Latest news with #DeepSeek-V3
Time of India
08-07-2025
- Business
- Time of India
Fingerprint scans, military experts and more: How and why ChatGPT-maker OpenAI is tightening security at the company
ChatGPT-maker OpenAI has overhauled its security operations to protect its valuable intellectual property from corporate espionage , a report has said. These enhanced measures come amid claims of the artificial intelligence (AI) giant being targeted by Chinese rivals like DeepSeek that garnered significant attention in early 2025 for its high-performing and cost-effective AI models, particularly its chatbot (DeepSeek-V3) and reasoning model (DeepSeek-R1). Citing sources close to the organisation, a report by The Financial Times claims that OpenAI began bolstering its security last year, the urgency intensified after Chinese AI startup DeepSeek released a rival model in January. Why OpenAI is tightening security of its data OpenAI claimed that DeepSeek had improperly copied its models using a technique known as "distillation" to create their AI system – an incident 'prompted OpenAI to be much more rigorous,' said one person close to its security team. The company is said to be 'aggressively' expanding its security personnel and practices, including its cybersecurity teams. How OpenAI is tightening security The company has implemented stricter controls on sensitive information and enhanced staff vetting in recent months. by Taboola by Taboola Sponsored Links Sponsored Links Promoted Links Promoted Links You May Like La mejor alarma arrasa en España, no vas a creer este precio Securitas Alarma Más información Undo Stricter policies, known as information "tenting," have been in place at their San Francisco offices since last summer. These policies significantly reduce the number of people who can access crucial information about technologies like algorithms and new products. For instance, during the development of their "Strawberry" (codenamed o1) model, staff had to confirm that other employees were part of the "Strawberry tent" before discussing the project in communal areas. Last October, the company hired Dane Stuckey as its new chief information security officer, who works alongside Matt Knight, OpenAI's vice-president of security products. Knight has been developing ways to leverage OpenAI's large language models to enhance its defenses against cyberattacks. Retired US Army General Paul Nakasone has also been appointed to OpenAI's board last year to help oversee its cybersecurity defenses. AI Masterclass for Students. Upskill Young Ones Today!– Join Now
AllAfrica
25-06-2025
- Business
- AllAfrica
DeepSeek gets Nvidia's high-end GPUs via Singapore: US official
DeepSeek, a Hangzhou-based artificial intelligence firm, has access to large volumes of Nvidia's high-end graphics processing units (GPUs) and supports China's military and intelligence operations, according to a United States official. An unnamed senior State Department official told Reuters in an interview that DeepSeek sought to use Southeast Asian shell companies to obtain high-end Nvidia chips, including H100 chips, which cannot be shipped to China under US rules. The report said the US official's comment showed that DeepSeek's fast-growing AI capabilities were exaggerated, as the company still relied heavily on US technology. The official warned that DeepSeek had provided user information to the Chinese government but had declined to comment on whether the US would implement further export controls or sanctions against DeepSeek. 'DeepSeek has willingly provided and will likely continue to support China's military and intelligence operations,' said the official. 'This effort goes above and beyond open-source access to DeepSeek's AI models.' An Nvidia spokesman told Reuters that the company's review indicates that DeepSeek used 'lawfully acquired' H800 products, not H100. According to the academic papers published by DeepSeek's researchers, the company used 2,048 Nvidia H800 chips to train its DeepSeek-V3 large language model (LLM). DeepSeek-V3 required 2.788 million H800 GPU hours for its complete training, meaning the total training time was about 56.7 days. DeepSeek also claimed that the training cost for its AI model was only US$5.58 million. By comparison, Meta spent US$500 million to train its Llama 3.1. A group of DeepSeek researchers said in a paper on January 22 that DeepSeek-R1's training used the 'distilled data' from Alibaba's Tongyi Qianwen (Qwen) and Llama. The 'distillation' method uses outputs from a larger AI model to train and improve a smaller one. However, US officials, including the incumbent Commerce Secretary Howard Lutnick, said DeepSeek could create its AI models 'dirt cheap' by purchasing many Nvidia chips and stealing data from Meta's open platform. In October 2022, the Biden administration banned the exports of Nvidia's A100 and H100 chips to China. In October 2023, it also banned the exports of the A800 and H800 chips to China. Many commentators have said that these export controls have a loophole that allows Chinese firms to access American high-end chips through third countries. In January this year, Alexandr Wang, chief executive of the US-based Scale AI, told CNBC that DeepSeek has 50,000 units of H100 chips, the most advanced Nvidia chips on the market. Wang did not provide any evidence or more details. 