Facebook-owner Meta has taken a significant leap in the AI domain by open-sourcing its cutting-edge language model, LLaMA 2.
This move brings a game-changing competitor to the field, directly challenging OpenAI’s GPT-4, which has been instrumental in powering tools like ChatGPT and Microsoft Bing.
During Microsoft’s Inspire event, Meta announced this pivotal decision, highlighting its strong alliance with Microsoft and expressing support for Azure and Windows platforms. As a result of this collaboration, Meta also revealed more details about the AI tools integrated into Microsoft’s 360 platform, along with their associated costs. Furthermore, in a groundbreaking partnership, Meta is teaming up with Qualcomm to bring LLaMa 2 to laptops, phones, and headsets by 2024, enabling AI-powered applications that function independently of cloud services.
According to Meta, the release of LLaMA 2 as an open-source model aims to empower businesses, startups, and researchers with greater access to AI tools, promoting a spirit of experimentation and collaboration within the community. LLaMA 2 boasts an impressive improvement over its predecessor, LLaMA 1, having been trained on 40 percent more data, including information sourced from publicly available online data sources. Meta asserts that LLaMA 2 stands out as a leader in reasoning, coding, proficiency, and knowledge tests when compared to other language models like Falcon and MPT.
The decision to open-source LLaMA also aligns with Meta’s commitment to enhancing safety and transparency in AI development. The company subjected the LLaMA 2 model to rigorous safety testing through red-teaming, a process involving the generation of adversarial prompts for fine-tuning the model’s behavior. This process was carried out both internally and externally. Meta is dedicated to disclosing its evaluation and tweaking procedures, fostering trust and accountability in AI research. To access LLaMA 2, developers can utilise Microsoft’s Azure platform. Additionally, Meta has plans to make LLaMA available through AWS, Hugging Face, and other providers.
In an official statement, Meta emphasised the value of an open approach in the development of AI models, particularly in the generative space where technology is rapidly evolving. By opening access to today’s AI models, Meta aims to empower developers and researchers to stress test them collectively, enabling faster identification and resolution of challenges.
While Meta initially introduced its LLaMA model in February, it faced a leak on 4chan shortly after the announcement. Now, with the release of LLaMA 2 under an open-source license, we can expect a surge of new AI-powered tools built upon this powerful model. Meta had received over 100,000 requests from researchers to use its first model, but the open-source LLaMA 2 is poised to reach an even wider audience, revolutionising the AI landscape.
