DeepSeek races to unveil new AI model amid China tech boom

#DeepSeek #races #unveil #model #China #tech #boom

The Chinese Startp Dippayek has caused the earthquake in the global equity markets, which has exceeded $ 1 trillion last month with the launch of its competitive price, R1 last month.

According to sources familiar with the firm, in response to its initial success, the company based in Hangzhou is now hurrying to release its successor, R2, which aims to initially start before the May date.

Although specific timelines are unknown, Dupic is hopeful that the R2 will enhance coding capabilities and increase the reasoning to many languages ​​ahead of English.

Details of the high -speed timeline for the release of R2 were not reported before. DiPsic did not respond to a request to comment on the story.

Competates are still digesting R1’s implications, built with less powerful NVIDIA chips, but is competitive with people created by hundreds of billion dollars in spending by tech giants based in the United States.

“The launch of the DipoCR2 model in the AI ​​industry can prove to be an important moment in the AI ​​industry,” said Vijayasamah Allogata, chief operating officer of the Indian tech services provider. He said that the success of Deepesic in creating a cost -effective AI model “would potentially encourage companies around the world to intensify their efforts … by breaking the throat of a few dominant players in this sector,” he said. Said

The R2 is likely to worry about the US government, which has identified the AI ​​leadership as a national priority. His release could further promote Chinese authorities and companies, of which dozens say they have begun to integrate the Deep Saches model into their products.

DiPsic is rarely known, whose founder Liang Wenfering became a billionaire through his quantity hedge fund high -flyer. Liang, described by the former employer as “low key and interviewed”, has not spoken to any media since July 2024.

Reuters interviewed a dozen former employees as well as the Quant Fund professionals about the information about Dippic and his parents’ high -flyer operations. It also reviewed state media articles, companies’ social media posts and research papers started in 2019.

He told a story of a company that worked like a more research lab than a profitable enterprise and was disgraced by the traditions of China’s High Pressure Tech Industry rating, even as many investors were responsible. It has become seen as the latest development. Appearance

New way

Liang was born in 1985 in a village in the southern province of Guangdong. He later obtained a communication engineering degree at Elite Jiang University.

His first job was running a research department at a smart imaging firm in Shanghai. His then boss, Zhou Choon, told state media February 9 that Liang had hired award -winning algorithm engineers and worked with “Flat Management Style”.

In Dippayek and High Flyer, Liang has also eliminated the same Chinese tech giants methods that are known for tough top downdown management, low salary for young employees and “996”-which a week to 9 am in the evening Works by 9am.

Liang opened his Beijing office at a distance from two of China’s most famous educational institutions, Songhwa University and Packing University.

According to two former employees, he regularly joined the technical details and was pleased to work together with General Z Inteent and recent graduates, which contains most of his workforce. He also generally described working for eight hours in a co -operation environment.

“Liang gave us control and treated us as experts,” said Benjamin Liu, a 26 -year -old researcher who left the company in September. He constantly asked questions and learned with us. ” “DiPsic allowed me to own the main parts of the pipeline, which was very interesting.”

Liang did not answer the questions sent through Deep Sak.

While Bido and other Chinese tech companies were running to create a Chat GPT version of GPT in 2023 and to take advantage of the global AI boom, Liang told the Chinese media outlet waves last year that this Had deliberately avoided spending too much on the app growth, instead focused on it. Improve the quality of the AI ​​model.

According to the three persons familiar with its compensation methods, both the depressic and the high -flyer are known to be openly paid. In a high flight, one of the rival fund manager who knew Liang, said, making 1.5 million yuan annually for senior data scientists, while rival fund manager, who knows Liang, among them One said.

According to two industry members, the larger was financed by the High Flyer, which became one of China’s most successful quantities funds, and even after the official crackdown in the sector, still tens of billions of yuan. Manages, according to two people in the industry.

Computing Power

Three people said that the success of the DPCAC with a low -cost AI model is based on high -flyer prolonged and investment in research and computing power.

Quant Fund was a former pioneer in AI trading, and a top executive said in 2020 that the top flyer was going to “All -In” on AI by re -investing 70 % of his income, mostly in AI research.