'We believe they have access to around 50,000 Hopper GPUs, which is not the same as 50,000 H100, as some have claimed,' a team of analysts led by Dylan Ratel wrote in a report on January 31. 'We believe DeepSeek has access to around 10,000 H800s and about 10,000 H100s. Furthermore, they have orders for many more H20's.' The team said DeepSeek's total server CapEx is about US$1.6 billion, plus an operational cost of US$944 million. On April 16, the US House Select Committee on the Chinese Communist Party (CCP) opened an investigation into Nvidia's sales across Asia after issuing a report claiming that DeepSeek had illicitly accessed the firm's chips to train its AI models. The latest Reuters report, citing three unnamed sources, said DeepSeek could still procure the H100 chips after the US banned Nvidia from selling them to China. The sources said DeepSeek obtained far below 50,000 units of Nvidia's high-end chips. Last December, The Information reported that Nvidia had asked Super Micro and Dell to audit their Southeast Asian customers to verify that they still possessed the Nvidia-powered servers they had bought. After DeepSeek released its low-cost DeepSeek-R1 model on January 20 this year, the Trump administration reportedly had started probing whether DeepSeek bought Nvidia's advanced chips through Singapore. 'Nvidia has stated that there is no reason to believe that DeepSeek obtained any export-controlled products from Singapore,' Singapore's Ministry of Trade and Industry said in a statement on February 1. 'We expect US companies, like Nvidia, to comply with US export controls and our domestic legislation,' it said. 'Our customs and law enforcement agencies will continue to work closely with their US counterparts.' In February, Singapore charged three men with fraud for allegedly helping ship Nvidia's high-end chips to DeepSeek in China in 2024. The accused include two Singaporeans, Aaron Woon Guo Jie, 41, and Alan Wei Zhaolun, 49. According to the court papers, the duo committed fraud by falsely declaring to US-based companies Super Micro Computer and Dell Technologies that the servers they purchased, possibly containing Nvidia chips under US export controls, would not be sent to unauthorized recipients. But eventually, the accused shipped these servers to Malaysia and potentially elsewhere. The third person charged was Chinese national Li Ming, 51, who claimed in 2023 that the imported items would be used by the Singapore-registered company Luxuriate Your Life Pte Ltd. The three defendants, if convicted, could face penalties of a jail term of up to 20 years or a fine, or both. Media reports said Wei, a naturalized Singaporean, spent more than two decades building his cloud solution businesses in Asia and owns or helps to manage at least 15 firms, including Aperia Cloud Services and A-Speed Infotech Pte. In March this year, Exsim, a Malaysian property and technology firm, said it mutually terminated its contract with Aperia Cloud Services for the Exsim Hyperscale Data Centre in Bukit Jalil and will source equipment from a 'Japanese Fortune 500 company.' On May 13 this year, the Trump administration issued guidance on protecting supply chains against diversion tactics. It said China used advanced US chips as part of its 'military modernization efforts to improve the speed and accuracy of its military decision making, planning, and logistics, as well as of its autonomous military systems, such as those used for cognitive electronic warfare, radar, signals intelligence, and jamming.' It now requires chip makers to follow a set of 'know your customer' (KYC) rules and raise red flags for possible transshipment of US high-end chips to China. The Financial Times reported on June 9 that the Trump administration might ease restrictions on selling chips to China in exchange for the country's key minerals. However, a meeting between US and Chinese officials in London on June 9-10 seemed to have focused on other things, such as allowing Chinese students to study in the US. Read: American AI chipmakers' eyes are on Saudi Arabia now, not China
