experimenting immediately, OpenAI said in an announcement Thursday morning.

  • Previously developers had to sit on a waiting list as OpenAI reviewed them before they could even get experimental access.
  • "We've added a lot of improvements across our API and added a number of safety features," says Peter Welinder, VP of products and partnerships at OpenAI. "We think a lot of value can come from getting more developers to build solutions to problems that they see in their environments."

Background: GPT-3 is a text-generating AI model that was trained on half a trillion words on the internet with 175 billion parameters — the values an AI aims to optimize during training — making it the largest and most sophisticated NLP model when it was released last summer.

  • GPT-3 can generate eerily human-sounding text in response to prompts, and can analyze and summarize the great unstructured data source that is written language.
  • OpenAI says that tens of thousands of developers have accessed GPT-3 through its API, building applications including question answering in customer feedback, support bots, and what Welinder says is "a lot of copywriting."

Yes, but: The ability to generate and tune human-like text at mass scale carries clear risks for misuse, especially in disinformation campaigns, though Welinder notes that the rate of misuse identified by OpenAI so far has "been really, really low."

How it works: Though the waiting list will be eliminated, developers will still need OpenAI's approval for production-scale applications, and the company has created automated tools to identify problematic text generated by GPT-3.

  • OpenAI has also updated its community guidelines, which ban hate, content that attempts to influence the political process, and all adult content excluding sex education or wellness.

Between the lines: The move marks another step in OpenAI's evolution from a nonprofit research center to a for-profit company partnered with Microsoft that is actively working to commercially license its products.

  • Ending the waiting list should help OpenAI grow that business, but Welinder also sees a practical value to "solving AI safety in practice," he says.
  • "We can start when the stakes are low in some sense, and then really iterate and build up the infrastructure tooling and organizational capacity for dealing with it."
Go deeper