Chinese start Deepseek released his flag that model R1 on January 20, Surprise of the Silicon Valley with the advanced model skills. R1 matches or exceeded The functionality of the one issued by Openai, Google and Meta – with a much smaller budget and without the latest chips.
Over the past week, the Deepseek application has proven popular to the public. It grew past chatgt in popularity, reaching No. 1 at the US Apple Apple store and inside The best Android free apps at Google Play store at the time of publication.
Deepseek's R1 release promoted questions whether billions of dollars of spending in previous years was worth it – and challenged the notion that the US leader is, for BBC.
He's Deepseek arrives while the US seems to spend on him. Last week, President Donald Trump announced a joint project with Openai, Oracle and Softbank called Stargate that carries up to $ 500 billion over the next four years in Data Centers and other infrastructure.
What is Deepseek?
Deepseek is a Chinese beginning of the one that creates Open the models he—To every developer can have access and build technology.
Deepseek is different from chatgpt because he declares his reasoning of the chain thinking before giving a response to a quick. Apple App Store and Google Play Store Views estimated that level of transparency, for Bloomberg.
It is free to download and use, although it requires users to register before they can access it.
Deepseek he. Photographs: Andrey Rudakov/Bloomberg through Getty Images
How much did Deepseek cost to develop?
In one paper Last month, Deepseek's researchers stated that they built and trained the model for under $ 6 million in just two months.
Connected: I co-founded an app I use Ai why I'm worried about my child's future
They stated that they used about 2,000 Nvidia H800 chips, which Nvidia adapted exclusively for China with lower data transfer rates, or slow speeds when compared to H100 chips used by US companies. In October, US Department of Trade stopped the sale of the H800 chip In China in order to prevent the entry into chips that can promote advances, especially for military purposes.
In contrast, Dario Amodei, General Director of SH.BA Startup Anthropic, said in July receiving $ 100 million to train he – And there are models today that cost closer $ 1 billion to train.
So how did Deepseek withdraw ahead of competition with fewer resources? Meta scientists' chief Yann Lecun stated in a post of topics On Saturday that Deepseek had “benefited from open research and open source”.
Connected: He's Head of Meta is in the middle of a public clash with Elon Musk: 'Secret impedes progress'
“They came up with new ideas and built them at the top of other people's work,” Lecun said. “Because their work has been published and an open source, everyone can benefit from it.”
Why is Deepseek so popular?
Deepseek is free and offers high line performance. Last week, scientific journal Nature published an article entitled, ”China's free, open model, Deepseek Thrills Scientist“The article showed that the R1 shows in some chemistry, mathematics and coding tasks were at the same time with one of Openai's most advanced models, The Model O1 Openai was released in September.
In addition to high performance, R1 is open weight, so researchers can study, reuse and build on it. It is not fully considered an open source because Deepseek did not do its public training data.
The model he has also received stellar ratings. Investor Marc Andreessen called it “one of the most amazing and impressive advances” he had “ever seen” in a Friday post on x While Microsoft Satya Nadella's CEO called it “super impressive“At the World Economic Forum last week in Switzerland.
Deepseek R1 is one of the most amazing and impressive advances I have ever seen – and as an open source, a deep gift for the world. ?
– Marc Andreessen ?? (@pmarca) January 24, 2025
What are Deepseek's effects on American technology shares?
Nvidia shares fell 13% behind the opening bell on Monday, wiping $ 465 billion from the Chipmaker market. Was the greatest decline in value in the US market history, for Bloomberg.
By NBC NewsMicrosoft and the alphabet fell 4% each while Meta fell nearly 2%.
NBC News also notes that the composition Nasdaq fell 3.4% in the opening bell, while the Dow fell 180 points and S&P 500 fell nearly 2%.
Connected: Why a prominent investor, 'Warren Buffett of Britain', is staying away from Nvidia shares