In 2020 and 2021, the High Flyer spent 1.2 billion yuan on two supercompoting AI clusters. The second cluster, firefish II, consisted of about 10,000 Nvidia A100 chips, used for training AI models.

The person with direct knowledge of the officials said that the depressic was not established at the time, so the collection of computing power drew the attention of Chinese securities regulators.

“Regulators wanted to know why they need so many chips?” The man said. “How will they use it? How will it have an impact on the market?”

Authorities decided not to intervene, in the move, which would be important for DPC’s fortunes: The United States banned the export of A100 chips to China in 2022, at which fireflier II was already at work.

According to a Chinese official thinking person, Beijing is now celebrating Deep Sak, but it has directed that they do not engage with the media without any approval.

The man said that the authorities had asked Liang to have a low profile because they were worried that too much hype in the media would draw unnecessary attention.

Along with the Chinese cabinet and the Ministry of Commerce, China’s Securities Regulators did not respond to requests for comment.

As one of the few companies with a large A100 cluster companies, two former employees said high -flyer and DPCAC succeeded in attracting some of China’s excellent research skills.

“The key advantage of wide (computing) resources is that it allows widespread experiences,” said former employee Liu.

Some Western AI businessmen, such as Scale AIC CEO Alexander Wang, have claimed that DiPsic has more than 50,000 high -end newcomers that are banned from exporting China. It has not presented evidence of the allegation or did not respond to Reuters’ requests to provide evidence.

Deep Sak has not responded to Wang’s claims. Two former employees attributed the company’s success to Liang’s more cost -effective AI architecture.

Its research papers suggest that Startup used techniques such as compound specialists (MOE) and multi -headlining attention (MLA).

The MOE technique divides an AI model into different fields of skill and triggers them with just one queries, such as more common architecture that use the entire model.

MLA Architecture allows a model to simultaneously take action on a piece of information, which helps to detect key details more efficiently.

Although rivals such as French Mistle have developed MOE -based models, Deep SEC was the first firm to rely on the architecture, while gaining equality with the more expensive models.

In early February, Burnstein Brokerage analysts estimated that the pricing of DPCAC was 20 to 40 times cheaper than the openings that Openi had charged for equal models.

For now, Western and Chinese tech giants have indicated plans to continue heavy AI expenses, but with the R1 and its first V3 model, Depsek’s success has indicated some of the strategies.

Openi has reduced prices this month, while Google’s Gemini has introduced the discounted levels of access. Since the launch of the R1, the Open has also released an O3-MINI model that relies on low computing power.

Adnan Masood, a UST services provider, told Reuters that the benchmark was operated in his laboratory, in which R1 often three times more token for argument as an Openi Scaledown Dowan model, Or used the data units processed by the AI ​​model.

China hugged Deep Sak

Even before the R1 focused globally, before that, there were signs that Deepesic had caught Beijing’s right. In January, Chinese media reported that Liang attended a meeting with Chinese Prime Minister Li Kiang in Beijing, before the leaders of well -known firms, as the nominated representative of the AI ​​sector.

On the competition of the cost of its models, Dhoom Dham, which followed, has made Beijing’s belief that he cannot compete with the United States, Chinese companies and government agencies embraced the Deep Scams at this pace at this pace. It is not presented to other firms.

At least 13 Chinese city governments and 10 energy companies say they have deployed DPESC to their system, while Tech Giants Lenovo, Bedo and Tennant, the owner of China’s largest social media app Vicht, Dippic’s models. I merged with your product.

Chinese leader Xi Jinping and Lee’s Chinese policy makers in Singapore’s Lee Koan Yu School of Public Policy said “Chinese leader Xi Jinping and Lee have indicated that they confirm Deep Sak.” “Now everyone only supports it.”

The Chinese throat comes when South Korea’s governments from Italy removed the DPS from the national App Stores, citing privacy concerns.

“If DiPsic becomes an AI -going model in Chinese state institutions, Western regulators increase restrictions on AI chips or software support,” said Stephen Woo, founder of AI expert and hedge fund Carthaj Capital. As another reason, “said Stephen Woo, an AI expert and founder of the Hedge Fund Carthej Capital.

There is a further challenge on Advanced AI chips that Liang has recognized.

“Our problem has never been financed,” he told the wives in July. “This is banned at high -end chips.”