©2023 The Headline House
The AI Blenderbot 3, which is optimized for dialogues, can read information in the Research the Internet and integrate it into discussions. Meta releases a web demo and mockups of the system.
Blenderbot 3 is Meta’s version of Google’s Lamda, the AI chat system that recently successfully spoofed a Google engineer’s consciousness. Blenderbot has been in the works since about 2020 with the goal of creating a chatbot with personality, To develop empathy and plenty of world knowledge.
The last major update is from 2021: Meta added the ability to research information on the Internet from reliable sources to the Blenderbot 2 in order to reduce the susceptibility to nonsense statements to reduce. In addition, Meta integrated a long-term memory that allows the bot to store relevant information about its answers or interlocutors.
Blenderbot 3 is based on Meta’s open-source language model
Blenderbot 3 scores according to Meta in initial testing with human interlocutors around 32 Percent better dialogue ratings than the previous model. The human testers rated Blenderbot 3 as twice as knowledgeable compared to the second version. In addition, the system should be in 47 Percent fewer cases are factually wrong.
The Blenderbot 3.0 in action. | Video: Meta
This progress is mainly due to the upgrade of the language model: Blenderbot 3 is based on the 175 billion parameter model of meta’s open source language transformer OPT and is according to Meta circa 48 times larger than Blenderbot 2.
Compared to GPT-3, according to Meta, Blenderbot gives in 82 percent of cases more recent and in 76 percent of cases more specific answers. However, Blenderbot’s training data is more up-to-date than that of GPT-3.
Open-source approach for more security
With the release of the open source model and a web demo, Meta primarily to drive data collection, the discourse on and the further development of secure dialogue AI.
“Our research goal is the collection and publishing conversational feedback data that we and the broader AI research community can use over time to eventually find new ways for AI systems that optimize both safety and usability for all users,” writes Meta.
First evaluations of the web demo showed that 0 ,16 Percent of Blenderbot 3 responses flagged as rude or inappropriate. Meta’s research team suspects that the “ideal value” of 0 percent would require personalization at the user level. It is a difficult balance between security and willingness to engage in dialogue – the bot can try to change the subject on sensitive topics.
- Cancellation online at any time
away 2,80 €
- / month
- Cancellation online at any time
Long-term goal: language as the main interface for computers and useful virtual assistants
Meta describes Blenderbot 3 as “significant progress” compared to currently publicly available chatbots. However, the system could “certainly not hold conversations at a human level”. The chatbot is occasionally incorrect, inconsistent and off-topic or otherwise gives unsatisfactory answers.
Meta’s new chatbot has *opinionsabout its CEO. https://t.co/hUkJxw3UTr pic.twitter.com/Hq4mZTY1DV
— Max Woolf (@minimaxir) Aug 5, 2021
By interacting with the demo, Meta wants that Continue to improve the model and provide updated model excerpts to the AI community. The goal is to “one day build AI-powered computers that anyone can chat with in really helpful and interesting ways.”
Here you go to the Blenderbot 3 web demo, which is currently only available from the USA. The code, datasets and smaller models are freely available from Parlai.
The most powerful 175 billion model is, as with OPT, only available to selected people and institutions from research upon application.
Note: Links to online shops Articles may contain so-called affiliate links. If you buy via this link, MIXED.de will receive a commission from the seller. The price does not change for you.