Dialogue-optimized AI Blenderbot 3 can search for information on the Internet and incorporate it into conversations. Meta publishes a web demo and models of the system to collaboratively improve its security.
Blenderbot 3 is the Meta version of lambda google, the AI chat system that recently succeeded in faking the consciousness of a Google engineer. Blenderbot has been in business since around 2020, with the goal of developing a conversational bot with character, empathy, and vast knowledge of the world.
The last major update dates back to 2021, when the Meta added to Blenderbot 2 the ability to search for information on the Internet from trusted sources to reduce its exposure to irrational data. Meta has also integrated a long-term memory that allows the bot to store relevant information about its answers or speaking partners.
Blenderbot 3 is based on the open source Meta language model
According to Meta, Blenderbot 3 scores 31 percent better in dialogue ratings than its predecessor in preliminary tests with its human counterparts. Blenderbot 3 is rated by human testers as twice as knowledgeable as the second version. Additionally, the system is said to be factually incorrect in 47 percent fewer cases.
This progress is mainly due to the upgrade of the language model: Blenderbot 3 is based on a 175 billion . parameter model OPT . Open Source Meta Language Converter Which, according to Meta, is about 58 times larger than Blenderbot 2.
Compared to GPT-3, Blenderbot provides more up-to-date answers 82 percent of the time and more specific answers 76 percent of the time, according to Meta. However, the Blenderbot training data is more recent than the GPT-3 data.
An open source approach for added security
With the release of an open source model and web demo, the primary goal of Meta is to drive data collection, discourse around and the development of artificial intelligence for secure conversation.
“Our research goal is to collect and release conversational feedback data that we and the broader AI research community can leverage over time to find new ways for conversational AI systems to improve security and engagement for everyone who uses them,” Meta writes.
Preliminary analysis of the web demo showed that 0.16 percent of Blenderbot 3’s responses were flagged as unfriendly or inappropriate. The “ideal” of 0 percent would likely require user-level customization, the Meta research team suspects. It will be a difficult balance between safety and willingness to engage in dialogue – the bot might try to change the subject on sensitive topics, for example.
Long-term goal: language is the main interface for computers and useful virtual assistants
Meta refers to Blenderbot 3 as a “big advance” over chatbots currently available to the public. However, the researchers said, the system is “definitely not at the human level” and sometimes gives answers that are incorrect, inconsistent, off-topic or otherwise unsatisfactory.
Meta’s new chatbot has opinions* about its CEO. pic.twitter.com/Hq4mZTY1DV
– Max Wolf (minimaxir) August 5, 2022
Through interactions with the demo, Meta hopes to continue to improve the model and provide updated sample snapshots for the AI community. The team’s goal is “artificial intelligence-powered computers that everyone can chat with in really useful and interesting ways.”
Head here to Blenderbot 3 webview, which is currently only available from the US. Code, datasets, and smaller models Available for free from Parlai.
The most powerful 175 billion model, as with the OPT, is only available to select individuals and institutions from the research community Upon request.
Note: Links to online stores in articles can be called affiliate links. If you buy through this link, MIXED gets a commission from the provider. For you the price does not change.
#Blenderbot #Meta #launches #powerful #dialogue #MIXEDnewscom