Industry stakeholders believe that such an approach may help Meta effectively rival OpenAI’s generative pre-trained transformer (GPT) AI models, widely considered the top foundational AI platform globally.
In an interview with Mint, Ragavan Srinivasan, vice-president of product and the head of its Llama models at Meta, said that the company’s open-source approach to AI will define its approach to generative AI for enterprises, going forward.
“Building AI models as open-source is good for developers, Meta and the world in general, as we’ve seen with our work on Pytorch and React. Rivals of the Llama foundational model are all proprietary, which means that enterprises have to develop their AI applications based on the terms defined by the companies making the models. As generative AI becomes integrated into the internet, we’d look to provide open intelligence in order to ensure that a rich developer and partner ecosystem builds around Llama,” Srinivasan said.
Srinivasan’s statement on the latest models, which natively support processing of queries and inferences in Hindi (among other languages), echoes Tuesday’s post by Zuckerberg on expanding its open-source approach. “We’re releasing Llama 3.1 405B, the first frontier-level open-source AI model, as well as new and improved Llama 3.1 70B and 8B models. In addition to having significantly better cost and performance relative to closed models, the fact that the 405B model is open will make it the best choice for fine-tuning and distilling smaller models,” the executive said in a blog post.
Industry stakeholders largely welcomed the move. Kashyap Kompella, chief executive of AI research and consultancy firm RPA2AI Research, said that Meta’s push for AI is multi-pronged.
“Meta is looking to be the open-source evangelist in the generative AI space, which can help it rival companies such as OpenAI, Microsoft, Google and Anthropic in offering one of the top five foundational AI models globally. Meta’s key strength here is that it can use its own models within its core social media services, which means that it does not necessarily need to monetize the models themselves. This is something that OpenAI cannot—since their core offering are the AI models, they will need to look at a direct monetization model instead,” he said.
Localized appeal
The local language support, meanwhile, is key to expanding localized appeal for its models. Srinivasan said that its current approach allows the Llama 3.1 models to be used as “plug and play” models for companies such as Sarvam, and compatriots such as CoRover etc—which are focused on building ‘Indianized’ generative AI applications.
“Enterprises can take our model, create its own entire training dataset based on their own proprietary data, and use it for themselves—thereby expanding the usability of our foundational model,” he added.
On 18 June, Mint reported a rising tug of war among Silicon Valley’s Big Tech majors to grab developer mind-share, which in turn could see a specific approach to generative AI thrive over others. While Apple isn’t currently building its own foundational model, the likes of Microsoft, OpenAI and Google are all looking to lure in developers by making their AI models as wide as possible in terms of support, access, pricing and relevance—which could be key especially for a large applications market such as India.
Zuckerberg and Srinivasan’s statements highlight an intention to develop the foundational Meta AI model “to become the standard for the long term”—an approach that Kompella said could help the company rival OpenAI’s dominance.
“Many organizations don’t want to depend on models they cannot run and control themselves. They don’t want closed model providers to be able to change their model, alter their terms of use, or even stop serving them entirely. They also don’t want to get locked into a single cloud that has exclusive rights to a model. Open source enables a broad ecosystem of companies with compatible toolchains that you can move between easily,” Zuckerberg’s post said.
Underlining this further, Ragavan said that a key strategy behind making Llama open-source is to expand developer support “without Meta having to line a single line of code.”
“In hardware optimization, developers can optimize to make it efficient in order to run cheaper versions of the AI models. Another example is for developers to integrate workflows of Nvidia or even AMD GPUs—expanding the scope of usage. For any of this, Meta could do so without having to write a single line of code,” he added.
Also Read: Amid push for Indian AI, myriad challenges pose concerns
For its Indic language approach, Meta worked with Peak XV (formerly Sequoia Capital India)-backed local language AI startup, Sarvam AI. Earlier this month, Pratyush Kumar, cofounder of Sarvam AI, told Mint that the company is building domain-specific sub-models based on foundational models by sourcing vernacular language data.
For instance, a certain future generative AI application by Sarvam could offer neonatal assistance to new mothers without the need for doctor visits—a factor that can help improve post-pregnancy hygiene and health in rural areas.
Also Read: Is Nvidia’s near-monopoly status dangerous for the AI industry?
