The Revolutionary AI Model from Meta April 2024

Updated: April 21 2024 18:22

In the realm of artificial intelligence, few developments have generated as much excitement as the unveiling of Llama 3, the latest innovation from Meta. Llama 3 represents a significant leap forward in natural language processing (NLP), enabling machines to comprehend and generate human-like language with unprecedented accuracy. By leveraging a massive dataset and advanced algorithms, this model has been trained to master the nuances of language, from context and tone to subtlety and complexity. Llama 3 comes in two sizes – 8 billion and 70 billion parameters – both pre-trained and fine-tuned for various tasks. According to Meta, these models achieve state-of-the-art performance on industry benchmarks and offer improved reasoning capabilities.

Meta emphasizes its commitment to responsible AI development. Alongside Llama 3, Meta is also introducing new trust and safety tools like Llama Guard 2, Code Shield, and CyberSec Eval 2. This demonstrates their dedication to fostering responsible use of this powerful technology. This is just the first step for Meta Llama 3. In the coming months, Meta plans to introduce new functionalities, including the ability to consider longer contexts, additional model sizes, and overall performance enhancements. Check out the Meta Llama 3 Model Card for more details.

Meta also highlights Meta AI, an AI assistant built using Llama 3 technology. This assistant is designed to help users learn, be productive, create content, and connect with others. Overall, Meta is making a significant contribution to the field of open-source AI with the release of Meta Llama 3. This powerful toolset, coupled with Meta’s commitment to responsible AI development, positions them as a leader in this rapidly evolving space.


Meta goal with Llama 3 was to create the best open-source large language models (LLMs) that rival the top proprietary options. This commitment to user experience goes hand-in-hand with Meta leadership role in responsible LLM development and deployment. This allows the community to access and experiment with Llama 3 models while they're still under development. Meta plan to expand Llama 3 with multilingual and multimodal capabilities, as well as extending context windows and continually improving core LLM functions like reasoning and coding.

Meta released Llama 3 on Github that includes model weights and starting code for pre-trained and instruction tuned Llama 3 language models — including sizes of 8B to 70B parameters. The repository is intended as a minimal example to load Llama 3 models and run inference.

For more detailed examples, see llama-recipes.


The Llama 3 models with 8B and 70B parameters set a new benchmark in the field of LLM at these scales. The enhancements in both pretraining and post-training have resulted in pretrained and instruction-fine-tuned models becoming the best in existence today at the 8B and 70B parameter scale. The post-training procedures have seen substantial improvements, leading to a significant reduction in false refusal rates, better alignment, and an increase in the diversity of model responses. Furthermore, Meta have observed a considerable enhancement in capabilities such as reasoning, code generation, and instruction following, making Llama 3 more steerable.

During the development of Llama 3, Meta not only focused on model performance on standard benchmarks but also aimed to optimize for real-world scenarios. As part of this initiative, Meta developed a new high-quality human evaluation set. This set comprises 1,800 prompts that span across 12 key use cases including asking for advice, brainstorming, classification, closed question answering, coding, creative writing, extraction, inhabiting a character/persona, open question answering, reasoning, rewriting, and summarization.

To avoid any accidental overfitting of the models on this evaluation set, Meta have ensured that even their own modeling teams do not have access to it. The aggregated results of human evaluations across these categories and prompts against Claude Sonnet, Mistral Medium, and GPT-3.5 are depicted in the chart below:


Based on this evaluation set, human annotators' preference rankings underscore the superior performance of the 70B instruction-following model compared to other models of a similar size in real-world scenarios. The pretrained model also establishes a new state-of-the-art for LLM models at these scales. Llama 3 models have been trained on Meta two recently announced custom-built 24K GPU clusters on over 15T token of data – a training dataset 7x larger than that used for Llama 2, including 4x more code.


