Meta Will Make Its AI Chatbot Available on The Internet for The General Public to Speak to
Meta hopes to gather feedback from the public by releasing the chatbot. BlenderBot chatters will be able to flag suspicious responses and Meta will publish their conversations and feedback later.
“We are committed to publicly releasing all data we collect in the demo in the hope that we can improve conversational Ai,” Kurt Shuster, a Meta research engineer who helped to create BlenderBot 3 told The Verge.
The release of prototype AI chatbots to the public has been risky for tech companies in the past.
Microsoft launched a chatbot called Tay, which learned from the public’s interactions. Twitter users quickly instructed Tay to recite a variety of racist and antisemitic statements. Microsoft took the bot offline within 24 hours.
Meta claims that the world of AI is changing a lot since Tay’s mishap and that BlenderBot offers safety rails to stop Meta from repeating Microsoft’s mistakes.
Mary Williamson, research engineer manager at Facebook AI Research, (FAIR), says that while Tay was created to learn in real-time from user interactions, BlenderBot can only do so statically. This means that it can remember what users say in a conversation and will retain this information via browser cookies if the user exits the program. However, this data will only help to improve the system.
Williamson told The Verge that “it’s just my opinion but that [Tay] episode was relatively unfortunate because it created a chatbot winter in which every institution was afraid of putting out public chatbots to research.”
Williamson states that the majority of chatbots used today are task-oriented and narrow. For example, customer service bots present users with a preprogrammed dialog tree that narrows down their query and then passes them on to a human agent who can do the actual job. Meta believes that the real prize lies in creating a system that is as natural and free-ranging as a human conversation.
Williamson says, “This lack in tolerance for bots saying unscrupulous things, in general, is regrettable.” “And that’s what we’re trying do is release it very responsibly and push research forward.”
Meta is publishing the code, training data, and smaller models. You can request access to the largest model with 175 billion parameters by filling out this form.