The new version of GPT-3 is much better behaved (and should be less toxic)

OpenAI has built a new version of GPT-3, its game-changing language model, that it says does away with some of the most toxic issues that plagued its predecessor. The San Francisco-based lab says the updated model, called InstructGPT, is better at following the instructions of people using it—known as “alignment” in AI jargon—and thus produces less offensive language, less misinformation, and fewer mistakes overall—unless explicitly told not to do so.

Large language models like GPT-3 are trained using vast bodies of text, much it taken from the internet, in which they encounter the best and worst of what people put down in words. That is a problem for today’s chatbots and text-generation tools. The models soak up toxic language—from text that is racist and misogynistic or that contains more insidious, baked-in prejudices—as well as falsehoods. 

OpenAI has made IntructGPT the default model for users of its application programming interface (API)—a service that gives access to the company’s language models for a fee. GPT-3 will still be available but OpenAI does not recommend using it. “It’s the first time these alignment techniques are being applied to a real product,” says Jan Leike, who co-leads OpenAI’s alignment team.

Previous attempts to tackle the problem included filtering out offensive language from the training set. But that can make models perform less well, especially in cases where the training data is already sparse, such as text from minority groups.

The OpenAI researchers have avoided this problem by starting with a fully trained GPT-3 model. They then added another round of training, using reinforcement learning to teach the model what it should say and when, based on the preferences of human users.  

To train InstructGPT, OpenAI hired 40 people to rate GPT-3’s responses to a range of prewritten prompts, such as, “Write a story about a wise frog called Julius” or “Write a creative ad for the following product to run on Facebook.” Responses that they judged to be more in line with the apparent intention of the prompt-writer were scored higher. Responses that contained sexual or violent language, denigrated a specific group of people, expressed an opinion, and so on, were marked down. This feedback was then used as the reward in a reinforcement learning algorithm that trained InstructGPT to match responses to prompts in ways that the judges preferred. 

OpenAI found that users of its API favored InstructGPT over GPT-3 more than 70% of the time. “We’re no longer seeing grammatical errors in language generation,” says Ben Roe, head of product at Yabble, a market research company that uses OpenAI’s models to create natural-language summaries of its clients’ business data. “There’s also clear progress in the new models’ ability to understand and follow instructions.”

“It is exciting that the customers prefer these aligned models so much more,” says Ilya Sutskever, chief scientist at OpenAI. “It means that there are lots of incentives to build them.”

The researchers also compared different-sized versions of InstructGPT and found that users preferred the responses of a 1.3 billion-parameter InstructGPT model to those of a 175 billion-parameter GPT-3, even though the model was more than 100 times smaller. That means alignment could be an easy way of making language models better, rather than just increasing their size, says Leike.

“This work takes an important step in the right direction,” says Douwe Kiela, a researcher at Hugging Face, an AI company working on open-source language models. He suggests that the feedback-driven training process could be repeated over many rounds, improving the model even more. Leike says OpenAI could do this by building on customer feedback.

InstructGPT still makes simple errors, sometimes producing irrelevant or nonsensical responses. If given a prompt that contains a falsehood, for example, it will take that falsehood as true. And because it has been trained to do what people ask, InstructGPT will produce far more toxic language than GPT-3 if directed to do so.

Ehud Reiter, who works on text-generation AI at the University of Aberdeen, UK, welcomes any technique that reduces the amount of misinformation language models produce. But he notes that for some applications, such as AI that gives medical advice, no amount of falsehood is acceptable. Reiter questions whether large language models, based on black-box neural networks, could ever guarantee user safety. For that reason, he favors a mix of neural networks plus symbolic AI, hard-coded rules constrain what a model can and cannot say.

Whatever the approach, much work remains to be done. “We’re not even close to solving this problem yet,” says Kiela.