- April 19, 2024
- AI Projects
Meta has released Llama 3, the latest update to its cutting-edge, freely available Large Language Model (LLM). This new model is touted as having outperformed existing best-in-class models such as the GPT-3.5 in practical applications, setting a new standard for performance and capability.
“When launching Llama 3, our goal was to create top-tier open models that would compete with today’s proprietary options,” Meta shared in the launch statement. The first versions of Llama 3 that have become available are equipped with 8 billion and 70 billion parameters. Meta indicates that it is in the process of developing even larger models with more than 400 billion parameters, which are expected to be published soon, along with detailed research results.
Development of Llama 3 took over two years, including significant dedication to collecting quality training data, improving distributed learning capabilities, refining the model architecture, and pioneering methods for fine-tuning instruction. The 70 billion Meta parameter model, refined for specific instructions, has shown superior performance compared to GPT-3.5, Claude, and other LLMs of similar scale in human judgment-based evaluations in various domains such as coding, logic, and creative composition. Similarly, the previous Meta model with 8 billion parameters also set new standards in general LLM assessment exercises.
“We strongly believe that these models are unmatched in their category,” said Meta. With an “open by default” policy, the tech giant aims to foster an AI collaborative development ecosystem. Llama 3 will be shared between major cloud service providers, simulation platforms, hardware manufacturers, and AI services.
Victor Botev of Iris.ai considers the launch of Meta Llama 3 a significant step against the background of the global tilt towards AI rules. “Meta’s choice to open source Llama 3 resonates with a broader call for responsible and ethical AI practices,” he commented. This step not only advocates for transparent development but also promotes joint learning and testing of different methodologies, contributing to the creation and application of regulations.
Meta also introduced an expanded collection of tools designed for AI security, including updated versions of Llama Guard for risk assessment, CyberSec Eval for detecting potential exploits, and a new Code Shield feature for eliminating dangerous code suggestions in real time. “However, the very fact of open source code does not automatically mean ethical responsibility in AI,” Botev added. “A multi-pronged strategy is essential to address issues such as data privacy, algorithm bias, and societal consequences that underpin future AI regulations around the world.”
The effectiveness of initiatives such as Llama 3 in encouraging oversight and collaborative efforts towards responsible AI depends on full ethical compliance throughout the AI development lifecycle and compliance with regulatory requirements. Meta continues to move in the right direction with the Llama series, but ethical AI requires continued commitment from all parties involved. Meta emphasizes a comprehensive strategy to ethically build and deploy Llama 3, despite extensive security testing. It advises developers to apply custom I/O filtering based on the specific needs of their applications.
With the integration of Llama 3, Meta AI is now the world’s leading artificial intelligence assistant, accessible via Facebook, Instagram, WhatsApp, Messenger, and the web, enhancing productivity, learning, creativity, and more. Meta is also moving toward introducing multimodal vision-enabled versions of Meta AI, starting with a preview of Ray-Ban’s Meta smart glasses.
While Llama 3’s achievements are notable, some skeptics question Meta’s motives for open source. However, the introduction of Meta Llama 3, especially after the recent Mistral AI milestone, further raises the bar for affordable LLMs in the AI landscape.
