On August 5, 2022, Meta released BlenderBot 3, a chatbot that can search the internet for information on practically any topic.

In addition, Meta also shares the BlenderBot 3 model, data, and source code with the scientific community to enhance conversational AI.

BlenderBot 3

We need to train artificial intelligence (AI) systems to adapt to our demands if we want to create systems that can interact with people in more intelligent, safer, and valuable ways. BlenderBot 3 can hold natural conversations with users, who may then give the model comments on how to improve its responses. To aid in the advancement of conversational AI research, the researchers have made the BlenderBot 3 model and model cards available to the scientific community.

The BlenderBot series has made strides in integrating conversational skills, such as personality, empathy, and expertise. It utilizes long-term memory and internet research to conduct meaningful interactions. In addition, BlenderBot 3 inherits these skills and provides superior performance. Because it is on the publicly available OPT-175B language model from Meta AI, it is approximately 58 times larger than BlenderBot 2.

Since it is well-known that all conversational AI chatbots occasionally imitate and generate hazardous, biased, or insulting comments, it is advisable to avoid using them. Therefore, the researchers have undertaken extensive study, co-organized workshops, and developed new safeguarding techniques for BlenderBot 3. Furthermore, BlenderBot can still make harsh or insulting comments despite our work, which is why we are collecting feedback to improve future chatbots.

Challenge

In their BlenderBot 3 demo, you can respond to each chat message by clicking the thumbs-up or thumbs-down symbols. Choosing a thumbs-down allows you to specify why you disliked the communication, whether it was irrelevant, nonsensical, unpleasant, or spam-like.

The researchers trained BlenderBot 3 with a significant amount of freely accessible language data to enhance its interactional capabilities. Their team assembled a large portion of the datasets used. For example, one recent dataset contains more than 20,000 discussions with individuals based on more than 1,000 conversational topics. Furthermore, the researchers trained BlenderBot 3 to learn from conversations and enhance the skills people value most, such as discussing nutritious recipes and locating child-friendly services in the city.

Tests and results

The researchers discovered that BlenderBot 3 performed 31% better on everyday tasks than its predecessors. Additionally, it has twice the information while making factual errors 47% less frequently. The researchers also discovered that only 0.16% of BlenderBot's comments to users were marked as impolite or inappropriate. Their study aims to gather and publish feedback data that the larger AI research community, and we can use later on. Researchers can then develop fresh ideas for making AI systems safer and more interesting for users.

Conclusion

Progress in AI depends significantly on how well the AI research community can build on the best technology. So, putting out chatbot models and datasets is the best way to learn how and why they work, what they can and can't do, and their limits.

Furthermore, BlenderBot 3 is a big step forward for public chatbots, but it's still not as good as a human. It's sometimes wrong, inconsistent, or off-topic. As more people try out their demo, the researchers plan to use the feedback they get to improve their models and share data that will help the AI community.