
A change that makes the company a bit more deserving of its name.
When OpenAI announced GPT-3 in May 2020 we were already awaiting the news. The model promised to meet the high expectations set by its older brother in 2019. The year before, OpenAI had published the source code of GPT-2 and it was a complete success for them both in terms of hype and results. From AI dungeon, an adventure video game with "infinite possibilities," to headlines in every tech news outlet.
And GPT-3 would be bigger, more versatile, and even qualitatively stronger.
OpenAI released GPT-3 in June 2020, but in contrast to GPT-2 – and to the deception of most -, they decided to set up a private API to filter who could use the system. With 175 billion parameters, it was the largest neural network at the time, capturing the attention of mass media, researchers, and AI businesses alike. People had to join a waitlist and patiently expect OpenAI to get back to them (many tried but almost no one got access). It was so infamously difficult to enter that people published posts explaining how they did it.
OpenAI seemed to be particularly worried about safety. Controlling who could access the model allowed the company to limit the potential harm. Yet, despite their efforts, the safety measures were ill-defined. One of the people who tried GPT-3 early on was Jason Rohrer, an indie video game developer. He built Project December as a chatbot service for users to play with customized instances of GPT-3. What he never expected was that one user, Joshua Barbeau, would imbue the personality of his late fiancee into the chatbot.
The story went viral and OpenAI decided to intervene.
After Rohrer refused to adhere to a set of strict requirements, they shut down Project December. Rohrer had unintendedly revealed a critical flaw in their safety system; OpenAI couldn’t effectively control GPT-3 users. Project December didn’t cause harm (on the contrary, it helped Joshua recover) but it could have. And it was in clear conflict with OpenAI’s use-case policies.
The company solved this conflict poorly, destroying Rohrer’s work and revealing both the flaws in their safety measures and a lack of willingness to acknowledge the potential good of unexpected use cases.
But now they seem to have fixed those weaknesses. No more unsafe ideas, no more unnecessary erasure of promising projects. And no more waiting.
How you can access GPT-3 – and what to expect
Yesterday, OpenAI published a blog post stating they’ve removed the waitlist for people to access the API. Now anyone can use GPT-3.
Accessing the playground
I’ve gone through the process so you know what to expect. First, they ask for the usual account details – name, password, email, and phone number verification. But they also ask for the intention of use, whether it’s to build a product or feature, for personal use, research, or for journalism and content creation purposes.
I know OpenAI doesn’t like people sharing GPT-3 completions on social media, blogs, or magazines, so I wanted to check whether they’d change accessibility depending on which use you selected. I entered with two different accounts, one for content creation and the other for personal use.
It wasn’t surprising to get this message in the first case: "We’re grateful for your interest in exploring our Technology, and our Communications team would be happy to help you. Please contact [email protected] with a short description of your project, and we’ll get in touch with next steps." But, in the second case, they granted me instant access to the playground. If you just want to test some ideas and play around with the model, I recommend selecting personal use.
Pricing and use
OpenAI charges per token – either prompted to or generated by GPT-3. (A token can be understood as a part of a word. It’s safe to assume a token equals 0.75 words.)
During the first three months, you have $18 of free credit available to use as you wish. In the case of the DaVinci model (the most powerful version of GPT-3), 1000 tokens cost $0.06. With $18 you’ll have 300K free tokens. You could write 4 full-length novels with those free $18 credits.
For those of you who have never tried the API, I recommend you to read the documentation carefully, play with the examples and use the presets (predefined themed prompts) to see how GPT-3 behaves. Prompting text to the system seems easy but can be counterintuitive. You’ll need some practice to get nice results.
And last but not least, you’ll have to keep in mind the safety guidelines.
Strict safety measures
The main reason OpenAI has only now decided to open the API is that they’ve tightened their safety policies; "our progress with safeguards makes it possible to remove the waitlist for GPT-3." Regardless of what you want to use GPT-3 for, I recommend carefully reading the use-case and content guidelines together with the FAQ and safety requirements (everything is in the documentation).
The use-case guidelines define three categories: Almost always approved, evaluated case-by-case, and disallowed. As you can tell, they’re very conservative in what they permit. I recommend reading the disallowed cases to make sure you’re not even close to those and then try to stay within the first category as much as you can. The content guidelines enumerate which categories are definitely prohibited. Those include hate, harassment, violence, self-harm, adult, political, spam, deception, and malware.
From what I’ve read, I get the feeling that they have now more control over what’s possible to do. This makes it safer to use GPT-3 freely as long as you adhere to the guidelines, but at the same time, it highly restricts the use cases. Still, I’d say it’s better to have controlled access to GTP-3 than not being allowed – at least for those of you who just want to play and experiment with the playground.
I’ve criticized OpenAI extensively these months but I find this news a step in the right direction for the company (although they haven’t changed in other key aspects like being Microsoft’s puppet). GPT-3 is still highly expensive for long-term research and product development but soon there’ll be (and there are) other options available – either coming from the open-source community or from direct competitors that want to grab a piece of the NLP market.
Now go have fun!
Subscribe to The Algorithmic Bridge. Bridging the gap between algorithms and people. A newsletter about the AI that matters to your life.
You can also support my work on Medium directly and get unlimited access by becoming a member using my referral link here! 🙂