Blenderbot 3 is Meta’s version of Google’s AI chat system Lambda, which recently impersonated the mind of a Google engineer. Since about 2020, people have been working on blenderbot to make a chatbot with personality, empathy, and a lot of world knowledge.
The last major update was in 2021 when Den Blenderbot 2 added the ability to find information on the Internet from reliable sources. This made it harder for people to say things that don’t make sense. Meta also added a long-term memory that the bot can use to keep track of important information about its answers or the people it talks to.
Blenderbot 3 is based on Meta’s open source language model
Meta says that Blenderbot 3 did 31% better in its first tests with people than its predecessor did in its first tests with people. Blenderbot 3 was twice as smart as Blenderbot 2, according to the people who tested it. Also, the system should be wrong about facts 47% less of the time.
This is mostly because the language model has been improved. Blenderbot 3 is based on Meta’s open source speech transformer OPT, which has 175 billion parameters and is about 58 times bigger than Blenderbot 2.
Meta says that Blenderbot gives more up-to-date answers 82% of the time and more specific answers 76% of the time when compared to GPT-3. But Blenderbot’s training data are more current than GPT-3’s.
Open-source approach for more security
Meta’s main goal with the release of the open source model and a web demo is to encourage more data collection, discussion, and development of secure dialogue AI.
“Our research goal is to collect and publish conversational feedback data that we and the larger AI research community can use over time to find new ways for AI systems to improve both security and usability for all users,” writes Meta.
First evaluations of the web demo showed that 0.16 per cent of Blenderbot 3 responses were flagged as unfriendly or inappropriate. The research team at Meta thinks that the “ideal value” of 0% would probably need to be tailored to each user. It is hard to find a good balance between security and willingness to talk. For example, the bot might try to change the subject when talking about sensitive topics.
Long-term goal: language as the main interface for computers and use virtual assistants
Meta says that Blenderbot 3 is a “significant improvement” over other chatbots that are available to the public. But the system “couldn’t have a conversation on a human level.” The chatbot sometimes gives answers that are wrong, inconsistent, off-topic, or otherwise not good enough.
Meta hopes that interactions with the demo will help it improve the model and give the AI community new model excerpts. The goal is to “one day build AI-powered computers that anyone can talk to in really helpful and interesting ways.”
Here’s a link to the Blenderbot 3 web demo, which is only available in the US right now. From Parlai, you can get the code, the datasets, and smaller models for free.
As with OPT, the most powerful 175 billion models are only available to a select group of people and institutions.