On Friday, Meta will test whether the new version of Blenderbot, an artificial intelligence chatbot, can handle the challenges of a real-world communication environment. On this day, the demo version of Blenderbot 3 with 175 billion parameters will go to conquer the open spaces of public access.
Meta has been working on the project since it first introduced the BlenderBot 1 chat app in 2020. Initially nothing more than an open source NLP experiment, by the following year BlenderBot 2 learned to memorize information from previous conversations and search the web for additional information about them. BlenderBot 3 can evaluate not only data from a conversation or the Internet, but also the people with whom it communicates. But AI will not be able to patent its inventions in the USA .
When a user registers an unsatisfactory system response (about 0.16 percent of them are recorded during the training process), Meta processes the user’s signal and enters it into the model to avoid repeating the error. The system also uses the Director algorithm, which first generates a response using the training data and then runs the response through a classifier to check if it matches a user-defined correct/incorrect scale.
“In order to generate a sentence, the language modeling and classifier mechanisms must agree,” the team wrote. “Using data on good and bad responses, we can train the classifier to block low-quality, toxic, contradictory, repetitive, or useless statements.” The system also uses a separate user comparison algorithm to detect untrustworthy or malicious replies by the interlocutor, teaching the system not to trust what that person is saying.
The BB3 is expected to speak more naturally than its predecessor, thanks in part to a significantly upgraded OPT-175B language model, nearly 60 times larger than the BB2 model. “We found that compared to BlenderBot 2, version 3 delivers a 31% improvement in the overall ranking of conversational tasks,” the team said. BlenderBot 3 is expected to provide answers that are 47% more correct, 82% more relevant, and 76% more specific.
The main problem in training AI chatbots is the sources of information. Initially, bots are trained in a carefully selected environment to avoid unwanted behavior as much as possible. At the training stage, possible topics are also limited. A chatbot released to the unlimited Internet is less predictable in its development – in communicating with people, it can touch on unwanted topics.
In addition to the quality of the intellectual chatbot, Meta could also improve the quality of moderation of the Ukrainian segment of Facebook. The Ministry of Digital Transformation applied to the company with such a request.
Blocking of Ukrainian bloggers on Facebook and Instagram — Ministry of Digital Transformation appealed to Meta with a request to improve the quality of moderation