© QINQIE99/Shutterstock.com
A new champion of artificial intelligence (AI) has just emerged, and it comes straight from the Middle Kingdom. Enough to make the hairs stand on end of Silicon Valley giants and American leaders…
A minimal training cost compared to OpenAI
It’s a real tidal wave. In just a few days, DeepSeek’s model, R1, has topped the list of free downloads on iPhones in the United States. It’s also one of the most downloaded productivity apps in the Google Play Store.
The startup, founded in 2023 by hedge fund manager Liang Wenfeng, focuses primarily on research despite commercialization. With this in mind, it has deployed its model in open source, allowing researchers to exploit its algorithm. The strength of its AI: it can learn and improve automatically without human supervision. Similarly, when it develops its answers, it supports its reasoning for the user. A feature that is highly praised by the latter.
The technical details of R1 have astounded researchers, and for good reason. It is said to be as efficient, or even more in some cases, than OpenAI's o1 model. All this while requiring “that” $5.6 million for its training, a fraction of the cost of comparable models developed by American giants.
And the young company is thinking big, with the ambition of soon achieving general artificial intelligence, that is, capable of performing or learning any cognitive task specific to humans.
200% Deposit Bonus up to €3,000 180% First Deposit Bonus up to $20,000
Screenshot of the DeepSeek app on the App Store. © Presse-citron
Chinese resilience
DeepSeek reportedly used 2,048 NVIDIA H800 chips. This GPU is much less efficient than the H100 used by companies across the Atlantic, but the startup is full of ultra-talented local engineers from the best universities in the country. They have thus been able to optimize the accelerators.< /p>
This is a real blow to the American giants, because DeepSeek’s feat proves that huge spending is not necessarily necessary to develop powerful models. This calls into question the investment strategies deployed within Silicon Valley.
Worse still, the startup demonstrates China’s ability to circumvent the heavy sanctions against it, with the US government banning its companies from supplying it with cutting-edge chips. In this context, its founder Liang Wenfeng has been elevated to the rank of national hero.
Chinese tech giants are also focusing on models that compete with ChatGPT and the like, but the competition will intensify further with the $500 billion Stargate project, which involves building a huge infrastructure in the United States to support OpenAI’s AI capabilities.
- DeepSeek, a young Chinese startup, is shaking up Silicon Valley with its open source R1 model.
- Capable of reasoning, it is comparable to OpenAI’s o1. But it cost drastically less.
- The model demonstrates China's ability to bypass US sanctions.
📍 To not miss any Presse-citron news, follow us on Google News and WhatsApp.
[ ]