Meta AI Research Labs have created a new state-of-the-art chatbot and are allowing members of the public to talk to the system to gather feedback on its capabilities.
The bot is called BlenderBot 3 and can be accessed on the web. (Although, at this point, it appears only US residents can do this.) BlenderBot 3 can engage in general chat, Meta says, but it also answers the kind of queries you might ask a digital assistant, “from talking about healthy health.” food recipes to find kid-friendly services in the city.”
The bot is a prototype and built on Meta’s earlier work with what are known as large language models, or LLMS, powerful but flawed text generation software of which OpenAI’s GPT-3 is the most well-known example. Like all LLMs, BlenderBot is initially trained on large sets of text data, which it mines for statistical patterns to generate language. Such systems have proven to be extremely flexible and have been put to a variety of uses, from generating code for programmers to helping authors write their next bestseller. However, these models also have serious flaws: they regurgitate bias in their training data and often fabricate answers to user questions (a big problem 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. Even more importantly, users can click on their responses to see where they got their information from. BlenderBot 3, in other words, you can cite your sources.
By releasing the chatbot to the general public, Meta wants to collect feedback on the various issues facing large language models. Users chatting with BlenderBot will be able to flag any suspicious responses from the system, and Meta says it has worked hard to “minimize the use of vulgar language, slurs, and culturally insensitive comments by bots.” Users will have to opt-in to have their data collected, and if so, their conversations and comments will be stored and then published by Meta for use by the general AI research community.
“We are committed to publishing all the data we collect in the demo in the hopes that we can improve conversational AI,” said Kurt Shuster, a research engineer at Meta who helped create BlenderBot 3. the edge.
Historically, releasing prototype AI chatbots to the public has been a risky move for tech companies. In 2016, Microsoft launched a chatbot named Tay on Twitter that learned from its interactions with the public. Unsurprisingly, Tay was soon trained by Twitter users to regurgitate a variety 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’s malfunction and that BlenderBot has all sorts of safety rails that should prevent Meta from repeating Microsoft’s mistakes.
Crucially, says Mary Williamson, research engineering manager at Facebook AI Research (FAIR), while Tay was designed to learn in real time from user interactions, BlenderBot is a static model. That means it is capable of remembering what users say within a conversation (and will even retain this information via browser cookies if a user exits the program and returns later), but this information will only be used to improve the system later.
It’s just my personal opinion, but that [Tay] The episode is relatively unfortunate, because it created this chatbot winter where every institution was afraid to bring out chatbots publicly for research,” says Williamson. the edge.
Williamson says that most chatbots in use today are limited and task-oriented. Think of customer service bots, for example, which often just present users with a pre-programmed dialog tree, narrowing down their query before passing it on to a human agent who can actually get the job done. The real prize is building a system that can carry out conversation as free and natural as a human, and Meta says the only way to achieve this is to allow bots to have free and natural conversations.
“This lack of tolerance towards bots saying useless things, in the broad sense of the word, is unfortunate,” says Williamson. “And what we’re trying to do is publish this very responsibly and push research forward.”
In addition to putting 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 has 175 billion parameters, through a form here.