Do not saw off the branch it is sitting on. This is the saying that Mark Zuckerberg must keep in mind as his AI transformation continues at full speed. For Meta, which has just unveiled two new language models from the Llama 3 family, this shift is both simple and perilous. Simple, because unlike many of its competitors, the American giant already has a thriving economic model. Meta is not going to sell pure AI, which is not easy to monetize, it will “augment” all its existing products (social networks, advertising system, etc.) thanks to it. It will thus be able to offer Internet users tools to create increasingly creative texts, images and music. For SMEs, AI capable of creating very effective advertising campaigns in just a few seconds.
This is certainly how we will consume generative artificial intelligence tomorrow: it will no longer be separated from our digital tools, but hybridized with them. The AI transition is, however, also perilous for Meta given the nature of its products: social networks and instant messaging. So many services focused on interaction between humans. Meta has an interest in proceeding with caution when administering its AI cocktail to its billions of users.
The way in which this will transform the face of Meta’s social networks can already be guessed. Instagram is experimenting, for example, with certain famous influencers a virtual avatar tool who maintains the discussion for them with their vast communities of subscribers. 100% artificial influencers will also multiply. We are already seeing fake accounts of young women generated by AI – often in suggestive poses. And artificial publications are starting to spread across platforms. Here, photos of children proudly holding a cake supposed to be their first pastry creation, an image actually generated by AI to collect likes and make an influential account. There, strange collections of surreal images of Jesus, entirely composed of shrimp, plastic bottles or even spaghetti. Publications which sometimes garner tens of millions of views, points out Stanford Internet Observatory, which analyzed 120 pages regularly posting images of this type. If the amount of artificial activity explodes, however, it could drown out that of authentic humans in the multitude. With the risk of tiring, in the long run, users of Meta networks.
Provided that Mark Zuckerberg’s group finds the right mix, this AI cocktail could however give it a new lease of life. The formidable real-time translation capabilities of new AI will transform social platforms into the Tower of Babel, where Japanese, Egyptians, Americans and other French people will not need to learn their respective languages to understand each other.
And AI is a good way to keep Internet users engaged by increasing their ability to find information directly on Facebook, Instagram or WhatsApp. No need, for example, to go out of these applications to look for a cooking recipe or a DIY tutorial. Tomorrow it will be enough to ask the integrated AI assistant to provide us with information on the subject. It is precisely this type of service that Meta AI, the new chatbot launched by the group in the United States and soon in other countries, offers. “With Llama 3 […] the main goal was to make Meta AI the smartest assistant offered for free. Because at the moment, I don’t believe that many people in the world are able to spend a lot to use this type of tool”, underlines Mark Zuckerberg in an interview with The Verge.
Meta has so far only revealed the mini version Llama 3 8B (8 billion parameters) and the medium version Llama 3 70B (70 billion parameters). But in the coming months, the group will present a more powerful version of 400 billion parameters. He must proceed with caution, because if ChatGPT has exceeded 100 million weekly users in record time, Meta has a phenomenal pool of 3 billion users. An AI deployment of this scale has “never been achieved before, not even by OpenAI,” Chris Cox, product manager for the American giant, told L’Express last week.
Controversy over AI censorship
To avoid highly flammable controversies, Meta has also worked hard on its rules of good conduct for AI. An extremely complex dosage. Google had the bitter experience of this in February, when Internet users sparked a lively controversy around images generated by its Gemini AI. This having represented historical actors (American founding fathers, Nazi German soldiers, etc.) by giving them multiple skin colors, some were quick to suspect an aversion on Google to representing white people. The explanation is much more rational: the generic image bases on which AIs are trained are very stereotypical – men are shown more in certain important functions than women, white people are more frequently visible, etc. All AI players are therefore groping to counterbalance this.
Google had therefore experimented with encouraging its AI to spontaneously present people of various skin colors when faced with a generic request. “Gemini […] has not correctly identified cases where this diversity is not relevant. Additionally, the model became too cautious: it refused to respond to certain innocuous queries, interpreting them as sensitive. These two elements led the model to behave inappropriately, thus generating embarrassing and erroneous images. We have therefore made the decision to temporarily suspend the generation of images of people, and we are working on an improved version, in order to offer a reliable and useful functionality, capable of generating precise images that meet the expectations of our users”, confided Joëlle Barral, director of AI research at Google DeepMind, in an exclusive interview given to L’Express at the end of February.
To avoid being accused of excessive caution or censorship, Meta claims to have worked a lot on cases where its AI refuses to respond to requests. “The public, I think, found [Llama 2] a little moralistic, a little reluctant to respond to certain requests. So we’ve worked a lot to reduce what we call ‘unfounded refusals’ to make the models more useful and responsive,” Nick Clegg, president of global affairs at Meta, told Reuters. Financial Times.
The next version of Llama 3 is eagerly awaited by the industry. Meta has made colossal investments in Nvidia’s AI chips, these famous H100s which are hard to find because everyone is snapping them up. While Elon Musk managed to buy 15,000 last year, Zuckerberg set the bar very high at the start of the year, announcing that he had ordered… 350,000! A record quantity which gives good reason to hope for the arrival of much more powerful models.
.