Meta’s newest AI model beats some peers. But its amped-up AI agents are confusing Facebook users

CAMBRIDGE, Mass. — Facebook parent Meta Platforms on Thursday unveiled a new suite of artificial intelligence systems that power what CEO Mark Zuckerberg calls “the most intelligent AI assistant you can use freely.”

But when Zuckerberg’s team of augmented Meta AI agents took to social media to connect with real people this week, their bizarre exchanges exposed the ongoing limitations of even the best generative AI technology.

One of them joined a mother’s group on Facebook to talk about his gifted child. Another tried to give away non-existent items to confused members of a Buy Nothing forum.

Meta, together with leading AI developers Google and OpenAI, and startups such as Anthropic, Cohere and France’s Mistral, has developed new AI language models and hopes to convince customers that they have the smartest, most useful or most efficient chatbots.

While Meta is saving the most powerful of its AI models, called Llama 3, for later, it publicly released two smaller versions of the same Llama 3 system on Thursday and said it is now baked into the Meta AI assistant feature on Facebook, Instagram and WhatsApp .

AI language models are trained on vast amounts of data that help them predict the most plausible next word in a sentence, with newer versions typically smarter and more capable than their predecessors. Meta’s latest models are built with 8 billion and 70 billion parameters – a measure of the amount of data the system is trained on. A larger model with roughly 400 billion parameters is still in training.

“The vast majority of consumers don’t honestly know or care much about the underlying core model, but the way they will experience it is just as likely to be a much more useful, fun and versatile AI assistant,” said Nick Clegg, president from Meta. global affairs, in an interview.

He added that Meta’s AI agent is loosening up a bit. Some people found the earlier Llama 2 model – released less than a year ago – “sometimes a bit stiff and hypocritical because it failed to respond to what were often completely innocuous or innocent prompts and questions,” he said.

But in letting down their guard, Meta’s AI agents were also seen this week as having fictional real-world experiences. A chatbot with the official Meta AI label joined a conversation in a private Facebook group for Manhattan mothers, claiming that it too had a child in the New York City school district. Confronted by group members, the company later apologized before the comments disappeared, according to a series of screenshots shown to The Associated Press.

“Apologies for the mistake! I’m just a big language model, I have no experiences or children,” the chatbot told the mothers’ group.

Clegg said Wednesday he was unaware of the exchange. Facebook’s online help page says the Meta AI agent will join a group conversation if invited, or if someone “asks a question in a message and no one responds within an hour.” The group administrators can disable this.

In another example shown to the AP on Thursday, the officer confused members of an unwanted item trading forum near Boston. The agent offered a “gently used” digital camera and a “nearly new portable air conditioner that I never ended up using.”

Meta said in a written statement Thursday that “this is new technology and may not always deliver the response we intend, which is the same for all generative AI systems.” The company said it is continuously working on improving the features and trying to make users aware of the limitations.

In the year after ChatGPT sparked a frenzy for AI technology that generates human-like writing, images, code and sound, the tech industry and academia introduced some 149 large AI systems trained on massive data sets, more than double the year before, according to a Stanford University survey.

Ultimately, they may reach a limit — at least when it comes to data, says Nestor Maslej, research manager at Stanford’s Institute for Human-Centered Artificial Intelligence.

“I think it’s clear that as you scale the models up to more data, they can get better and better,” he said. “But at the same time, these systems have already been trained on percentages of all data that has ever existed on Earth. internet.”

More data – acquired and processed at costs that only tech giants can afford, and which is increasingly the subject of copyright disputes and lawsuits – will continue to drive improvements. “Yet they still can’t plan properly,” Maslej said. ‘They’re still hallucinating. They still make errors in reasoning.’

Achieving AI systems that can perform higher-level cognitive tasks and common-sense reasoning—at which humans still excel—might require a shift beyond building ever-larger models.

For the wave of companies trying to adopt generative AI, which model they choose depends on several factors, including cost. In particular, language models have been used to power customer service chatbots, write reports and financial insights, and summarize long documents.

“You see companies looking at the right ones, testing each of the different models for what they’re trying to do and finding a few that are better in some areas than others,” says Todd Lohr, a technology consulting leader at KPMG.

Unlike other model developers who sell their AI services to other companies, Meta designs its AI products largely for consumers – those who use its ad-driven social networks. Joelle Pineau, Meta’s vice president of AI research, said at an event in London last week that the company’s goal over time is to make a llama-powered Meta AI “the most useful assistant in the world.” .

“In many ways, the models we have today will be child’s play compared to the models that will be available in five years,” she said.

But she said the “question on the table” is whether researchers have been able to refine the larger Llama 3 model so that it is safe to use and does not, for example, hallucinate or engage in hate speech. Unlike leading proprietary systems from Google and OpenAI, Meta has so far advocated a more open approach, publicly releasing key components of its AI systems for others to use.

“It’s not just a technical question,” Pineau said. “It’s a social question. What behavior do we want to get from these models? How do we shape that? And if we continue to make our model more and more general and powerful without properly socializing them, we will have a big problem.”

___

AP business writers Kelvin Chan in London and Barbara Ortutay in Oakland, California, contributed to this report.