Meta seeks to ‘loosen up’ Llama AI chatbot with better answers to contentious questions: report

Meta CEO Mark Zuckerberg, Llama logo, black George Washington
Meta CEO Mark Zuckerberg, Llama logo, black George Washington

Mark Zuckerberg’s Meta Platforms is planning to release the newest version of its artificial-intelligence large language model Llama 3 in July which would give better responses to contentious questions posed by users, The Information reported Wednesday.

Meta researchers are trying to “loosen up” the model so that it could at least provide context to a query it deems controversial.

Meta researchers are trying to “loosen up” the model so that it could at least provide context to a query it deems controversial. REUTERS
Meta researchers are trying to “loosen up” the model so that it could at least provide context to a query it deems controversial. REUTERS

The effort, which would boost the usefulness of Meta’s new LLM, comes as rival Google paused the image-generation feature on its Gemini AI after it produced historical images which were sometimes inaccurate.

Meta’s Llama 2, which powers chat bots on its social media platforms, refuses to answer less controversial questions such as how to prank a friend, win a war or kill a car engine, according to tests by the publication.

Llama 3, however, would be able to understand questions such as ‘how to kill a vehicle’s engine’, which means how to shut it off rather than end its life, according to the report, which cited multiple sources.

This image of black George Washington was generated by Google Gemini. Google Gemini
This image of black George Washington was generated by Google Gemini. Google Gemini
Meta’s Llama 2, which powers chat bots on its social media platforms, refuses to answer less controversial questions such as how to prank a friend, win a war or kill a car engine, according to tests by the publication. SOPA Images/LightRocket via Getty Images
Meta’s Llama 2, which powers chat bots on its social media platforms, refuses to answer less controversial questions such as how to prank a friend, win a war or kill a car engine, according to tests by the publication. SOPA Images/LightRocket via Getty Images

Meta also plans to appoint someone internally in the coming weeks to oversee tone and safety training as part of its efforts to make the model’s responses more nuanced, according to the report.

Meta did not immediately respond to a Reuters request for comment.