more than five years later Microsoft’s truly monumental Taye debacle, the incident remains a stark reminder of how quickly artificial intelligence can be disrupted after exposure to the internet’s potent toxicity, and a warning against building robots without a sufficiently robust behavioral tether. On Friday, Meta’s AI research arm will see if its latest version of Blenderbot AI can withstand the horrors of the internet with a public demo of its 175 billion-parameter Blenderbot 3.
A major hurdle currently facing chatbot technologies (and the natural language processing algorithms that drive them) is procurement. Chatbots have traditionally been trained in highly curated environments – because otherwise you always get a Taye – but this ultimately limits the topics it can discuss to the specific topics available in the lab. Instead, you could have a chatbot pull information from the internet to access a broad range of topics, but could and could at some point be completely Nazismized.
“It is not possible for a researcher to predict or simulate every dialogue scenario in a research setting alone,” the Meta AI researchers wrote in a blog post on Friday. “The field of AI is still far from truly intelligent AI systems that can understand, engage and chat with us just like other humans. To build models better suited to real-world environments, chatbots need to move from being diverse, to ‘in the wild’ ‘ people with a wide range of perspectives.”
Meta has been working hard to solve this problem It debuted the BlenderBot 1 chat app in 2020What started out as just an open-source NLP experiment, by its second year, BlenderBot 2 had learned to remember information it had discussed in previous conversations, as well as more details on how to search the internet for a specific topic. BlenderBot 3 takes these capabilities a step further, evaluating not only the data it pulls from the web, but also the people it talks to.
When a user records an unsatisfactory response from the system — currently hovering around 0.16% of all training responses — Meta feeds the user’s feedback into the model to avoid repeating errors. The system also employs the Director algorithm, which first uses the training data to generate a response, and then runs the response through a classifier to check if it fits within the range of right and wrong defined by user feedback.
“To generate a sentence, the language modeling and classifier mechanisms must be aligned,” the team wrote. “Using data indicating good or bad responses, we can train a classifier to penalize low-quality, toxic, contradictory or repeated statements, and often unhelpful statements.” The system also employs a separate user-weighting algorithm to detect responses from human conversationalists Unreliable or malicious responses – essentially teaching the system not to believe what that person said.
“Our live, interactive, public demo Enables BlenderBot 3 to learn from organic interactions with a variety of people,” the team wrote. “We encourage adults in the United States to try demonstrations, engage in natural conversations about topics of interest, and share their responses to help advance research.
The BB3 is expected to be more natural and chatty than its predecessor, in part due 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 improved by 31% in the overall score on the dialogue task, as assessed by human judgment,” the team said. “It was also considered twice as knowledgeable, while actually being incorrect 47% less time. Compared to GPT3, on topical questions, it was found to be 82% of the time and more specific 76% of the time latest time.”
All products featured by Engadget are selected by our editorial team, independent of our parent company. Some of our stories include affiliate links. We may receive an affiliate commission if you purchase through one of these links.