Open data, open compute, open models — these are the key ingredients for India’s success in the global artificial intelligence race, and there is no way forward but for the Indian government to step up and invest to create these enabling conditions, according to experts and industry representatives.
Abhishek Singh, president and CEO, National eGovernance Division, and managing director and CEO, Digital India Corporation, said that once key pieces like compute and access to Prasar Bharati content are solved, maybe Indian startups will start building foundational models.
"As long as we are dependent on the foundation models of OpenAI or other western models, there will be biases that will not really be suited for the Indian situation,” Singh said.
The government is the biggest holder of data, he said, adding that it is working on the India Datasets Platform to enable API-based data sharing and the National Data Governance Policy to put in place data management offices in every government department. Data must be made available with privacy preservation and randomisation tools so that researchers and startups can leverage it, Singh said.
India already has many startups developing AI solutions in agriculture, health care, education and so on, Singh said. The government’s role will be to support them in scaling up, bring them to the level of maturity and enable them to explore global markets.
DATA-DRIVEN DREAM
Indian industry too ‘salivates’ over all the data housed in Doordarshan, said Debjani Ghosh, president of Indian IT industry apex body NASSCOM.
"It's like we are sitting on the most precious treasure trove, and we are not using it,” Ghosh said, adding that high-value open datasets are a must.
Late last year, Ola founder Bhavish Aggarwal-backed Krutrim AI launched India's first foundational large language model. The multilingual model is trained on two trillion tokens, comprehends 22 Indian languages and can generate content in 10 languages.
But so far, most Indian AI innovators continue to depend on LLMs made in the west. Even Sarvam AI’s Hindi LLM OpenHathi, apart from being an only seven-billion parameter model, was built on top of Meta’s Llama-2 model and not from scratch.
SHOW ME THE MONEY
And Indian industry is unlikely to put money into building a large and open source foundational model given the costs involved and compute power required.
"I don't think in the industry, anyone’s looking to build an organic, large GPT, and for obvious reasons,” Ghosh said. "That is something the government will have to drive. That is something that has to be one of the government’s big investments, and they have to do it quickly.”
She added that the industry’s interest lies in building small- or medium-use cases and vertical solutions for domains. In the absence of an Indian equivalent of OpenAI’s GPT, businesses will opt for models openly available.
Even so, funding for Indian AI startups is not sufficient. “If you look at an average Indian startup today that's building for AI, building AI solutions, they will get maybe $5-10 million, at best,” Ghosh said. "The same capabilities in the US, we are talking about anything from $30 to $50 million.”
The message from investors is that if you want the big funds, you have to move somewhere else, she added. There is a need for deep and patient capital to build Indian use cases, Ghosh said. “That’s where the government can definitely play a role through some sort of incentivisation or scheme, which will enable them leapfrog a little faster,” Ghosh said, adding that this is needed especially for small and medium companies to deploy solutions at scale.
BUILDING BLOCKS
The government has, in recent months, come up with a solution for the compute crunch AI companies face.
ET reported on January 16 that the government may allot up to Rs 1O.OOO crore to set up supercomputing and quantum computing hubs in collaboration with the private industry under public-private partnerships. It plans to create a 30,000 GPU (graphic processing unit) cluster in a digital public infrastructure model to make available high performance computing to startups, micro, small and medium enterprises (MSMEs) and other companies, either on lease or on a 'compute-asa-service' basis, officials said.
"The first big bottleneck is that all of the AI that we know today, large language models, are largely being produced by private sector companies, majority of which are in the US,” Rahul Matthan, technology policy lawyer and partner, Trilegal, told ET.
And this poses a geopolitical risk, he argued. API-based access to foreign models creates a funding constraint that Indian companies may not be able to afford. If they opt for open source models, there maybe constraints in terms of quality.
While there has been a lot of conversation on the risks AI poses, particularly deepfakes, industry leaders believe this should not distract from the positives.
Regulation should only be enabling, Matthan reiterated, and there is no need for separate regulation of AI. Existing frameworks are sufficient to address the possible harms that the new technology may bring.
The focus should be on identifying and prioritising three or four sectors, such as manufacturing, logistics, agriculture, etc, which are critical to India’s GDP growth, and deploy AI there to boost productivity and India’s global competitiveness, Ghosh concluded.
By Annapurna Roy & Surabhi Agarwal
https://economictimes.indiatimes.com/epaper/delhicapital/2024/feb/18/tech-tonic/the-aidea-of-india/articleshow/107787272.cms
Comments