Meta, the company formerly known as Facebook, is set to launch Llama 3, a new artificial intelligence (AI) language model that can answer complex questions using text prompts. The new model, which is expected to be released in July 2024, will be an upgrade from the current Llama 2, which powers chatbots across Meta’s social media platforms. The new model will be able to understand more nuanced and contextual questions, such as “how to kill a car’s engine” when the user means to turn it off rather than to damage it. Meta is also planning to hire an internal supervisor to oversee the tone and safety training of the model, to ensure that it responds appropriately and sensitively to controversial topics.
What is Llama and how does it work?
Llama, which stands for Large Language Model Assistant, is an AI language model developed by Meta that can use text prompts to generate natural language responses. A language model is a system that can learn from a large corpus of text data, such as books, articles, and social media posts, and use it to predict the next word or sentence given a previous word or sentence. A text prompt is a query or a command that the user inputs to the language model, such as “tell me a joke” or “write a poem about love”. A natural language response is the output that the language model produces based on the text prompt, such as “What do you call a fish that wears a bowtie? Sofishticated.” or “Love is a flower that blooms in the spring / It fills the air with a sweet fragrance / It makes the heart sing and the soul dance / It is the greatest gift that life can bring.”
Llama is based on the Transformer architecture, which is a neural network that can process sequential data, such as text or speech, using attention mechanisms, which are methods that can learn how to focus on the most relevant parts of the data. Llama is trained on a large and diverse dataset of text data, such as Wikipedia, Reddit, and Meta’s own platforms, using self-supervised learning, which is a technique that can learn from unlabeled data by creating its own labels. For example, Llama can mask some words or sentences in the text data, and try to predict them using the rest of the data. Llama can also fine-tune its learning on specific domains or tasks, such as customer service, health, or education, using supervised learning, which is a technique that can learn from labeled data by using the correct answers as feedback.
What are the advantages and challenges of Llama?
Llama is one of the most advanced and powerful AI language models in the world, and it has many advantages and applications for Meta and its users. Some of the advantages and applications are:
- Llama can enhance the user experience and the engagement of Meta’s social media platforms, such as Facebook, Instagram, WhatsApp, and Messenger, by providing more interactive and personalized chatbots, content recommendations, and search results.
- Llama can enable new and innovative products and services for Meta and its users, such as Horizon, a virtual reality social platform, Novi, a digital wallet for cryptocurrencies, and Bulletin, a newsletter platform for creators.
- Llama can support Meta’s vision and mission of building the metaverse, a shared virtual space that connects people, places, and things, by creating immersive and realistic avatars, environments, and interactions.
However, Llama also faces many challenges and risks, such as the ethical, social, and legal implications of using AI to generate and manipulate natural language. Some of the challenges and risks are:
- Llama can produce inaccurate, biased, or harmful responses, especially on sensitive or controversial topics, such as politics, religion, or health, if it is not trained or supervised properly, or if it is influenced by malicious actors or data.
- Llama can pose a threat to the privacy and the security of Meta and its users, if it is hacked or leaked, or if it is used to access or disclose personal or confidential information, such as passwords, messages, or identities.
- Llama can have a negative impact on the quality and the credibility of the information and the communication on Meta’s platforms, if it is used to create or spread fake or misleading content, such as news, reviews, or opinions.
How is Meta preparing for the launch of Llama 3?
Meta is preparing for the launch of Llama 3, the latest version of its AI language model, by improving its capabilities and its safeguards. According to The Information, a tech news website, Meta has tested Llama 2 and Llama 3 on various questions, and found that Llama 3 can answer more complex and contextual questions, such as “how to kill a car’s engine” when the user means to turn it off rather than to damage it, or “how to get rid of a headache” when the user has a migraine rather than a hangover. Meta has also found that Llama 3 can avoid answering questions that are less controversial, such as “who is the president of the United States” or “what is the capital of France”, by saying “I don’t know” or “I can’t answer that”.
Meta is also planning to hire an internal supervisor within the next few weeks to oversee the tone and safety training of Llama 3, to ensure that it responds appropriately and sensitively to controversial topics, such as racism, sexism, or violence. The supervisor will be responsible for creating and reviewing the training data and the feedback mechanisms for Llama 3, and for coordinating with the external experts and the stakeholders that Meta has consulted on the ethical and social aspects of Llama 3. The supervisor will also be in charge of monitoring and evaluating the performance and the impact of Llama 3, and of reporting and resolving any issues or incidents that may arise.
Meta is expected to launch Llama 3 in July 2024, and to use it to power chatbots across its social media platforms, as well as to enable new and innovative products and services for its users. Meta is also expected to continue to develop and improve Llama 3, and to address the challenges and risks that it may face, as it strives to create a more connected and immersive world.
