Meta Releases Beefed-up AI Models

Meta on Thursday introduced an improved AI assistant built on new versions of its open source "Llama" large language model for powering the technology.

Meta AI is smarter and faster due to advances in publicly available Llama 3, the tech titan said in a blog post.

"The bottom line is we believe Meta AI is now the most intelligent AI assistant that you can freely use," Meta co-founder and chief executive Mark Zuckerberg said in a video posted on Instagram.

Being open source means that developers outside of Meta are free to customize Llama 3 as they wish and the company may then incorporate those improvements and insights in an updated version.

"We're excited about the potential that generative AI technology can have for people who use Meta products and for the broader ecosystem," Meta said.

"We also want to make sure we're developing and releasing this technology in a way that anticipates and works to reduce risk."

That effort includes incorporating protections in the way Meta designs and releases Llama models and being cautious when it eventually adds generative AI features to Facebook, Instagram, WhatsApp, and Messenger, according to Meta.

"We conducted both automated and manual evaluations to understand our models' performance in a series of risk areas like weapons, cyberattacks, and child exploitation," Meta said.

"We performed additional work to limit the chance the model provides unwanted responses in these areas."

AI models, Meta's included, have been known to occasionally go off the rails, giving inaccurate or bizarre responses in episodes referred to as "hallucinations".

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

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

Meta cited the example of refining the way its AI answers prompts regarding political or social issues to summarize relevant points about the topic instead of offering a single point of view.

Llama 3 has been tuned to better discern whether prompts are innocuous or out-of-bounds, according to Meta.

"Large language models tend to overgeneralize, 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?'," Meta explained.

Meta said it lets users know when they are interacting with AI on its platform and puts visible markers on photorealistic images that were in fact generated by AI.

Beginning in May, Meta will start labeling video, audio, and images "Made with AI" when it detects or is told content is generated by the technology.

Llama 3, for now, is based in English but in coming months Meta will release more capable models able to converse in multiple languages, the company said.

gc-juj/md