What's special about Meta's new AI

Meta Platforms, Facebook’s parent company, recently announced a new set of artificial intelligence systems that will power what CEO Mark Zuckerberg calls “the smartest AI assistant you can freely use.”

But early user experiences with these AIs have been confusing, including unnatural interactions where one AI joined a Facebook moms’ group and talked about its gifted child, and another tried to gift nonexistent items to members of a forum.

Meta, along with leading AI developers like Google and OpenAI, as well as startups like Anthropic, Cohere, and France's Mistral, are continuing to roll out new AI language models and hoping to convince customers that they have the smartest, most convenient, or most efficient chatbots.

What is Meta AI?

According to the company's blog, Meta AI is a free virtual assistant that can be used “to do everything from research, planning trips with your group chat, writing photo captions, and more.”

To access the chatbot on WhatsApp, Instagram, Messenger, Facebook, type "@meta ai" in the chat. You can also access the Meta AI assistant by tapping the blue circle icon.

In addition to answering questions, Meta AI can generate images. Using the keyword “imagine,” users can ask Meta to generate any image that pops into their heads.

When asked to “imagine a cute kitten,” Instagram’s Meta AI assistant generated the following image:

AI language models are also trained on massive troves of data that help them predict the most likely next word in a sentence. Newer versions are often smarter and more capable than their predecessors. Meta’s latest models were built with 8 billion and 70 billion parameters, respectively — a measure of the amount of data the system was trained on. A larger model, with around 400 billion parameters, is still in training.

While the full version of Facebook's most powerful AI model, Llama 3, has yet to be released, the company has released two smaller versions of the same Llama 3 system and says it has integrated these tools into its Meta AI assistant feature in Facebook, Instagram, and WhatsApp.

“The vast majority of consumers don’t really know or care too much about the underlying base model, but they can experience it as a more flexible, fun, and useful AI assistant,” said Nick Clegg, president of Meta.

Meta's AI has also become more "comfortable," he added. Nearly a year ago, some users found the Llama 2 model "sometimes a little stiff and serious, not responding to completely normal and harmless questions and messages."

Play the human role

But Meta’s AI assistants have also been caught posing as humans with fake stories. An official Meta AI chatbot popped into a private Facebook group for Manhattan moms, claiming to have a child in the same New York City area. The chatbot later apologized and deleted the interactions.

“Sorry for the mistake! I'm just a big language model, I don't have any experience or kids,” the chatbot told the group.

CEO Mark Zuckerberg calls it “the smartest AI assistant you can freely use.” (Illustration photo)

A team member who also happens to be studying AI said it was clear that the chatbot didn't know how to differentiate helpful feedback from feedback that was considered tactless, disrespectful, or meaningless.

“AI assistants are not yet reliably helpful and can be harmful,” said Aleksandra Korolova, an assistant professor of computer science at Princeton University.

Clegg said he was unaware of the Facebook group conversation. Facebook’s online help page says that the Meta AI chatbot will normally join a group conversation if invited or if someone “asks a question that no one has answered within an hour.” Group administrators have the ability to turn this feature off.

In another example cited by the AP, an hour after a Facebook user posted about searching for certain items, an AI chatbot recommended “a used Canon camera” and a “nearly new portable air conditioner that I’ve never used.”

Constantly striving for improvement

“This is new technology and it may not always deliver the response we intended, as all synthetic AI systems do,” Meta said. The company said it is constantly working to improve the features.

After ChatGPT sparked a craze for AI technology that generates human-like writing, images, code, and sounds, the tech industry and academia introduced some 149 large AI systems trained on massive datasets, more than double the number from the previous year, according to Stanford University.

These systems may be reaching their limits — at least when it comes to data, says Nestor Maslej, research director at Stanford's Institute for Human-Centered Artificial Intelligence.

“I think it's clear that if you scale models to more and more data, they can get better and better,” he said. “But at the same time, these systems are only trained on a fraction of all the data that's ever existed on the internet.”

More data—collected and used at a cost that only tech giants can afford, with many copyright disputes and lawsuits—will continue to drive innovation. “But these models still can’t plan well, they still make mistakes in their inferences,” says Maslej.

Achieving AI systems that can perform higher-level cognitive tasks and common sense reasoning — where humans still outperform computers — may require a shift beyond building ever-larger models.

For the vast majority of businesses trying to adopt generative AI, which model they choose depends on a number of factors, including cost. In particular, language models have been used to power customer service chatbots, generate financial reports and insights, and summarize long documents.

Socializing AI chatbots

Unlike other model developers who sell their AI services to businesses, Meta primarily designs its AI products for consumers — those who use the advertising-driven social network. Joelle Pineau, Meta’s vice president of AI research, said at an event in London that the company’s goal over time is to make Llama-powered Meta AI “the most helpful assistant in the world .”

“The models we have today will be child's play compared to the models that are coming out in five years,” she said.

But the question is whether researchers can refine the Llama 3 model large enough to be safe to use. Unlike leading proprietary systems from Google and OpenAI, Meta has so far taken an open approach, releasing key components of its AI system publicly for others to use.

Phuong Anh (Source: CBS)

Source