“Could three Indian engineers with $ 10 million build something similar to Openai?” The video sequence is experiencing a renewed popularity on social networks this month. We see Sam Altman questioned by an audience of business leaders and investors in New Delhi in 2023. His answer is scathing: “No hope of competing with us in the training of foundation models, you should not even to try.”
The scene had aroused a certain excitement in the country’s tech community. “Dear Sam, from a business manager to another … Challenge accepted”, had immediately retorted CP Gurnani, the director of Tech Mahindra, one of the largest IT service groups, in a post broadcast on X. And yet, eighteen months later, it was the publication of an open source model by the Chinese Deepseek who tickles American domination in the sector, hitherto deemed absolute. The company was created in May 2023 in Hangzhou and it claims to have spent only $ 5 million in computing power for the training of its model, where Openai spent 100 million for GPT-4.
Why was India not able to achieve this feat? The question has been animating for a few weeks the country which co -chases the AI action Summit organized in Paris from February 6 to 11. Exactly like that of Deepseek, the Sutra Dual2l model of Twoai, the most advanced Indian start-up in the sector, is based on an open source model made available to all by the Meta group of Mark Zuckerberg. It nevertheless displays performance much lower than that built by Chinese from this same base. And Deepseek is only the tip of the iceberg with regard to artificial intelligence in China. We should also mention Qwen (Alibaba), Sensetime, Minimax, Kimi and Duobao (Bytedance).
The Indian market muffled by the Americans
Many Indian voices call to follow the scheme established by the country’s spatial research organization, the ISRO. She managed to bring India into the space race by sending the Mars orbit of the Mangalyan satellite in March, in 2013, and ten years later, the Chandrayaan-3 probe on the Moon. Having cost $ 74 and $ 75 million respectively, these exploits were greeted as an ode to ingenuity and frugality.
The construction of extended language models is actually quite easy. The fundamental science behind these LLM has existed since 2017. Improvements, since then, have been available via research articles. Deepseek, who refined an approach called “mixture of experts” -several specialized submodels work together in order to answer a question, instead of a single large model taking care of everything -has publicly detailed its works.
Beyond the model, access to data is key. OPENAI accused Deepseek of having used data generated by his AI without permission. This is not lacking in salt, Openai being prosecuted by various media, especially Indians, which accuse him of having used their content without authorization to train his models.
The Indians also point to the difficulty of finding the right specialists. It is true that the best engineers and researchers in the country emigrate to the United States. But Deepseek proved that it was not essential to recruit superstars of artificial intelligence, hiring young graduates from local universities, sometimes with specialties other than IT or mathematics. The breeding ground for Indian talents is deep enough to compete.
Finally, the money is there. In March 2024, the government launched the Indiaai mission, allocating $ 1.5 billion for the next five years and creating a cluster of 19,000 fleas which will be made available to local start-ups.
The most plausible explanation of this delay in AI is the absence of a protected market, the same reason why India does not have its own Google or Facebook. As in Europe, local businesses are stifled by American service providers because they are cheaper and better. Conversely, Chinese players benefit from a huge market from which Openai and Anthropic are banned. India is known to be the second openai market in number of subscribers after the United States. Sam Altman plans to go there in early February and meet Prime Minister Modi, as he had already done in 2023. The opportunity to reiterate his declaration of then?
.