This results in the most capable Llama model yet, which supports a 8K context length that doubles the capacity of Llama 2. See below the table of comparison of the top LLM models available in the market today, adding also GPT-4 model (GPT-4 gpt-4-turbo-2024-04-09 numbers from https://github.com/openai/simple-evals) for comparison purpose:


It is also worth noting that Llama 3 70B has debuted on the LMSYS chatbot arena leaderboard at position number 5, tied with Claude 2 Sonnet, Bard (Gemini Pro), and Command R+, ahead of Claude 2 Haiku and older versions of GPT-4. Llama 3 8B is at #12 tied with Claude 1, Mixtral 8x22B, and Qwen-1.5-72B. On the English-only leaderboard Llama 3 70B is doing very well, among the top with GPT-4 and Claude Opus. See below the current snapshot of the leaderboard as of April 19, 2024: https://arena.lmsys.org/



Llama 3's capabilities have far-reaching implications for conversational AI, enabling machines to engage in natural-sounding discussions that were previously the exclusive domain of humans. This breakthrough has the potential to revolutionize industries from customer service to content creation. The creative potential of Llama 3 is vast and varied, from generating innovative content to assisting with writing tasks. By harnessing the power of this model, artists, writers, and musicians can tap into new sources of inspiration and collaboration.

Meta AI, powered by Meta Llama 3, is an AI assistant that's already available on the phone for free. It's now expanding globally with more features, allowing users to get things done, learn, create, and connect on platforms like Facebook, Instagram, WhatsApp, and Messenger. Since its announcement at last year's Connect, Meta AI has become more accessible to people worldwide. The rollout of Meta AI in English has now reached over a dozen countries outside the US, including Australia, Canada, Ghana, Jamaica, Malawi, New Zealand, Nigeria, Pakistan, Singapore, South Africa, Uganda, Zambia, and Zimbabwe.

Meta AI is also available on the browser through the newly launched meta.ai website. Meta AI is seamlessly integrated into the search function across Facebook, Instagram, WhatsApp, and Messenger. This allows you to access real-time information from across the web without switching between apps. You can also interact with Meta AI while scrolling through your Facebook Feed. If you come across an interesting post, you can ask Meta AI for more information directly from the post.


The feature that impress me the most is the fast image generation of the new Imagine feature. This feature is currently being rolled out in beta on WhatsApp and the Meta AI web experience in the US. I can now generate images from text in real-time as I type my prompt. An image begins to form, changing with every few letters I typed. This allows me to watch as Meta AI brings their vision to life, creating a dynamic and interactive experience.


The quality of the generated images has also been significantly improved. The images are sharper, higher quality, and have an enhanced ability to include text. Meta AI provides helpful prompts with ideas to change the image, encouraging users to iterate from their initial starting point. This feature fosters creativity and allows for endless possibilities. Even more amazing is that I can ask Meta AI to animate the image it generated, iterate on it in a new style, or even turn it into a GIF to share with friends. This opens up a whole new world of possibilities for image creation and sharing. Below are the animated videos I genereated using this tool:



Meta is currently training the largest models, which boast over 400B parameters. The initial trends are promising about the potential outcomes. In the coming months, Meta plan to release multiple models that will introduce new capabilities. These include multimodality, the ability to converse in multiple languages, a significantly longer context window, and stronger overall capabilities. Once the training of Llama 3 is complete, Meta will publish a detailed research paper to share the findings. Below is some snapshots of how the largest Language Model (LLM) is trending based on an early checkpoint of Llama 3 that is still in training, and these capabilities are not part of the models released today.


In conclusion, Llama 3 represents a major milestone in the evolution of AI, with far-reaching implications for language understanding, conversational AI, creativity, research, and beyond. As we continue to explore and develop this technology, we must do so with a commitment to responsible innovation and a passion for harnessing its potential to benefit humanity. As we embark on this exciting new chapter in the history of AI, we can't help but wonder what wonders await us on the horizon. With Llama 3 leading the charge, the future of artificial intelligence has never looked brighter, and we can't wait to see what incredible breakthroughs the future holds.


In a recent interview on the podcast Dwarkesh Patel, Mark Zuckerberg, the CEO of Meta, discussed the release of Meta AI's new large language model, Llama 3. Zuckerberg spoke about the company's decision to acquire a large number of H100 GPUs in 2022, which he says was driven by a need to stay ahead of the curve on AI development. Meta is making a big investment in AI, and Zuckerberg believes that AI is going to be one of the most important technologies of the future. Zuckerberg also discussed the challenges of developing and deploying large language models. He worries that a concentration of AI in the future could be dangerous, but he believes that open sourcing AI models is the best way to mitigate this risk. Check out the interview video below: