Meta‘s latest open-source AI model, Llama 3.1, is set to significantly aid Indian startups by providing synthetic data generation capabilities, addressing the challenge of limited real-world quality data in local languages.
Ragavan Srinivasan, Meta’s Vice President of Product Management, highlighted that Llama 3.1 serves as a general base model customizable by companies and developers to capture the nuances of languages and cultures, such as Hindi, Kannada, or Marathi.
The new Llama 3.1 model, boasting 405 billion parameters, aims to close the performance gap with top closed-source models like OpenAI’s GPT-4 and Anthropic’s Claude 3.5 Sonnet. Trained on 15 trillion tokens using 16,000 of Nvidia’s H100 GPUs, Llama 3.1 supports synthetic data generation for fine-tuning and understanding linguistic subtleties.
Meta has also updated its commercial license, allowing developers to generate synthetic data from Llama 3.1, which can be used to train other proprietary models. This model distillation capability enables large models to transfer intelligence elements to smaller models without extensive real-world data training.
Experts believe this strategic move by Meta could help customers optimize price and performance needs based on specific applications built on Llama’s foundation. Hemant Mohapatra, a partner at Lightspeed, remarked, “Meta made a strategic bet. You can now achieve the specificity of small language models (SLMs) for narrow use cases while benefiting from Llama’s broad knowledge base.”
Despite the high cost of deploying the 405B model for large-scale enterprise use, Meta anticipates that most organizations will opt for smaller versions, such as the 70B and 8B models, hosted on the cloud. Srinivasan expects greater efficiency and optimization in the coming months, a typical progression for open-source software.
Llama’s unique model distillation capability is a standout feature, difficult to find among proprietary closed vendors. Llama’s various versions have recorded 300 million downloads globally, with India ranking among the top 3-4 markets for Meta models.
While Meta AI assistant for consumer use is gaining traction, rivalling the popularity of ChatGPT, experts caution that surpassing ChatGPT‘s usage is a challenging feat. Arun Chandrasekaran, a distinguished vice president and analyst at Gartner, noted, “ChatGPT continues to be the most popular conversational AI chatbot due to its higher accuracy, multi-modality, and extensibility.” He added that while Meta AI’s usage might grow within apps like Instagram, WhatsApp, and Facebook, OpenAI maintains a significant mindshare and an impressive paid customer base as a versatile, general-purpose chatbot.
