Meta AI introduced a new version of the smart BlenderBot. The chatbot has been given more human functionality and thus has become more realistic in conversations.
BlenderBot was developed to have more realistic conversations with people than chatbots can currently do. Especially if it could be done in a smarter, safer, and more viable way. The necessary AI models must therefore be able to adapt to the changing needs of people.
The first version of the chatbot used artificial intelligence for empathy, knowledge, and personality. In version 2, this functionality has been extended with a long-term memory capacity. This allowed BlenderBot to have interactive and sophisticated conversations in almost any area.
BlenderBot3
In the recently released third version of BlenderBot, functionality has been added that remembers and learns from past interactions and optimizes itself accordingly. The chatbot roams the Internet for information on almost any topic, but it also evolves through natural conversations and comments with real people.
BlenderBot3 is the first chatbot with 175 billion parameters, complete with model weights, code, data sets, and model cards. The chatbot is based on the already giant OPT-175B language model of social media. This AI model is 58 times the size of the model supported by BlenderBot2.
Applied ML Techniques
The chatbot is powered by two advanced machine learning technologies from Meta AI; SeeKeR and Director. These techniques allow BlenderBot3 to learn from interactions and feedback. SeeKeR is a Modular Language Model (LM) that uses a search engine to stay relevant and up to date.
Output is a new paradigm that helps the chatbot generate answers based on language and classification models. The first mechanism helps BlenderBot3 to get the most relevant and streamlined answers based on training data. The second mechanism, the classifier, tells a model what is right or wrong in an answer based on human feedback. To get an answer, both the language model and the classifier must agree that everything is correct.
Prevent abuse
Another feature of BlenderBot3 that Meta AI pays a lot of attention to is that the chatbot is highly trained to prevent abuse. When a chatbot learns from human conversations and comments, the bot can be manipulated into uttering malicious comments or language.
To prevent this possibility of manipulation, learning algorithms that distinguish between beneficial and harmful feedback have been added. These algorithms filter notes, extracting comments that are considered malicious or suspicious. These algorithms take into account the behavior of (human) users throughout the entire conversation. In this way, the algorithms learn to trust those users with good intentions and distrust those who have negative intentions. In this way, BlenderBot3 should become more secure and smarter.
Trial version available
A beta version of BlenderBot3 has now been released in the US. The AI department of the social media giant is inviting everyone to engage with the smart chatbot so that it can learn and train itself more. In addition, due to the many interactions, the chatbot will also be able to point to the comments itself, such as thumbs up or down.
Meta AI also indicates that it will share data from BlenderBot3 interactions and conversations. The base model and model cards were also shared with science to advance conversational research on AI.
advice: Google AI LaMDA will be self-aware, experts express heavy criticism