Mint
23-06-2025
- Business
- Mint
US accuses Chinese AI firm DeepSeek of aiding Beijing's military and dodging chip export rules
The United States has accused Chinese artificial intelligence firm DeepSeek of aiding Beijing's military and intelligence services, with a senior U.S. official claiming the company has also attempted to bypass export restrictions to acquire advanced American semiconductor technology. The allegations, disclosed in an interview withReuters, mark a significant escalation in Washington's scrutiny of Chinese AI firms amid an ongoing technological rivalry and broader trade tensions between the two global powers. According to the U.S. State Department official, who spoke on condition of anonymity, DeepSeek has not only collaborated with China's military and intelligence sectors but has also tried to obtain restricted Nvidia chips through front companies based in Southeast Asia. 'DeepSeek has willingly provided and is likely to continue supporting China's military and intelligence operations,' the official stated. 'This goes well beyond simple open-source AI model access.' The Hangzhou-based startup shocked the global tech community earlier this year by claiming its AI reasoning models, DeepSeek-V3 and DeepSeek-R1, matched or exceeded those developed by U.S. giants like OpenAI and Meta, at a significantly lower cost. The company said it had spent just $5.58 million on computing power to train its models, a figure that has drawn scepticism from AI researchers who believe the actual costs were likely far higher. The U.S. official further alleged that DeepSeek is sharing user data and analytics with China's surveillance systems. While Chinese law mandates that firms must comply with government data requests, this explicit claim, if true, could raise serious privacy concerns for DeepSeek's tens of millions of daily global users. U.S. lawmakers have previously warned that DeepSeek transmits American user data through infrastructure linked to China Mobile, a state-owned telecom operator. DeepSeek has so far remained silent on questions related to its privacy policies and alleged data-sharing practices. Adding to concerns, DeepSeek has reportedly been cited over 150 times in procurement records linked to China's People's Liberation Army (PLA) and associated defence entities. The U.S. official claimed the firm has supplied technology to PLA research bodies, thoughReuters was unable to independently verify these procurement links. Despite U.S. restrictions on the export of Nvidia's high-end H100 chips to China since 2022, DeepSeek has reportedly acquired a substantial quantity of them. The U.S. official alleged that the company used shell entities in Southeast Asia to access the chips and is attempting to utilise regional data centres to remotely operate the hardware. While the official declined to confirm whether DeepSeek had successfully evaded these controls, the suggestion that one of China's most high-profile AI firms could be skirting U.S. restrictions is likely to trigger further investigations. Responding toReuters, Nvidia said it does not support any firm violating export controls or appearing on U.S. entity lists. 'With current export regulations, we are effectively excluded from the China data centre market, now served mainly by domestic players such as Huawei,' the company said. Reuters reported that DeepSeek had obtained H100 chips, others disputed the scale of its holdings. Nvidia stated its own review found DeepSeek had used H800 processors, a less powerful, export-compliant variant, rather than H100s. Earlier this year, Singaporean authorities charged three individuals with fraud in a case linked to the illicit transfer of Nvidia chips to DeepSeek. Meanwhile, Malaysian officials confirmed last week that they are investigating whether a Chinese company is using Nvidia-equipped servers in the country to train large language models, potentially in violation of domestic regulations. Despite mounting concerns, DeepSeek has not yet been added to any U.S. trade blacklist, and there is no public evidence that Nvidia knowingly facilitated any military-related work by the company. Under U.S. rules, companies must refrain from exporting advanced chips to Chinese firms that are either blacklisted or involved in the development of weapons of mass destruction. Nonetheless, the developments are likely to increase pressure on Washington to tighten its monitoring of Chinese tech firms, particularly those with potential military links. (With inputs from Reuters)
Time of India
10-06-2025
- Business
- Time of India
