(Reuters) – Meta Platforms plans to release the latest version of Llama 3, an artificial intelligence large-scale language model that provides better answers to controversial questions from users, in July, The Information reported on Wednesday. .
Meta-researchers are trying to “relax” the model so that it can at least provide context for queries that appear to be controversial.
The effort to increase the usefulness of Meta’s new LLM comes after rival Google suspended Gemini AI’s image generation capabilities after it produced sometimes inaccurate historical images.
The magazine’s tests show that Meta’s Llama 2, which powers chatbots on the company’s social media platforms, teaches less controversial topics such as how to prank a friend, how to win a war, and how to turn off a car’s engine. He refuses to answer any questions he doesn’t have.
However, the report, which cites multiple sources, says Rama 3 can understand questions such as “how to shut off a vehicle’s engine,” or how to stop the engine without ending its lifespan.
Meta also plans to appoint someone internally in the coming weeks to oversee tone and safety training as part of an effort to make model responses more nuanced, according to the report. That’s what it means.
Mehta did not immediately respond to a Reuters request for comment.
(Reporting by Priyanka.G and Yuvraj Malik in Bengaluru; Editing by Shailesh Kuber)