ByteDance advances DeepSeek work in AI reasoning with open-source project led by intern
Advertisement
DAPO, or Decoupled Clip and Dynamic Sampling Policy Optimisation, is a scalable reinforcement learning algorithm that helps a large language model (LLM) achieve better complex reasoning behaviour such as self-verification and iterative refinement, according to a research paper published earlier this week by ByteDance and Tsinghua University's Institute for AI Industry Research.
The algorithm outperformed the reinforcement learning approach in DeepSeek's R1 reasoning model, scoring 50 points in the American Invitational Mathematics Examination (AIME) 2024 using Alibaba Group Holding's Qwen2.5-32B base model, compared with 47 points attained by R1 when applying the same Alibaba model, the paper showed. Alibaba owns the South China Morning Post.
Notably, DAPO achieved the better result with 50 per cent fewer training steps.
TikTok owner ByteDance has invested heavily in artificial intelligence. Photo: Digitimes
The achievement drew positive academic and industry comments. Google DeepMind engineer Philipp Schmid, who shared the project on X, said the new method was 'better than' DeepSeek's 'group relative policy optimisation (GRPO)' in reinforcement learning. GRPO is one of DeepSeek's training methods that enables a model to learn by comparing different actions and making updates with a 'group' of observations.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
South China Morning Post
6 hours ago
- South China Morning Post
Hong Kong is poised to catch the global AI wave
In the first wave of the internet and e-commerce, Hong Kong was slow to adopt the new technology. With the rise of artificial intelligence (AI), the city is faring much better this time. No doubt mainland China's emergence as a global tech superpower has a lot to do with it as it gives both direction and incentive for local officials, universities and firms on how to deploy their ample capital resources, research capabilities and scientific talent. The sudden appearance of DeepSeek , with its ingenious 'less is more' approach, has also served as an inspiration to the nation. According to a global survey by fintech giant Finastra, 38 per cent of Hong Kong's financial institutions have adopted generative AI, well ahead of the global average of 26 per cent. Top-tier researchers, both local and those from the mainland, have helped push our universities to the forefront of innovation, along with government support. That has translated into our high university rankings in data science and AI. This year's QS World University Rankings placed the Hong Kong University of Science and Technology, the University of Hong Kong and the Chinese University of Hong Kong in the top 20. Even Cyberport, long looked upon as glorified real estate, is finally fulfilling its original mandate. Its AI Supercomputing Centre – which only began operations in December – is a cornerstone of the government's strategy to foster a local AI ecosystem and has already achieved more than 90 per cent utilisation. It has achieved in months what it was trying to do over many years. AI development is building up a momentum of its own.
HKFP
13 hours ago
- HKFP
China denies asking firms to collect data illegally after new EU probe on TikTok
Beijing denied on Friday asking firms to 'illegally' collect and store users' personal information, after an Irish regulator helping the European Union regulate data privacy began investigating Chinese social media giant TikTok. 'The Chinese government attaches great importance to and protects data privacy and security in accordance with the law,' foreign ministry spokeswoman Mao Ning said. Beijing 'has never and will never require companies or individuals to illegally collect or store data', Mao said. 'We hope that the European side will respect the market economy and fair competition, and provide a fair, just and non-discriminatory business environment for companies from all countries,' she told a regular news conference. The social media giant has been in the crosshairs of Western governments for years over fears that personal data could be used by China for espionage or propaganda purposes. However, TikTok has insisted that it has never received any requests from Chinese authorities for European users' data. TikTok was fined 530 million euros (US$620 million) in May by the Data Protection Commission over sending personal data to China, although the Chinese social media giant had insisted this data was only accessed remotely. TikTok, which has 1.5 billion users worldwide, is a division of Chinese tech giant ByteDance.
South China Morning Post
a day ago
- South China Morning Post
Robinhood CEO's AI maths start-up valued at US$875 million, just shy of unicorn status
Harmonic AI, an artificial intelligence start-up co-founded by Robinhood Markets chief executive officer Vlad Tenev, has raised US$100 million in funding to tackle a problem that has sometimes confounded AI models: maths. The series B funding round was led by Kleiner Perkins, with participation from Sequoia Capital, Index Ventures and Paradigm. The deal valued the AI start-up at US$875 million, said Tenev, who serves as the company's executive chairman, a non-operating role. Harmonic's CEO is Tudor Achim, who previously led autonomous driving start-up Founded in 2023 by Tenev and Achim, the Palo Alto, California-based start-up aims to build AI systems that can solve complex maths problems, creating what the company refers to as mathematical superintelligence. Harmonic plans to make its flagship AI model, Aristotle, available to researchers and the general public later this year. 'The near-term goal is to build an AI that solves maths problems at a level that is superior to any human,' Tenev said. 'The ultimate goal would be to solve major unsolved mathematical problems and expand that to problems in physics and computer science.'
