DeepSeek unveils new AI reasoning method as anticipation for its next-gen model rises
artificial intelligence (AI) start-up
DeepSeek has introduced a novel approach to improving the reasoning capabilities of large language models (LLMs), as the public awaits the release of the company's next-generation model.
Advertisement
In collaboration with researchers from Tsinghua University, DeepSeek developed a technique that combines methods referred to as generative reward modelling (GRM) and self-principled critique tuning, according to a paper published on Friday. The dual approach aims to enable LLMs to deliver better and faster results to general queries.
The resulting DeepSeek-GRM models outperformed existing methods, having 'achieved competitive performance' with strong public reward models, the researchers wrote. Reward modelling is a process that guides an LLM towards human preferences.
DeepSeek intended to make the GRM models open source, according to the researchers, but they did not give a timeline.
The academic paper, published on the online scientific paper repository arXiv, comes amid speculation about the start-up's next move following the global attention garnered by the firm's V3 foundation model and R1 reasoning model.
Advertisement
Reuters reported last month that DeepSeek-R2, the successor to R1, could be released as soon as this month, as the company rushes to capitalise on its rising profile. The release of DeepSeek-R1 rocked the global tech community with its cost-efficient performance that rivalled leading models.
DeepSeek has remained tight-lipped about the rumoured R2 release. It has not commented on the matter through official public channels, but a customer service account denied the report in a group chat with business clients, Chinese media outlets reported last month.
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
RTHK
28 minutes ago
- RTHK
No 'backdoors' in our chips, says Nvidia
No 'backdoors' in our chips, says Nvidia Nvidia says it "does not have 'backdoors' in our chips that would give anyone a remote way to access or control them". File photo: Reuters Nvidia chips do not contain "backdoors" that allow remote access, the US tech giant has said, after Beijing summoned company representatives to discuss "serious security issues". The California-based company is a world-leading producer of AI semiconductors and this month became the first company to hit US$4 trillion in market value. But it has become entangled in trade tensions between China and the United States, and Washington effectively restricts which chips Nvidia can export to China on national security grounds. "Cybersecurity is critically important to us. Nvidia does not have 'backdoors' in our chips that would give anyone a remote way to access or control them," Nvidia said on Thursday. A key issue has been Chinese access to the "H20" – a less powerful version of Nvidia's AI processing units that the company developed specifically for export to China. Nvidia said this month it would resume H20 sales to China after Washington pledged to remove licensing curbs that had halted exports. But the tech giant still faces obstacles – US lawmakers have proposed plans to require Nvidia and other manufacturers of advanced AI chips to include built-in location tracking capabilities. Beijing's top internet regulator said on Thursday it had summoned Nvidia representatives to discuss recently discovered "serious security issues" involving the H20. The Cyberspace Administration of China said it had asked Nvidia to "explain the security risks of vulnerabilities and backdoors in its H20 chips sold to China and submit relevant supporting materials". (AFP)
AllAfrica
an hour ago
- AllAfrica
US-China trade talks threaten to explode over Russia oil
The United States has urged China to halt its purchases of oil and gas from Russia, threatening it with secondary sanctions, including possible 100% tariffs. The warning came after the US reached trade deals with the European Union and Japan that set tariffs at 15%. In meetings in Stockholm, Sweden, earlier this week, US Treasury Secretary Scott Bessent expressed to trade negotiators led by Chinese Vice Premier He Lifeng Washington's displeasure with China's continued purchases of sanctioned Russian as well as Iranian oil. Bessent also complained about China's sales of over US$15 billion worth of dual-use technology goods to Russia that have reputedly supported Moscow's war against Ukraine. In a briefing on Wednesday (July 30), Bessent told the media that Chinese officials responded to his call by stating that China is a sovereign nation with its own energy needs and that oil purchases would be based on the country's internal policies. 'The Chinese take their sovereignty very seriously. We don't want to impede their sovereignty, so they'd like to pay a 100% tariff,' Bessent said, sarcastically. He told the press that US and Chinese officials did not strike a deal to extend their 90-day tariff truce, which is set to expire on August 12. He said that US President Donald Trump would decide whether the US would extend the deadline, typically by 90 days, or reimpose punishing tariffs. During his presidential election campaign, Trump vowed to end the war in Ukraine within 24 hours of becoming president. Since Trump took office on January 20, he has tried to bring Ukrainian President Volodymyr Zelensky and Russian President Vladimir Putin to the negotiating table. On March 14, Trump admitted that he was being 'a little bit sarcastic' when he said he could end the Ukraine war within a day. In early July, reports said he came to the realization that Putin had no intention of engaging in negotiations. On July 14, Trump set a 50-day deadline for Moscow to end the war, or the US will send billions of weapons to Ukraine, impose a 100% tariff on Russian goods, and also on products coming from countries that bought Russian oil. On July 22, Bessent said that he would discuss China's oil purchases from Russia and Iran with Chinese officials. On Monday, Trump said Putin's new deadline to end the war would be 10 or 12 days. That would be a few days before the US-China tariff truce ends on August 12. The world's spotlight is now on whether Beijing will distance itself from Russia to avoid 100% US tariffs. 'China will take energy supply measures that are right for China based on our national interests. Tariff wars have no winners,' Guo Jiakun, a spokesperson of the Chinese Foreign Ministry, said Wednesday. 