© Deepseek
The announcement had the effect of a bomb in Silicon Valley. On January 20, 2024, the Chinese startup DeepSeek unveiled its new artificial intelligence models. The performance of its systems astounded experts and caused an earthquake in the financial markets. American tech giants lost billions on the stock market while investors wondered: how can a virtually unknown company compete with OpenAI and Google ?
DeepSeek was not born in the posh neighborhoods of Shanghai or Beijing, but in Hangzhou, a Chinese metropolis known for its technological dynamism. Its founder, Liang Wenfeng, 40 years old, has an atypical profile. This mathematician by training first created an investment fund based on AI in 2015, before launching a charitable fund in 2020.
This financial expertise combined with a bold technological vision forges DeepSeek's DNA. Liang Wenfeng understands that the key to success lies in optimizing resources rather than in the race for raw power. An approach that contrasts with that of the American giants.
DeepSeek's strength lies in its ability to do more with less. While OpenAI uses over 16,000 Nvidia chips to train its models, the Chinese startup achieves similar performance with just 2,000 GPUs. This technical feat drastically reduces costs: DeepSeek spends about $6 million on computing power alone.
This efficiency is based on an innovative approach called “mixture of experts”. DeepSeek engineers have developed a method to distribute data analysis among several specialized models, while minimizing losses related to information transfers. An optimization that revolutionizes established AI paradigms.
The secret lies in the way DeepSeek organizes the dialogue between its different specialized models. Rather than processing all the information in a single massive system, the company has created an intelligent distributed architecture that only solicits the components necessary for each task.
The DeepSeek-V3 base model matches ChatGPT on many tasks: answering questions, solving logic problems and generating code. The free conversational interface already attracts millions of users. In France, the DeepSeek mobile application is at the top of downloads on iOS and Android.
Even more impressive, the DeepSeek R1 model dedicated to reasoning is on a par with GPT-4. A performance praised by Sam Altman himself, the CEO of OpenAI, on social networks. The Chinese startup also offers free access to this advanced model, unlike its competitors who reserve it for premium subscriptions.
Comparative tests conducted by independent experts confirm these performances. On standard industry benchmarks, DeepSeek obtains scores remarkably close to those of the best American models. This feat is all the more notable given that the company reached this level in less than a year of development.
DeepSeek is also disrupting established economic models. Its general public interface remains entirely free, while its prices for businesses are 10 to 40 times lower than those of OpenAI. This aggressive pricing policy is based on the frugality of its models in computing resources.
The company is also banking on open source, freely sharing the code of its technologies. This strategy allows other players to develop their own applications based on DeepSeek, creating a dynamic ecosystem. An approach that contrasts with the industrial secret jealously guarded by the American giants.
Startups with limited resources can now exploit cutting-edge models without massive investments in infrastructure. A change that could redistribute the cards in the digital economy.
DeepSeek’s success questions the effectiveness of American restrictions on the export of chips to China. Paradoxically, these limitations seem to have stimulated Chinese innovation, pushing companies to optimize their resources. The country is thus developing unique expertise in frugal AI.
This technological advancement is part of a race for technological supremacy between China and the United States. DeepSeek’s ability to develop high-performance models with limited resources could redistribute the cards in the AI industry.
This newcomer is also showing the tip of its code a few days after the grand announcement of the $500 billion Stargate project by US President Donald Trump.
200% Deposit Bonus up to €3,000 180% First Deposit Bonus up to $20,000Unlike Western companies that integrate teams dedicated to ethical issues, the Chinese startup's practices in this area remain opaque. Its chatbot notably applies certain censorship rules in accordance with Chinese regulations.
The company's privacy policy specifies that user data can be analyzed and stored on Chinese servers for an unlimited period.
DeepSeek's handling of sensitive questions illustrates these issues. The system systematically avoids certain political or historical subjects, reflecting the constraints of the Chinese context.
It is still important to know that it is possible to run DeepSeek locally, eliminating the risks associated with transmitting data over the network and remote servers.
15:40OpenAI tackles DeepSeek
After very positive statements, OpenAI is changing its tune. The American AI giant accuses DeepSeek of having used part of its work to develop its artificial intelligence. An accusation with the overtones of a publicity stunt, the technique used by DeepSeek to perfect its AI being used by most companies in the sector.
More details
08:07French Tech also impacted by the DeepSeek phenomenon
The rapid arrival of DeepSeek sounds like a godsend for French start-ups. The Chinese company's open-source model allows French Tech companies to develop or improve their development tools faster than expected. All this for free.
Read the full article
28/01Apple, the only brand to resist the DeepSeek assault
While all American AI companies have experienced a stock market crash, Apple has resisted. Why?? Because its internally developed model does not require all the power and investment colossal of its competitors. This choice of home made finally seems to bear fruit, even if the speed of development of Apple Intelligence is slower than that of competitors.
More d& #8217;infos
28/01DeepSeek launches Janus, its image generator competing with DALL-E
In full rise, DeepSeek announces its own image generator. Called Janus and already available for trial, it positions itself as a direct competitor to DALL-E, its counterpart used by ChatGPT. The head of OpenAI welcomed the arrival of DeepSeek, describing its work as impressive. He is looking forward to this competition, which will push OpenAI to improve in the months and years to come.
More info
01/27DeepSeek victim of a cyberattack
After the buzz, the hard blow. The Chinese AI startup DeepSeek had to suspend registrations for its service, due to a cyberattack that disrupted its operation. DeepSeek explains that its open-source model, although less expensive, may be the cause of this targeted attack.
Learn more
01/27But who is DeepSeek ?
The Chinese start-up DeepSeek has shaken up the AI world by unveiling a model that is just as powerful as ChatGPT but requires less computing power and therefore less investment. Enough to call into question the billions of dollars distributed in recent years. Investors have also weighed down the major American AI companies, led by Nvidia. But who is DeepSeek really?? We dug a little deeper to find out more about this Chinese start-up.
Read the article
📍 To not miss any Presse-citron news, follow us on Google News and WhatsApp.
[ ]