Rednote joins wave of Chinese firms releasing open-source AI models
HighlightsChina's Rednote, also known as Xiaohongshu, has released an open-source large language model called joining other Chinese tech firms in making their artificial intelligence models freely available. The open-source strategy of Chinese companies contrasts with the proprietary approach of many U.S. tech giants like OpenAI and Google, although some American firms, including Meta, have also released open-source models. Rednote's new model performs comparably to Alibaba's Qwen 2.5 series in coding tasks but lags behind more advanced models such as DeepSeek-V3. China's Rednote , one of the country's most popular social media platforms, has released an open-source large language model , joining a wave of Chinese tech firms making their artificial intelligence models freely available. The approach contrasts with many U.S. tech giants like OpenAI and Google, which have kept their most advanced models proprietary, though some American firms including Meta have also released open-source models. Open sourcing allows Chinese companies to demonstrate their technological capabilities, build developer communities and spread influence globally at a time when the US has sought to stymie China's tech progress with export restrictions on advanced semiconductors. Rednote's model, called is available for download on developer platform Hugging Face. A company technical paper describing it was uploaded on Friday. In coding tasks, the model performs comparably to Alibaba 's Qwen 2.5 series, though it trails more advanced models such as DeepSeek-V3, the technical paper said. RedNote, also known by its Chinese name Xiaohongshu, is an Instagram-like platform where users share photos, videos, text posts and live streams. The platform gained international attention earlier this year when some U.S. users flocked to the app amid concerns over a potential TikTok ban. The company has invested in large language model development since 2023, not long after OpenAI's release of ChatGPT in late 2022. It has accelerated its AI efforts in recent months, launching Diandian, an AI-powered search application that helps users find content on Xiaohongshu's main platform. Other companies that are pursuing an open-source approach include Alibaba which launched Qwen 3 , an upgraded version of its model in April. Earlier this year, startup DeepSeek released its low-cost R1 model as open-source software, shaking up the global AI industry due to its competitive performance despite being developed at a fraction of the cost of Western rivals.
Indian Express
10-06-2025
- Business
- Indian Express
Rednote joins wave of Chinese firms releasing open-source AI models
China's Rednote, one of the country's most popular social media platforms, has released an open-source large language model, joining a wave of Chinese tech firms making their artificial intelligence models freely available. The approach contrasts with many U.S. tech giants like OpenAI and Google, which have kept their most advanced models proprietary, though some American firms including Meta have also released open-source models. Open sourcing allows Chinese companies to demonstrate their technological capabilities, build developer communities and spread influence globally at a time when the U.S. has sought to stymie China's tech progress with export restrictions on advanced semiconductors. Rednote's model, called is available for download on developer platform Hugging Face. A company technical paper describing it was uploaded on Friday. In coding tasks, the model performs comparably to Alibaba's Qwen 2.5 series, though it trails more advanced models such as DeepSeek-V3, the technical paper said. RedNote, also known by its Chinese name Xiaohongshu, is an Instagram-like platform where users share photos, videos, text posts and live streams. The platform gained international attention earlier this year when some U.S. users flocked to the app amid concerns over a potential TikTok ban. The company has invested in large language model development since 2023, not long after OpenAI's release of ChatGPT in late 2022. It has accelerated its AI efforts in recent months, launching Diandian, an AI-powered search application that helps users find content on Xiaohongshu's main platform. Other companies that are pursuing an open-source approach include Alibaba which launched Qwen 3, an upgraded version of its model in April. Earlier this year, startup DeepSeek released its low-cost R1 model as open-source software, shaking up the global AI industry due to its competitive performance despite being developed at a fraction of the cost of Western rivals.
