More than half a decade after Microsoft’s Truly Monumental Taye Debacle, the incident remains a stark reminder of how quickly an AI can become corrupted after exposure to powerful internet toxicity and a warning against building bots without sufficiently robust behavioral attachments. On Friday, Meta’s AI Research division will see if its latest iteration of Blenderbot AI can withstand the horrors of the interwebs with the public demo release of its 175 billion parameter Blenderbot 3.
A major hurdle currently facing chatbot technologies (and the natural language processing algorithms that power them) is that of sourcing. Traditionally, chatbots are trained in highly organized environments – otherwise you invariably get a Taye – but this ends up limiting the topics he can discuss to the specific ones available in the lab. Conversely, you can have the chatbot pull information from the internet to gain access to a wide range of topics, but it could go Nazi at some point, and probably will.
“Researchers cannot predict or simulate all conversational scenarios in search contexts alone,” Meta AI researchers wrote in a blog post on Friday. “The field of AI is still a long way from truly intelligent AI systems that can understand, interact and converse with us like other humans can. In order to create models that are more adaptable to real-world environments, chatbots need to learn from a diverse, broad perspective with people ‘in nature’.”
Meta has been working to resolve the issue since he first introduced BlenderBot 1 chat app in 2020. Initially little more than an experiment in open source NLP, by the following year BlenderBot 2 had learned both to remember information it had discussed in previous conversations and to search the Internet for additional details on a given topic. . BlenderBot 3 takes these capabilities even further by evaluating not only the data it pulls from the web, but also the people it talks to.
When a user logs an unsatisfactory response from the system, which currently accounts for approximately 0.16% of all training responses, Meta feeds the user’s feedback back into the model to prevent them from repeating the error. The system also uses the Director algorithm which first generates an answer using training data and then runs the answer through a classifier to check if it falls within a scale of good and bad defined by feedback of the user.
“To generate a sentence, the modeling and classification mechanisms of the language must agree,” the team wrote. “By using data that indicates right and wrong answers, we can train the classifier to penalize low-quality, toxic, contradictory, or repetitive statements, and statements that are generally unnecessary.” The system also uses a separate user-weighting algorithm to detect unreliable or malicious responses from the human interlocutor, essentially teaching the system not to trust what that person has to say.
“Our live, interactive, public demo allows BlenderBot 3 to learn organic interactions with all kinds of people,” the team wrote. “We encourage adults in the US to try the demo, have natural conversations about topics of interest, and share their answers to advance the research.”
BB3 should speak more naturally and conversationally than its predecessor, thanks in part to its massive upgrade OPT-175B language model, which is nearly 60 times larger than the BB2 model. “We found that, compared to BlenderBot 2, BlenderBot 3 provides a 31% improvement in overall score on conversational tasks as assessed by human judgments,” the team said. “He is also rated twice as informed, while being factually incorrect 47% less of the time. Compared to GPT3, on topical issues, he is more up-to-date 82% of the time and more specific 76% time.”