The AI research labs at Meta have created a new state-of-the-art chatbot and allow audience members to talk to the system in order to gather feedback about its capabilities.
The bot is called BlenderBot 3 and can be accessed on the web. (Although, for now, only US residents seem to be able to do that.) BlenderBot 3 can participate in a public chat, Meta says, but it also answers the kind of queries you might ask your digital assistant,” From talking about health and food recipes to finding kid-friendly amenities in the city.”
The bot is a prototype and builds on previous Meta work with what is known as Large Language Models, or LLMS – a powerful but flawed script generation program of which OpenAI’s GPT-3 is a widely known example. Like all LLMs, BlenderBot is initially trained on extensive datasets of text, which are then drilled into statistical patterns in order to generate a language. These systems have proven to be very flexible and have been put to a host of uses, from creating code for programmers to helping authors write their next bestsellers. However, these models also have serious flaws: they rejuvenate biases in their training data and often invent answers to users’ questions (a big problem if they’re useful as digital assistants).
This last issue is something Meta specifically wants to test with BlenderBot. A great advantage of a chatbot is that it is able to search the internet to talk about specific topics. More importantly, users can then click on their responses to see where they got their information from. In other words, BlenderBot 3 can cite its sources.
By launching the chatbot to the general public, Meta wants to gather feedback on the various issues facing the large language models. Users who speak with BlenderBot will be able to report any suspicious responses from the system, and Meta says it has worked hard to “reduce bots’ use of vulgar language, slurs, and culturally insensitive comments.” Users will have to sign up for their data to be collected, and if so, their conversations and comments will be stored and later published by Meta for use by the general AI research community.
“We are committed to making all the data we collect in the demo publicly available in the hope that we can improve the AI for conversations,” said Kurt Schuster, a research engineer at Meta who helped create BlenderBot 3. the edge.
Historically, releasing prototypes of AI chatbots to the public has been a risky move for tech companies. In 2016, Microsoft released a Twitter chat bot named Tay that learned from his interactions with the audience. Somewhat predictably, Twitter users quickly trained Tay to spew out a slew of racist, anti-Semitic, and misogynistic statements. In response, Microsoft took the bot offline less than 24 hours later.
Meta says that the world of AI has changed a lot since Tay crashed and that BlenderBot has all kinds of safety bars that should prevent the Meta from repeating Microsoft’s mistakes.
Crucially, says Mary Williamson, director of research engineering at Facebook AI Research (FAIR), while Tay is designed to learn in real time from user interactions, BlenderBot is a consistent model. This means that it is able to remember what users say in the conversation (and will keep this information via browser cookies if the user exits the program and comes back later) but this data will only be used to improve the system further.
“It’s only my personal opinion, but this [Tay] The episode is relatively unfortunate, because it created this chatbot winter where every organization was afraid to put public chatbots up for research,” says Williamson. the edge.
Williamson says that most chatbots in use today are narrow and task-oriented. Think of customer service bots, for example, which often only present users with a pre-programmed dialog tree, narrowing down their query before handing them over to a human agent who can actually get the job done. The real prize is building a system that can have free and natural conversation like a human, and Mita says the only way to achieve this is to allow bots to have free and natural conversations.
“The lack of tolerance for bots that say unhelpful things, in their broad sense, is unfortunate,” Williamson says. “And what we’re trying to do is issue this with great responsibility and push the research forward.”
In addition to placing BlenderBot 3 on the web, Meta also publishes the underlying code, training dataset, and smaller model variants. Researchers can request access to the largest model, which contains 175 billion variables, through a model here.