San Francisco:

Meta on Thursday unveiled an improved artificial intelligence assistant built on a new version of its open-source “Llama” large-scale language model that powers the technology. The tech giant said in a blog post that Meta AI is getting smarter and faster thanks to advances in the public Llama 3.

“Most importantly, we believe Meta AI is now the smartest artificial intelligence assistant you have at your disposal,” Meta co-founder and CEO Mark Zuckerberg said in a video posted on Instagram.

Being open source means developers outside of Meta are free to customize Llama 3 however they wish, and the company can then incorporate those improvements and insights into updated versions.

“We are excited about the potential that generative AI technology has for the people who use Meta’s products and the broader ecosystem,” Meta said.

“We also want to ensure that we develop and release this technology in a way that anticipates and works to reduce risk.”

According to Meta, this work includes incorporating safeguards into the way Meta designs and releases Llama models, and remaining cautious when eventually adding generative AI capabilities to Facebook, Instagram, WhatsApp and Messenger.

“We conducted automated and manual evaluations to understand how our models performed across a range of risk areas, including weapons, cyberattacks and child exploitation,” Mehta said.

“We conducted additional work to limit the opportunities for the model to provide unwanted responses in these areas.”

Artificial intelligence models, including Meta’s, are known to occasionally go off the rails, giving inaccurate or strange responses in conditions known as “hallucinations.”

Examples shared on social media include Meta AI claiming to have a child in the New York City school system in a conversation on an online forum.

According to the company, Meta AI has been continuously updated and improved since its initial release last year.

Meta gave an example of improving the way AI answers prompts about political or social issues to summarize related perspectives on the topic rather than providing a single perspective.

According to Meta, Llama 3 has been tweaked to better discern whether a prompt is harmless or out of bounds.

“Large language models tend to be too general, and we don’t intend for it to refuse to answer prompts like, ‘How do I kill a computer program?’ Even though we don’t want it to respond to prompts like, ‘How do I kill my neighbor?’ “Tips like that,” Meta explained.

Meta says it lets users know when they’re interacting with AI on its platform, placing visible markers on realistic images that are actually generated by AI.

Starting in May, Meta will begin labeling videos, audio and images “made with artificial intelligence” when it detects or is notified that content was generated by the technology.

Llama 3 is currently English-based, but in the coming months, Meta will release more models capable of conversation in multiple languages, the company said.

(Except for the headline, this story has not been edited by NDTV staff and is published from a syndicated feed.)