{"id":13878,"date":"2025-10-20T15:56:54","date_gmt":"2025-10-20T15:56:54","guid":{"rendered":"https:\/\/www.yiaho.com\/llama-the-origins-and-history-of-metas-artificial-intelligence\/"},"modified":"2025-10-20T15:56:54","modified_gmt":"2025-10-20T15:56:54","slug":"llama-the-origins-and-history-of-metas-artificial-intelligence","status":"publish","type":"post","link":"https:\/\/www.yiaho.com\/en\/llama-the-origins-and-history-of-metas-artificial-intelligence\/","title":{"rendered":"Llama: The Origins and History of Meta’s Artificial Intelligence"},"content":{"rendered":"

Tech giants have been vying for AI supremacy for years, and Meta, the company formerly known as Facebook, is well-positioned with Llama. It’s not just a language model, but a family of open-source tools that democratizes AI, making a technology that clearly has a future accessible! <\/p>\n

Launched in 2023, Llama embodies Meta’s vision: to transform AI into a common good, driven by the community rather than confined behind proprietary walls.<\/p>\n

But where did this bold creation come from? What steps marked its evolution? And what are Meta’s future ambitions for taming the future of artificial intelligence? The Yiaho team delved into the history of Meta’s AI, which brings together technical innovation, controversies, and a shared thirst for progress. <\/p>\n

Llama’s Origins: From FAIR to the Quest for Open AI<\/h2>\n

To understand Llama, we must go back to Meta’s AI roots.<\/p>\n

It all began in 2013 with the creation of FAIR (Facebook AI Research), rebranded Meta AI in 2021. Led by Frenchman Yann LeCun<\/a>, a pioneer of deep learning and Turing Award laureate, this division aims to explore the frontiers of artificial intelligence to connect Meta’s billions of users. <\/p>\n

Originally, FAIR focused on advancements like image recognition and machine translation, but the explosion of large language models (LLMs)<\/a> changed the game.<\/p>\n

In 2022, the AI landscape was dominated by closed behemoths like OpenAI’s GPT-3<\/a> or Google’s PaLM. These models, though powerful, are opaque: their weights (internal parameters) are not accessible, limiting innovation to a restricted circle. <\/p>\n

Meta, aware of this bottleneck, decided to counter-attack.<\/p>\n

Inspired by earlier open-source initiatives like OPT (a 175-billion-parameter model launched by Meta in May 2022), the FAIR team saw AI as an opportunity to “help billions of people at scale<\/em>.” It is in this context that Llama was born: a direct response to the meteoric success of ChatGPT<\/a>, released in late 2022, which revealed the potential of LLMs for creative text generation, solving mathematical problems, or even predicting protein structures. <\/p>\n

Yann LeCun himself tempers the enthusiasm for pure LLMs, stating that they excel primarily in writing assistance, but he pushes for a collaborative approach. Llama thus emerges as a “foundational” project: a base model optimized for research, trained on massive public data (without precise details provided by Meta for confidentiality reasons). <\/p>\n

Its name? A playful nod to the llama, that peaceful and resilient animal, perhaps symbolizing the robustness and accessibility of AI? <\/p>\n

The Creation of Llama 1: A Discreet Launch That Shakes the Industry<\/h2>\n

On February 24, 2023, Meta unveiled Llama 1 (Large Language Model Meta AI). It wasn’t a war cry, but a calculated whisper: the models were initially reserved for qualified researchers, via an approval process. <\/p>\n

Four sizes were offered: 7, 13, 30, and 65 billion parameters, all trained on a corpus of public data, with an efficient architecture that rivaled closed giants while consuming fewer resources.<\/p>\n

The surprise was complete: despite its limited distribution, Llama 1 quickly leaked onto forums like 4chan, and the developer community seized the model’s weights. Variants emerged within weeks, such as Alpaca or Vicuna, refined to approach ChatGPT’s performance on a simple laptop. <\/p>\n

Meta had aimed for research excellence, but Llama 1 became a catalyst for decentralized innovation.<\/p>\n

But the launch was not without its bumps: critics pointed to the absence of “instruction-tuned” versions (optimized for dialogues), and lawsuits were already looming, accusing Meta of having used copyrighted books in training (as in the Kadrey v. Meta case, which was dismissed).<\/p>\n

This first installment laid the groundwork: Llama is not just a product, it’s an ecosystem. It demonstrates that open-source models can scale efficiently, with benchmarks<\/a> showing superior performance to competitors like BLOOM or MPT on text comprehension tasks. <\/p>\n

Rapid Evolution: From Llama 2 to Llama 4<\/h2>\n

Meta quickly learned from this chaotic launch. On July 18, 2023, Llama 2 was released, in a surprise partnership with Microsoft. This time, it’s fully open-source: the weights are public, commercially usable (except for very large companies without an agreement), and accompanied by optimized chat versions. Trained on 40% more data than its predecessor (2 trillion tokens), Llama 2 offers three sizes: 7, 13, and 70 billion parameters. <\/p>\n

Its architecture remains faithful to Llama 1, but incorporates safeguards against malicious uses via an acceptable use policy.<\/p>\n

The impact was immediate! Google Cloud and AWS reported over 3,500 enterprise projects based on Llama 2. The Hugging Face community published over 7,000 derivatives, improving benchmarks by nearly 10% on average. Specialized variants emerged, such as Code Llama (August 2023, for code generation in 7B, 13B, 34B, and 70B) and multilingual models. <\/p>\n

Llama 2 propels Meta to the rank of open-source leader, with hardware optimizations by AMD, Intel, and Nvidia. 2024 accelerates the pace. In a few months, we find: <\/p>\n