Meta’s AI research labs have created a new state-of-the-art chatbot and let the public test it.BlenderBot 3 has been released to public users in the United States. Meta believes BlenderBot 3 can participate in regular chats and answer questions from the digital assistant, such as identifying child-friendly locations.
BlenderBot 3 chats and answers Google-like queries
The bot is a prototype based on Meta’s previous work with Large Language Models (LLMS). BlenderBot was trained on massive text data sets to find statistical patterns and produce language. Such algorithms have been used to generate code for programmers and help writers overcome mental block. These models replicate biases in their training data and often generate solutions to user queries (a concern if they are to be effective as digital assistants).
Meta wants BlenderBot to test this problem. The chatbot can search the web for specified topics. Users can click on his answers to find out where he got their information from. BlenderBot 3 uses quotation marks.
Meta tries to gather information about the enormous difficulties of the language model by publishing a chatbot. BlenderBot users can report questionable answers, and Meta has tried to “minimize bots’ use of foul language, insults, and culturally incorrect expressions.” If users agree, Meta will save their discussions and comments for AI researchers.
Meta Research Engineer Kurt Schuster, who helped design BlenderBot 3, told The Verge:
How AI development has benefited BlenderBot 3 over the years
Tech companies generally shy away from releasing prototype AI chatbots to the public. Tey, Microsoft’s Twitter chatbot, learned through public interactions in 2016. Twitter users have taught Tay racist, anti-Semitic and sexist things. Microsoft removed the bot after 24 hours.
Meta claims the AI evolved after Tay’s glitch, and that BlenderBot includes safety rails to prevent a repeat.
BlenderBot is a static model, says Mary Williamson, head of research engineering at Facebook AI Research (FAIR). It may remember what users say in chat (and will store this information via browser cookies if the user leaves and returns), but this data will only be used to improve the system later.
“That’s just my opinion, but that [Tay] The incident is serious because it caused this chatbot to crash,” Williamson told The Verge.
Williamson believes that most chatbots are task-oriented. Consider customer service bots that offer consumers a pre-programmed conversation tree before handing them off to a human representative. Meta argues that the only way to design a system that can have real free conversations like humans is to allow bots.
Williamson thinks it’s unfortunate that bots can’t say anything constructive. “We are responsibly posting this for further research. »
Meta also releases the BlenderBot 3 source, training database, and smaller model versions. Researchers can request the 175 billion parameter model here.
Reference : fun-academy