'Coercion and pressuring cannot solve problems. China will firmly safeguard its own sovereignty, security, and development interests.' Chinese commentators said the China-Russia-Iran alliance will not be broken by US pressure. Chen Fei, an associate professor of international relations at Central China Normal University, says in an article that the United States' wish to break China's ties with Russia and Iran is unlikely to succeed. 'As one of the world's largest crude oil importers, China maintains close energy cooperation with Russia and Iran,' he writes. 'The US tried to include this issue in the Sino-US trade negotiation framework and intended to exert pressure on China, attempting to use energy issues as a bargaining chip. This is a strategic test of Sino-Russian and Sino-Iranian relations.' He opines the US played this 'energy card' because it faced an unfavorable political and economic situation within its country. 'China will not change its foreign policy and the direction of international cooperation due to unilateral pressure from the US. It will continue to work hand in hand with like-minded countries to respond to global challenges jointly,' he adds. Xin Qiang, a professor at China's Fudan University, said China would not allow the US to make non-tariff issues the focus of the trade negotiations or bundle trade matters with non-trade issues. A Shandong-based columnist argues that the US cannot intervene in China's purchases of Russian and Iranian oil products, as all these transactions are settled in renminbi, rendering US sanctions a 'paper tiger.' Citing Customs data, the writer says that Russia has been China's top crude oil supplier for 20 consecutive months, delivering 2.16 million barrels daily via pipeline, approximately 600,000 barrels more than those from Saudi Arabia. He says Iran also ships 1.8 million barrels of crude oil to China daily at peak times. 'The US wants to use secondary sanctions to force China to stop buying Russian and Iranian oil and shift to buy its shale oil,' he says. 'China has not bought shale oil from the US for three consecutive months. The US seemed tough but exposed its anxiety about falling oil exports.' He says the US will only fuel its inflation if it imposes an additional 100% tariff on Chinese imports. He says that European countries, including Germany and France, will not follow the United States' call to reduce oil and gas imports from Russia and Iran, respectively. Trump said he decided to shorten Putin's deadline to end the Ukraine war from 50 days to 10-12 days because he saw no progress being made toward achieving a ceasefire. His decision followed two significant developments. On July 23, Trump announced that the US had agreed to a 'massive' trade deal with Japan where Tokyo committed to invest $550 billion in the US, while the US will only impose a 15% tariff on Japanese imports. On July 27, Trump and European Commission President Ursula von der Leyen met in Scotland and agreed on a similar 15% US tariff on all EU goods. 'I think that the Chinese were surprised by the magnitude of the Japan deal, and by the magnitude and the terms of the European deal,' Bessent said Wednesday. 'I believe they were in more of a mood for a wide-ranging discussion. The discussion centered on the two economies.' He said the US side expressed its concern about Chinese overcapacity globally and its impact on other countries within the next few years. He said he would not be surprised if Europe, Canada and Australia followed in the United States' footsteps and imposed anti-dumping tariffs on Chinese goods at some point. The fact that the Trump administration had reached trade deals with the EU, Japan and some ASEAN countries before the meeting with the Chinese side has created some pressure on Beijing, said Yang Shuiqing, a researcher at the Institute of American Studies at the Chinese Academy of Social Sciences. He said that under their trade deals with Washington, the EU and Japan will open their markets to US companies, while their tariffs will hurt Chinese firms' competitiveness. Additionally, the EU and Japan have pledged to increase investment in the US, potentially leading to a reduction in investment in China. Yang said that, fortunately, China has a 'rare earth card' to play, making America unable to ignore its demands. Yuyuan Tantian, a social media account affiliated with the state-owned China Central Television, reported on Wednesday that Chinese Commerce Minister Wang Wentao met with representatives from approximately 10 American companies in Beijing. It stated that Wang discussed business opportunities for US firms in China. Some other Chinese commentators also believed that the US would not escalate its trade war with China in the short term, as Trump does not want to jeopardize his chance to visit China on September 3, when China celebrates the 80th anniversary of its victory in the War against Japanese Aggression and the World Anti-Fascist War. Read: China-EU play nice despite widening cracks
South China Morning Post
17 hours ago
- South China Morning Post
DeepSeek founder shares best paper award at top global AI research conference
A research paper co-authored by Liang Wenfeng, founder of Chinese artificial intelligence start-up DeepSeek, was honoured with the best paper award at the Association for Computational Linguistics (ACL) conference in Vienna, Austria, widely recognised as the premier global conference for AI researchers. The paper, titled 'Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention,' was published on February 27, with Liang listed as one of 15 authors. The 'native sparse attention' mechanism is a core improvement that underpins the high efficiency and low-cost performance of DeepSeek's AI models. The paper's win comes as Chinese scientists and researchers are outperforming US peers in basic research in the field of computational linguistics and natural language processing. At this year's ACL conference, more than half of the first-named authors on accepted papers originated from China, up from less than 30 per cent last year. The US ranked second, with 14 per cent of first-named authors, according to ACL data. Among the four best papers recognised by ACL, two author teams were from China. They included Liang's DeepSeek team and Yang Yaodong's team from Peking University. An undated photo of DeepSeek's Liang Wenfeng. Photo: Weibo Yang, an assistant professor at the Institute of Artificial Intelligence and chief scientist of the Peking University-PsiBot Joint Laboratory, led research that explored a possible mechanism explaining the fragility of alignment in language models, attributed to the elasticity of language models.
