Artificial intelligence will take over the networks: Meta launches its model on WhatsApp, Facebook and Instagram | Technologies

This Thursday, Meta announced its great success in the competitive artificial intelligence (AI) sector. Users of all of its apps, some of the most popular on the planet, will have an assistant to whom they can ask questions, ask for advice or ask for drawings. WhatsApp, Instagram, Facebook and Messenger will actively encourage you to ask Meta AI questions. Meta will also create a separate app and page to consult with your assistant for free, just like you can now do with ChatGPT or Google’s Gemini.

The movement is Meta’s big leap in artificial intelligence. This Thursday the company also presented Llama 3, a new version of its model that will delight all visitors. Meta’s decision will popularize access to AI in an unusual way. Its applications are used by more than 3 billion people around the world. While the success of OpenAI’s ChatGPT has been overwhelming, this new level of access is significant. 11% of Spaniards say they use ChatGPT at least once a day.

The assistant will not only take questions and requests, but also create illustrations and animate them in real time to create something similar to a GIF file, as shown in the company’s demos. It will work similarly to others already on the market, with the notable novelty that image creation is instantaneous: while the user writes, the machine adds or removes details from the drawing.

The launch of Meta IA will for now be available only in English and in 14 countries such as the USA, Canada, Australia, Nigeria, Pakistan or South Africa, among others. At the moment, not a single European or Latin American country. When contacted by EL PAÍS, the company states that it has “nothing to share” about the arrival of this news in Spanish or in Spanish-speaking countries. In his presentation message, Mehta explains that “the models we ran today were refined to produce results in English only.”

The company, however, already warns that its results in other languages ​​will be worse: “To prepare for the next multilingual use cases, more than 5% of the Llama 3 pre-training dataset consists of high-quality data in languages. in addition to English, covering more than 30 languages. However, we do not expect the same level of performance in these languages ​​as in English.”

It’s difficult to predict how much and how people will use these in-app virtual assistants. In September, Meta launched 28 artificial intelligence chatbots in the US with profiles such as Kendall Jenner, Paris Hilton or rapper Snoop Dogg. They were not successful. Snoop Dogg has only 14,000 followers, but his Instagram account has 86 million. But the emergence of artificial intelligence can provide new opportunities for users.

And like the rest of the previous releases in this field, Llama 3 will not be exempt from the bugs and hallucinations that plague its competitors. There are those who have already found a black Italian fascist soldier, as happened a few weeks ago when Google’s AI drew several black Nazis. Given that these tools will be available on a mass scale, across all networks and on a planetary scale, we are already warned about the problem of misinformation and infotoxication that they may ultimately generate.

Another new feature of the virtual assistant is that it can insert itself into group conversations, presumably to encourage them. This week, a Princeton University professor posted a screenshot of a private parent Facebook group where Meta AI says her child is in a New York City public school and shares her child’s experiences with teachers.

The meta algorithm puts it in first place. Facebook’s help page says Meta AI will join a group conversation regardless of whether it’s invited or if someone “asks a question in a message and no one responds within an hour.” This makes it clear that beyond the user experience of the bot, the clear goal is to attract attention and increase user interaction with each application.

Meta has already had to apologize at least a couple of times for his bot’s comments or replies. In both cases, the company gave an excuse related to the known errors of these bots: “This is new technology and will not always return the expected response, which is the same thing that happens with all generative artificial intelligence systems,” the company said. to the AP agency.

These mistakes and outbreaks will be inevitable. Meta has refined its model to make it less shy about answering questions on sensitive topics. “If someone asks about a controversial political issue, our goal is for Meta AI not to offer a single opinion, but to summarize relevant views on the topic,” the company says. But, he continues, “if someone asks for a perspective on a problem, we want to respect that person’s intent and give Meta AI an answer to the specific question.”

In addition to expanding AI’s reach to hundreds of millions of people, Meta is also trying to break into the OpenAI, Google, and Microsoft fields with its free help site Meta.ai. According to Zuckerberg, the company continues to train an improved version of Llama. If it remains free, it could give Meta a competitive advantage over paid versions of OpenAI (ChatGPT 4) and possibly Google (which will be exploring the possibility of launching a premium search engine).

At the presentation of this launch on his Instagram account, Mark Zuckerberg, executive president of Meta, appeared with an appearance and a network that was much commented on on the networks. Zuckerberg said Llama 3 will be “the smartest AI assistant that’s free to use.” All major corporations and startups representatives of the sector (French Mistral, Anthropic or Perplexity) try to convince users that their assistants or models are the ones who best and most quickly solve any problem. At this point, it remains difficult to distinguish the unique traits of each participant, and there is no standardized system to allow for rigorous comparisons.

The company didn’t want to openly explain what type of content it trained its model on. The lack of new material for these programs is one of the biggest debates in the sector, and Meta has admitted that it used AI-generated language (also called “synthetic data”) for training. So far, experiments with non-human language to expand the data have not been successful: “We also use synthetic data for training in areas such as programming, reasoning and large contexts. For example, we use synthetic data to create longer documents for training,” explains Meta’s presentation note.

You can follow El Pais Technology V Facebook And X or register here to receive our weekly newsletter.