consultantsreview logo

Consultants Review Magazine

Llama to Support Indian Startups Working on Indic Language Models

By Consultants Review Team Thursday, 25 July 2024

The biggest open-source artificial intelligence model from Meta, Llama 3.1, has the ability to generate synthetic data. This will help Indian startups who might be having trouble finding real-world quality data in their native tongues. A top executive stated as much.

Vice president of product management at Meta Ragavan Srinivasan said, "We think of Llama as this general base model, that companies and developers like (Indian AI startup) Sarvam should be able to customize to bring the nuances of language and culture."

A large-scale general-purpose language model finds it difficult to capture such subtlety, but it can produce synthetic data that can be refined to fully comprehend the subtleties of languages like Marathi, Kannada, or Hindi.

With the release of Llama 3.1 405B, the biggest open-source AI model to date, Meta asserts that it has closed the performance gap with leading closed-source models, like Anthropic's Claude 3.5 Sonnet and Open AI's GPT-4o. The most recent version of Llama, which contains 405 billion parameters, was trained using 16,000 of Nvidia's incredibly costly H100 GPUs on a whopping 15 trillion coins.

Additionally, it declared that Meta's commercial license will be updated to let developers to create synthetic data from Llama 405B, which may then be utilized to train and distill other proprietary models. The capacity of a big model to impart some of its wisdom to a smaller model without requiring it to be trained on enormous amounts of real-world data is known as "model distillation."

Experts stated that, based on applications developed on top of Llama's core, Meta's new offering might assist clients in meeting their specific price/performance goals and optimizing use.

"Meta placed a really wise wager...Hemant Mohapatra, a partner at Lightspeed, stated, "You can now make use of the depth of knowledge represented by the Llama LLM while getting the specificity of SLMs (small language models) for specific use cases. "One could see how Meta/Llama could become the central technological layer in a world where there are millions of SLMs."

However, one difficulty is the cost of implementing the 405B model for large-scale corporate applications. Thus, according to Meta's Srinivasan, the majority of organizations will likely adopt the smaller versions, which are housed in the cloud and are 70B and 8B.

In contrast to other closed-source competitors like GPT-4o Mini or Gemini Flash, he stated, "I would expect a lot more efficiency and optimization over the course of the next few months." This is common with software that is open-source.

In addition to being expensive, he said, model distillation is a highly special offering that is difficult to get in closed, proprietary providers.

According to the firm, 300 million copies of all Llama versions have been downloaded worldwide to date. India ranks in the top 3–4 markets for Meta models, according to Srinivasan.

In the meanwhile, ChatGPT's popularity is being challenged by the Meta AI assistant for consumer usage. According to a post by CEO Mark Zuckerburg, Meta AI assistant usage is expected to surpass ChatGPT usage by the end of 2024, only months after its introduction. However, experts believe that's a difficult process.

"Due to its superior accuracy, versatility, and expandability, ChatGPT remains the most widely used conversational AI chatbot," stated Arun Chandrasekaran, senior vice president and analyst at Gartner. "While the use of Meta AI within apps (like Instagram, WhatsApp, and Facebook) may increase, OpenAI remains a general-purpose and versatile chatbot with a sizable mindshare and an impressive paid customer base."

Current Issue