OpenAI reports that they have designed a new version of their amazing GPT-3 AI program, now called “InstructGPT,” which is “more truthful and less toxic” than predecessor programs and also “much better at following instructions,” and which “generate[s] sexual and violent content” less often.
Language models like GPT-3 have a broad range of uses, which include chatbot programs. Anyone who has ever spoken to a chatbot will confirm that they frequently spout nonsense, sometimes smutty or violent nonsense (encouraging a user to kill himself, for example).
This is because, according to OpenAI, their language program is “trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants. In other words, these models aren’t aligned with their users.”
So chatbots predict the most likely next word in a sentence to simulate intelligent conversation; they don’t actually understand what they are saying.
InstructGPT fixes this problem by using reinforcement learning from human feedback, in which users rank responses, fine-tuning the ‘bot’s behavior. Luka uses a version of this method for its popular Replika app, although Luka’s language model uses less data than OpenAI’s GPT program.
Audere Magazine regularly publishes articles about technology and VR. Read more here.
Image by Simis1/Pixabay