Meta presents BlenderBot 3, a ‘chatbot’ capable of talking about “almost any topic”

MADRID, Aug. 8 (Portaltic/EP) –

Meta Platforms has released a demo Blender Bot 3the new generation of its ‘chatbot’, which is capable of talking about “almost any subject” and is now available on its website for residents of the United States.

A ‘chatbot’ is a virtual assistant that communicates with users of a website through text messages to generate a conversation. Its work is usually linked to offering support and help, and its popularity stands out especially among millennial users and Generation Z, according to a recent report by the Boston Consulting Group.

Given this context, Meta has announced the launch of a public demonstration of BlenderBot 3, the third generation of its virtual assistant, which the company has highlighted is capable of talking about “almost any topic”, as it has pointed out in a statement posted on his Artificial Intelligence (AI) blog.

In order to do so, BlenderBot 3 incorporates all of the abilities of its predecessors, including Internet search or long-term memory, as well as human-like features, such as “personality and empathy”.

To perfect its service, Meta has announced that it has completed the training of its assistant with a new database consisting of more than 20,000 conversations with people on more than 1,000 topics of conversationto be able to interact with people in a “smarter, safer and more useful” way.

The technology has also highlighted the ability of BlenderBot 3 to learn on the fly and improve the quality of its service. That way, when one of your answers is “unsatisfactory“For the user, the platform collects their feedback on it to refine its conversational model.

This Meta Helper combines two recently developed machine learning techniques, SeeKeR and Director, to build conversational models that can learn from interactions and feedback.

In this sense, the developer has commented that Director is capable of generating responses based on two mechanisms: language modeling and classification. The first provides this ‘chatbot’ with the most relevant answers, based on the data collected during its training.

The classification mechanism then helps you distinguish between a correct and incorrect response based on human feedback. Thus, to generate a response phrase, both mechanisms must agree. and complement each other.

From this combination, Meta trains its classification system so that it penalizes the phrases that BlenderBot 3 offers and are considered as “toxic, contradictory, repetitive or of low quality”.

In this way, the company develops new learning algorithms that aim to distinguish between useful responses and examples that may be harmful to users.


Likewise, the AI ​​laboratory explains that this method, which takes into account human behavior during conversations, learns to trust some users and improves the learning of its ‘chatbot’, compared to other conventional training methods.

Finally, BlenderBot 3 also incorporates an update to Meta techniques to encourage safer dialogue, thanks to features like allowing the assistant to respond to human comments on “challenging conversations” to encourage a more “civilized” conversation.

58 TIMES MORE POWERFUL THAN THE PREVIOUS GENERATION

Compared to its predecessors, Meta has highlighted the debut in BlenderBot 3, which integrates its own language model OPT-175Baspires to be up to 58 times more powerful than the one used in the previous generation, BlenderBot 2.

In this sense, BlenderBot 3 has improved its response to conversational tasks by 31 percent and makes 47 percent fewer mistakes and only 0.16 percent of its responses have been rated by users as “rude or inappropriate”.

Despite these advances, Meta has recognized that there are still tasks where it can improve, such as incorrect answers, with 1.1 percent of its testers reporting that their answers were incorrect.

For this reason, the technology developer hopes to collect as much information as possible in this first demonstration open to the American publicin order to perfect the tasks of your assistant.

Looking to the future, the assistant’s website indicates that it hopes to take this initiative to a greater number of countries in addition to the United States, as it has pointed out on the website of the virtual assistant in which it currently works.

In addition, the company has announced that it will share the BlenderBot 3 model, as well as its data and development code, with the scientific community, to fuel the growth of this field of AI.

Related articles

UVA mortgage loans under the magnifying glass

There are already ten banks that offer UVA mortgage loans for first or second homes. This competition between banking entities promises to revolutionize...

The scientific study of the May 19 supercar confirms that we are experiencing an...

At dawn on Sunday, May 19, cosmic chance had a very special encounter in store for us. The night...

Ecuador – First Division: Libertad vs Deportivo Cuenca Date 14

Starting at 7:00 p.m. next Monday, May 27, Deportivo Cuenca visits Libertad at the Reina del Cisne stadium, for the duel corresponding...