Meta’s AI research labs have created a new state-of-the-art chatbot and let members of the public talk to the system to gather feedback on its capabilities.
The bot is called BlenderBot 3 and is accessible via the internet. (Although at this point it seems like only U.S. residents can do this.) BlenderBot 3 can make general purpose talk, Meta says, but also answers the kinds of questions you might ask a digital assistant, “from talking about healthy eating recipes.” to finding child-friendly facilities in the city.”
The bot is a prototype and built on Meta’s previous work with what is known as large language models or LLMS – powerful but flawed text generation software of which OpenAI’s GPT-3 is the most famous example. Like all LLMs, BlenderBot is initially trained on huge data sets of text, which it searches for statistical patterns to generate language. Such systems have proven to be extremely flexible and have been used for a variety of purposes, from generating code for programmers to helping authors write their next bestseller. However, these models also have serious flaws: they outbreaks of prejudice in their training data and often devise answers to questions from users (a big deal if they’re going to be useful as digital assistants).
This last issue is something Meta wants to test specifically with BlenderBot. A great feature of the chatbot is that it is able to search the internet to talk about specific topics. More importantly, users can then click on the answers to see where the information is coming from. In other words, BlenderBot 3 can cite its sources.
By releasing the chatbot to the general public, Meta aims to gather feedback on the various issues facing major language models. Users chatting with BlenderBot can flag suspicious responses from the system, and Meta says it has worked hard to “minimize the bots’ use of vulgar language, defamation, and culturally insensitive comments.” Users must choose to have their data collected, and if so, their conversations and feedback will be saved and later published by Meta to be used by the general AI research community.
“We are committed to publicly releasing all the data we collect in the demo in the hopes that we can improve conversational AI,” Kurt Shuster, a research engineer at Meta who helped create BlenderBot 3, said. The edge.
Releasing prototype AI chatbots to the public has historically been a risky move for technology companies. In 2016, Microsoft released a chatbot called Tay on Twitter that learned from its interactions with the public. Somewhat predictably, Twitter users soon coached Tay in uttering a range of racist, anti-Semitic, and misogynistic statements. In response, Microsoft pulled the bot offline less than 24 hours later.
Meta says the world of AI has changed a lot since Tay’s failure and that BlenderBot has all kinds of safety rails to keep Meta from repeating Microsoft’s mistakes.
Crucially, said Mary Williamson, research engineering manager at Facebook AI Research (FAIR), while Tay is designed to learn from user interactions in real time, BlenderBot is a static model. That means it is able to remember what users say in a conversation (and will even save this information via browser cookies if a user leaves the program and returns later), but this data will only be used to further improve the system.
“It’s just my personal opinion, but that” [Tay] episode is relatively unfortunate, because it created this chatbot winter where every institution was afraid to turn off public chatbots for investigation,” Williamson tells WebMD. The edge.
Williamson says most chatbots in use today are narrow and task-oriented. For example, consider customer service bots, who often simply present users with a pre-programmed dialogue tree, refining their query before handing it over to a human agent who can actually get the job done. The real prize is building a system that can have a conversation as free and natural as a human’s, and Meta says the only way to achieve this is to let bots have free and natural conversations.
“This lack of tolerance for bots saying useless things, in the broad sense of it, is unfortunate,” Williamson says. “And what we’re trying to do is release this in a very responsible way and move the research forward.”
In addition to putting BlenderBot 3 on the web, Meta is also publishing the underlying code, training dataset and smaller model variants. Researchers can request access to the largest model, which has 175 billion parameters, via a form here.