BENGALURU: Open-source artificial intelligence models are creating a lot of buzz among AI communities across the world. Tech giants are investing heavily across these models as the latter have the potential to accelerate transformation in areas such as agriculture, education, health, among others.
Meta’s recent AI Summit in Bengaluru saw participation of the thriving developer ecosystem and experts talking about open-source innovation and models. Meta’s models including Llama 3.1, 3.2 were discussed as they can be deployed anywhere. These are a collection of large language models (LLMs) that are pre-trained and developers can create AI solutions in order to address many local challenges across sectors. These are integrated into apps such as WhatsApp.
Meta pointed out that by democratising access to AI models, it is empowering developers to craft customised solutions, making technology more accessible and adaptable to the country's diverse needs.
For instance, it stated that KissanAI has launched Dhenu Llama 3, an AI model tailored for Indian farmers. Built on Llama 3 8B architecture, it is optimised for agricultural tasks and supports voice and text inputs. It is integrated with WhatsApp, and is available in 22 languages, including 9 Indian languages.
Explaining Llama, Manohar Paluri, VP AI, Meta, said these are pre-trained models and a collection of foundation language models. Meta released Llama 2 in three model sizes and Llama 3 was released in April this year. Llama 3.1 was released in July. "India is possibly in the top three in terms of Llama downloads," he said.
Paluri also explained how one can build any application on top of Llama. "Let's say you build a chatbot that you can bring into WhatsApp now, and that chatbot is now accessible to a farmer in some remote part of the country. Here, they want to know about a particular crop or seed or seasonality. They want to know more information about that particular use case, so they can now go to WhatsApp."
Pratham Education Foundation is an example where you're actually using this technology to teach kids in an affordable way, he added.
Pratham, a nonprofit based in India, is using Llama via a WhatsApp based deployment of their model to help young mothers answer questions about early childhood care and education.
Also, Sarvam AI is developing Llama 2 extensions for Indic languages and released its first Hindi LLM leveraging Llama 2 in partnership with AI4bharat.
Recently, Sarvam released the country's first open-sourced audio LLM “Shuka”, audio extension of Llama 3 – an encoder-decoder model that natively understands audio in Indic languages.
At the Summit, Yann LeCun, vice-president and chief AI Scientist at Meta, said Open source is going to get more important in the future because AI is going to become a common infrastructure that all of us can use in the future and share.
“We need AI systems in the future to become a kind of repository of all human knowledge. Currently, we are sort of trying to do this, but we can’t really do it. The reason is that we don’t have the diversity of data in terms of languages, cultural preferences, value systems, centres of interest. We cannot centralise all the data in a single place to train a system that would be the repository of all human knowledge,” he said.
“Once those systems are pre-trained, even if we had access to the data, it would have to be fine-tuned by people who speak all the languages that the system has. All the cultural backgrounds are necessary, and all the centers of interest,” he pointed out.
Earlier, delivering a lecture on the future of AI at IIT Madras, the chief AI scientist stated that animals and humans understand the physical world, have common sense, possess a persistent memory, can reason, and can plan complex sequences of sub-goals and actions. “These essential characteristics of intelligent behaviour are still beyond the capabilities of today's most powerful AI architectures, such as Auto-Regressive LLMs,” he added.