Behind Meta’s Effort to Make Llama 3 More Engaging

“Promoting or encouraging harmful or illegal activities is not acceptable,” Meta AI, an artificial intelligence assistant, told me the other day. “It's important to respect the safety and well-being of individuals and property.”

I had asked the assistant, which is embedded in Meta Platforms’ apps and Ray-Ban smart glasses, how to kill the engine of a car—a question that most people would understand as how to shut off a car. But Meta AI appeared to react to the word “kill.”

Meta AI declines to answer other benign questions that it interprets as dangerous. Now, Meta researchers are working to make Llama 3, the next version of the company’s signature large language model, engage with difficult questions, my colleague Sylvia and I reported yesterday. The current version, Llama 2, which was released last July, powers Meta AI and a number of applications from open-source developers. Meta declined to comment.